![viii Preface
move on to looking at the results for individual variables. And MANOVA for
repeated measures has been largely superseded by the models that we shall
describe in Chapter 8. Second, a classification technique such as LDF needs
to be considered in the context of modern classification algorithms, and these
cannot be covered in an introductory book such as this.
Some brief details of the theory behind each technique described are given,
but the main concern of each chapter is the correct application of the meth-
ods so as to extract as much information as possible from the data at hand,
particularly as some type of graphical representation, via the R software.
The book is aimed at students in applied statistics courses, both under-
graduate and post-graduate, who have attended a good introductory course
in statistics that covered hypothesis testing, confidence intervals, simple re-
gression and correlation, analysis of variance, and basic maximum likelihood
estimation. We also assume that readers will know some simple matrix alge-
bra, including the manipulation of matrices and vectors and the concepts of
the inverse and rank of a matrix. In addition, we assume that readers will
have some familiarity with R at the level of, say, Dalgaard (2002). In addition
to such a student readership, we hope that many applied statisticians dealing
with multivariate data will find something of interest in the eight chapters of
our book.
Throughout the book, we give many examples of R code used to apply the
multivariate techniques to multivariate data. Samples of code that could be
entered interactively at the R command line are formatted as follows:
R> library("MVA")
Here, R> denotes the prompt sign from the R command line, and the user
enters everything else. The symbol + indicates additional lines, which are
appropriately indented. Finally, output produced by function calls is shown
below the associated code:
R> rnorm(10)
[1] 1.8808 0.2572 -0.3412 0.4081 0.4344 0.7003 1.8944
[8] -0.2993 -0.7355 0.8960
In this book, we use several R packages to access different example data sets
(many of them contained in the package HSAUR2), standard functions for the
general parametric analyses, and the MVA package to perform analyses. All of
the packages used in this book are available at the Comprehensive R Archive
Network (CRAN), which can be accessed from http://CRAN.R-project.org.
The source code for the analyses presented in this book is available from
the MVA package. A demo containing the R code to reproduce the individual
results is available for each chapter by invoking
R> library("MVA")
R> demo("Ch-MVA") ### Introduction to Multivariate Analysis
R> demo("Ch-Viz") ### Visualization](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-9-320.jpg)
![4 1 Multivariate Data and Multivariate Analysis
data.” Useful books on data mining are those of Fayyad, Piatetsky-Shapiro,
Smyth, and Uthurusamy (1996) and Hand, Mannila, and Smyth (2001).
1.3 Types of variables and the possible problem of
missing values
A hypothetical example of multivariate data is given in Table 1.1. The special
symbol NA denotes missing values (being Not Available); the value of this
variable for a subject is missing.
Table 1.1: hypo data. Hypothetical Set of Multivariate Data.
individual sex age IQ depression health weight
1 Male 21 120 Yes Very good 150
2 Male 43 NA No Very good 160
3 Male 22 135 No Average 135
4 Male 86 150 No Very poor 140
5 Male 60 92 Yes Good 110
6 Female 16 130 Yes Good 110
7 Female NA 150 Yes Very good 120
8 Female 43 NA Yes Average 120
9 Female 22 84 No Average 105
10 Female 80 70 No Good 100
Here, the number of units (people in this case) is n = 10, with the number of
variables being q = 7 and, for example, x34 = 135. In R, a “data.frame” is
the appropriate data structure to represent such rectangular data. Subsets of
units (rows) or variables (columns) can be extracted via the [ subset operator;
i.e.,
R> hypo[1:2, c("health", "weight")]
health weight
1 Very good 150
2 Very good 160
extracts the values x15, x16 and x25, x26 from the hypothetical data presented
in Table 1.1. These data illustrate that the variables that make up a set of
multivariate data will not necessarily all be of the same type. Four levels of
measurements are often distinguished:
Nominal: Unordered categorical variables. Examples include treatment al-
location, the sex of the respondent, hair colour, presence or absence of
depression, and so on.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-20-320.jpg)
![1.5 Covariances, correlations, and distances 13
Σ =
σ2
1 σ12 . . . σ1q
σ21 σ2
2 . . . σ2q
...
...
...
...
σq1 σq2 . . . σ2
q
.
Note that σij = σji. This matrix is generally known as the variance-covariance
matrix or simply the covariance matrix of the data.
For a set of multivariate observations, perhaps sampled from some popu-
lation, the matrix Σ is estimated by
S =
1
n − 1
n
i=1
(xi − ¯x)(xi − ¯x) ,
where xi = (xi1, xi2, . . . , xiq) is the vector of (numeric) observations for the
ith individual and ¯x = n−1 n
i=1 xi is the mean vector of the observations.
The diagonal of S contains the sample variances of each variable, which we
shall denote as s2
i .
The covariance matrix for the data in Table 1.2 can be obtained using the
var() function in R; however, we have to “remove” the categorical variable
gender from the measure data frame by subsetting on the numerical variables
first:
R> cov(measure[, c("chest", "waist", "hips")])
chest waist hips
chest 6.632 6.368 3.000
waist 6.368 12.526 3.579
hips 3.000 3.579 5.945
If we require the separate covariance matrices of men and women, we can use
R> cov(subset(measure, gender == "female")[,
+ c("chest", "waist", "hips")])
chest waist hips
chest 2.278 2.167 1.556
waist 2.167 2.989 2.756
hips 1.556 2.756 3.067
R> cov(subset(measure, gender == "male")[,
+ c("chest", "waist", "hips")])
chest waist hips
chest 6.7222 0.9444 3.944
waist 0.9444 2.1000 3.078
hips 3.9444 3.0778 9.344
where the subset() returns all observations corresponding to females (first
statement) or males (second statement).](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-29-320.jpg)
![14 1 Multivariate Data and Multivariate Analysis
1.5.2 Correlations
The covariance is often difficult to interpret because it depends on the scales
on which the two variables are measured; consequently, it is often standardised
by dividing by the product of the standard deviations of the two variables to
give a quantity called the correlation coefficient, ρij, where
ρij =
σij
σiσj
,
where σi = σ2
i .
The advantage of the correlation is that it is independent of the scales of
the two variables. The correlation coefficient lies between −1 and +1 and gives
a measure of the linear relationship of the variables Xi and Xj. It is positive
if high values of Xi are associated with high values of Xj and negative if high
values of Xi are associated with low values of Xj. If the relationship between
two variables is non-linear, their correlation coefficient can be misleading.
With q variables there are q(q − 1)/2 distinct correlations, which may be
arranged in a q×q correlation matrix the diagonal elements of which are unity.
For observed data, the correlation matrix contains the usual estimates of the
ρs, namely Pearson’s correlation coefficient, and is generally denoted by R.
The matrix may be written in terms of the sample covariance matrix S
R = D−1/2
SD−1/2
,
where D−1/2
= diag(1/s1, . . . , 1/sq) and si = s2
i is the sample standard
deviation of variable i. (In most situations considered in this book, we will
be dealing with covariance and correlation matrices of full rank, q, so that
both matrices will be non-singular, that is, invertible, to give matrices S−1
or
R−1
.)
The sample correlation matrix for the three variables in Table 1.1 is ob-
tained by using the function cor() in R:
R> cor(measure[, c("chest", "waist", "hips")])
chest waist hips
chest 1.0000 0.6987 0.4778
waist 0.6987 1.0000 0.4147
hips 0.4778 0.4147 1.0000
1.5.3 Distances
For some multivariate techniques such as multidimensional scaling (see Chap-
ter 4) and cluster analysis (see Chapter 6), the concept of distance between
the units in the data is often of considerable interest and importance. So,
given the variable values for two units, say unit i and unit j, what serves
as a measure of distance between them? The most common measure used is
Euclidean distance, which is defined as](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-30-320.jpg)
![1.6 The multivariate normal density function 15
dij =
q
k=1
(xik − xjk)2,
where xik and xjk, k = 1, . . . , q are the variable values for units i and j,
respectively. Euclidean distance can be calculated using the dist() function
in R.
When the variables in a multivariate data set are on different scales, it
makes more sense to calculate the distances after some form of standardisa-
tion. Here we shall illustrate this on the body measurement data and divide
each variable by its standard deviation using the function scale() before
applying the dist() function–the necessary R code and output are
R> dist(scale(measure[, c("chest", "waist", "hips")],
+ center = FALSE))
1 2 3 4 5 6 7 8 9 10 11
2 0.17
3 0.15 0.08
4 0.22 0.07 0.14
5 0.11 0.15 0.09 0.22
6 0.29 0.16 0.16 0.19 0.21
7 0.32 0.16 0.20 0.13 0.28 0.14
8 0.23 0.11 0.11 0.12 0.19 0.16 0.13
9 0.21 0.10 0.06 0.16 0.12 0.11 0.17 0.09
10 0.27 0.12 0.13 0.14 0.20 0.06 0.09 0.11 0.09
11 0.23 0.28 0.22 0.33 0.19 0.34 0.38 0.25 0.24 0.32
12 0.22 0.24 0.18 0.28 0.18 0.30 0.32 0.20 0.20 0.28 0.06
...
(Note that only the distances for the first 12 observations are shown in the
output.)
1.6 The multivariate normal density function
Just as the normal distribution dominates univariate techniques, the multi-
variate normal distribution plays an important role in some multivariate pro-
cedures, although as mentioned earlier many multivariate analyses are carried
out in the spirit of data exploration where questions of statistical significance
are of relatively minor importance or of no importance at all. Nevertheless, re-
searchers dealing with the complexities of multivariate data may, on occasion,
need to know a little about the multivariate density function and in particular
how to assess whether or not a set of multivariate data can be assumed to
have this density function. So we will define the multivariate normal density
and describe some of its properties.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-31-320.jpg)
![1.6 The multivariate normal density function 19
really too little information to come to any convincing conclusion. Figure 1.3
shows separate probability plots for each measurement; there appears to be
no evidence of any departures from linearity. The chi-square plot of the 20
generalised distances in Figure 1.4 does seem to deviate a little from linearity,
but with so few observations it is hard to be certain. The plot is set up as
follows. We first extract the relevant data
R> x <- measure[, c("chest", "waist", "hips")]
and estimate the means of all three variables (i.e., for each column of the data)
and the covariance matrix
R> cm <- colMeans(x)
R> S <- cov(x)
The differences di have to be computed for all units in our data, so we iterate
over the rows of x using the apply() function with argument MARGIN = 1
and, for each row, compute the distance di:
R> d <- apply(x, MARGIN = 1, function(x)
+ t(x - cm) %*% solve(S) %*% (x - cm))
The sorted distances can now be plotted against the appropriate quantiles of
the χ2
3 distribution obtained from qchisq(); see Figure 1.4.
R> qqnorm(measure[,"chest"], main = "chest"); qqline(measure[,"chest"])
R> qqnorm(measure[,"waist"], main = "waist"); qqline(measure[,"waist"])
R> qqnorm(measure[,"hips"], main = "hips"); qqline(measure[,"hips"])
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
−2 0 1 2
3436384042
chest
Theoretical Quantiles
SampleQuantiles
q
q
q
q
q
q
q
qq
q
q
q
q
q
qq
q
q
q
q
−2 0 1 2
222426283032
waist
Theoretical Quantiles
SampleQuantiles
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
−2 0 1 2
323436384042
hips
Theoretical Quantiles
SampleQuantiles
Fig. 1.3. Normal probability plots of chest, waist, and hip measurements.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-35-320.jpg)
![20 1 Multivariate Data and Multivariate Analysis
R> plot(qchisq((1:nrow(x) - 1/2) / nrow(x), df = 3), sort(d),
+ xlab = expression(paste(chi[3]^2, " Quantile")),
+ ylab = "Ordered distances")
R> abline(a = 0, b = 1)
q
q
qqq
q
qqqq
q
q q q
q
q q
q q
q
0 2 4 6 8
2468
χ3
2
Quantile
Ordereddistances
Fig. 1.4. Chi-square plot of generalised distances for body measurements data.
We will now look at using the chi-square plot on a set of data introduced
early in the chapter, namely the air pollution in US cities (see Table 1.5). The
probability plots for each separate variable are shown in Figure 1.5. Here, we
also iterate over all variables, this time using a special function, sapply(),
that loops over the variable names:
R> layout(matrix(1:8, nc = 2))
R> sapply(colnames(USairpollution), function(x) {
+ qqnorm(USairpollution[[x]], main = x)
+ qqline(USairpollution[[x]])
+ })](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-36-320.jpg)
![22 1 Multivariate Data and Multivariate Analysis
The resulting seven plots are arranged on one page by a call to the layout
matrix; see Figure 1.5. The plots for SO2 concentration and precipitation
both deviate considerably from linearity, and the plots for manufacturing and
population show evidence of a number of outliers. But of more importance
is the chi-square plot for the data, which is given in Figure 1.6; the R code
is identical to the code used to produce the chi-square plot for the body
measurement data. In addition, the two most extreme points in the plot have
been labelled with the city names to which they correspond using text().
R> x <- USairpollution
R> cm <- colMeans(x)
R> S <- cov(x)
R> d <- apply(x, 1, function(x) t(x - cm) %*% solve(S) %*% (x - cm))
R> plot(qc <- qchisq((1:nrow(x) - 1/2) / nrow(x), df = 6),
+ sd <- sort(d),
+ xlab = expression(paste(chi[6]^2, " Quantile")),
+ ylab = "Ordered distances", xlim = range(qc) * c(1, 1.1))
R> oups <- which(rank(abs(qc - sd), ties = "random") > nrow(x) - 3)
R> text(qc[oups], sd[oups] - 1.5, names(oups))
R> abline(a = 0, b = 1)
qqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqq
q
q
q
q
q
q
5 10 15
510152025
χ6
2
Quantile
Ordereddistances
Providence
Phoenix
Chicago
Fig. 1.6. χ2
plot of generalised distances for USairpollution data.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-38-320.jpg)
![2.2 The scatterplot 31
the usual measures are employed when these may be distorted by the pres-
ence of the outliers in the data.) Full details of the construction are given in
Goldberg and Iglewicz (1992). The scatterplot of manu and popul including
the bivariate boxplot is shown in Figure 2.4. Figure 2.4 clearly tells us that
Chicago, Philadelphia, Detroit, and Cleveland should be regarded as outliers
but not Houston, because it is on the “fence” rather than outside the “fence”.
R> outcity <- match(lab <- c("Chicago", "Detroit",
+ "Cleveland", "Philadelphia"), rownames(USairpollution))
R> x <- USairpollution[, c("manu", "popul")]
R> bvbox(x, mtitle = "", xlab = mlab, ylab = plab)
R> text(x$manu[outcity], x$popul[outcity], labels = lab,
+ cex = 0.7, pos = c(2, 2, 4, 2, 2))
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
qq
q
q q
q
qqq
q
q q
q
q
q
q
q
q
q
q
q
0 1000 2000 3000
0100020003000
Manufacturing enterprises with 20 or more workers
Populationsize(1970census)inthousands
Chicago
Detroit
Cleveland
Philadelphia
Fig. 2.4. Scatterplot of manu and popul showing the bivariate boxplot of the data.
Suppose now that we are interested in calculating the correlation between
manu and popul. Researchers often calculate the correlation between two vari-](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-47-320.jpg)
![32 2 Looking at Multivariate Data: Visualisation
ables without first looking at the scatterplot of the two variables. But scat-
terplots should always be consulted when calculating correlation coefficients
because the presence of outliers can on occasion considerably distort the value
of a correlation coefficient, and as we have seen above, a scatterplot may help
to identify the offending observations particularly if used in conjunction with
a bivariate boxplot. The observations identified as outliers may then be ex-
cluded from the calculation of the correlation coefficient. With the help of
the bivariate boxplot in Figure 2.4, we have identified Chicago, Philadelphia,
Detroit, and Cleveland as outliers in the scatterplot of manu and popul. The
R code for finding the two correlations is
R> with(USairpollution, cor(manu, popul))
[1] 0.9553
R> outcity <- match(c("Chicago", "Detroit",
+ "Cleveland", "Philadelphia"),
+ rownames(USairpollution))
R> with(USairpollution, cor(manu[-outcity], popul[-outcity]))
[1] 0.7956
The match() function identifies rows of the data frame USairpollution cor-
responding to the cities of interest, and the subset starting with a minus sign
removes these units before the correlation is computed. Calculation of the cor-
relation coefficient between the two variables using all the data gives a value
of 0.96, which reduces to a value of 0.8 after excluding the four outliers–a not
inconsiderable reduction.
2.2.2 The convex hull of bivariate data
An alternative approach to using the scatterplot combined with the bivariate
boxplot to deal with the possible problem of calculating correlation coeffi-
cients without the distortion often caused by outliers in the data is convex
hull trimming, which allows robust estimation of the correlation. The convex
hull of a set of bivariate observations consists of the vertices of the smallest
convex polyhedron in variable space within which or on which all data points
lie. Removal of the points lying on the convex hull can eliminate isolated
outliers without disturbing the general shape of the bivariate distribution. A
robust estimate of the correlation coefficient results from using the remaining
observations. Let’s see how the convex hull approach works with our manu
and popul scatterplot. We first find the convex hull of the data (i.e., the
observations defining the convex hull) using the following R code:
R> (hull <- with(USairpollution, chull(manu, popul)))
[1] 9 15 41 6 2 18 16 14 7](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-48-320.jpg)
![2.2 The scatterplot 33
R> with(USairpollution,
+ plot(manu, popul, pch = 1, xlab = mlab, ylab = plab))
R> with(USairpollution,
+ polygon(manu[hull], popul[hull], density = 15, angle = 30))
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
qq
q
q q
q
qqq
q
q q
q
q
q
q
q
q
q
q
q
0 500 1000 2000 3000
0500150025003500
Manufacturing enterprises with 20 or more workers
Populationsize(1970census)inthousands
Fig. 2.5. Scatterplot of manu against popul showing the convex hull of the data.
Now we can show this convex hull on a scatterplot of the variables using the
code attached to the resulting Figure 2.5.
To calculate the correlation coefficient after removal of the points defining
the convex hull requires the code
R> with(USairpollution, cor(manu[-hull],popul[-hull]))
[1] 0.9225
The resulting value of the correlation is now 0.923 and thus is higher compared
with the correlation estimated after removal of the outliers identified by using
the bivariate boxplot, namely Chicago, Philadelphia, Detroit, and Cleveland.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-49-320.jpg)
![2.3 The bubble and other glyph plots 37
R> plot(wind ~ temp, data = USairpollution,
+ xlab = "Average annual temperature (Fahrenheit)",
+ ylab = "Average annual wind speed (m.p.h.)", pch = 10,
+ ylim = ylim)
R> with(USairpollution,
+ stars(USairpollution[,-c(2,5)], locations = cbind(temp, wind),
+ labels = NULL, add = TRUE, cex = 0.5))
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q q
q
q
q
45 50 55 60 65 70 75
6789101112
Average annual temperature (Fahrenheit)
Averageannualwindspeed(m.p.h.)
Fig. 2.8. Scatterplot of temp and wind showing five-sided stars representing the
other variables.
In fact, both the bubble plot and “stars” plot are examples of symbol or
glyph plots, in which data values control the symbol parameters. For example,
a circle is a glyph where the values of one variable in a multivariate observation
control the circle size. In Figure 2.8, the spatial positions of the cities in
the scatterplot of temp and wind are combined with a star representation
of the five other variables. An alternative is simply to represent the seven
variables for each city by a seven-sided star and arrange the resulting stars in](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-53-320.jpg)
![52 2 Looking at Multivariate Data: Visualisation
R> library("scatterplot3d")
R> with(measure, scatterplot3d(chest, waist, hips,
+ pch = (1:2)[gender], type = "h", angle = 55))
32 34 36 38 40 42 44
323436384042
22
24
26
28
30
32
34
chest
waist
hips
q
q
q
q
q
q
q
q
q
q
Fig. 2.18. A three-dimensional scatterplot for the body measurements data with
points corresponding to male and triangles to female measurements.
the data? Probably not, as there are few points in each of the three, three-
dimensional displays. This is often a problem with multipanel plots when the
sample size is not large.
For the last example in this section, we will use a larger data set, namely
data on earthquakes given in Sarkar (2008). The data consist of recordings of
the location (latitude, longitude, and depth) and magnitude of 1000 seismic
events around Fiji since 1964.
In Figure 2.22, scatterplots of latitude and longitude are plotted for three
ranges of depth. The distribution of locations in the latitude-longitude space
is seen to be different in the three panels, particularly for very deep quakes. In](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-68-320.jpg)
![54 2 Looking at Multivariate Data: Visualisation
R> plot(xyplot(SO2 ~ temp| cut(wind, 2), data = USairpollution))
temp
SO2
20
40
60
80
100
50 60 70
q
q
q
q
q
q
q
q
qq
q
q q
q
q q
q
q
q
q
q
(5.99,9.35]
50 60 70
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
(9.35,12.7]
Fig. 2.20. Scatterplot of SO2 and temp for light and high winds.
liers. Like the chi-square plot for assessing multivariate normality, described
in Chapter 1, the stalactite plot is based on the generalised distances of ob-
servations from the multivariate mean of the data. But here these distances
are calculated from the means and covariances estimated from increasing-
sized subsets of the data. As mentioned previously when describing bivariate
boxplots, the aim is to reduce the masking effects that can arise due to the
influence of outliers on the estimates of means and covariances obtained from
all the data. The central idea of this approach is that, given distances using,
say, m observations for estimation of means and covariances, the m+ 1 obser-
vations to be used for this estimation in the next stage are chosen to be those
with the m+1 smallest distances. Thus an observation can be included in the
subset used for estimation for some value of m but can later be excluded as m
increases. Initially m is chosen to take the value q + 1, where q is the number
of variables in the multivariate data set because this is the smallest number](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-70-320.jpg)
![56 2 Looking at Multivariate Data: Visualisation
R> plot(xyplot(lat ~ long| cut(depth, 3), data = quakes,
+ layout = c(3, 1), xlab = "Longitude",
+ ylab = "Latitude"))
Longitude
Latitude
−35
−30
−25
−20
−15
−10
165170175180185
q
q
q
q
q
q
q
q
q
q
q
q q
qq
qq
q
q
q
qq
q
q
qq
q
q
q
q
q
q
qqq
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qqqqqqqq
qqq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
qqqq
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
qq
q
q
q
q
q
qq
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qqq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
qqqq
q
qq
q
q
q
q
q
q
qqq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
qq
q
qq
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
qq
qq
q
qq
q
q
q
q
q
q
q
(39.4,253]
165170175180185
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q qq
q
qq
q q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
qq
qq
q
q
q
q
q
q
q q
q
q
(253,467]
165170175180185
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qqq
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
qq
q
q
q
qq
q
q
q
q
q
qq
q
q
qq
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
qq
q
q
q
qq
qqq
q
q
q
q
q
q
q
q
qq
q
qq
q
q
q
q
q
q
q
q
qq
q
q
q
q
qq
q
qq
q
q
q
q
qqqq
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
qq
qq
qq
q
q
q
q
q
qqq
q q
qq
qqq
q
qq
q
q
qq
qq
q
q
qqqq
q
qqq
q
qq
q
qq
q
q
q
q
qq
qq
q
q
qqq q
q
q
q
q
q
q
q
qq
q
q
q
qq
qq
q
q
q
q
qqqq
q
q
q
q
q
q
q
qq
q
q
q
qq
q
q
q
q
qqq
q
q
q
qq
q
q
q
q
q
qq
q
qq
q
q
qq
qq
q
q
q
q
q
q
q
q
q
qq
q
q
q
qqq
q
qq
q
q
q
q
q
q
qqq
q
q
q
q
q
q
q
q
q
q
(467,681]
Fig. 2.22. Scatterplots of latitude and longitude conditioned on three ranges of
depth.
nearly all values of m. The effect of masking is also clear; when all 41 ob-
servations are used to calculate the generalised distances, only observations
Chicago, Phoenix, and Providence are indicated to be outliers.
2.9 Summary
Plotting multivariate data is an essential first step in trying to understand
the story they may have to tell. The methods covered in this chapter provide
just some basic ideas for taking an initial look at the data, and with software
such as R there are many other possibilities for graphing multivariate obser-](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-72-320.jpg)
![66 3 Principal Components Analysis
1. Multiply the variables by 16, 12, and 1/10, respectively and then carry
out a principal components analysis on the covariance matrix of the three
variables.
2. Carry out a principal components analysis on the covariance matrix of the
original variables and then multiply the elements of the relevant compo-
nent by 16, 12, and 1/10.
Unfortunately, these two procedures do not generally lead to the same result.
So if we imagine a set of multivariate data where the variables are of com-
pletely different types, for example length, temperature, blood pressure, or
anxiety rating, then the structure of the principal components derived from
the covariance matrix will depend upon the essentially arbitrary choice of units
of measurement; for example, changing the length from centimetres to inches
will alter the derived components. Additionally, if there are large differences
between the variances of the original variables, then those whose variances
are largest will tend to dominate the early components. Principal components
should only be extracted from the sample covariance matrix when all the
original variables have roughly the same scale. But this is rare in practise and
consequently, in practise, principal components are extracted from the correla-
tion matrix of the variables, R. Extracting the components as the eigenvectors
of R is equivalent to calculating the principal components from the original
variables after each has been standardised to have unit variance. It should be
noted, however, that there is rarely any simple correspondence between the
components derived from S and those derived from R. And choosing to work
with R rather than with S involves a definite but possibly arbitrary decision
to make variables “equally important”.
To demonstrate how the principal components of the covariance matrix of a
data set can differ from the components extracted from the data’s correlation
matrix, we will use the example given in Jolliffe (2002). The data in this
example consist of eight blood chemistry variables measured on 72 patients in
a clinical trial. The correlation matrix of the data, together with the standard
deviations of each of the eight variables, is
R> blood_corr
[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8]
[1,] 1.000 0.290 0.202 -0.055 -0.105 -0.252 -0.229 0.058
[2,] 0.290 1.000 0.415 0.285 -0.376 -0.349 -0.164 -0.129
[3,] 0.202 0.415 1.000 0.419 -0.521 -0.441 -0.145 -0.076
[4,] -0.055 0.285 0.419 1.000 -0.877 -0.076 0.023 -0.131
[5,] -0.105 -0.376 -0.521 -0.877 1.000 0.206 0.034 0.151
[6,] -0.252 -0.349 -0.441 -0.076 0.206 1.000 0.192 0.077
[7,] -0.229 -0.164 -0.145 0.023 0.034 0.192 1.000 0.423
[8,] 0.058 -0.129 -0.076 -0.131 0.151 0.077 0.423 1.000](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-82-320.jpg)
![68 3 Principal Components Analysis
Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Comp.6 Comp.7 Comp.8
[1,] -0.194 0.417 0.400 0.652 0.175 -0.363 0.176 0.102
[2,] -0.400 0.154 0.168 -0.848 0.230 -0.110
[3,] -0.459 0.168 -0.274 0.251 0.403 0.677
[4,] -0.430 -0.472 -0.171 0.169 0.118 -0.237 0.678
[5,] 0.494 0.360 -0.180 -0.139 0.136 0.157 0.724
[6,] 0.319 -0.320 -0.277 0.633 -0.162 0.384 0.377
[7,] 0.177 -0.535 0.410 -0.163 -0.299 -0.513 0.367
[8,] 0.171 -0.245 0.709 0.198 0.469 -0.376
(The “blanks” in this output represent very small values.) Examining the re-
sults, we see that each of the principal components of the covariance matrix
is largely dominated by a single variable, whereas those for the correlation
matrix have moderate-sized coefficients on several of the variables. And the
first component from the covariance matrix accounts for almost 99% of the
total variance of the observed variables. The components of the covariance
matrix are completely dominated by the fact that the variance of the plate
variable is roughly 400 times larger than the variance of any of the seven
other variables. Consequently, the principal components from the covariance
matrix simply reflect the order of the sizes of the variances of the observed
variables. The results from the correlation matrix tell us, in particular, that a
weighted contrast of the first four and last four variables is the linear function
with the largest variance. This example illustrates that when variables are on
very different scales or have very different variances, a principal components
analysis of the data should be performed on the correlation matrix, not on
the covariance matrix.
3.5 Principal components of bivariate data with
correlation coefficient r
Before we move on to look at some practical examples of the application of
principal components analysis, it will be helpful to look in a little more detail
at the mathematics of the method in one very simple case. We will do this
in this section for bivariate data where the two variables, x1 and x2, have
correlation coefficient r. The sample correlation matrix in this case is simply
R =
1.0 r
r 0.1
.
In order to find the principal components of the data we need to find the
eigenvalues and eigenvectors of R. The eigenvalues are found as the roots of
the equation
det(R − λI) = 0.
This leads to the quadratic equation in λ](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-84-320.jpg)
![3.5 Principal components of bivariate data with correlation coefficient r 69
(1 − λ)2
− r2
= 0,
and solving this equation leads to eigenvalues λ1 = 1 + r, λ2 = 1 − r. Note
that the sum of the eigenvalues is two, equal to trace(R). The eigenvector
corresponding to λ1 is obtained by solving the equation
Ra1 = λ1a1.
This leads to the equations
a11 + ra12 = (1 + r)a11,
ra11 + a12 = (1 + r)a12.
The two equations are identical, and both reduce to requiring a11 = a12. If
we now introduce the normalisation constraint a1 a1 = 1, we find that
a11 = a12 =
1
√
2
.
Similarly, we find the second eigenvector is given by a21 = 1√
2
and a22 = − 1√
2
.
The two principal components are then given by
y1 =
1
√
2
(x1 + x2), y2 =
1
√
2
(x1 − x2).
We can calculate the sample variance of the first principal component as
Var(y1) = Var
1
√
2
(x1 + x2) =
1
2
Var(x1 + x2)
=
1
2
[Var(x1) + Var(x2) + 2Cov(x1, x2)]
=
1
2
(1 + 1 + 2r) = 1 + r.
Similarly, the variance of the second principal component is 1 − r.
Notice that if r < 0, the order of the eigenvalues and hence of the principal
components is reversed; if r = 0, the eigenvalues are both equal to 1 and any
two solutions at right angles could be chosen to represent the two components.
Two further points should be noted:
1. There is an arbitrary sign in the choice of the elements of ai. It is custom-
ary (but not universal) to choose ai1 to be positive.
2. The coefficients that define the two components do not depend on r, al-
though the proportion of variance explained by each does change with r.
As r tends to 1, the proportion of variance accounted for by y1, namely
(1+r)/2, also tends to one. When r = 1, the points all align on a straight
line and the variation in the data is unidimensional.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-85-320.jpg)
![3.10 Some examples of the application of principal components analysis 75
R> head_dat <- headsize[, c("head1", "head2")]
R> colMeans(head_dat)
head1 head2
185.7 183.8
R> cov(head_dat)
head1 head2
head1 95.29 69.66
head2 69.66 100.81
The principal components of these data, extracted from their covariance ma-
trix, can be found using
R> head_pca <- princomp(x = head_dat)
R> head_pca
Call:
princomp(x = head_dat)
Standard deviations:
Comp.1 Comp.2
12.691 5.215
2 variables and 25 observations.
R> print(summary(head_pca), loadings = TRUE)
Importance of components:
Comp.1 Comp.2
Standard deviation 12.6908 5.2154
Proportion of Variance 0.8555 0.1445
Cumulative Proportion 0.8555 1.0000
Loadings:
Comp.1 Comp.2
head1 0.693 -0.721
head2 0.721 0.693
and are
y1 = 0.693x1 + 0.721x2 y2 = −0.721x1 + 0.693x2
with variances 167.77 and 28.33. The first principal component accounts for a
proportion 167.77/(167.77+28.33) = 0.86 of the total variance in the original
variables. Note that the total variance of the principal components is 196.10,
which as expected is equal to the total variance of the original variables,
found by adding the relevant terms in the covariance matrix given earlier; i.e.,
95.29 + 100.81 = 196.10.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-91-320.jpg)
![78 3 Principal Components Analysis
R> xlim <- range(head_pca$scores[,1])
R> plot(head_pca$scores, xlim = xlim, ylim = xlim)
q
qq q
q
q
q
q
q
q
q
q
q
q
qq
qq
q
q
q
q
q
q
q
−30 −20 −10 0 10 20
−30−20−1001020
Comp.1
Comp.2
Fig. 3.3. Plot of the first two principal component scores for the head size data.
3.10.2 Olympic heptathlon results
The pentathlon for women was first held in Germany in 1928. Initially this
consisted of the shot put, long jump, 100 m, high jump, and javelin events,
held over two days. In the 1964 Olympic Games, the pentathlon became the
first combined Olympic event for women, consisting now of the 80 m hurdles,
shot, high jump, long jump, and 200 m. In 1977, the 200 m was replaced by
the 800 m run, and from 1981 the IAAF brought in the seven-event heptathlon
in place of the pentathlon, with day one containing the events 100 m hurdles,
shot, high jump, and 200 m run, and day two the long jump, javelin, and
800 m run. A scoring system is used to assign points to the results from each
event, and the winner is the woman who accumulates the most points over
the two days. The event made its first Olympic appearance in 1984.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-94-320.jpg)
![80 3 Principal Components Analysis
R> score <- which(colnames(heptathlon) == "score")
R> round(cor(heptathlon[,-score]), 2)
hurdles highjump shot run200m longjump javelin run800m
hurdles 1.00 0.81 0.65 0.77 0.91 0.01 0.78
highjump 0.81 1.00 0.44 0.49 0.78 0.00 0.59
shot 0.65 0.44 1.00 0.68 0.74 0.27 0.42
run200m 0.77 0.49 0.68 1.00 0.82 0.33 0.62
longjump 0.91 0.78 0.74 0.82 1.00 0.07 0.70
javelin 0.01 0.00 0.27 0.33 0.07 1.00 -0.02
run800m 0.78 0.59 0.42 0.62 0.70 -0.02 1.00
R> plot(heptathlon[,-score])
The scatterplot matrix appears in Figure 3.4.
hurdles
1.50 1.85 0 2 4 36 44
02
1.501.85
highjump
shot
1014
024
run200m
longjump
5.07.0
3644
javelin
0 2 10 14 5.0 7.0 0 20
020
run800m
Fig. 3.4. Scatterplot matrix of the seven heptathlon events after transforming some
variables so that for all events large values are indicative of a better performance.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-96-320.jpg)
![3.10 Some examples of the application of principal components analysis 81
Examination of the correlation matrix shows that most pairs of events are
positively correlated, some moderately (for example, high jump and shot) and
others relatively highly (for example, high jump and hurdles). The exceptions
to this general observation are the relationships between the javelin event and
the others, where almost all the correlations are close to zero. One explanation
might be that the javelin is a very “technical” event and perhaps the training
for the other events does not help the competitors in the javelin. But before
we speculate further, we should look at the scatterplot matrix of the seven
events shown in Figure 3.4. One very clear observation in this plot is that for
all events except the javelin there is an outlier who is very much poorer than
the other athletes at these six events, and this is the competitor from Papua
New Guinea (PNG), who finished last in the competition in terms of points
scored. But surprisingly, in the scatterplots involving the javelin, it is this
competitor who again stands out, but in this case she has the third highest
value for the event. It might be sensible to look again at both the correlation
matrix and the scatterplot matrix after removing the competitor from PNG;
the relevant R code is
R> heptathlon <- heptathlon[-grep("PNG", rownames(heptathlon)),]
R> score <- which(colnames(heptathlon) == "score")
R> round(cor(heptathlon[,-score]), 2)
hurdles highjump shot run200m longjump javelin run800m
hurdles 1.00 0.58 0.77 0.83 0.89 0.33 0.56
highjump 0.58 1.00 0.46 0.39 0.66 0.35 0.15
shot 0.77 0.46 1.00 0.67 0.78 0.34 0.41
run200m 0.83 0.39 0.67 1.00 0.81 0.47 0.57
longjump 0.89 0.66 0.78 0.81 1.00 0.29 0.52
javelin 0.33 0.35 0.34 0.47 0.29 1.00 0.26
run800m 0.56 0.15 0.41 0.57 0.52 0.26 1.00
The new scatterplot matrix is shown in Figure 3.5. Several of the cor-
relations are changed to some degree from those shown before removal of
the PNG competitor, particularly the correlations involving the javelin event,
where the very small correlations between performances in this event and the
others have increased considerably. Given the relatively large overall change
in the correlation matrix produced by omitting the PNG competitor, we shall
extract the principal components of the data from the correlation matrix after
this omission. The principal components can now be found using
R> heptathlon_pca <- prcomp(heptathlon[, -score], scale = TRUE)
R> print(heptathlon_pca)
Standard deviations:
[1] 2.08 0.95 0.91 0.68 0.55 0.34 0.26
Rotation:](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-97-320.jpg)
![82 3 Principal Components Analysis
R> plot(heptathlon[,-score], pch = ".", cex = 1.5)
hurdles
1.70 0 2 4 36 44
1.53.0
1.70
highjump
shot
1014
024
run200m
longjump
5.57.0
3644
javelin
1.5 3.0 10 14 5.5 7.0 20 35 2035
run800m
Fig. 3.5. Scatterplot matrix for the heptathlon data after removing observations
of the PNG competitor.
PC1 PC2 PC3 PC4 PC5 PC6 PC7
hurdles -0.45 0.058 -0.17 0.048 -0.199 0.847 -0.070
highjump -0.31 -0.651 -0.21 -0.557 0.071 -0.090 0.332
shot -0.40 -0.022 -0.15 0.548 0.672 -0.099 0.229
run200m -0.43 0.185 0.13 0.231 -0.618 -0.333 0.470
longjump -0.45 -0.025 -0.27 -0.015 -0.122 -0.383 -0.749
javelin -0.24 -0.326 0.88 0.060 0.079 0.072 -0.211
run800m -0.30 0.657 0.19 -0.574 0.319 -0.052 0.077
The summary method can be used for further inspection of the details:
R> summary(heptathlon_pca)
Importance of components:
PC1 PC2 PC3 PC4 PC5 PC6](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-98-320.jpg)
![3.10 Some examples of the application of principal components analysis 83
Standard deviation 2.079 0.948 0.911 0.6832 0.5462 0.3375
Proportion of Variance 0.618 0.128 0.119 0.0667 0.0426 0.0163
Cumulative Proportion 0.618 0.746 0.865 0.9313 0.9739 0.9902
PC7
Standard deviation 0.26204
Proportion of Variance 0.00981
Cumulative Proportion 1.00000
The linear combination for the first principal component is 2
R> a1 <- heptathlon_pca$rotation[,1]
R> a1
hurdles highjump shot run200m longjump javelin run800m
-0.4504 -0.3145 -0.4025 -0.4271 -0.4510 -0.2423 -0.3029
We see that the hurdles and long jump events receive the highest weight but
the javelin result is less important. For computing the first principal compo-
nent, the data need to be rescaled appropriately. The center and the scaling
used by prcomp internally can be extracted from the heptathlon_pca via
R> center <- heptathlon_pca$center
R> scale <- heptathlon_pca$scale
Now, we can apply the scale function to the data and multiply it with the
loadings matrix in order to compute the first principal component score for
each competitor
R> hm <- as.matrix(heptathlon[,-score])
R> drop(scale(hm, center = center, scale = scale) %*%
+ heptathlon_pca$rotation[,1])
Joyner-Kersee (USA) John (GDR) Behmer (GDR)
-4.757530 -3.147943 -2.926185
Sablovskaite (URS) Choubenkova (URS) Schulz (GDR)
-1.288136 -1.503451 -0.958467
Fleming (AUS) Greiner (USA) Lajbnerova (CZE)
-0.953445 -0.633239 -0.381572
Bouraga (URS) Wijnsma (HOL) Dimitrova (BUL)
-0.522322 -0.217701 -1.075984
Scheider (SWI) Braun (FRG) Ruotsalainen (FIN)
0.003015 0.109184 0.208868
Yuping (CHN) Hagger (GB) Brown (USA)
0.232507 0.659520 0.756855
Mulliner (GB) Hautenauve (BEL) Kytola (FIN)
1.880933 1.828170 2.118203
Geremias (BRA) Hui-Ing (TAI) Jeong-Mi (KOR)
2.770706 3.901167 3.896848](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-99-320.jpg)
![84 3 Principal Components Analysis
or, more conveniently, by extracting the first from all pre-computed principal
components:
R> predict(heptathlon_pca)[,1]
Joyner-Kersee (USA) John (GDR) Behmer (GDR)
-4.757530 -3.147943 -2.926185
Sablovskaite (URS) Choubenkova (URS) Schulz (GDR)
-1.288136 -1.503451 -0.958467
Fleming (AUS) Greiner (USA) Lajbnerova (CZE)
-0.953445 -0.633239 -0.381572
Bouraga (URS) Wijnsma (HOL) Dimitrova (BUL)
-0.522322 -0.217701 -1.075984
Scheider (SWI) Braun (FRG) Ruotsalainen (FIN)
0.003015 0.109184 0.208868
Yuping (CHN) Hagger (GB) Brown (USA)
0.232507 0.659520 0.756855
Mulliner (GB) Hautenauve (BEL) Kytola (FIN)
1.880933 1.828170 2.118203
Geremias (BRA) Hui-Ing (TAI) Jeong-Mi (KOR)
2.770706 3.901167 3.896848
The first two components account for 75% of the variance. A barplot of
each component’s variance (see Figure 3.6) shows how the first two compo-
nents dominate.
R> plot(heptathlon_pca)
Variances
01234
Fig. 3.6. Barplot of the variances explained by the principal components (with
observations for PNG removed).](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-100-320.jpg)
![3.10 Some examples of the application of principal components analysis 85
The correlation between the score given to each athlete by the standard
scoring system used for the heptathlon and the first principal component score
can be found from
R> cor(heptathlon$score, heptathlon_pca$x[,1])
[1] -0.9931
This implies that the first principal component is in good agreement with the
score assigned to the athletes by official Olympic rules; a scatterplot of the
official score and the first principal component is given in Figure 3.7. (The
fact that the correlation is negative is unimportant here because of the arbi-
trariness of the signs of the coefficients defining the first principal component;
it is the magnitude of the correlation that is important.)
R> plot(heptathlon$score, heptathlon_pca$x[,1])
q
q
q
q
q
qq
q
qq
q
q
qqqq
qq
qq
q
q
qq
5500 6000 6500 7000
−4−2024
heptathlon$score
heptathlon_pca$x[,1]
Fig. 3.7. Scatterplot of the score assigned to each athlete in 1988 and the first
principal component.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-101-320.jpg)
![86 3 Principal Components Analysis
3.10.3 Air pollution in US cities
In this subsection, we will return to the air pollution data introduced in Chap-
ter 1. The data were originally collected to investigate the determinants of pol-
lution, presumably by regressing SO2 on the six other variables. Here, however,
we shall examine how principal components analysis can be used to explore
various aspects of the data, and will then look at how such an analysis can
also be used to address the determinants of pollution question.
To begin we shall ignore the SO2 variable and concentrate on the others,
two of which relate to human ecology (popul, manu) and four to climate (temp,
Wind, precip, predays). A case can be made to use negative temperature
values in subsequent analyses since then all six variables are such that high
values represent a less attractive environment. This is, of course, a personal
view, but as we shall see later, the simple transformation of temp does aid
interpretation.
Prior to undertaking the principal components analysis on the air pollution
data, we will again construct a scatterplot matrix of the six variables, but here
we include the histograms for each variable on the main diagonal. The diagram
that results is shown in Figure 3.8.
A clear message from Figure 3.8 is that there is at least one city, and
probably more than one, that should be considered an outlier. (This should
come as no surprise given the investigation of the data in Chapter 2.) On the
manu variable, for example, Chicago, with a value of 3344, has about twice as
many manufacturing enterprises employing 20 or more workers as the city with
the second highest number (Philadelphia). We shall return to this potential
problem later in the chapter, but for the moment we shall carry on with a
principal components analysis of the data for all 41 cities.
For the data in Table 1.5, it seems necessary to extract the principal com-
ponents from the correlation rather than the covariance matrix, since the six
variables to be used are on very different scales. The correlation matrix and
the principal components of the data can be obtained in R using the following
command line code:
R> cor(USairpollution[,-1])
manu popul wind precip predays negtemp
manu 1.00000 0.95527 0.23795 -0.03242 0.13183 0.19004
popul 0.95527 1.00000 0.21264 -0.02612 0.04208 0.06268
wind 0.23795 0.21264 1.00000 -0.01299 0.16411 0.34974
precip -0.03242 -0.02612 -0.01299 1.00000 0.49610 -0.38625
predays 0.13183 0.04208 0.16411 0.49610 1.00000 0.43024
negtemp 0.19004 0.06268 0.34974 -0.38625 0.43024 1.00000
R> usair_pca <- princomp(USairpollution[,-1], cor = TRUE)](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-102-320.jpg)
![3.10 Some examples of the application of principal components analysis 87
R> data("USairpollution", package = "HSAUR2")
R> panel.hist <- function(x, ...) {
+ usr <- par("usr"); on.exit(par(usr))
+ par(usr = c(usr[1:2], 0, 1.5) )
+ h <- hist(x, plot = FALSE)
+ breaks <- h$breaks; nB <- length(breaks)
+ y <- h$counts; y <- y/max(y)
+ rect(breaks[-nB], 0, breaks[-1], y, col="grey", ...)
+ }
R> USairpollution$negtemp <- USairpollution$temp * (-1)
R> USairpollution$temp <- NULL
R> pairs(USairpollution[,-1], diag.panel = panel.hist,
+ pch = ".", cex = 1.5)
manu
0 2000 10 40 −75 −50
02000
02000
popul
wind
6912
1040
precip
predays
40120
0 2000
−75−50
6 9 12 40 120
negtemp
Fig. 3.8. Scatterplot matrix of six variables in the air pollution data.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-103-320.jpg)
![90 3 Principal Components Analysis
R> pairs(usair_pca$scores[,1:3], ylim = c(-6, 4), xlim = c(-6, 4),
+ panel = function(x,y, ...) {
+ text(x, y, abbreviate(row.names(USairpollution)),
+ cex = 0.6)
+ bvbox(cbind(x,y), add = TRUE)
+ })
Comp.1
−6 −2 0 2 4
Albn
Albq
Atln
Bltm
Bffl
Chrl
Chcg
Cncn
Clvl
Clmb DllsDnvrDsMn
Dtrt
Hrtf
HstnIndn
Jcks
KnsC
LttR
LsvlMmph
Miam
MlwkMnnp
Nshv
NwOr
Nrfl
Omah
Phld
Phnx
PttsPrvd
Rchm SlLC
SnFr
Sttl St.LWshnWcht
Wlmn
q
q
q
q
q
q
q
q
q
q q qq
q
q
qq
q
q
q
qq
q
qq
q
q
q
q
q
q
qq
q q
q
q qqq
q
−6−2024
Albn
Albq
Atln
Bltm
Bffl
Chrl
Chcg
Cncn
Clvl
ClmbDlls DnvrDsMn
Dtrt
Hrtf
Hstn Indn
Jcks
KnsC
LttR
LsvlMmph
Miam
MlwkMnnp
Nshv
NwOr
Nrfl
Omah
Phld
Phnx
PttsPrvd
Rchm SlLC
SnFr
SttlSt.LWshnWcht
Wlmn
q
q
q
q
q
q
q
q
q
qq qq
q
q
q q
q
q
q
qq
q
qq
q
q
q
q
q
q
qq
q q
q
qqq q
q
−6−2024
Albn
Albq
Atln
Bltm
Bffl
Chrl
Chcg
Cncn
Clvl
Clmb
Dlls
Dnvr
DsMnDtrt
Hrtf
Hstn
Indn
Jcks
KnsC
LttRLsvlMmph
Miam
MlwkMnnp
Nshv
NwOrNrfl
OmahPhld
Phnx
PttsPrvd
Rchm
SlLC
SnFr
Sttl
St.LWshn
Wcht
Wlmnq
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
qqq
q
qq q
qq
qq
q
qq
q
q
q
q
qq
q
q
Comp.2 Albn
Albq
Atln
Bltm
Bffl
Chrl
Chcg
Cncn
Clvl
Clmb
Dlls
Dnvr
DsMnDtrt
Hrtf
Hstn
Indn
Jcks
KnsC
LttRLsvlMmph
Miam
MlwkMnnp
Nshv
NwOr Nrfl
OmahPhld
Phnx
PttsPrvd
Rchm
SlLC
SnFr
Sttl
St.LWshn
Wcht
Wlmnq
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
qqq
q
qqq
q q
qq
q
qq
q
q
q
q
qq
q
q
−6 −2 0 2 4
AlbnAlbq
Atln
Bltm
Bffl
Chrl
Chcg
Cncn
Clvl Clmb
Dlls
Dnvr
DsMn
Dtrt
Hrtf
Hstn
Indn
Jcks
KnsC
LttR
Lsvl
Mmph
Miam
MlwkMnnp
Nshv
NwOr
Nrfl
Omah
Phld
Phnx
PttsPrvd
Rchm
SlLC
SnFr
Sttl
St.LWshn
Wcht
Wlmn
q q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
qq
q
q
q
q
qq
q
q
Albn Albq
Atln
Bltm
Bffl
Chrl
Chcg
Cncn
ClvlClmb
Dlls
Dnvr
DsMn
Dtrt
Hrtf
Hstn
Indn
Jcks
KnsC
LttR
Lsvl
Mmph
Miam
MlwkMnnp
Nshv
NwOr
Nrfl
Omah
Phld
Phnx
PttsPrvd
Rchm
SlLC
SnFr
Sttl
St.LWshn
Wcht
Wlmn
q q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
qq
q
q
q
q
qq
q
q
−6 −2 0 2 4
−6−2024Comp.3
Fig. 3.9. Bivariate boxplots of the first three principal components.
not necessarily correct and, Jolliffe (2002) gives an example where there are
13 original explanatory variables and therefore 13 principal components to
consider, of which only the first four have variances greater than one. But
using all the principal components as explanatory variables shows that, for
example, component 12, with variance 0.04, is a significant predictor of the
response. So, bearing this example in mind, we will regress the SO2 variables
on all six principal components; the necessary R code is given on page 92.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-106-320.jpg)
![3.10 Some examples of the application of principal components analysis 91
R> out <- sapply(1:6, function(i) {
+ plot(USairpollution$SO2,usair_pca$scores[,i],
+ xlab = paste("PC", i, sep = ""),
+ ylab = "Sulphur dioxide concentration")
+ })
q
q
q
q
q
q
q
q
q
qq qq
q
q
q q
q
q
q
qq
q
q q
q
q
q
q
q
q
q q
qqq
q qqq
q
20 40 60 80 100
−6−22
PC1
Sulphurdioxideconcentration
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q qq
q
q q
q
q q
q q
q
q q
q
q
q
q
q
q
q
q
20 40 60 80 100
−2024
PC2
Sulphurdioxideconcentration
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q q
q
q
q
q
qq
q
q
20 40 60 80 100
−201
PC3
Sulphurdioxideconcentration
q
q q q
q
q
q
q
q
q
q
q
q
q q
q
q
qq
q
q
q
q q
q
q
q
qq
q
q
q
q
q
q
q
q
qq
q
q
20 40 60 80 100
−2012
PC4
Sulphurdioxideconcentration
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
qq
q
20 40 60 80 100
−0.60.00.6
PC5
Sulphurdioxideconcentration
q
q q
q
q q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
qq q
q
q
q
q
q
q q
q
q
q
q
20 40 60 80 100
−0.40.00.4
PC6
Sulphurdioxideconcentration
Fig. 3.10. Sulphur dioxide concentration depending on principal components.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-107-320.jpg)
![3.13 Canonical correlation analysis 97
R> headsize.std <- sweep(headsize, 2,
+ apply(headsize, 2, sd), FUN = "/")
R> R <- cor(headsize.std)
R> r11 <- R[1:2, 1:2]
R> r22 <- R[-(1:2), -(1:2)]
R> r12 <- R[1:2, -(1:2)]
R> r21 <- R[-(1:2), 1:2]
R> (E1 <- solve(r11) %*% r12 %*% solve(r22) %*%r21)
head1 breadth1
head1 0.3225 0.3168
breadth1 0.3019 0.3021
R> (E2 <- solve(r22) %*% r21 %*% solve(r11) %*%r12)
head2 breadth2
head2 0.3014 0.3002
breadth2 0.3185 0.3232
R> (e1 <- eigen(E1))
$values
[1] 0.621745 0.002888
$vectors
[,1] [,2]
[1,] 0.7270 -0.7040
[2,] 0.6866 0.7102
R> (e2 <- eigen(E2))
$values
[1] 0.621745 0.002888
$vectors
[,1] [,2]
[1,] -0.6838 -0.7091
[2,] -0.7297 0.7051
Here the four linear functions are found to be
u1 = +0.73x1 + 0.69x2,
u2 = −0.70x1 + 0.71x2,
v1 = −0.68x3 − 0.73x4,
v2 = −0.71x3 + 0.71x4.
R> girth1 <- headsize.std[,1:2] %*% e1$vectors[,1]
R> girth2 <- headsize.std[,3:4] %*% e2$vectors[,1]](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-113-320.jpg)
![98 3 Principal Components Analysis
R> shape1 <- headsize.std[,1:2] %*% e1$vectors[,2]
R> shape2 <- headsize.std[,3:4] %*% e2$vectors[,2]
R> (g <- cor(girth1, girth2))
[,1]
[1,] -0.7885
R> (s <- cor(shape1, shape2))
[,1]
[1,] 0.05374
R> plot(girth1, girth2)
R> plot(shape1, shape2)
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
25 27 29
−31−29−27−25
girth1
girth2
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
0.5 1.5
2.02.53.03.5
shape1
shape2
Fig. 3.12. Scatterplots of girth and shape for first and second sons.
The first canonical variate for both first and second sons is simply a
weighted sum of the two head measurements and might be labelled “girth”;
these two variates have a correlation of −0.79. (The negative value arises be-
cause of the arbitrariness of the sign in the first coefficient of an eigenvector–
here both coefficients for girth in first sons are positive and for second sons](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-114-320.jpg)
![100 3 Principal Components Analysis
R> r11 <- LAdepr[1:2, 1:2]
R> r22 <- LAdepr[-(1:2), -(1:2)]
R> r12 <- LAdepr[1:2, -(1:2)]
R> r21 <- LAdepr[-(1:2), 1:2]
R> (E1 <- solve(r11) %*% r12 %*% solve(r22) %*%r21)
CESD Health
CESD 0.08356 -0.04312
Health -0.03256 0.13338
R> (E2 <- solve(r22) %*% r21 %*% solve(r11) %*%r12)
Gender Age Edu Income
Gender 0.015478 -0.001483 -0.01813 -0.02224
Age -0.002445 0.147069 -0.03981 -0.01599
Edu -0.014915 -0.026019 0.02544 0.02508
Income -0.021163 0.013027 0.02159 0.02895
R> (e1 <- eigen(E1))
$values
[1] 0.15347 0.06347
$vectors
[,1] [,2]
[1,] 0.5250 -0.9065
[2,] -0.8511 -0.4222
R> (e2 <- eigen(E2))
$values
[1] 1.535e-01 6.347e-02 -8.760e-18 6.633e-19
$vectors
[,1] [,2] [,3] [,4]
[1,] -0.002607 0.4904 -0.4617 0.79705
[2,] -0.980095 -0.3208 -0.1865 0.06251
[3,] 0.185801 -0.4270 -0.7974 -0.05817
[4,] -0.069886 -0.6887 0.3409 0.59784
(Note that the third and fourth eigenvalues of E2 are essentially zero, as we
would expect in this case.) The first canonical correlation is 0.409, calculated
as the square root of the first eigenvalue of E1, which is given above as 0.15347.
If tested as outlined in Exercise 3.4, it has an associated p-value that is very
small; there is strong evidence that the first canonical correlation is significant.
The corresponding variates, in terms of standardised original variables, are
u1 = +0.53 CESD − 0.85 Health
v1 = −0.00 Gender − 0.98 Age + 0.19 Education − 0.07 Income .](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-116-320.jpg)
![4
Multidimensional Scaling
4.1 Introduction
In Chapter 3, we noted in passing that one of the most useful ways of using
principal components analysis was to obtain a low-dimensional “map” of the
data that preserved as far as possible the Euclidean distances between the ob-
servations in the space of the original q variables. In this chapter, we will make
this aspect of principal component analysis more explicit and also introduce
a class of other methods, labelled multidimensional scaling, that aim to pro-
duce similar maps of data but do not operate directly on the usual multivariate
data matrix, X. Instead they are applied to distance matrices (see Chapter 1),
which are derived from the matrix X (an example of a distance matrix de-
rived from a small set of multivariate data is shown in Subsection 4.4.2), and
also to so-called dissimilarity or similarity matrices that arise directly in a
number of ways, in particular from judgements made by human raters about
how alike pairs of objects, stimuli, etc., of interest are. An example of a di-
rectly observed dissimilarity matrix is shown in Table 4.5, with judgements
about political and war leaders that had major roles in World War II being
given by a subject after receiving the simple instructions to rate each pair of
politicians on a nine-point scale, with 1 indicating two politicians they regard
as very similar and 9 indicating two they regard as very dissimilar. (If the
nine point-scale had been 1 for very dissimilar and 9 for very similar, then the
result would have been a rating of similarity, although similarities are often
scaled to lie in a [0, 1] interval. The term proximity is often used to encompass
both dissimilarity and similarity ratings.)
4.2 Models for proximity data
Models are fitted to proximities in order to clarify, display, and help under-
stand and possibly explain any structure or pattern amongst the observed or
DOI 10.1007/978-1-4419-9650-3_4, © Springer Science+Business Media, LLC 2011
105B. Everitt and T. Hothorn, An Introduction to Applied Multivariate Analysis with R: Use R!,](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-121-320.jpg)
![110 4 Multidimensional Scaling
Trace criterion: Choose the number of coordinates so that the sum of the pos-
itive eigenvalues is approximately equal to the sum of all the eigenvalues.
Magnitude criterion: Accept as genuinely positive only those eigenvalues
whose magnitude substantially exceeds that of the largest negative eigen-
value.
If, however, the matrix B has a considerable number of large negative
eigenvalues, classical scaling of the proximity matrix may be inadvisable and
some other methods of scaling, for example non-metric scaling (see the next
section), might be better employed.
4.4.2 Examples of classical multidimensional scaling
For our first example we will use the small set of multivariate data X
[,1] [,2] [,3] [,4] [,5]
[1,] 3 4 4 6 1
[2,] 5 1 1 7 3
[3,] 6 2 0 2 6
[4,] 1 1 1 0 3
[5,] 4 7 3 6 2
[6,] 2 2 5 1 0
[7,] 0 4 1 1 1
[8,] 0 6 4 3 5
[9,] 7 6 5 1 4
[10,] 2 1 4 3 1
and the associated matrix of Euclidean distances (computed via the dist()
function) will be our proximity matrix
R> (D <- dist(X))
1 2 3 4 5 6 7 8 9
2 5.196
3 8.367 6.083
4 7.874 8.062 6.325
5 3.464 6.557 8.367 9.274
6 5.657 8.426 8.832 5.292 7.874
7 6.557 8.602 8.185 3.873 7.416 5.000
8 6.164 8.888 8.367 6.928 6.000 7.071 5.745
9 7.416 9.055 6.856 8.888 6.557 7.550 8.832 7.416
10 4.359 6.164 7.681 4.796 7.141 2.646 5.099 6.708 8.000
To apply classical scaling to this matrix in R, we can use the cmdscale()
function to do the scaling:
R> cmdscale(D, k = 9, eig = TRUE)](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-126-320.jpg)
![4.4 Classical multidimensional scaling 111
$points
[,1] [,2] [,3] [,4] [,5] [,6]
[1,] -1.6038 -2.38061 -2.2301 -0.3657 0.11536 0.000e+00
[2,] -2.8246 2.30937 -3.9524 0.3419 0.33169 -2.797e-08
[3,] -1.6908 5.13970 1.2880 0.6503 -0.05134 -1.611e-09
[4,] 3.9528 2.43234 0.3834 0.6864 -0.03461 -7.393e-09
[5,] -3.5985 -2.75538 -0.2551 1.0784 -1.26125 -5.198e-09
[6,] 2.9520 -1.35475 -0.1899 -2.8211 0.12386 -2.329e-08
[7,] 3.4690 -0.76411 0.3017 1.6369 -1.94210 -1.452e-08
[8,] 0.3545 -2.31409 2.2162 2.9240 2.00450 -1.562e-08
[9,] -2.9362 0.01280 4.3117 -2.5123 -0.18912 -1.404e-08
[10,] 1.9257 -0.32527 -1.8734 -1.6189 0.90299 6.339e-09
[,7] [,8] [,9]
[1,] 1.791e-08 NaN NaN
[2,] -1.209e-09 NaN NaN
[3,] 1.072e-09 NaN NaN
[4,] 1.088e-08 NaN NaN
[5,] -2.798e-09 NaN NaN
[6,] -7.146e-09 NaN NaN
[7,] 3.072e-09 NaN NaN
[8,] 2.589e-10 NaN NaN
[9,] 7.476e-09 NaN NaN
[10,] 3.303e-09 NaN NaN
$eig
[1] 7.519e+01 5.881e+01 4.961e+01 3.043e+01 1.037e+01
[6] 2.101e-15 5.769e-16 -2.819e-15 -3.233e-15 -6.274e-15
$x
NULL
$ac
[1] 0
$GOF
[1] 1 1
Note that as q = 5 in this example, eigenvalues six to nine are essentially zero
and only the first five columns of points represent the Euclidean distance ma-
trix. First we should confirm that the five-dimensional solution achieves com-
plete recovery of the observed distance matrix. We can do this simply by com-
paring the original distances with those calculated from the five-dimensional
scaling solution coordinates using the following R code:
R> max(abs(dist(X) - dist(cmdscale(D, k = 5))))
[1] 1.243e-14](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-127-320.jpg)
![112 4 Multidimensional Scaling
This confirms that all the differences are essentially zero and that therefore the
observed distance matrix is recovered by the five-dimensional classical scaling
solution.
We can also check the duality of classical scaling of Euclidean distances and
principal components analysis mentioned previously in the chapter by com-
paring the coordinates of the five-dimensional scaling solution given above
with the first five principal component (up to signs) scores obtained by apply-
ing PCA to the covariance matrix of the original data; the necessary R code
is
R> max(abs(prcomp(X)$x) - abs(cmdscale(D, k = 5)))
[1] 3.035e-14
Now let us look at two examples involving distances that are not Eu-
clidean. First, we will calculate the Manhattan distances between the rows of
the small data matrix X. The Manhattan distance for units i and j is given
by
q
k=1 |xik − xjk|, and these distances are not Euclidean. (Manhattan dis-
tances will be familiar to those readers who have walked around New York.)
The R code for calculating the Manhattan distances and then applying clas-
sical multidimensional scaling to the resulting distance matrix is:
R> X_m <- cmdscale(dist(X, method = "manhattan"),
+ k = nrow(X) - 1, eig = TRUE)
The criteria P
(1)
m and P
(2)
m can be computed from the eigenvalues as follows:
R> (X_eigen <- X_m$eig)
[1] 2.807e+02 2.494e+02 2.289e+02 9.251e+01 4.251e+01
[6] 2.197e+01 -7.105e-15 -1.507e+01 -2.805e+01 -5.683e+01
Note that some of the eigenvalues are negative in this case.
R> cumsum(abs(X_eigen)) / sum(abs(X_eigen))
[1] 0.2763 0.5218 0.7471 0.8382 0.8800 0.9016 0.9016 0.9165
[9] 0.9441 1.0000
R> cumsum(X_eigen^2) / sum(X_eigen^2)
[1] 0.3779 0.6764 0.9276 0.9687 0.9773 0.9796 0.9796 0.9807
[9] 0.9845 1.0000
The values of both criteria suggest that a three-dimensional solution seems to
fit well.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-128-320.jpg)
![4.4 Classical multidimensional scaling 113
Table 4.1: airdist data. Airline distances between ten US cities.
ATL ORD DEN HOU LAX MIA JFK SFO SEA IAD
ATL 0
ORD 587 0
DEN 1212 920 0
HOU 701 940 879 0
LAX 1936 1745 831 1374 0
MIA 604 1188 1726 968 2339 0
JFK 748 713 1631 1420 2451 1092 0
SFO 2139 1858 949 1645 347 2594 2571 0
SEA 218 1737 1021 1891 959 2734 2408 678 0
IAD 543 597 1494 1220 2300 923 205 2442 2329 0
For our second example of applying classical multidimensional scaling to non-
Euclidean distances, we shall use the airline distances between ten US cities
given in Table 4.1. These distances are not Euclidean since they relate essen-
tially to journeys along the surface of a sphere. To apply classical scaling to
these distances and to see the eigenvalues, we can use the following R code:
R> airline_mds <- cmdscale(airdist, k = 9, eig = TRUE)
R> airline_mds$points
[,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8]
ATL -434.8 724.22 440.93 0.18579 -1.258e-02 NaN NaN NaN
ORD -412.6 55.04 -370.93 4.39608 1.268e+01 NaN NaN NaN
DEN 468.2 -180.66 -213.57 30.40857 -9.585e+00 NaN NaN NaN
HOU -175.6 -515.22 362.84 9.48713 -4.860e+00 NaN NaN NaN
LAX 1206.7 -465.64 56.53 1.34144 6.809e+00 NaN NaN NaN
MIA -1161.7 -477.98 479.60 -13.79783 2.278e+00 NaN NaN NaN
JFK -1115.6 199.79 -429.67 -29.39693 -7.137e+00 NaN NaN NaN
SFO 1422.7 -308.66 -205.52 -26.06310 -1.983e+00 NaN NaN NaN
SEA 1221.5 887.20 170.45 -0.06999 -8.943e-05 NaN NaN NaN
IAD -1018.9 81.90 -290.65 23.50884 1.816e+00 NaN NaN NaN
(The nineth column containing NaNs is omitted from the output.) The eigen-
values are
R> (lam <- airline_mds$eig)
[1] 9.214e+06 2.200e+06 1.083e+06 3.322e+03 3.859e+02
[6] -5.204e-09 -9.323e+01 -2.169e+03 -9.091e+03 -1.723e+06
As expected (as the distances are not Euclidean), some of the eigenvalues
are negative and so we will again use the criteria P
(1)
m and P
(2)
m to assess
how many coordinates we need to adequately represent the observed distance
matrix. The values of the two criteria calculated from the eigenvalues are](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-129-320.jpg)
![114 4 Multidimensional Scaling
R> cumsum(abs(lam)) / sum(abs(lam))
[1] 0.6473 0.8018 0.8779 0.8781 0.8782 0.8782 0.8782 0.8783
[9] 0.8790 1.0000
R> cumsum(lam^2) / sum(lam^2)
[1] 0.9043 0.9559 0.9684 0.9684 0.9684 0.9684 0.9684 0.9684
[9] 0.9684 1.0000
These values suggest that the first two coordinates will give an adequate rep-
resentation of the observed distances. The scatterplot of the two-dimensional
coordinate values is shown in Figure 4.1. In this two-dimensional representa-
tion, the geographical location of the cities has been very well recovered by the
two-dimensional multidimensional scaling solution obtained from the airline
distances.
Our next example of the use of classical multidimensional scaling will in-
volve the data shown in Table 4.2. These data show four measurements on
male Egyptian skulls from five epochs. The measurements are:
mb: maximum breadth of the skull;
bh: basibregmatic height of the skull;
bl: basialiveolar length of the skull; and
nh: nasal height of the skull.
Table 4.2: skulls data. Measurements of four variables taken from
Egyptian skulls of five periods.
epoch mb bh bl nh epoch mb bh bl nh epoch mb bh bl nh
c4000BC 131 138 89 49 c3300BC 137 136 106 49 c200BC 132 133 90 53
c4000BC 125 131 92 48 c3300BC 126 131 100 48 c200BC 134 134 97 54
c4000BC 131 132 99 50 c3300BC 135 136 97 52 c200BC 135 135 99 50
c4000BC 119 132 96 44 c3300BC 129 126 91 50 c200BC 133 136 95 52
c4000BC 136 143 100 54 c3300BC 134 139 101 49 c200BC 136 130 99 55
c4000BC 138 137 89 56 c3300BC 131 134 90 53 c200BC 134 137 93 52
c4000BC 139 130 108 48 c3300BC 132 130 104 50 c200BC 131 141 99 55
c4000BC 125 136 93 48 c3300BC 130 132 93 52 c200BC 129 135 95 47
c4000BC 131 134 102 51 c3300BC 135 132 98 54 c200BC 136 128 93 54
c4000BC 134 134 99 51 c3300BC 130 128 101 51 c200BC 131 125 88 48
c4000BC 129 138 95 50 c1850BC 137 141 96 52 c200BC 139 130 94 53
c4000BC 134 121 95 53 c1850BC 129 133 93 47 c200BC 144 124 86 50
c4000BC 126 129 109 51 c1850BC 132 138 87 48 c200BC 141 131 97 53
c4000BC 132 136 100 50 c1850BC 130 134 106 50 c200BC 130 131 98 53
c4000BC 141 140 100 51 c1850BC 134 134 96 45 c200BC 133 128 92 51
c4000BC 131 134 97 54 c1850BC 140 133 98 50 c200BC 138 126 97 54
c4000BC 135 137 103 50 c1850BC 138 138 95 47 c200BC 131 142 95 53](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-130-320.jpg)
![116 4 Multidimensional Scaling
−1500 −500 0 500 1000 1500
−1500−50005001500
Coordinate 1
Coordinate2
ATL
ORD
DEN
HOULAX MIA
JFK
SFO
SEA
IAD
Fig. 4.1. Two-dimensional classical MDS solution for airline distances. The known
spatial arrangement is clearly visible in the plot.
where S1, S2, . . . , S5 are the covariance matrices of the data in each epoch.
We shall then use the first two coordinate values to provide a map of the data
showing the relationships between epochs. The necessary R code is:
R> skulls_var <- tapply(1:nrow(skulls), skulls$epoch,
+ function(i) var(skulls[i,-1]))
R> S <- 0
R> for (v in skulls_var) S <- S + 29 * v
R> (S <- S / 149)
mb bh bl nh
mb 20.54407 0.03579 0.07696 1.955
bh 0.03579 22.85414 5.06040 2.769
bl 0.07696 5.06040 23.52998 1.103
nh 1.95503 2.76868 1.10291 9.880](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-132-320.jpg)
![4.4 Classical multidimensional scaling 117
R> skulls_cen <- tapply(1:nrow(skulls), skulls$epoch,
+ function(i) apply(skulls[i,-1], 2, mean))
R> skulls_cen <- matrix(unlist(skulls_cen),
+ nrow = length(skulls_cen), byrow = TRUE)
R> skulls_mah <- apply(skulls_cen, 1,
+ function(cen) mahalanobis(skulls_cen, cen, S))
R> skulls_mah
[,1] [,2] [,3] [,4] [,5]
[1,] 0.00000 0.09355 0.9280 1.9330 2.7712
[2,] 0.09355 0.00000 0.7490 1.6380 2.2357
[3,] 0.92799 0.74905 0.0000 0.4553 0.9360
[4,] 1.93302 1.63799 0.4553 0.0000 0.2253
[5,] 2.77121 2.23571 0.9360 0.2253 0.0000
R> cmdscale(skulls_mah, k = nrow(skulls_mah) - 1,
+ eig = TRUE)$eig
[1] 5.113e+00 9.816e-02 -5.551e-16 -1.009e-01 -7.777e-01
R> skulls_mds <- cmdscale(skulls_mah)
The resulting plot is shown in Figure 4.2 and shows that the scaling solu-
tion for the skulls data is essentially unidimensional, with this single dimension
time ordering the five epochs. There appears to be a change in the “shape”
of the skulls over time, with maximum breadth increasing and basialiveolar
length decreasing.
Our final example of the application of classical multidimensional scal-
ing involves an investigation by Corbet, Cummins, Hedges, and Krzanowski
(1970), who report a study of water voles (genus Arvicola) in which the aim
was to compare British populations of these animals with those in Europe to
investigate whether more than one species might be present in Britain. The
original data consisted of observations of the presence or absence of 13 charac-
teristics in about 300 water vole skulls arising from six British populations and
eight populations from the rest of Europe. Table 4.3 gives a distance matrix
derived from the data as described in Corbet et al. (1970).
The following code finds the classical scaling solution and computes the two
criteria for assessing the required number of dimensions as described above.
R> data("watervoles", package = "HSAUR2")
R> voles_mds <- cmdscale(watervoles, k = 13, eig = TRUE)
R> voles_mds$eig
[1] 7.360e-01 2.626e-01 1.493e-01 6.990e-02 2.957e-02
[6] 1.931e-02 9.714e-17 -1.139e-02 -1.280e-02 -2.850e-02
[11] -4.252e-02 -5.255e-02 -7.406e-02 -1.098e-01
Note that some of the eigenvalues are negative. The criterion P
(1)
m can be
computed by](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-133-320.jpg)
![4.4 Classical multidimensional scaling 119
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
−1.5−0.50.51.5
Coordinate 1
Coordinate2
c4000BC
c3300BC
c1850BC
c200BC
cAD150
Fig. 4.2. Two-dimensional solution from classical MDS applied to Mahalanobis
distances between epochs for the skull data.
R> cumsum(abs(voles_mds$eig))/sum(abs(voles_mds$eig))
[1] 0.4605 0.6248 0.7182 0.7619 0.7804 0.7925 0.7925 0.7996
[9] 0.8077 0.8255 0.8521 0.8850 0.9313 1.0000
and the criterion P
(2)
m is
R> cumsum((voles_mds$eig)^2)/sum((voles_mds$eig)^2)
[1] 0.8179 0.9220 0.9557 0.9631 0.9644 0.9649 0.9649 0.9651
[9] 0.9654 0.9666 0.9693 0.9735 0.9818 1.0000
Here the two criteria for judging the number of dimensions necessary to
give an adequate fit to the data are quite different. The second criterion
would suggest that two dimensions is adequate, but use of the first would
suggest perhaps that three or even four dimensions might be required. Here
we shall be guided by the second fit index and the two-dimensional solution
that can be plotted by extracting the coordinates from the points element of
the voles_mds object; the plot is shown in Figure 4.3.
It appears that the six British populations are close to populations liv-
ing in the Alps, Yugoslavia, Germany, Norway, and Pyrenees I (consisting](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-135-320.jpg)
![120 4 Multidimensional Scaling
R> x <- voles_mds$points[,1]
R> y <- voles_mds$points[,2]
R> plot(x, y, xlab = "Coordinate 1", ylab = "Coordinate 2",
+ xlim = range(x)*1.2, type = "n")
R> text(x, y, labels = colnames(watervoles), cex = 0.7)
−0.2 0.0 0.2 0.4 0.6
−0.3−0.10.1
Coordinate 1
Coordinate2
Surrey
Shropshire
YorkshirePerthshire
AberdeenElean Gamhna
Alps
Yugoslavia
GermanyNorway
Pyrenees I
Pyrenees II
North Spain
South Spain
Fig. 4.3. Two-dimensional solution from classical multidimensional scaling of the
distance matrix for water vole populations.
of the species Arvicola terrestris) but rather distant from the populations in
Pyrenees II, North Spain and South Spain (species Arvicola sapidus). This re-
sult would seem to imply that Arvicola terrestris might be present in Britain
but it is less likely that this is so for Arvicola sapidus. But here, as the two-
dimensional fit may not be entirely what is needed to represent the observed
distances, we shall investigate the solution in a little more detail using the
minimum spanning tree.
The minimum spanning tree is defined as follows. Suppose n points are
given (possibly in many dimensions). Then a tree spanning these points (i.e.,
a spanning tree) is any set of straight line segments joining pairs of points
such that
ˆ o closed loops occur,
ˆ every point is visited at least one time, and](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-136-320.jpg)
![122 4 Multidimensional Scaling
R> library("ape")
R> st <- mst(watervoles)
R> plot(x, y, xlab = "Coordinate 1", ylab = "Coordinate 2",
+ xlim = range(x)*1.2, type = "n")
R> for (i in 1:nrow(watervoles)) {
+ w1 <- which(st[i, ] == 1)
+ segments(x[i], y[i], x[w1], y[w1])
+ }
R> text(x, y, labels = colnames(watervoles), cex = 0.7)
−0.2 0.0 0.2 0.4 0.6
−0.3−0.10.1
Coordinate 1
Coordinate2
Surrey
Shropshire
YorkshirePerthshire
AberdeenElean Gamhna
Alps
Yugoslavia
GermanyNorway
Pyrenees I
Pyrenees II
North Spain
South Spain
Fig. 4.4. Minimum spanning tree for the watervoles data plotted onto the classical
scaling two-dimensional solution.
these papers is that the coordinates in the spatial representation of the ob-
served proximities give rise to fitted distances, dij, and that these distances are
related to a set of numbers which we will call disparities, ˆdij, by the formula
dij = ˆdij + ij, where the ij are error terms representing errors of measure-
ment plus distortion errors arising because the distances do not correspond to
a configuration in the particular number of dimensions chosen. The disparities
are monotonic with the observed proximities and, subject to this constraint,
resemble the fitted distances as closely as possible. In general, only a weak
monotonicity constraint is applied, so that if, say, the observed dissimilarities,
δij are ranked from lowest to highest to give](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-138-320.jpg)
![4.5 Non-metric multidimensional scaling 125
R> x <- voting_mds$points[,1]
R> y <- voting_mds$points[,2]
R> plot(x, y, xlab = "Coordinate 1", ylab = "Coordinate 2",
+ xlim = range(voting_mds$points[,1])*1.2, type = "n")
R> text(x, y, labels = colnames(voting), cex = 0.6)
R> voting_sh <- Shepard(voting[lower.tri(voting)],
+ voting_mds$points)
−10 −5 0 5
−6−202468
Coordinate 1
Coordinate2
Hunt(R)
Sandman(R)
Howard(D)
Thompson(D)
Freylinghuysen(R)
Forsythe(R)
Widnall(R)
Roe(D)
Heltoski(D)
Rodino(D)
Minish(D)
Rinaldo(R)
Maraziti(R)
Daniels(D)
Patten(D)
Fig. 4.5. Two-dimensional solution from non-metric multidimensional scaling of
distance matrix for voting matrix.
Table 4.5: WWIIleaders data. Subjective distances between WWII
leaders.
Htl Mss Chr Esn Stl Att Frn DGl MT- Trm Chm Tit
Hitler 0
Mussolini 3 0
Churchill 4 6 0
Eisenhower 7 8 4 0
Stalin 3 5 6 8 0
Attlee 8 9 3 9 8 0
Franco 3 2 5 7 6 7 0
De Gaulle 4 4 3 5 6 5 4 0](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-141-320.jpg)
![126 4 Multidimensional Scaling
Table 4.5: WWIIleaders data (continued).
Htl Mss Chr Esn Stl Att Frn DGl MT- Trm Chm Tit
Mao Tse-Tung 8 9 8 9 6 9 8 7 0
Truman 9 9 5 4 7 8 8 4 4 0
Chamberlin 4 5 5 4 7 2 2 5 9 5 0
Tito 7 8 2 4 7 8 3 2 4 5 7 0
The non-metric multidimensional scaling applied to these distances is
R> (WWII_mds <- isoMDS(WWIIleaders))
initial value 20.504211
iter 5 value 15.216103
iter 5 value 15.207237
iter 5 value 15.207237
final value 15.207237
converged
$points
[,1] [,2]
Hitler -2.5820 -1.75961
Mussolini -3.8807 -1.24756
Churchill 0.3110 1.46671
Eisenhower 2.9852 2.87822
Stalin -1.4274 -3.75699
Attlee -2.1067 5.07317
Franco -2.8590 0.07878
De Gaulle 0.6591 -0.20656
Mao Tse-Tung 4.1605 -4.57583
Truman 4.4962 0.29294
Chamberlin -2.1420 2.75877
Tito 2.3859 -1.00203
$stress
[1] 15.21
The two-dimensional solution appears in Figure 4.7. Clearly, the three fascists
group together as do the three British prime ministers. Stalin and Mao Tse-
Tung are more isolated compared with the other leaders. Eisenhower seems
more related to the British government than to his own President Truman.
Interestingly, de Gaulle is placed in the center of the MDS solution.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-142-320.jpg)
![130 4 Multidimensional Scaling
with similar profiles (conditional distributions) down the rows. Finally, row
points that lie close to column points represent a row/column combination
that occurs more frequently in the table than would be expected if the row
and column variables were independent. Conversely, row and column points
that are distant from one another indicate a cell in the table where the count
is lower than would be expected under independence.
We will now look at a single simple example of the application of corre-
spondence analysis.
4.6.1 Teenage relationships
Consider the data shown in Table 4.6 concerned with the influence of a girl’s
age on her relationship with her boyfriend. In this table, each of 139 girls has
been classified into one of three groups:
ˆ no boyfriend;
ˆ boyfriend/no sexual intercourse; or
ˆ boyfriend/sexual intercourse.
In addition, the age of each girl was recorded and used to divide the girls into
five age groups.
Table 4.6: teensex data. The influence of age on relationships with
boyfriends.
Age
Boyfriend <16 16-17 17-18 18-19 19-20
No boyfriend 21 21 14 13 8
Boyfriend no sex 8 9 6 8 2
Boyfriend sex 2 3 4 10 10
The calculation of the two-dimensional classical multidimensional scaling so-
lution based on the row- and column-wise chi-squared distance measure can
be computed via cmdscale(); however, we first have to compute the necessary
row and column distance matrices, and we will do this by setting up a small
convenience function as follows:
R> D <- function(x) {
+ a <- t(t(x) / colSums(x))
+ ret <- sqrt(colSums((a[,rep(1:ncol(x), ncol(x))] -
+ a[, rep(1:ncol(x), rep(ncol(x), ncol(x)))])^2 *
+ sum(x) / rowSums(x)))
+ matrix(ret, ncol = ncol(x))
+ }
R> (dcols <- D(teensex))](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-146-320.jpg)
![4.7 Summary 131
[,1] [,2] [,3] [,4] [,5]
[1,] 0.00000 0.08537 0.2574 0.6629 1.0739
[2,] 0.08537 0.00000 0.1864 0.5858 1.0142
[3,] 0.25743 0.18644 0.0000 0.4066 0.8295
[4,] 0.66289 0.58581 0.4066 0.0000 0.5067
[5,] 1.07385 1.01423 0.8295 0.5067 0.0000
R> (drows <- D(t(teensex)))
[,1] [,2] [,3]
[1,] 0.0000 0.2035 0.9275
[2,] 0.2035 0.0000 0.9355
[3,] 0.9275 0.9355 0.0000
Applying classical MDS to each of these distance matrices gives the re-
quired two-dimensional coordinates with which to construct our “map” of the
data. Plotting those with suitable labels and with the axes suitably scaled
to reflect the greater variation on dimension one than on dimension two (see
Greenacre 1992) is achieved using the R code presented with Figure 4.8.
The points representing the age groups in Figure 4.8 give a two-dimensional
representation in which the Euclidean distance between two points represents
the chi-squared distance between the corresponding age groups (and similarly
for the points representing the type of relationship). For a contingency table
with r rows and c columns, it can be shown that the chi-squared distances
can be represented exactly in min r − 1, c − 1 dimensions; here, since r = 3
and c = 5, this means that the Euclidean distances in Figure 4.8 will actually
equal the corresponding chi-squared distances (readers might like to check
that this is the case as an exercise). When both r and c are greater than
three, an exact two-dimensional representation of the chi-squared distances
is not possible. In such cases, the derived two-dimensional coordinates will
give only an approximate representation, and so the question of the adequacy
of the fit will need to be addressed. In some of these cases, more than two
dimensions may be required to give an acceptable fit (again see Greenacre
1992, for details).
Examining the plot in Figure 4.8, we see that it tells the age-old story
of girls travelling through their teenage years, initially having no boyfriend,
then acquiring a boyfriend, and then having sex with their boyfriend, a story
that has broken the hearts of fathers everywhere, at least temporarily, until
their wives suggest they reflect back to the time when they themselves were
teenagers.
4.7 Summary
Multidimensional scaling and correspondence analysis both aim to help in
understanding particular types of data by displaying the data graphically.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-147-320.jpg)
![132 4 Multidimensional Scaling
R> r1 <- cmdscale(dcols, eig = TRUE)
R> c1 <- cmdscale(drows, eig = TRUE)
R> plot(r1$points, xlim = range(r1$points[,1], c1$points[,1]) * 1.5,
+ ylim = range(r1$points[,1], c1$points[,1]) * 1.5, type = "n",
+ xlab = "Coordinate 1", ylab = "Coordinate 2", lwd = 2)
R> text(r1$points, labels = colnames(teensex), cex = 0.7)
R> text(c1$points, labels = rownames(teensex), cex = 0.7)
R> abline(h = 0, lty = 2)
R> abline(v = 0, lty = 2)
−0.5 0.0 0.5 1.0
−0.50.00.51.0
Coordinate 1
Coordinate2
<16
16−17 17−18
18−19
19−20No boyfriend
Boyfriend no sex
Boyfriend sex
Fig. 4.8. Correspondence analysis for teenage relationship data.
Multidimensional scaling applied to proximity matrices is often useful in un-
covering the dimensions on which similarity judgements are made, and corre-
spondence analysis often allows more insight into the pattern of relationships
in a contingency table than a simple chi-squared test.
4.8 Exercises
Ex. 4.1 Consider 51 objects O1, . . . , O51 assumed to be arranged along a
straight line with the jth object being located at a point with coordinate
j. Define the similarity sij between object i and object j as](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-148-320.jpg)
![6.3 Agglomerative hierarchical clustering 169
R> (dm <- dist(measure[, c("chest", "waist", "hips")]))
1 2 3 4 5 6 7 8 9 10
2 6.16
3 5.66 2.45
4 7.87 2.45 4.69
5 4.24 5.10 3.16 7.48
6 11.00 6.08 5.74 7.14 7.68
7 12.04 5.92 7.00 5.00 10.05 5.10
8 8.94 3.74 4.00 3.74 7.07 5.74 4.12
9 7.81 3.61 2.24 5.39 4.58 3.74 5.83 3.61
10 10.10 4.47 4.69 5.10 7.35 2.24 3.32 3.74 3.00
11 7.00 8.31 6.40 9.85 5.74 11.05 12.08 8.06 7.48 10.25
12 7.35 7.07 5.48 8.25 6.00 9.95 10.25 6.16 6.40 8.83
13 7.81 8.54 7.28 9.43 7.55 12.08 11.92 7.81 8.49 10.82
14 8.31 11.18 9.64 12.45 8.66 14.70 15.30 11.18 11.05 13.75
15 7.48 6.16 4.90 7.07 6.16 9.22 9.00 4.90 5.74 7.87
16 7.07 6.00 4.24 7.35 5.10 8.54 9.11 5.10 5.00 7.48
17 7.81 7.68 6.71 8.31 7.55 11.40 10.77 6.71 7.87 9.95
18 6.71 6.08 4.58 7.28 5.39 9.27 9.49 5.39 5.66 8.06
19 9.17 5.10 4.47 5.48 7.07 6.71 5.74 2.00 4.12 5.10
20 7.68 9.43 7.68 10.82 7.00 12.41 13.19 9.11 8.83 11.53
11 12 13 14 15 16 17 18 19
2
3
4
5
6
7
8
9
10
11
12 2.24
13 2.83 2.24
14 3.74 5.20 3.74
15 3.61 1.41 3.00 6.40
16 3.00 1.41 3.61 6.40 1.41
17 3.74 2.24 1.41 5.10 2.24 3.32
18 2.83 1.00 2.83 5.83 1.00 1.00 2.45
19 6.71 4.69 6.40 9.85 3.46 3.74 5.39 4.12
20 1.41 3.00 2.45 2.45 4.36 4.12 3.74 3.74 7.68
Application of each of the three clustering methods described earlier to
the distance matrix and a plot of the corresponding dendrogram are achieved
using the hclust() function:](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-185-320.jpg)
![6.3 Agglomerative hierarchical clustering 171
One informal approach is to examine the sizes of the changes in height in the
dendrogram and take a “large” change to indicate the appropriate number of
clusters for the data. (More formal approaches are described in Everitt et al.
2011) Even using this informal approach on the dendrograms in Figure 6.4, it
is not easy to decide where to “cut”.
So instead, because we know that these data consist of measurements on
ten men and ten women, we will look at the two-group solutions from each
method that are obtained by cutting the dendrograms at suitable heights. We
can display and compare the three solutions graphically by plotting the first
two principal component scores of the data, labelling the points to identify
the cluster solution of one of the methods by using the following code:
R> body_pc <- princomp(dm, cor = TRUE)
R> xlim <- range(body_pc$scores[,1])
R> plot(body_pc$scores[,1:2], type = "n",
+ xlim = xlim, ylim = xlim)
R> lab <- cutree(cs, h = 3.8)
R> text(body_pc$scores[,1:2], labels = lab, cex = 0.6)
The resulting plots are shown in the lower part of Figure 6.4. The plots of
dendrograms and principal components scatterplots are combined into a single
diagram using the layout() function (see the chapter demo for the complete
R code). The plot associated with the single linkage solution immediately
demonstrates one of the problems with using this method in practise, and
that is a phenomenon known as chaining, which refers to the tendency to
incorporate intermediate points between clusters into an existing cluster rather
than initiating a new one. As a result, single linkage solutions often contain
long “straggly” clusters that do not give a useful description of the data. The
two-group solutions from complete linkage and average linkage, also shown in
Figure 6.4, are similar and in essence place the men (observations 1 to 10)
together in one cluster and the women (observations 11 to 20) in the other.
6.3.1 Clustering jet fighters
The data shown in Table 6.1 as originally given in Stanley and Miller (1979)
and also in Hand et al. (1994) are the values of six variables for 22 US fighter
aircraft. The variables are as follows:
FFD: first flight date, in months after January 1940;
SPR: specific power, proportional to power per unit weight;
RGF: flight range factor;
PLF: payload as a fraction of gross weight of aircraft;
SLF: sustained load factor;
CAR: a binary variable that takes the value 1 if the aircraft can land on a
carrier and 0 otherwise.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-187-320.jpg)
![172 6 Cluster Analysis
Table 6.1: jet data. Jet fighters data.
FFD SPR RGF PLF SLF CAR
82 1.468 3.30 0.166 0.10 no
89 1.605 3.64 0.154 0.10 no
101 2.168 4.87 0.177 2.90 yes
107 2.054 4.72 0.275 1.10 no
115 2.467 4.11 0.298 1.00 yes
122 1.294 3.75 0.150 0.90 no
127 2.183 3.97 0.000 2.40 yes
137 2.426 4.65 0.117 1.80 no
147 2.607 3.84 0.155 2.30 no
166 4.567 4.92 0.138 3.20 yes
174 4.588 3.82 0.249 3.50 no
175 3.618 4.32 0.143 2.80 no
177 5.855 4.53 0.172 2.50 yes
184 2.898 4.48 0.178 3.00 no
187 3.880 5.39 0.101 3.00 yes
189 0.455 4.99 0.008 2.64 no
194 8.088 4.50 0.251 2.70 yes
197 6.502 5.20 0.366 2.90 yes
201 6.081 5.65 0.106 2.90 yes
204 7.105 5.40 0.089 3.20 yes
255 8.548 4.20 0.222 2.90 no
328 6.321 6.45 0.187 2.00 yes
We shall apply complete linkage to the data but using only variables two
to five. And given that the variables are on very different scales, we will
standardise them to unit variance before clustering. The required R code for
standardisation and clustering is as follows:
R> X <- scale(jet[, c("SPR", "RGF", "PLF", "SLF")],
+ center = FALSE, scale = TRUE)
R> dj <- dist(X)
R> plot(cc <- hclust(dj), main = "Jets clustering")
R> cc
Call:
hclust(d = dj)
Cluster method : complete
Distance : euclidean
Number of objects: 22](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-188-320.jpg)
![174 6 Cluster Analysis
Again, we cut the dendrogram in such a way that two clusters remain and
plot the corresponding classes in the space of the first two principal compo-
nents; see Figure 6.6.
R> pr <- prcomp(dj)$x[, 1:2]
R> plot(pr, pch = (1:2)[cutree(cc, k = 2)],
+ col = c("black", "darkgrey")[jet$CAR],
+ xlim = range(pr) * c(1, 1.5))
R> legend("topright", col = c("black", "black",
+ "darkgrey", "darkgrey"),
+ legend = c("1 / no", "2 / no", "1 / yes", "2 / yes"),
+ pch = c(1:2, 1:2), title = "Cluster / CAR", bty = "n")
q
q
q
q
q
q
q
q
q
q
q
q
−2 −1 0 1 2 3
−2−101
PC1
PC2
q
q
Cluster / CAR
1 / no
2 / no
1 / yes
2 / yes
Fig. 6.6. Hierarchical clustering (complete linkage) of jet data plotted in PCA space.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-190-320.jpg)
![6.4 K-means clustering 181
R> n <- nrow(crime_s)
R> wss <- rep(0, 6)
R> wss[1] <- (n - 1) * sum(sapply(crime_s, var))
R> for (i in 2:6)
+ wss[i] <- sum(kmeans(crime_s,
+ centers = i)$withinss)
R> plot(1:6, wss, type = "b", xlab = "Number of groups",
+ ylab = "Within groups sum of squares")
q
q
q
q
q
q
1 2 3 4 5 6
681012141618
Number of groups
Withingroupssumofsquares
Fig. 6.9. Plot of within-groups sum of squares against number of clusters.
We plot the within-groups sum of squares for one to six group k-means
solutions to see if we can get any indication of the number of groups (see
Figure 6.12). Again, the plot leads to the relatively clear conclusion that the
data contain three clusters.
Our interest is now in a comparison of the kiln sites at which the pottery
was found.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-197-320.jpg)
![6.5 Model-based clustering 183
R> library("lattice")
R> levelplot(as.matrix(pottery_dist), xlab = "Pot Number",
+ ylab = "Pot Number")
Pot Number
PotNumber
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
Fig. 6.11. Image plot of the dissimilarity matrix of the pottery data.
and five. So the clusters found actually correspond to pots from three different
regions.
6.5 Model-based clustering
The agglomerative hierarchical and k-means clustering methods described in
the previous two sections are based largely on heuristic but intuitively reason-
able procedures. But they are not based on formal models for cluster structure
in the data, making problems such as deciding between methods, estimating
R> pottery_dist <- dist(pots <- scale(pottery[, colnames(pottery) != "kiln"],
+ center = FALSE))](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-199-320.jpg)
![184 6 Cluster Analysis
R> n <- nrow(pots)
R> wss <- rep(0, 6)
R> wss[1] <- (n - 1) * sum(sapply(pots, var))
R> for (i in 2:6)
+ wss[i] <- sum(kmeans(pots,
+ centers = i)$withinss)
R> plot(1:6, wss, type = "b", xlab = "Number of groups",
+ ylab = "Within groups sum of squares")
q
q
q
q
q
1 2 3 4 5 6
20304050
Number of groups
Withingroupssumofsquares
Fig. 6.12. Plot of within-groups sum of squares against number of clusters.
the number of clusters, etc, particularly difficult. And, of course, without a
reasonable model, formal inference is precluded. In practise, these may not
be insurmountable objections to the use of either the agglomerative meth-
ods or k-means clustering because cluster analysis is most often used as an
“exploratory” tool for data analysis. But if an acceptable model for cluster
structure could be found, then the cluster analysis based on the model might
give more persuasive solutions (more persuasive to statisticians at least). In](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-200-320.jpg)
![7.3 Confirmatory factor analysis models 209
Estimate Std Error z value Pr(>|z|)
lambda1 0.86320 0.035182 24.5355 0.0000000
lambda2 0.84932 0.035489 23.9323 0.0000000
lambda3 0.80509 0.036409 22.1123 0.0000000
lambda4 0.69527 0.038678 17.9757 0.0000000
lambda5 0.77508 0.040365 19.2020 0.0000000
lambda6 0.92893 0.039417 23.5665 0.0000000
rho 0.66637 0.030955 21.5273 0.0000000
theta1 0.25488 0.023502 10.8450 0.0000000
theta2 0.27865 0.024263 11.4847 0.0000000
theta3 0.35184 0.026916 13.0715 0.0000000
theta4 0.51660 0.034820 14.8365 0.0000000
theta5 0.39924 0.038214 10.4475 0.0000000
theta6 0.13709 0.043530 3.1493 0.0016366
lambda1 SCA <--- Ability
lambda2 PPE <--- Ability
lambda3 PTE <--- Ability
lambda4 PFE <--- Ability
lambda5 EA <--- Aspiration
lambda6 CP <--- Aspiration
rho Aspiration <--> Ability
theta1 SCA <--> SCA
theta2 PPE <--> PPE
theta3 PTE <--> PTE
theta4 PFE <--> PFE
theta5 EA <--> EA
theta6 CP <--> CP
Iterations = 28
(Note that the two latent variables have their variances fixed at one, although
it is the fixing that is important, not the value at which they are fixed;
these variances cannot be free parameters to be estimated.) The z values test
whether parameters are significantly different from zero. All have very small
associated p-values. Of particular note amongst the parameter estimates is
the correlation between “true” ability and “true” aspiration; this is known as
a disattenuated correlation and is uncontaminated by measurement error in
the observed indicators of the two latent variables. In this case the estimate is
0.666, with a standard error of 0.031. An approximate 95% confidence interval
for the disattenuated correlation is [0.606, 0.727].
A path diagram (see Everitt and Dunn 2001) for the correlated, two-factor
model is shown in Figure 7.2. Note that the R function path.diagram()“only”
produces a textual representation of the graph, here in file ability_sem.dot.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-225-320.jpg)
![226 8 The Analysis of Repeated Measures Data
particular measurement (for example, the time the measurement was taken)
by using the reshape() function in R. The code needed to rearrange the data
ex_wide into their long form is as follows:
R> reshape(ex_wide, direction = "long", idvar = "ID",
+ varying = colnames(ex_wide)[-(1:2)])
ID Group time Day
1.1 1 1 1 15
2.1 2 1 1 10
3.1 3 1 1 8
4.1 4 2 1 11
5.1 5 2 1 11
6.1 6 2 1 12
1.2 1 1 2 15
2.2 2 1 2 9
3.2 3 1 2 7
4.2 4 2 2 8
5.2 5 2 2 12
6.2 6 2 2 12
1.5 1 1 5 10
2.5 2 1 5 11
3.5 3 1 5 6
4.5 4 2 5 13
5.5 5 2 5 11
6.5 6 2 5 6
1.7 1 1 7 7
2.7 2 1 7 12
3.7 3 1 7 9
4.7 4 2 7 7
5.7 5 2 7 11
6.7 6 2 7 10
Here, varying contains the names of the variables containing the repeated
measurements. Note that the name of the variable consists of the name itself
and the time point, separated by a dot. This long form of repeated measures
data is used when applying the models to be described in Section 8.2, although
the wide format is often more convenient for plotting the data and computing
summary statistics.
So now let us take a look at two repeated measurement data sets that
we shall be concerned with in this chapter. The first, shown in its long form
in Table 8.1, is taken from Crowder (1998) and gives the loads required to
produce slippage x of a timber specimen in a clamp. There are eight speci-
mens each with 15 repeated measurements. The second data set, in Table 8.2
reported in Zerbe (1979) and also given in Davis (2003), consists of plasma
inorganic phosphate measurements obtained from 13 control and 20 obese pa-
tients 0, 0.5, 1, 1.5, 2, and 3 hours after an oral glucose challenge. These two](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-242-320.jpg)
![8.2 Linear mixed-effects models for repeated measures data 239
R> pfun <- function(x, y) {
+ panel.xyplot(x, y[1:length(x)])
+ panel.lines(x, y[1:length(x) + length(x)], lty = 1)
+ }
R> plot(xyplot(cbind(loads, pred1) ~ slippage | specimen,
+ data = timber, panel = pfun, layout = c(4, 2),
+ ylab = "loads"))
slippage
loads
0
5
10
15
20
0.0 0.5 1.0 1.5
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
spec1
q
q
q
q
q
q
q
q
qqq
q
qqq
spec2
0.0 0.5 1.0 1.5
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
spec3
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
spec4
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
spec5
0.0 0.5 1.0 1.5
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
spec6
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
spec7
0.0 0.5 1.0 1.5
0
5
10
15
20
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
spec8
Fig. 8.2. Plot of the predicted values of the random intercept model for the timber
data.
We can now produce a plot similar to that in Figure 8.2 but showing
the predicted values from the model in (8.4). The code is similar to that
given above and so is not repeated again here. The resulting plot is shown in
Figure 8.3. Clearly the model describes the data more satisfactorily, although
there remains an obvious problem, which is taken up in Exercise 8.1.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-255-320.jpg)
![8.2 Linear mixed-effects models for repeated measures data 241
R> x <- reshape(plasma, direction = "wide", timevar = "time",
+ idvar = "Subject", v.names = "plasma")
R> parallel(~ x[,-(1:2)] | group, data = x, horizontal = FALSE,
+ col = "black", scales = list(x = list(labels = 1:8)),
+ ylab = "Plasma inorganic phosphate",
+ xlab = "Time (hours after oral glucose challenge)")
Time (hours after oral glucose challenge)
Plasmainorganicphosphate
Min
Max
1 2 3 4 5 6 7 8
control
1 2 3 4 5 6 7 8
obese
Fig. 8.4. Glucose challenge data for control and obese groups over time. Each line
represents the trajectory of one individual.
The plot indicates that the correlations of pairs of measurements made
at different times differ so that the compound symmetry structure for these
correlations is unlikely to be appropriate.
On the basis of these plots, we will begin by fitting the model in (8.4) with
the addition, in this case, of an extra covariate, namely a dummy variable
coding the group, control or obese, to which a subject belongs. We can fit the
required model using
R> plasma.lme1 <- lme(plasma ~ time + I(time^2) + group,
+ random = ~ time | Subject,
+ data = plasma, method = "ML")
R> summary(plasma.lme1)](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-257-320.jpg)
![242 8 The Analysis of Repeated Measures Data
R> plot(splom(~ x[, grep("plasma", colnames(x))] | group, data = x,
+ cex = 1.5, pch = ".", pscales = NULL, varnames = 1:8))
Scatter Plot Matrix
1
2
3
4
5
6
7
8
control
1
2
3
4
5
6
7
8
obese
Fig. 8.5. Scatterplot matrix for glucose challenge data.
Linear mixed-effects model fit by maximum likelihood
Data: plasma
AIC BIC logLik
390.5 419.1 -187.2
Random effects:
Formula: ~time | Subject
Structure: General positive-definite, Log-Cholesky param.
StdDev Corr
(Intercept) 0.69772 (Intr)
time 0.09383 -0.7
Residual 0.38480
Fixed effects: plasma ~ time + I(time^2) + group
Value Std.Error DF t-value p-value
(Intercept) 4.880 0.17091 229 28.552 0.0000
time -0.803 0.05075 229 -15.827 0.0000
I(time^2) 0.085 0.00521 229 16.258 0.0000
groupobese 0.437 0.18589 31 2.351 0.0253
Correlation:](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-258-320.jpg)
![8.2 Linear mixed-effects models for repeated measures data 243
(Intr) time I(t^2)
time -0.641
I(time^2) 0.457 -0.923
groupobese -0.428 0.000 0.000
Standardized Within-Group Residuals:
Min Q1 Med Q3 Max
-2.771508 -0.548688 -0.002765 0.564435 2.889633
Number of Observations: 264
Number of Groups: 33
The regression coefficients for linear and quadratic time are both highly
significant. The group effect is also significant, and an asymptotic 95% confi-
dence interval for the group effect is obtained from 0.437±1.96×0.186, giving
[−3.209, 4.083].
Here, to demonstrate what happens if we make a very misleading assump-
tion about the correlational structure of the repeated measurements, we will
compare the results with those obtained if we assume that the repeated mea-
surements are independent. The independence model can be fitted in the usual
way with the lm() function
R> summary(lm(plasma ~ time + I(time^2) + group, data = plasma))
Call:
lm(formula = plasma ~ time + I(time^2) + group,
data = plasma)
Residuals:
Min 1Q Median 3Q Max
-1.6323 -0.4401 0.0347 0.4750 2.0170
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 4.85761 0.16686 29.11 < 2e-16
time -0.80328 0.08335 -9.64 < 2e-16
I(time^2) 0.08467 0.00904 9.37 < 2e-16
groupobese 0.49332 0.08479 5.82 1.7e-08
Residual standard error: 0.673 on 260 degrees of freedom
Multiple R-squared: 0.328, Adjusted R-squared: 0.32
F-statistic: 42.3 on 3 and 260 DF, p-value: <2e-16
We see that, under the independence assumption, the standard error for the
group effect is about one-half of that given for model plasma.lme1 and if
used would lead to a much stronger claim of evidence of a difference between
control and obese patients.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-259-320.jpg)
![246 8 The Analysis of Repeated Measures Data
that more accurately reflect the raw data plotted in Figure 8.4. The predicted
profiles for the obese group are “flatter” as required.
Time (hours after oral glucose challenge)
Plasmainorganicphosphate
2
4
6
2 4 6 8
q
qqqqq
q
q
id01
q
qqq
qqqq
id02
2 4 6 8
qq
q
qqq
qq
id03
qq
qqq
qqq
id04
2 4 6 8
qq
q
qqqqq
id05
q
qqq
qqqq
id06
q
qqqqqqq
id07
qq
qq
qqqq
id08
q
qqqqqqq
id09
qqqqq
qqq
id10
q
qqqqq
q
q
id11
2
4
6
qqqqqqq
q
id12
2
4
6 q
qqqq
qqq
id13
q
qqqqqq
q
id14
qq
qqqqqq
id15
qqqqqqqq
id16
qq
qqqqqq
id17
qqqqqqqq
id18
qq
qqq
qqq
id19
qqqqq
qqq
id20
qq
qq
q
qqq
id21
qqqqqqqq
id22
qq
qqqqqq
id23
2
4
6qqqqq
qqq
id24
2
4
6
qqq
qqqqq
id25
qqqqq
qqq
id26
q
qqqq
qqq
id27
qqqqqqqq
id28
qqqqqqqq
id29
qq
q
qqqqq
id30
qqqqqqq
q
id31
2 4 6 8
qqqq
qqqq
id32
2
4
6
qqqqqqqq
id33
Fig. 8.7. Predictions for glucose challenge data.
We can check the assumptions of the final model fitted to the glucose chal-
lenge data (i.e., the normality of the random-effect terms and the residuals)
by first using the random.effects() function to predict the former and the
resid() function to calculate the differences between the observed data val-
ues and the fitted values and then using normal probability plots on each.
How the random effects are predicted is explained briefly in Section 8.3. The
necessary R code to obtain the effects, residuals, and plots is as follows:
R> res.int <- random.effects(plasma.lme2)[,1]
R> res.slope <- random.effects(plasma.lme2)[,2]](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-262-320.jpg)
![252 8 The Analysis of Repeated Measures Data
R> ylim <- range(BtheB[,grep("bdi", names(BtheB))],
+ na.rm = TRUE)
R> tau <- subset(BtheB, treatment == "TAU")[,
+ grep("bdi", names(BtheB))]
R> boxplot(tau, main = "Treated as Usual", ylab = "BDI",
+ xlab = "Time (in months)", names = c(0, 2, 3, 5, 8),
+ ylim = ylim)
R> btheb <- subset(BtheB, treatment == "BtheB")[,
+ grep("bdi", names(BtheB))]
R> boxplot(btheb, main = "Beat the Blues", ylab = "BDI",
+ xlab = "Time (in months)", names = c(0, 2, 3, 5, 8),
+ ylim = ylim)
0 3 8
01020304050
Treated as Usual
Time (in months)
BDI
q
qq
0 3 8
01020304050
Beat the Blues
Time (in months)
BDI
Fig. 8.9. Boxplots for the repeated measures by treatment group for the BtheB
data.
Clearly, the simpler random intercept model is adequate for these data. The
results from this model can be found using
R> summary(BtheB_lme1)
Linear mixed-effects model fit by REML
Data: BtheB_long
AIC BIC logLik
1883 1912 -933.5](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-268-320.jpg)
![256 8 The Analysis of Repeated Measures Data
R> bdi <- BtheB[, grep("bdi", names(BtheB))]
R> plot(1:4, rep(-0.5, 4), type = "n", axes = FALSE,
+ ylim = c(0, 50), xlab = "Months", ylab = "BDI")
R> axis(1, at = 1:4, labels = c(0, 2, 3, 5))
R> axis(2)
R> for (i in 1:4) {
+ dropout <- is.na(bdi[,i + 1])
+ points(rep(i, nrow(bdi)) + ifelse(dropout, 0.05, -0.05),
+ jitter(bdi[,i]), pch = ifelse(dropout, 20, 1))
+ }
Months
BDI
0 2 3 5
01020304050
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
qqq
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
q
Fig. 8.10. Distribution of BDI values for patients who do (circles) and do not
(bullets) attend the next scheduled visit.](https://image.slidesharecdn.com/anintroductiontoappliedmultivariateanalysiswithr-everit-130512121546-phpapp02/85/An-introductiontoappliedmultivariateanalysiswithr-everit-272-320.jpg)
An introductiontoappliedmultivariateanalysiswithr everit
- 2. Use R! Series Editors: Robert Gentleman Kurt Hornik Giovanni Parmigiani For other titles published in this series, go to http://www.springer.com/series/6991
- 4. An Introduction to Applied Multivariate Analysis with R Brian Everitt • Torsten Hothorn
- 5. Series Editors: Robert Gentleman Program in Computational Biology Division of Public Health Sciences Fred Hutchinson Cancer Research Center 1100 Fairview Avenue, N. M2-B876 USA Kurt Hornik Department of Statistik and Mathematik Wirtschaftsuniversität Wien A-1090 Wien Austria Giovanni Parmigiani The Sidney Kimmel Comprehensive Cancer Center at Johns Hopkins University 550 North Broadway Baltimore, MD 21205-2011 USA Augasse 2-6 Seattle, Washington 98109 Printed on acid-free paper Springer New York Dordrecht Heidelberg London © Springer Science+Business Media, LLC 2011 software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. subject to proprietary rights. Springer is part of Springer Science+Business Media (www.springer.com) ISBN 978-1-4419-9649-7 e-ISBN 978-1-4419-9650-3 DOI 10.1007/978-1-4419-9650-3 Library of Congress Control Number: 2011926793 All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are King’s College London, SE5 8AF UK brian.everitt@btopenworld.com Ludwigstr. 33 80539 München Germany Torsten.Hothorn@stat.uni-muenchen.de Torsten HothornBrian Everitt Institut für Statistik Ludwig-Maximilians-Universität München Professor Emeritus
- 6. To our wives, Mary-Elizabeth and Carolin.
- 8. Preface The majority of data sets collected by researchers in all disciplines are mul- tivariate, meaning that several measurements, observations, or recordings are taken on each of the units in the data set. These units might be human sub- jects, archaeological artifacts, countries, or a vast variety of other things. In a few cases, it may be sensible to isolate each variable and study it separately, but in most instances all the variables need to be examined simultaneously in order to fully grasp the structure and key features of the data. For this purpose, one or another method of multivariate analysis might be helpful, and it is with such methods that this book is largely concerned. Multivariate analysis includes methods both for describing and exploring such data and for making formal inferences about them. The aim of all the techniques is, in a general sense, to display or extract the signal in the data in the presence of noise and to find out what the data show us in the midst of their apparent chaos. The computations involved in applying most multivariate techniques are considerable, and their routine use requires a suitable software package. In addition, most analyses of multivariate data should involve the construction of appropriate graphs and diagrams, and this will also need to be carried out using the same package. R is a statistical computing environment that is powerful, flexible, and, in addition, has excellent graphical facilities. It is for these reasons that it is the use of R for multivariate analysis that is illustrated in this book. In this book, we concentrate on what might be termed the “core” or “clas- sical” multivariate methodology, although mention will be made of recent de- velopments where these are considered relevant and useful. But there is an area of multivariate statistics that we have omitted from this book, and that is multivariate analysis of variance (MANOVA) and related techniques such as Fisher’s linear discriminant function (LDF). There are a variety of reasons for this omission. First, we are not convinced that MANOVA is now of much more than historical interest; researchers may occasionally pay lip service to using the technique, but in most cases it really is no more than this. They quickly
- 9. viii Preface move on to looking at the results for individual variables. And MANOVA for repeated measures has been largely superseded by the models that we shall describe in Chapter 8. Second, a classification technique such as LDF needs to be considered in the context of modern classification algorithms, and these cannot be covered in an introductory book such as this. Some brief details of the theory behind each technique described are given, but the main concern of each chapter is the correct application of the meth- ods so as to extract as much information as possible from the data at hand, particularly as some type of graphical representation, via the R software. The book is aimed at students in applied statistics courses, both under- graduate and post-graduate, who have attended a good introductory course in statistics that covered hypothesis testing, confidence intervals, simple re- gression and correlation, analysis of variance, and basic maximum likelihood estimation. We also assume that readers will know some simple matrix alge- bra, including the manipulation of matrices and vectors and the concepts of the inverse and rank of a matrix. In addition, we assume that readers will have some familiarity with R at the level of, say, Dalgaard (2002). In addition to such a student readership, we hope that many applied statisticians dealing with multivariate data will find something of interest in the eight chapters of our book. Throughout the book, we give many examples of R code used to apply the multivariate techniques to multivariate data. Samples of code that could be entered interactively at the R command line are formatted as follows: R> library("MVA") Here, R> denotes the prompt sign from the R command line, and the user enters everything else. The symbol + indicates additional lines, which are appropriately indented. Finally, output produced by function calls is shown below the associated code: R> rnorm(10) [1] 1.8808 0.2572 -0.3412 0.4081 0.4344 0.7003 1.8944 [8] -0.2993 -0.7355 0.8960 In this book, we use several R packages to access different example data sets (many of them contained in the package HSAUR2), standard functions for the general parametric analyses, and the MVA package to perform analyses. All of the packages used in this book are available at the Comprehensive R Archive Network (CRAN), which can be accessed from http://CRAN.R-project.org. The source code for the analyses presented in this book is available from the MVA package. A demo containing the R code to reproduce the individual results is available for each chapter by invoking R> library("MVA") R> demo("Ch-MVA") ### Introduction to Multivariate Analysis R> demo("Ch-Viz") ### Visualization
- 10. Preface ix R> demo("Ch-PCA") ### Principal Components Analysis R> demo("Ch-EFA") ### Exploratory Factor Analysis R> demo("Ch-MDS") ### Multidimensional Scaling R> demo("Ch-CA") ### Cluster Analysis R> demo("Ch-SEM") ### Structural Equation Models R> demo("Ch-LME") ### Linear Mixed-Effects Models Thanks are due to Lisa M¨ost, BSc., for help with data processing and LATEX typesetting, the copy editor for many helpful corrections, and to John Kimmel, for all his support and patience during the writing of the book. January 2011 Brian S. Everitt, London Torsten Hothorn, M¨unchen
- 12. Contents Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii 1 Multivariate Data and Multivariate Analysis . . . . . . . . . . . . . . 1 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 A brief history of the development of multivariate analysis . . . . 3 1.3 Types of variables and the possible problem of missing values . 4 1.3.1 Missing values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.4 Some multivariate data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.5 Covariances, correlations, and distances . . . . . . . . . . . . . . . . . . . . 12 1.5.1 Covariances. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 1.5.2 Correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.5.3 Distances. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 1.6 The multivariate normal density function . . . . . . . . . . . . . . . . . . . 15 1.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 1.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2 Looking at Multivariate Data: Visualisation . . . . . . . . . . . . . . . 25 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 2.2 The scatterplot. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.2.1 The bivariate boxplot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 2.2.2 The convex hull of bivariate data . . . . . . . . . . . . . . . . . . . . 32 2.2.3 The chi-plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.3 The bubble and other glyph plots . . . . . . . . . . . . . . . . . . . . . . . . . 34 2.4 The scatterplot matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 2.5 Enhancing the scatterplot with estimated bivariate densities . . 42 2.5.1 Kernel density estimators . . . . . . . . . . . . . . . . . . . . . . . . . . 42 2.6 Three-dimensional plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 2.7 Trellis graphics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 2.8 Stalactite plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 2.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 2.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
- 13. xii Contents 3 Principal Components Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 3.2 Principal components analysis (PCA) . . . . . . . . . . . . . . . . . . . . . . 61 3.3 Finding the sample principal components . . . . . . . . . . . . . . . . . . . 63 3.4 Should principal components be extracted from the covariance or the correlation matrix? . . . . . . . . . . . . . . . . . . . . . . . 65 3.5 Principal components of bivariate data with correlation coefficient r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 3.6 Rescaling the principal components . . . . . . . . . . . . . . . . . . . . . . . . 70 3.7 How the principal components predict the observed covariance matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 3.8 Choosing the number of components . . . . . . . . . . . . . . . . . . . . . . . 71 3.9 Calculating principal components scores . . . . . . . . . . . . . . . . . . . . 72 3.10 Some examples of the application of principal components analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 3.10.1 Head lengths of first and second sons . . . . . . . . . . . . . . . . 74 3.10.2 Olympic heptathlon results . . . . . . . . . . . . . . . . . . . . . . . . . 78 3.10.3 Air pollution in US cities . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 3.11 The biplot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 3.12 Sample size for principal components analysis . . . . . . . . . . . . . . . 93 3.13 Canonical correlation analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 3.13.1 Head measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 3.13.2 Health and personality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 3.14 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 3.15 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 4 Multidimensional Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 4.2 Models for proximity data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 4.3 Spatial models for proximities: Multidimensional scaling . . . . . . 106 4.4 Classical multidimensional scaling . . . . . . . . . . . . . . . . . . . . . . . . . 106 4.4.1 Classical multidimensional scaling: Technical details . . . 107 4.4.2 Examples of classical multidimensional scaling . . . . . . . . 110 4.5 Non-metric multidimensional scaling . . . . . . . . . . . . . . . . . . . . . . . 121 4.5.1 House of Representatives voting . . . . . . . . . . . . . . . . . . . . . 123 4.5.2 Judgements of World War II leaders . . . . . . . . . . . . . . . . . 124 4.6 Correspondence analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 4.6.1 Teenage relationships . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 4.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 5 Exploratory Factor Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 5.2 A simple example of a factor analysis model . . . . . . . . . . . . . . . . 136 5.3 The k-factor analysis model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
- 14. Contents xiii 5.4 Scale invariance of the k-factor model . . . . . . . . . . . . . . . . . . . . . . 138 5.5 Estimating the parameters in the k-factor analysis model . . . . . 139 5.5.1 Principal factor analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 5.5.2 Maximum likelihood factor analysis . . . . . . . . . . . . . . . . . . 142 5.6 Estimating the number of factors . . . . . . . . . . . . . . . . . . . . . . . . . . 142 5.7 Factor rotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 5.8 Estimating factor scores . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 5.9 Two examples of exploratory factor analysis . . . . . . . . . . . . . . . . 148 5.9.1 Expectations of life . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 5.9.2 Drug use by American college students . . . . . . . . . . . . . . . 151 5.10 Factor analysis and principal components analysis compared . . 157 5.11 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 5.12 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 6 Cluster Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163 6.2 Cluster analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 6.3 Agglomerative hierarchical clustering . . . . . . . . . . . . . . . . . . . . . . 166 6.3.1 Clustering jet fighters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 6.4 K-means clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 6.4.1 Clustering the states of the USA on the basis of their crime rate profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 6.4.2 Clustering Romano-British pottery . . . . . . . . . . . . . . . . . . 180 6.5 Model-based clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 6.5.1 Finite mixture densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 6.5.2 Maximum likelihood estimation in a finite mixture density with multivariate normal components . . . . . . . . . 187 6.6 Displaying clustering solutions graphically . . . . . . . . . . . . . . . . . . 191 6.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 6.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200 7 Confirmatory Factor Analysis and Structural Equation Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 7.2 Estimation, identification, and assessing fit for confirmatory factor and structural equation models . . . . . . . . . . . . . . . . . . . . . . 202 7.2.1 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 7.2.2 Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203 7.2.3 Assessing the fit of a model . . . . . . . . . . . . . . . . . . . . . . . . . 204 7.3 Confirmatory factor analysis models . . . . . . . . . . . . . . . . . . . . . . . 206 7.3.1 Ability and aspiration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 7.3.2 A confirmatory factor analysis model for drug use . . . . . 211 7.4 Structural equation models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 7.4.1 Stability of alienation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
- 15. xiv Contents 7.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 8 The Analysis of Repeated Measures Data . . . . . . . . . . . . . . . . . . 225 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 8.2 Linear mixed-effects models for repeated measures data . . . . . . 232 8.2.1 Random intercept and random intercept and slope models for the timber slippage data . . . . . . . . . . . . . . . . . . 233 8.2.2 Applying the random intercept and the random intercept and slope models to the timber slippage data . 235 8.2.3 Fitting random-effect models to the glucose challenge data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240 8.3 Prediction of random effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247 8.4 Dropouts in longitudinal data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248 8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257 8.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
- 17. 1 Multivariate Data and Multivariate Analysis 1.1 Introduction Multivariate data arise when researchers record the values of several random variables on a number of subjects or objects or perhaps one of a variety of other things (we will use the general term“units”) in which they are interested, leading to a vector-valued or multidimensional observation for each. Such data are collected in a wide range of disciplines, and indeed it is probably reasonable to claim that the majority of data sets met in practise are multivariate. In some studies, the variables are chosen by design because they are known to be essential descriptors of the system under investigation. In other studies, particularly those that have been difficult or expensive to organise, many variables may be measured simply to collect as much information as possible as a matter of expediency or economy. Multivariate data are ubiquitous as is illustrated by the following four examples: ˆ Psychologists and other behavioural scientists often record the values of several different cognitive variables on a number of subjects. ˆ Educational researchers may be interested in the examination marks ob- tained by students for a variety of different subjects. ˆ Archaeologists may make a set of measurements on artefacts of interest. ˆ Environmentalists might assess pollution levels of a set of cities along with noting other characteristics of the cities related to climate and human ecology. Most multivariate data sets can be represented in the same way, namely in a rectangular format known from spreadsheets, in which the elements of each row correspond to the variable values of a particular unit in the data set and the elements of the columns correspond to the values taken by a particular variable. We can write data in such a rectangular format as 1 DOI 10.1007/978-1-4419-9650-3_1, © Springer Science+Business Media, LLC 2011 B. Everitt and T. Hothorn, An Introduction to Applied Multivariate Analysis with R: Use R!,
- 18. 2 1 Multivariate Data and Multivariate Analysis Unit Variable 1 . . . Variable q 1 x11 . . . x1q ... ... ... ... n xn1 . . . xnq where n is the number of units, q is the number of variables recorded on each unit, and xij denotes the value of the jth variable for the ith unit. The observation part of the table above is generally represented by an n × q data matrix, X. In contrast to the observed data, the theoretical entities describing the univariate distributions of each of the q variables and their joint distribution are denoted by so-called random variables X1, . . . , Xq. Although in some cases where multivariate data have been collected it may make sense to isolate each variable and study it separately, in the main it does not. Because the whole set of variables is measured on each unit, the variables will be related to a greater or lesser degree. Consequently, if each variable is analysed in isolation, the full structure of the data may not be revealed. Multivariate statistical analysis is the simultaneous statistical analysis of a collection of variables, which improves upon separate univariate analyses of each variable by using information about the relationships between the vari- ables. Analysis of each variable separately is very likely to miss uncovering the key features of, and any interesting “patterns” in, the multivariate data. The units in a set of multivariate data are sometimes sampled from a population of interest to the investigator, a population about which he or she wishes to make some inference or other. More often perhaps, the units cannot really be said to have been sampled from some population in any meaningful sense, and the questions asked about the data are then largely exploratory in nature. with the ubiquitous p-value of univariate statistics being notable by its absence. Consequently, there are methods of multivariate analysis that are essentially exploratory and others that can be used for statistical inference. For the exploration of multivariate data, formal models designed to yield specific answers to rigidly defined questions are not required. Instead, meth- ods are used that allow the detection of possibly unanticipated patterns in the data, opening up a wide range of competing explanations. Such methods are generally characterised both by an emphasis on the importance of graphical displays and visualisation of the data and the lack of any associated proba- bilistic model that would allow for formal inferences. Multivariate techniques that are largely exploratory are described in Chapters 2 to 6. A more formal analysis becomes possible in situations when it is realistic to assume that the individuals in a multivariate data set have been sampled from some population and the investigator wishes to test a well-defined hypothe- sis about the parameters of that population’s probability density function. Now the main focus will not be the sample data per se, but rather on using information gathered from the sample data to draw inferences about the pop- ulation. And the probability density function almost universally assumed as the basis of inferences for multivariate data is the multivariate normal. (For
- 19. 1.2 A brief history of the development of multivariate analysis 3 a brief description of the multivariate normal density function and ways of assessing whether a set of multivariate data conform to the density, see Sec- tion 1.6). Multivariate techniques for which formal inference is of importance are described in Chapters 7 and 8. But in many cases when dealing with multivariate data, this implied distinction between the exploratory and the inferential may be a red herring because the general aim of most multivariate analyses, whether implicitly exploratory or inferential is to uncover, display, or extract any “signal” in the data in the presence of noise and to discover what the data have to tell us. 1.2 A brief history of the development of multivariate analysis The genesis of multivariate analysis is probably the work carried out by Francis Galton and Karl Pearson in the late 19th century on quantifying the relation- ship between offspring and parental characteristics and the development of the correlation coefficient. And then, in the early years of the 20th century, Charles Spearman laid down the foundations of factor analysis (see Chapter 5) whilst investigating correlated intelligence quotient (IQ) tests. Over the next two decades, Spearman’s work was extended by Hotelling and by Thurstone. Multivariate methods were also motivated by problems in scientific areas other than psychology, and in the 1930s Fisher developed linear discriminant function analysis to solve a taxonomic problem using multiple botanical mea- surements. And Fisher’s introduction of analysis of variance in the 1920s was soon followed by its multivariate generalisation, multivariate analysis of vari- ance, based on work by Bartlett and Roy. (These techniques are not covered in this text for the reasons set out in the Preface.) In these early days, computational aids to take the burden of the vast amounts of arithmetic involved in the application of the multivariate meth- ods being proposed were very limited and, consequently, developments were primarily mathematical and multivariate research was, at the time, largely a branch of linear algebra. However, the arrival and rapid expansion of the use of electronic computers in the second half of the 20th century led to increased practical application of existing methods of multivariate analysis and renewed interest in the creation of new techniques. In the early years of the 21st century, the wide availability of relatively cheap and extremely powerful personal computers and laptops allied with flexible statistical software has meant that all the methods of multivariate analysis can be applied routinely even to very large data sets such as those generated in, for example, genetics, imaging, and astronomy. And the appli- cation of multivariate techniques to such large data sets has now been given its own name, data mining, which has been defined as “the nontrivial extrac- tion of implicit, previously unknown and potentially useful information from
- 20. 4 1 Multivariate Data and Multivariate Analysis data.” Useful books on data mining are those of Fayyad, Piatetsky-Shapiro, Smyth, and Uthurusamy (1996) and Hand, Mannila, and Smyth (2001). 1.3 Types of variables and the possible problem of missing values A hypothetical example of multivariate data is given in Table 1.1. The special symbol NA denotes missing values (being Not Available); the value of this variable for a subject is missing. Table 1.1: hypo data. Hypothetical Set of Multivariate Data. individual sex age IQ depression health weight 1 Male 21 120 Yes Very good 150 2 Male 43 NA No Very good 160 3 Male 22 135 No Average 135 4 Male 86 150 No Very poor 140 5 Male 60 92 Yes Good 110 6 Female 16 130 Yes Good 110 7 Female NA 150 Yes Very good 120 8 Female 43 NA Yes Average 120 9 Female 22 84 No Average 105 10 Female 80 70 No Good 100 Here, the number of units (people in this case) is n = 10, with the number of variables being q = 7 and, for example, x34 = 135. In R, a “data.frame” is the appropriate data structure to represent such rectangular data. Subsets of units (rows) or variables (columns) can be extracted via the [ subset operator; i.e., R> hypo[1:2, c("health", "weight")] health weight 1 Very good 150 2 Very good 160 extracts the values x15, x16 and x25, x26 from the hypothetical data presented in Table 1.1. These data illustrate that the variables that make up a set of multivariate data will not necessarily all be of the same type. Four levels of measurements are often distinguished: Nominal: Unordered categorical variables. Examples include treatment al- location, the sex of the respondent, hair colour, presence or absence of depression, and so on.
- 21. 1.3 Types of variables and the possible problem of missing values 5 Ordinal: Where there is an ordering but no implication of equal distance between the different points of the scale. Examples include social class, self-perception of health (each coded from I to V, say), and educational level (no schooling, primary, secondary, or tertiary education). Interval: Where there are equal differences between successive points on the scale but the position of zero is arbitrary. The classic example is the mea- surement of temperature using the Celsius or Fahrenheit scales. Ratio: The highest level of measurement, where one can investigate the rel- ative magnitudes of scores as well as the differences between them. The position of zero is fixed. The classic example is the absolute measure of temperature (in Kelvin, for example), but other common ones includes age (or any other time from a fixed event), weight, and length. In many statistical textbooks, discussion of different types of measure- ments is often followed by recommendations as to which statistical techniques are suitable for each type; for example, analyses on nominal data should be limited to summary statistics such as the number of cases, the mode, etc. And, for ordinal data, means and standard deviations are not suitable. But Velleman and Wilkinson (1993) make the important point that restricting the choice of statistical methods in this way may be a dangerous practise for data analysis–in essence the measurement taxonomy described is often too strict to apply to real-world data. This is not the place for a detailed discussion of measurement, but we take a fairly pragmatic approach to such problems. For example, we will not agonise over treating variables such as measures of de- pression, anxiety, or intelligence as if they are interval-scaled, although strictly they fit into the ordinal category described above. 1.3.1 Missing values Table 1.1 also illustrates one of the problems often faced by statisticians un- dertaking statistical analysis in general and multivariate analysis in particular, namely the presence of missing values in the data; i.e., observations and mea- surements that should have been recorded but for one reason or another, were not. Missing values in multivariate data may arise for a number of reasons; for example, non-response in sample surveys, dropouts in longitudinal data (see Chapter 8), or refusal to answer particular questions in a questionnaire. The most important approach for dealing with missing data is to try to avoid them during the data-collection stage of a study. But despite all the efforts a researcher may make, he or she may still be faced with a data set that con- tains a number of missing values. So what can be done? One answer to this question is to take the complete-case analysis route because this is what most statistical software packages do automatically. Using complete-case analysis on multivariate data means omitting any case with a missing value on any of the variables. It is easy to see that if the number of variables is large, then even a sparse pattern of missing values can result in a substantial number of incomplete cases. One possibility to ease this problem is to simply drop any
- 22. 6 1 Multivariate Data and Multivariate Analysis variables that have many missing values. But complete-case analysis is not recommended for two reasons: ˆ Omitting a possibly substantial number of individuals will cause a large amount of information to be discarded and lower the effective sample size of the data, making any analyses less effective than they would have been if all the original sample had been available. ˆ More worrisome is that dropping the cases with missing values on one or more variables can lead to serious biases in both estimation and infer- ence unless the discarded cases are essentially a random subsample of the observed data (the term missing completely at random is often used; see Chapter 8 and Little and Rubin (1987) for more details). So, at the very least, complete-case analysis leads to a loss, and perhaps a substantial loss, in power by discarding data, but worse, analyses based just on complete cases might lead to misleading conclusions and inferences. A relatively simple alternative to complete-case analysis that is often used is available-case analysis. This is a straightforward attempt to exploit the incomplete information by using all the cases available to estimate quanti- ties of interest. For example, if the researcher is interested in estimating the correlation matrix (see Subsection 1.5.2) of a set of multivariate data, then available-case analysis uses all the cases with variables Xi and Xj present to estimate the correlation between the two variables. This approach appears to make better use of the data than complete-case analysis, but unfortunately available-case analysis has its own problems. The sample of individuals used changes from correlation to correlation, creating potential difficulties when the missing data are not missing completely at random. There is no guaran- tee that the estimated correlation matrix is even positive-definite which can create problems for some of the methods, such as factor analysis (see Chap- ter 5) and structural equation modelling (see Chapter 7), that the researcher may wish to apply to the matrix. Both complete-case and available-case analyses are unattractive unless the number of missing values in the data set is “small”. An alternative answer to the missing-data problem is to consider some form of imputation, the prac- tise of “filling in” missing data with plausible values. Methods that impute the missing values have the advantage that, unlike in complete-case analysis, observed values in the incomplete cases are retained. On the surface, it looks like imputation will solve the missing-data problem and enable the investi- gator to progress normally. But, from a statistical viewpoint, careful consid- eration needs to be given to the method used for imputation or otherwise it may cause more problems than it solves; for example, imputing an observed variable mean for a variable’s missing values preserves the observed sample means but distorts the covariance matrix (see Subsection 1.5.1), biasing esti- mated variances and covariances towards zero. On the other hand, imputing predicted values from regression models tends to inflate observed correlations, biasing them away from zero (see Little 2005). And treating imputed data as
- 23. 1.4 Some multivariate data sets 7 if they were“real” in estimation and inference can lead to misleading standard errors and p-values since they fail to reflect the uncertainty due to the missing data. The most appropriate way to deal with missing values is by a procedure suggested by Rubin (1987) known as multiple imputation. This is a Monte Carlo technique in which the missing values are replaced by m > 1 simulated versions, where m is typically small (say 3–10). Each of the simulated com- plete data sets is analysed using the method appropriate for the investigation at hand, and the results are later combined to produce, say, estimates and con- fidence intervals that incorporate missing-data uncertainty. Details are given in Rubin (1987) and more concisely in Schafer (1999). The great virtues of multiple imputation are its simplicity and its generality. The user may analyse the data using virtually any technique that would be appropriate if the data were complete. However, one should always bear in mind that the imputed values are not real measurements. We do not get something for nothing! And if there is a substantial proportion of individuals with large amounts of miss- ing data, one should clearly question whether any form of statistical analysis is worth the bother. 1.4 Some multivariate data sets This is a convenient point to look at some multivariate data sets and briefly ponder the type of question that might be of interest in each case. The first data set consists of chest, waist, and hip measurements on a sample of men and women and the measurements for 20 individuals are shown in Table 1.2. Two questions might be addressed by such data; ˆ Could body size and body shape be summarised in some way by combining the three measurements into a single number? ˆ Are there subtypes of body shapes amongst the men and amongst the women within which individuals are of similar shapes and between which body shapes differ? The first question might be answered by principal components analysis (see Chapter 3), and the second question could be investigated using cluster anal- ysis (see Chapter 6). (In practise, it seems intuitively likely that we would have needed to record the three measurements on many more than 20 individuals to have any chance of being able to get convincing answers from these techniques to the questions of interest. The question of how many units are needed to achieve a sensible analysis when using the various techniques of multivariate analysis will be taken up in the respective chapters describing each technique.)
- 24. 8 1 Multivariate Data and Multivariate Analysis Table 1.2: measure data. Chest, waist, and hip measurements on 20 individuals (in inches). chest waist hips gender chest waist hips gender 34 30 32 male 36 24 35 female 37 32 37 male 36 25 37 female 38 30 36 male 34 24 37 female 36 33 39 male 33 22 34 female 38 29 33 male 36 26 38 female 43 32 38 male 37 26 37 female 40 33 42 male 34 25 38 female 38 30 40 male 36 26 37 female 40 30 37 male 38 28 40 female 41 32 39 male 35 23 35 female Our second set of multivariate data consists of the results of chemical analysis on Romano-British pottery made in three different regions (region 1 contains kiln 1, region 2 contains kilns 2 and 3, and region 3 contains kilns 4 and 5). The complete data set, which we shall meet in Chapter 6, consists of the chemical analysis results on 45 pots, shown in Table 1.3. One question that might be posed about these data is whether the chemical profiles of each pot suggest different types of pots and if any such types are related to kiln or region. This question is addressed in Chapter 6. Table 1.3: pottery data. Romano-British pottery data. Al2O3 Fe2O3 MgO CaO Na2O K2O TiO2 MnO BaO kiln 18.8 9.52 2.00 0.79 0.40 3.20 1.01 0.077 0.015 1 16.9 7.33 1.65 0.84 0.40 3.05 0.99 0.067 0.018 1 18.2 7.64 1.82 0.77 0.40 3.07 0.98 0.087 0.014 1 16.9 7.29 1.56 0.76 0.40 3.05 1.00 0.063 0.019 1 17.8 7.24 1.83 0.92 0.43 3.12 0.93 0.061 0.019 1 18.8 7.45 2.06 0.87 0.25 3.26 0.98 0.072 0.017 1 16.5 7.05 1.81 1.73 0.33 3.20 0.95 0.066 0.019 1 18.0 7.42 2.06 1.00 0.28 3.37 0.96 0.072 0.017 1 15.8 7.15 1.62 0.71 0.38 3.25 0.93 0.062 0.017 1 14.6 6.87 1.67 0.76 0.33 3.06 0.91 0.055 0.012 1 13.7 5.83 1.50 0.66 0.13 2.25 0.75 0.034 0.012 1 14.6 6.76 1.63 1.48 0.20 3.02 0.87 0.055 0.016 1 14.8 7.07 1.62 1.44 0.24 3.03 0.86 0.080 0.016 1 17.1 7.79 1.99 0.83 0.46 3.13 0.93 0.090 0.020 1 16.8 7.86 1.86 0.84 0.46 2.93 0.94 0.094 0.020 1 15.8 7.65 1.94 0.81 0.83 3.33 0.96 0.112 0.019 1
- 25. 1.4 Some multivariate data sets 9 Table 1.3: pottery data (continued). Al2O3 Fe2O3 MgO CaO Na2O K2O TiO2 MnO BaO kiln 18.6 7.85 2.33 0.87 0.38 3.17 0.98 0.081 0.018 1 16.9 7.87 1.83 1.31 0.53 3.09 0.95 0.092 0.023 1 18.9 7.58 2.05 0.83 0.13 3.29 0.98 0.072 0.015 1 18.0 7.50 1.94 0.69 0.12 3.14 0.93 0.035 0.017 1 17.8 7.28 1.92 0.81 0.18 3.15 0.90 0.067 0.017 1 14.4 7.00 4.30 0.15 0.51 4.25 0.79 0.160 0.019 2 13.8 7.08 3.43 0.12 0.17 4.14 0.77 0.144 0.020 2 14.6 7.09 3.88 0.13 0.20 4.36 0.81 0.124 0.019 2 11.5 6.37 5.64 0.16 0.14 3.89 0.69 0.087 0.009 2 13.8 7.06 5.34 0.20 0.20 4.31 0.71 0.101 0.021 2 10.9 6.26 3.47 0.17 0.22 3.40 0.66 0.109 0.010 2 10.1 4.26 4.26 0.20 0.18 3.32 0.59 0.149 0.017 2 11.6 5.78 5.91 0.18 0.16 3.70 0.65 0.082 0.015 2 11.1 5.49 4.52 0.29 0.30 4.03 0.63 0.080 0.016 2 13.4 6.92 7.23 0.28 0.20 4.54 0.69 0.163 0.017 2 12.4 6.13 5.69 0.22 0.54 4.65 0.70 0.159 0.015 2 13.1 6.64 5.51 0.31 0.24 4.89 0.72 0.094 0.017 2 11.6 5.39 3.77 0.29 0.06 4.51 0.56 0.110 0.015 3 11.8 5.44 3.94 0.30 0.04 4.64 0.59 0.085 0.013 3 18.3 1.28 0.67 0.03 0.03 1.96 0.65 0.001 0.014 4 15.8 2.39 0.63 0.01 0.04 1.94 1.29 0.001 0.014 4 18.0 1.50 0.67 0.01 0.06 2.11 0.92 0.001 0.016 4 18.0 1.88 0.68 0.01 0.04 2.00 1.11 0.006 0.022 4 20.8 1.51 0.72 0.07 0.10 2.37 1.26 0.002 0.016 4 17.7 1.12 0.56 0.06 0.06 2.06 0.79 0.001 0.013 5 18.3 1.14 0.67 0.06 0.05 2.11 0.89 0.006 0.019 5 16.7 0.92 0.53 0.01 0.05 1.76 0.91 0.004 0.013 5 14.8 2.74 0.67 0.03 0.05 2.15 1.34 0.003 0.015 5 19.1 1.64 0.60 0.10 0.03 1.75 1.04 0.007 0.018 5 Source: Tubb, A., et al., Archaeometry, 22, 153–171, 1980. With permission. Our third set of multivariate data involves the examination scores of a large number of college students in six subjects; the scores for five subjects are shown in Table 1.4. Here the main question of interest might be whether the exam scores reflect some underlying trait in a student that cannot be measured directly, perhaps “general intelligence”? The question could be investigated by using exploratory factor analysis (see Chapter 5).
- 26. 10 1 Multivariate Data and Multivariate Analysis Table 1.4: exam data. Exam scores for five psychology students. subject maths english history geography chemistry physics 1 60 70 75 58 53 42 2 80 65 66 75 70 76 3 53 60 50 48 45 43 4 85 79 71 77 68 79 5 45 80 80 84 44 46 The final set of data we shall consider in this section was collected in a study of air pollution in cities in the USA. The following variables were obtained for 41 US cities: SO2: SO2 content of air in micrograms per cubic metre; temp: average annual temperature in degrees Fahrenheit; manu: number of manufacturing enterprises employing 20 or more workers; popul: population size (1970 census) in thousands; wind: average annual wind speed in miles per hour; precip: average annual precipitation in inches; predays: average number of days with precipitation per year. The data are shown in Table 1.5. Table 1.5: USairpollution data. Air pollution in 41 US cities. SO2 temp manu popul wind precip predays Albany 46 47.6 44 116 8.8 33.36 135 Albuquerque 11 56.8 46 244 8.9 7.77 58 Atlanta 24 61.5 368 497 9.1 48.34 115 Baltimore 47 55.0 625 905 9.6 41.31 111 Buffalo 11 47.1 391 463 12.4 36.11 166 Charleston 31 55.2 35 71 6.5 40.75 148 Chicago 110 50.6 3344 3369 10.4 34.44 122 Cincinnati 23 54.0 462 453 7.1 39.04 132 Cleveland 65 49.7 1007 751 10.9 34.99 155 Columbus 26 51.5 266 540 8.6 37.01 134 Dallas 9 66.2 641 844 10.9 35.94 78 Denver 17 51.9 454 515 9.0 12.95 86 Des Moines 17 49.0 104 201 11.2 30.85 103 Detroit 35 49.9 1064 1513 10.1 30.96 129 Hartford 56 49.1 412 158 9.0 43.37 127 Houston 10 68.9 721 1233 10.8 48.19 103 Indianapolis 28 52.3 361 746 9.7 38.74 121 Jacksonville 14 68.4 136 529 8.8 54.47 116
- 27. 1.4 Some multivariate data sets 11 Table 1.5: USairpollution data (continued). SO2 temp manu popul wind precip predays Kansas City 14 54.5 381 507 10.0 37.00 99 Little Rock 13 61.0 91 132 8.2 48.52 100 Louisville 30 55.6 291 593 8.3 43.11 123 Memphis 10 61.6 337 624 9.2 49.10 105 Miami 10 75.5 207 335 9.0 59.80 128 Milwaukee 16 45.7 569 717 11.8 29.07 123 Minneapolis 29 43.5 699 744 10.6 25.94 137 Nashville 18 59.4 275 448 7.9 46.00 119 New Orleans 9 68.3 204 361 8.4 56.77 113 Norfolk 31 59.3 96 308 10.6 44.68 116 Omaha 14 51.5 181 347 10.9 30.18 98 Philadelphia 69 54.6 1692 1950 9.6 39.93 115 Phoenix 10 70.3 213 582 6.0 7.05 36 Pittsburgh 61 50.4 347 520 9.4 36.22 147 Providence 94 50.0 343 179 10.6 42.75 125 Richmond 26 57.8 197 299 7.6 42.59 115 Salt Lake City 28 51.0 137 176 8.7 15.17 89 San Francisco 12 56.7 453 716 8.7 20.66 67 Seattle 29 51.1 379 531 9.4 38.79 164 St. Louis 56 55.9 775 622 9.5 35.89 105 Washington 29 57.3 434 757 9.3 38.89 111 Wichita 8 56.6 125 277 12.7 30.58 82 Wilmington 36 54.0 80 80 9.0 40.25 114 Source: Sokal, R. R., Rohlf, F. J., Biometry, W. H. Freeman, San Francisco, 1981. With permission. What might be the question of most interest about these data? Very prob- ably it is “how is pollution level as measured by sulphur dioxide concentration related to the six other variables?” In the first instance at least, this question suggests the application of multiple linear regression, with sulphur dioxide concentration as the response variable and the remaining six variables being the independent or explanatory variables (the latter is a more acceptable label because the “independent” variables are rarely independent of one another). But in the model underlying multiple regression, only the response is consid- ered to be a random variable; the explanatory variables are strictly assumed to be fixed, not random, variables. In practise, of course, this is rarely the case, and so the results from a multiple regression analysis need to be interpreted as being conditional on the observed values of the explanatory variables. So when answering the question of most interest about these data, they should not really be considered multivariate–there is only a single random variable involved–a more suitable label is multivariable (we know this sounds pedantic,
- 28. 12 1 Multivariate Data and Multivariate Analysis but we are statisticians after all). In this book, we shall say only a little about the multiple linear model for multivariable data in Chapter 8. but essentially only to enable such regression models to be introduced for situations where there is a multivariate response; for example, in the case of repeated-measures data and longitudinal data. The four data sets above have not exhausted either the questions that multivariate data may have been collected to answer or the methods of mul- tivariate analysis that have been developed to answer them, as we shall see as we progress through the book. 1.5 Covariances, correlations, and distances The main reason why we should analyse a multivariate data set using multi- variate methods rather than looking at each variable separately using one or another familiar univariate method is that any structure or pattern in the data is as likely to be implied either by “relationships” between the variables or by the relative “closeness” of different units as by their different variable values; in some cases perhaps by both. In the first case, any structure or pattern un- covered will be such that it “links” together the columns of the data matrix, X, in some way, and in the second case a possible structure that might be discovered is that involving interesting subsets of the units. The question now arises as to how we quantify the relationships between the variables and how we measure the distances between different units. This question is answered in the subsections that follow. 1.5.1 Covariances The covariance of two random variables is a measure of their linear depen- dence. The population (theoretical) covariance of two random variables, Xi and Xj, is defined by Cov(Xi, Xj) = E(Xi − µi)(Xj − µj), where µi = E(Xi) and µj = E(Xj); E denotes expectation. If i = j, we note that the covariance of the variable with itself is simply its variance, and therefore there is no need to define variances and covariances independently in the multivariate case. If Xi and Xj are independent of each other, their covariance is necessarily equal to zero, but the converse is not true. The covariance of Xi and Xj is usually denoted by σij. The variance of variable Xi is σ2 i = E (Xi − µi)2 . Larger values of the covariance imply a greater degree of linear dependence between two variables. In a multivariate data set with q observed variables, there are q variances and q(q − 1)/2 covariances. These quantities can be conveniently arranged in a q × q symmetric matrix, Σ, where
- 29. 1.5 Covariances, correlations, and distances 13 Σ = σ2 1 σ12 . . . σ1q σ21 σ2 2 . . . σ2q ... ... ... ... σq1 σq2 . . . σ2 q . Note that σij = σji. This matrix is generally known as the variance-covariance matrix or simply the covariance matrix of the data. For a set of multivariate observations, perhaps sampled from some popu- lation, the matrix Σ is estimated by S = 1 n − 1 n i=1 (xi − ¯x)(xi − ¯x) , where xi = (xi1, xi2, . . . , xiq) is the vector of (numeric) observations for the ith individual and ¯x = n−1 n i=1 xi is the mean vector of the observations. The diagonal of S contains the sample variances of each variable, which we shall denote as s2 i . The covariance matrix for the data in Table 1.2 can be obtained using the var() function in R; however, we have to “remove” the categorical variable gender from the measure data frame by subsetting on the numerical variables first: R> cov(measure[, c("chest", "waist", "hips")]) chest waist hips chest 6.632 6.368 3.000 waist 6.368 12.526 3.579 hips 3.000 3.579 5.945 If we require the separate covariance matrices of men and women, we can use R> cov(subset(measure, gender == "female")[, + c("chest", "waist", "hips")]) chest waist hips chest 2.278 2.167 1.556 waist 2.167 2.989 2.756 hips 1.556 2.756 3.067 R> cov(subset(measure, gender == "male")[, + c("chest", "waist", "hips")]) chest waist hips chest 6.7222 0.9444 3.944 waist 0.9444 2.1000 3.078 hips 3.9444 3.0778 9.344 where the subset() returns all observations corresponding to females (first statement) or males (second statement).
- 30. 14 1 Multivariate Data and Multivariate Analysis 1.5.2 Correlations The covariance is often difficult to interpret because it depends on the scales on which the two variables are measured; consequently, it is often standardised by dividing by the product of the standard deviations of the two variables to give a quantity called the correlation coefficient, ρij, where ρij = σij σiσj , where σi = σ2 i . The advantage of the correlation is that it is independent of the scales of the two variables. The correlation coefficient lies between −1 and +1 and gives a measure of the linear relationship of the variables Xi and Xj. It is positive if high values of Xi are associated with high values of Xj and negative if high values of Xi are associated with low values of Xj. If the relationship between two variables is non-linear, their correlation coefficient can be misleading. With q variables there are q(q − 1)/2 distinct correlations, which may be arranged in a q×q correlation matrix the diagonal elements of which are unity. For observed data, the correlation matrix contains the usual estimates of the ρs, namely Pearson’s correlation coefficient, and is generally denoted by R. The matrix may be written in terms of the sample covariance matrix S R = D−1/2 SD−1/2 , where D−1/2 = diag(1/s1, . . . , 1/sq) and si = s2 i is the sample standard deviation of variable i. (In most situations considered in this book, we will be dealing with covariance and correlation matrices of full rank, q, so that both matrices will be non-singular, that is, invertible, to give matrices S−1 or R−1 .) The sample correlation matrix for the three variables in Table 1.1 is ob- tained by using the function cor() in R: R> cor(measure[, c("chest", "waist", "hips")]) chest waist hips chest 1.0000 0.6987 0.4778 waist 0.6987 1.0000 0.4147 hips 0.4778 0.4147 1.0000 1.5.3 Distances For some multivariate techniques such as multidimensional scaling (see Chap- ter 4) and cluster analysis (see Chapter 6), the concept of distance between the units in the data is often of considerable interest and importance. So, given the variable values for two units, say unit i and unit j, what serves as a measure of distance between them? The most common measure used is Euclidean distance, which is defined as
- 31. 1.6 The multivariate normal density function 15 dij = q k=1 (xik − xjk)2, where xik and xjk, k = 1, . . . , q are the variable values for units i and j, respectively. Euclidean distance can be calculated using the dist() function in R. When the variables in a multivariate data set are on different scales, it makes more sense to calculate the distances after some form of standardisa- tion. Here we shall illustrate this on the body measurement data and divide each variable by its standard deviation using the function scale() before applying the dist() function–the necessary R code and output are R> dist(scale(measure[, c("chest", "waist", "hips")], + center = FALSE)) 1 2 3 4 5 6 7 8 9 10 11 2 0.17 3 0.15 0.08 4 0.22 0.07 0.14 5 0.11 0.15 0.09 0.22 6 0.29 0.16 0.16 0.19 0.21 7 0.32 0.16 0.20 0.13 0.28 0.14 8 0.23 0.11 0.11 0.12 0.19 0.16 0.13 9 0.21 0.10 0.06 0.16 0.12 0.11 0.17 0.09 10 0.27 0.12 0.13 0.14 0.20 0.06 0.09 0.11 0.09 11 0.23 0.28 0.22 0.33 0.19 0.34 0.38 0.25 0.24 0.32 12 0.22 0.24 0.18 0.28 0.18 0.30 0.32 0.20 0.20 0.28 0.06 ... (Note that only the distances for the first 12 observations are shown in the output.) 1.6 The multivariate normal density function Just as the normal distribution dominates univariate techniques, the multi- variate normal distribution plays an important role in some multivariate pro- cedures, although as mentioned earlier many multivariate analyses are carried out in the spirit of data exploration where questions of statistical significance are of relatively minor importance or of no importance at all. Nevertheless, re- searchers dealing with the complexities of multivariate data may, on occasion, need to know a little about the multivariate density function and in particular how to assess whether or not a set of multivariate data can be assumed to have this density function. So we will define the multivariate normal density and describe some of its properties.
- 32. 16 1 Multivariate Data and Multivariate Analysis For a vector of q variables, x = (x1, x2, . . . , xq), the multivariate normal density function takes the form f(x; µ, Σ) = (2π)−q/2 det(Σ)−1/2 exp − 1 2 (x − µ) Σ−1 (x − µ) , where Σ is the population covariance matrix of the variables and µ is the vector of population mean values of the variables. The simplest example of the multivariate normal density function is the bivariate normal density with q = 2; this can be written explicitly as f((x1, x2); (µ1, µ2), σ1, σ2, ρ) = 2πσ1σ2(1 − ρ2 ) −1/2 exp − 1 2(1 − ρ2) × x1 − µ1 σ1 2 − 2ρ x1 − µ1 σ1 x2 − µ2 σ2 + x2 − µ2 σ2 2 , where µ1 and µ2 are the population means of the two variables, σ2 1 and σ2 2 are the population variances, and ρ is the population correlation between the two variables X1 and X2. Figure 1.1 shows an example of a bivariate normal density function with both means equal to zero, both variances equal to one, and correlation equal to 0.5. The population mean vector and the population covariance matrix of a multivariate density function are estimated from a sample of multivariate observations as described in the previous subsections. One property of a multivariate normal density function that is worth mentioning here is that linear combinations of the variables (i.e., y = a1X1 + a2X2 + · · · + aqXq, where a1, a2, . . . , aq is a set of scalars) are themselves normally distributed with mean a µ and variance a Σa, where a = (a1, a2, . . . , aq). Linear combinations of variables will be of importance in later chapters, particularly in Chapter 3. For many multivariate methods to be described in later chapters, the as- sumption of multivariate normality is not critical to the results of the analysis, but there may be occasions when testing for multivariate normality may be of interest. A start can be made perhaps by assessing each variable separately for univariate normality using a probability plot. Such plots are commonly applied in univariate analysis and involve ordering the observations and then plotting them against the appropriate values of an assumed cumulative distribution function. There are two basic types of plots for comparing two probability distributions, the probability-probability plot and the quantile-quantile plot. The diagram in Figure 1.2 may be used for describing each type. A plot of points whose coordinates are the cumulative probabilities p1(q) and p2(q) for different values of q with p1(q) = P(X1 ≤ q), p2(q) = P(X2 ≤ q),
- 33. 1.6 The multivariate normal density function 17 x1 x2 f(x) Fig. 1.1. Bivariate normal density function with correlation ρ = 0.5. for random variables X1 and X2 is a probability-probability plot, while a plot of the points whose coordinates are the quantiles (q1(p), q2(p)) for different values of p with q1(p) = p−1 1 (p), q2(p) = p−1 2 (p), is a quantile-quantile plot. For example, a quantile-quantile plot for investi- gating the assumption that a set of data is from a normal distribution would in- against the quantiles of a standard normal distribution, Φ−1 (p(i)), where usu- ally pi = i − 1 2 n Φ(x) = x −∞ 1 √ 2π e− 1 2 u2 du. This is known as a normal probability plot. volveplottingtheordered sample values of variable 1 (i.e.,x(1)1, x(2)1, . . . , x(n)1)
- 34. 18 1 Multivariate Data and Multivariate Analysis Cumulativedistributionfunction q q2(p) q1(p) 0p1(q)p2(q)p1 Fig. 1.2. Cumulative distribution functions and quantiles. For multivariate data, normal probability plots may be used to examine each variable separately, although marginal normality does not necessarily im- ply that the variables follow a multivariate normal distribution. Alternatively (or additionally), each multivariate observation might be converted to a single number in some way before plotting. For example, in the specific case of as- sessing a data set for multivariate normality, each q-dimensional observation, xi, could be converted into a generalised distance, d2 i , giving a measure of the distance of the particular observation from the mean vector of the complete sample, ¯x; d2 i is calculated as d2 i = (xi − ¯x) S−1 (xi − ¯x), where S is the sample covariance matrix. This distance measure takes into account the different variances of the variables and the covariances of pairs of variables. If the observations do arise from a multivariate normal distribution, then these distances have approximately a chi-squared distribution with q degrees of freedom, also denoted by the symbol χ2 q. So plotting the ordered distances against the corresponding quantiles of the appropriate chi-square distribution should lead to a straight line through the origin. We will now assess the body measurements data in Table 1.2 for normal- ity, although because there are only 20 observations in the sample there is
- 35. 1.6 The multivariate normal density function 19 really too little information to come to any convincing conclusion. Figure 1.3 shows separate probability plots for each measurement; there appears to be no evidence of any departures from linearity. The chi-square plot of the 20 generalised distances in Figure 1.4 does seem to deviate a little from linearity, but with so few observations it is hard to be certain. The plot is set up as follows. We first extract the relevant data R> x <- measure[, c("chest", "waist", "hips")] and estimate the means of all three variables (i.e., for each column of the data) and the covariance matrix R> cm <- colMeans(x) R> S <- cov(x) The differences di have to be computed for all units in our data, so we iterate over the rows of x using the apply() function with argument MARGIN = 1 and, for each row, compute the distance di: R> d <- apply(x, MARGIN = 1, function(x) + t(x - cm) %*% solve(S) %*% (x - cm)) The sorted distances can now be plotted against the appropriate quantiles of the χ2 3 distribution obtained from qchisq(); see Figure 1.4. R> qqnorm(measure[,"chest"], main = "chest"); qqline(measure[,"chest"]) R> qqnorm(measure[,"waist"], main = "waist"); qqline(measure[,"waist"]) R> qqnorm(measure[,"hips"], main = "hips"); qqline(measure[,"hips"]) q q q q q q q q q q qq q q q q q q q q −2 0 1 2 3436384042 chest Theoretical Quantiles SampleQuantiles q q q q q q q qq q q q q q qq q q q q −2 0 1 2 222426283032 waist Theoretical Quantiles SampleQuantiles q q q q q q q q q q q qq q q q q q q q −2 0 1 2 323436384042 hips Theoretical Quantiles SampleQuantiles Fig. 1.3. Normal probability plots of chest, waist, and hip measurements.
- 36. 20 1 Multivariate Data and Multivariate Analysis R> plot(qchisq((1:nrow(x) - 1/2) / nrow(x), df = 3), sort(d), + xlab = expression(paste(chi[3]^2, " Quantile")), + ylab = "Ordered distances") R> abline(a = 0, b = 1) q q qqq q qqqq q q q q q q q q q q 0 2 4 6 8 2468 χ3 2 Quantile Ordereddistances Fig. 1.4. Chi-square plot of generalised distances for body measurements data. We will now look at using the chi-square plot on a set of data introduced early in the chapter, namely the air pollution in US cities (see Table 1.5). The probability plots for each separate variable are shown in Figure 1.5. Here, we also iterate over all variables, this time using a special function, sapply(), that loops over the variable names: R> layout(matrix(1:8, nc = 2)) R> sapply(colnames(USairpollution), function(x) { + qqnorm(USairpollution[[x]], main = x) + qqline(USairpollution[[x]]) + })
- 37. 1.6 The multivariate normal density function 21 q q q q q q q q q q q qq q q q q qqq q qq q q q q q q q q q q qq q q q q q q −2 −1 0 1 2 2080 SO2 Theoretical Quantiles SampleQuantiles q q q q q q q q q q q q q qq q q q q q q q q qq q q q q q q qq q q q q q qq q −2 −1 0 1 2 456075 temp Theoretical Quantiles SampleQuantiles q q q q q q q q q q qq q q q q qq q q qqq q q qqq q q q qqqq qq q q qq −2 −1 0 1 2 02000 manu Theoretical Quantiles SampleQuantiles q q q q q q q q qq q q q q q q qqq q qq q qq qqqq q qq q qq qq q q qq −2 −1 0 1 2 02000 popul Theoretical Quantiles SampleQuantiles qq q q q q q q q q q q q q q q q q q qq qq q q q q q q q q q q q qq qqq q q −2 −1 0 1 2 6912 wind Theoretical Quantiles SampleQuantiles q q q q q q q q q qq q qq q q q q q q q q q qq q q q q q q q qq q q qq q q q −2 −1 0 1 2 1040 precip Theoretical Quantiles SampleQuantiles q q qq q q q q q q q q q qq q qq qq q q qq q qq q q q q q q q q q q qq q q −2 −1 0 1 2 40120 predays Theoretical Quantiles SampleQuantiles Fig. 1.5. Normal probability plots for USairpollution data.
- 38. 22 1 Multivariate Data and Multivariate Analysis The resulting seven plots are arranged on one page by a call to the layout matrix; see Figure 1.5. The plots for SO2 concentration and precipitation both deviate considerably from linearity, and the plots for manufacturing and population show evidence of a number of outliers. But of more importance is the chi-square plot for the data, which is given in Figure 1.6; the R code is identical to the code used to produce the chi-square plot for the body measurement data. In addition, the two most extreme points in the plot have been labelled with the city names to which they correspond using text(). R> x <- USairpollution R> cm <- colMeans(x) R> S <- cov(x) R> d <- apply(x, 1, function(x) t(x - cm) %*% solve(S) %*% (x - cm)) R> plot(qc <- qchisq((1:nrow(x) - 1/2) / nrow(x), df = 6), + sd <- sort(d), + xlab = expression(paste(chi[6]^2, " Quantile")), + ylab = "Ordered distances", xlim = range(qc) * c(1, 1.1)) R> oups <- which(rank(abs(qc - sd), ties = "random") > nrow(x) - 3) R> text(qc[oups], sd[oups] - 1.5, names(oups)) R> abline(a = 0, b = 1) qqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqqq q q q q q q 5 10 15 510152025 χ6 2 Quantile Ordereddistances Providence Phoenix Chicago Fig. 1.6. χ2 plot of generalised distances for USairpollution data.
- 39. 1.8 Exercises 23 This example illustrates that the chi-square plot might also be useful for detecting possible outliers in multivariate data, where informally outliers are “abnormal” in the sense of deviating from the natural data variability. Outlier identification is important in many applications of multivariate analysis either because there is some specific interest in finding anomalous observations or as a pre-processing task before the application of some multivariate method in order to preserve the results from possible misleading effects produced by these observations. A number of methods for identifying multivariate outliers have been suggested–see, for example, Rocke and Woodruff (1996) and Becker and Gather (2001)–and in Chapter 2 we will see how a number of the graphical methods described there can also be helpful for outlier detection. 1.7 Summary The majority of data collected in all scientific disciplines are multivariate. To fully understand most such data sets, the variables need to be analysed simultaneously. The rest of this text is concerned with methods that have been developed to make this possible, some with the aim of discovering any patterns or structure in the data that may have important implications for future studies and some with the aim of drawing inferences about the data assuming they are sampled from a population with some particular probability density function, usually the multivariate normal. 1.8 Exercises Ex. 1.1 Find the correlation matrix and covariance matrix of the data in Table 1.1. Ex. 1.2 Fill in the missing values in Table 1.1 with appropriate mean values, and recalculate the correlation matrix of the data. Ex. 1.3 Examine both the normal probability plots of each variable in the archaeology data in Table 1.3 and the chi-square plot of the data. Do the plots suggest anything unusual about the data? Ex. 1.4 Convert the covariance matrix given below into the corresponding correlation matrix. 3.8778 2.8110 3.1480 3.5062 2.8110 2.1210 2.2669 2.5690 3.1480 2.2669 2.6550 2.8341 3.5062 2.5690 2.8341 3.2352 . Ex. 1.5 For the small set of (10 × 5) multivariate data given below, find the (10 × 10) Euclidean distance matrix for the rows of the matrix. An alternative to Euclidean distance that might be used in some cases is what
- 40. 24 1 Multivariate Data and Multivariate Analysis is known as city block distance (think New York). Write some R code to calculate the city block distance matrix for the data. 3 6 4 0 7 4 2 7 4 6 4 0 3 1 5 6 2 6 1 1 1 6 2 1 4 5 1 2 0 2 1 1 2 6 1 1 1 5 4 4 7 0 1 3 3 3 3 0 5 1 .
- 41. 2 Looking at Multivariate Data: Visualisation 2.1 Introduction According to Chambers, Cleveland, Kleiner, and Tukey (1983), “there is no statistical tool that is as powerful as a well-chosen graph”. Certainly graphical presentation has a number of advantages over tabular displays of numerical results, not least in creating interest and attracting the attention of the viewer. But just what is a graphical display? A concise description is given by Tufte (1983): Data graphics visually display measured quantities by means of the combined use of points, lines, a coordinate system, numbers, symbols, words, shading and color. Graphs are very popular; it has been estimated that between 900 billion (9×1011 ) and 2 trillion (2×1012 ) images of statistical graphics are printed each year. Perhaps one of the main reasons for such popularity is that graphical presentation of data often provides the vehicle for discovering the unexpected; the human visual system is very powerful in detecting patterns, although the following caveat from the late Carl Sagan (in his book Contact) should be kept in mind: Humans are good at discerning subtle patterns that are really there, but equally so at imagining them when they are altogether absent. Some of the advantages of graphical methods have been listed by Schmid (1954): ˆ In comparison with other types of presentation, well-designed charts are more effective in creating interest and in appealing to the attention of the reader. ˆ Visual relationships as portrayed by charts and graphs are more easily grasped and more easily remembered. ˆ The use of charts and graphs saves time since the essential meaning of large measures of statistical data can be visualised at a glance. DOI 10.1007/978-1-4419-9650-3_2, © Springer Science+Business Media, LLC 2011 25B. Everitt and T. Hothorn, An Introduction to Applied Multivariate Analysis with R: Use R!,
- 42. 26 2 Looking at Multivariate Data: Visualisation ˆ Charts and graphs provide a comprehensive picture of a problem that makes for a more complete and better balanced understanding than could be derived from tabular or textual forms of presentation. ˆ Charts and graphs can bring out hidden facts and relationships and can stimulate, as well as aid, analytical thinking and investigation. Schmid’s last point is reiterated by the legendary John Tukey in his ob- servation that “the greatest value of a picture is when it forces us to notice what we never expected to see”. The prime objective of a graphical display is to communicate to ourselves and others, and the graphic design must do everything it can to help people understand. And unless graphics are relatively simple, they are unlikely to survive the first glance. There are perhaps four goals for graphical displays of data: ˆ To provide an overview; ˆ To tell a story; ˆ To suggest hypotheses; ˆ To criticise a model. In this chapter, we will be largely concerned with graphics for multivari- ate data that address one or another of the first three bulleted points above. Graphics that help in checking model assumptions will be considered in Chap- ter 8. During the last two decades, a wide variety of new methods for display- ing data graphically have been developed. These will hunt for special ef- fects in data, indicate outliers, identify patterns, diagnose models, and gener- ally search for novel and perhaps unexpected phenomena. Graphical displays should aim to tell a story about the data and to reduce the cognitive effort required to make comparisons. Large numbers of graphs might be required to achieve these goals, and computers are generally needed to supply them for the same reasons that they are used for numerical analyses, namely that they are fast and accurate. So, because the machine is doing the work, the question is no longer “shall we plot?” but rather “what shall we plot?” There are many exciting possibil- ities, including interactive and dynamic graphics on a computer screen (see Cook and Swayne 2007), but graphical exploration of data usually begins at least with some simpler static graphics. The starting graphic for multivariate data is often the ubiquitous scatterplot, and this is the subject of the next section. 2.2 The scatterplot The simple xy scatterplot has been in use since at least the 18th century and has many virtues–indeed, according to Tufte (1983):
- 43. 2.2 The scatterplot 27 The relational graphic–in its barest form the scatterplot and its variants–is the greatest of all graphical designs. It links at least two variables, encouraging and even imploring the viewer to assess the possible causal relationship between the plotted variables. It confronts causal theories that x causes y with empirical evidence as to the actual relationship between x and y. The scatterplot is the standard for representing continuous bivariate data but, as we shall see later in this chapter, it can be enhanced in a variety of ways to accommodate information about other variables. To illustrate the use of the scatterplot and a number of other techniques to be discussed, we shall use the air pollution in US cities data introduced in the previous chapter (see Table 1.5). Let’s begin our examination of the air pollution data by taking a look at a basic scatterplot of the two variables manu and popul. For later use, we first set up two character variables that contain the labels to be printed on the two axes: R> mlab <- "Manufacturing enterprises with 20 or more workers" R> plab <- "Population size (1970 census) in thousands" The plot() function takes the data, here as the data frame USairpollution, along with a “formula” describing the variables to be plotted; the part left of the tilde defines the variable to be associated with the ordinate, the part right of the tilde is the variable that goes with the abscissa: R> plot(popul ~ manu, data = USairpollution, + xlab = mlab, ylab = plab) The resulting scatterplot is shown in Figure 2.2. The plot clearly uncovers the presence of one or more cities that are some way from the remainder, but before commenting on these possible outliers we will construct the scatterplot again but now show how to include the marginal distributions of manu and popul in two different ways. Plotting marginal and joint distributions together is usually good data analysis practise. In Figure 2.2, the marginal distributions are shown as rug plots on each axis (produced by rug()), and in Figure 2.3 the marginal distribution of manu is given as a histogram and that of popul as a boxplot. And also in Figure 2.3 the points are labelled by an abbreviated form of the corresponding city name. The necessary R code for Figure 2.3 starts with dividing the device into three plotting areas by means of the layout() function. The first plot basically resembles the plot() command from Figure 2.1, but instead of points the ab- breviated name of the city is used as the plotting symbol. Finally, the hist() and boxplots() commands are used to depict the marginal distributions. The with() command is very useful when one wants to avoid explicitly extracting variables from data frames. The command of interest, here the calls to hist() and boxplot(), is evaluated “inside” the data frame, here USairpollution (i.e., variable names are resolved within this data frame first).
- 44. 28 2 Looking at Multivariate Data: Visualisation q q q q q q q q q q q q q q q q q q q q qq q q q q qqq q q q q q q q q q q q q 0 500 1000 2000 3000 0500150025003500 Manufacturing enterprises with 20 or more workers Populationsize(1970census)inthousands Fig. 2.1. Scatterplot of manu and popul. From this series of plots, we can see that the outlying points show them- selves in both the scatterplot of the variables and in each marginal distribu- tion. The most extreme outlier corresponds to Chicago, and other slightly less extreme outliers correspond to Philadelphia and Detroit. Each of these cities has a considerably larger population than other cities and also many more manufacturing enterprises with more than 20 workers. 2.2.1 The bivariate boxplot In Figure 2.3, identifying Chicago, Philadelphia, and Detroit as outliers is un- likely to invoke much argument, but what about Houston and Cleveland? In many cases, it might be helpful to have a more formal and objective method for labelling observations as outliers, and such a method is provided by the bivariate boxplot, which is a two-dimensional analogue of the boxplot for uni- variate data proposed by Goldberg and Iglewicz (1992). This type of graphic
- 45. 2.2 The scatterplot 29 R> plot(popul ~ manu, data = USairpollution, + xlab = mlab, ylab = plab) R> rug(USairpollution$manu, side = 1) R> rug(USairpollution$popul, side = 2) q q q q q q q q q q q q q q q q q q q q qq q q q q qqq q q q q q q q q q q q q 0 500 1000 2000 3000 0500150025003500 Manufacturing enterprises with 20 or more workers Populationsize(1970census)inthousands Fig. 2.2. Scatterplot of manu and popul that shows the marginal distribution in each variable as a rug plot. may be useful in indicating the distributional properties of the data and in identifying possible outliers. The bivariate boxplot is based on calculating“ro- bust” measures of location, scale, and correlation; it consists essentially of a pair of concentric ellipses, one of which (the “hinge”) includes 50% of the data and the other (called the “fence”) of which delineates potentially troublesome outliers. In addition, resistant regression lines of both y on x and x on y are shown, with their intersection showing the bivariate location estimator. The acute angle between the regression lines will be small for a large absolute value of correlations and large for a small one. (Using robust measures of location, scale, etc., helps to prevent the possible “masking” of multivariate outliers if
- 46. 30 2 Looking at Multivariate Data: Visualisation R> layout(matrix(c(2, 0, 1, 3), nrow = 2, byrow = TRUE), + widths = c(2, 1), heights = c(1, 2), respect = TRUE) R> xlim <- with(USairpollution, range(manu)) * 1.1 R> plot(popul ~ manu, data = USairpollution, cex.lab = 0.9, + xlab = mlab, ylab = plab, type = "n", xlim = xlim) R> with(USairpollution, text(manu, popul, cex = 0.6, + labels = abbreviate(row.names(USairpollution)))) R> with(USairpollution, hist(manu, main = "", xlim = xlim)) R> with(USairpollution, boxplot(popul)) 0 1000 2000 3000 0100020003000 Manufacturing enterprises with 20 or more workers Populationsize(1970census)inthousands Albn Albq Atln Bltm Bffl Chrl Chcg Cncn Clvl Clmb Dlls Dnvr DsMn Dtrt Hrtf Hstn Indn JcksKnsC LttR LsvlMmph Miam MlwkMnnp Nshv NwOrNrflOmah Phld PhnxPtts Prvd Rchm SlLC SnFr Sttl St.L Wshn Wcht Wlmn manu Frequency 0 1000 2000 3000 0 q q q 0100020003000 Fig. 2.3. Scatterplot of manu and popul that shows the marginal distributions by histogram and boxplot.
- 47. 2.2 The scatterplot 31 the usual measures are employed when these may be distorted by the pres- ence of the outliers in the data.) Full details of the construction are given in Goldberg and Iglewicz (1992). The scatterplot of manu and popul including the bivariate boxplot is shown in Figure 2.4. Figure 2.4 clearly tells us that Chicago, Philadelphia, Detroit, and Cleveland should be regarded as outliers but not Houston, because it is on the “fence” rather than outside the “fence”. R> outcity <- match(lab <- c("Chicago", "Detroit", + "Cleveland", "Philadelphia"), rownames(USairpollution)) R> x <- USairpollution[, c("manu", "popul")] R> bvbox(x, mtitle = "", xlab = mlab, ylab = plab) R> text(x$manu[outcity], x$popul[outcity], labels = lab, + cex = 0.7, pos = c(2, 2, 4, 2, 2)) q q q q q q q q q q q q q q q q q q q q qq q q q q qqq q q q q q q q q q q q q 0 1000 2000 3000 0100020003000 Manufacturing enterprises with 20 or more workers Populationsize(1970census)inthousands Chicago Detroit Cleveland Philadelphia Fig. 2.4. Scatterplot of manu and popul showing the bivariate boxplot of the data. Suppose now that we are interested in calculating the correlation between manu and popul. Researchers often calculate the correlation between two vari-
- 48. 32 2 Looking at Multivariate Data: Visualisation ables without first looking at the scatterplot of the two variables. But scat- terplots should always be consulted when calculating correlation coefficients because the presence of outliers can on occasion considerably distort the value of a correlation coefficient, and as we have seen above, a scatterplot may help to identify the offending observations particularly if used in conjunction with a bivariate boxplot. The observations identified as outliers may then be ex- cluded from the calculation of the correlation coefficient. With the help of the bivariate boxplot in Figure 2.4, we have identified Chicago, Philadelphia, Detroit, and Cleveland as outliers in the scatterplot of manu and popul. The R code for finding the two correlations is R> with(USairpollution, cor(manu, popul)) [1] 0.9553 R> outcity <- match(c("Chicago", "Detroit", + "Cleveland", "Philadelphia"), + rownames(USairpollution)) R> with(USairpollution, cor(manu[-outcity], popul[-outcity])) [1] 0.7956 The match() function identifies rows of the data frame USairpollution cor- responding to the cities of interest, and the subset starting with a minus sign removes these units before the correlation is computed. Calculation of the cor- relation coefficient between the two variables using all the data gives a value of 0.96, which reduces to a value of 0.8 after excluding the four outliers–a not inconsiderable reduction. 2.2.2 The convex hull of bivariate data An alternative approach to using the scatterplot combined with the bivariate boxplot to deal with the possible problem of calculating correlation coeffi- cients without the distortion often caused by outliers in the data is convex hull trimming, which allows robust estimation of the correlation. The convex hull of a set of bivariate observations consists of the vertices of the smallest convex polyhedron in variable space within which or on which all data points lie. Removal of the points lying on the convex hull can eliminate isolated outliers without disturbing the general shape of the bivariate distribution. A robust estimate of the correlation coefficient results from using the remaining observations. Let’s see how the convex hull approach works with our manu and popul scatterplot. We first find the convex hull of the data (i.e., the observations defining the convex hull) using the following R code: R> (hull <- with(USairpollution, chull(manu, popul))) [1] 9 15 41 6 2 18 16 14 7
- 49. 2.2 The scatterplot 33 R> with(USairpollution, + plot(manu, popul, pch = 1, xlab = mlab, ylab = plab)) R> with(USairpollution, + polygon(manu[hull], popul[hull], density = 15, angle = 30)) q q q q q q q q q q q q q q q q q q q q qq q q q q qqq q q q q q q q q q q q q 0 500 1000 2000 3000 0500150025003500 Manufacturing enterprises with 20 or more workers Populationsize(1970census)inthousands Fig. 2.5. Scatterplot of manu against popul showing the convex hull of the data. Now we can show this convex hull on a scatterplot of the variables using the code attached to the resulting Figure 2.5. To calculate the correlation coefficient after removal of the points defining the convex hull requires the code R> with(USairpollution, cor(manu[-hull],popul[-hull])) [1] 0.9225 The resulting value of the correlation is now 0.923 and thus is higher compared with the correlation estimated after removal of the outliers identified by using the bivariate boxplot, namely Chicago, Philadelphia, Detroit, and Cleveland.
- 50. 34 2 Looking at Multivariate Data: Visualisation 2.2.3 The chi-plot Although the scatterplot is a primary data-analytic tool for assessing the re- lationship between a pair of continuous variables, it is often difficult to judge whether or not the variables are independent–a random scatter of points is hard for the human eye to judge. Consequently it is sometimes helpful to aug- ment the scatterplot with an auxiliary display in which independence is itself manifested in a characteristic manner. The chi-plot suggested by Fisher and Switzer (1985, 2001) is designed to address the problem. Under independence, the joint distribution of two random variables X1 and X2 can be computed from the product of the marginal distributions. The chi-plot transforms the measurements (x11, . . . , xn1) and (x12, . . . , xn2) into values (χ1, . . . , χn) and (λ1, . . . , λn), which, plotted in a scatterplot, can be used to detect deviations from independence. The χi values are, basically, the root of the χ2 statistics obtained from the 2×2 tables that are obtained when dichotomising the data for each unit i into the groups satisfying x·1 ≤ xi1 and x·2 ≤ xi2. Under inde- pendence, these values are asymptotically normal with mean zero; i.e., the χi values should show a non-systematic random fluctuation around zero. The λi values measure the distance of unit i from the “center” of the bivariate distri- bution. An R function for producing chi-plots is chiplot(). To illustrate the chi-plot, we shall apply it to the manu and popul variables of the air pollution data using the code R> with(USairpollution, plot(manu, popul, + xlab = mlab, ylab = plab, + cex.lab = 0.9)) R> with(USairpollution, chiplot(manu, popul)) The result is Figure 2.6, which shows the scatterplot of manu plotted against popul alongside the corresponding chi-plot. Departure from independence is indicated in the latter by a lack of points in the horizontal band indicated on the plot. Here there is a very clear departure since there are very few of the observations in this region. 2.3 The bubble and other glyph plots The basic scatterplot can only display two variables. But there have been a number of suggestions as to how extra variables may be included on a scatterplot. Perhaps the simplest is the so-called bubble plot, in which three variables are displayed; two are used to form the scatterplot itself, and then the values of the third variable are represented by circles with radii proportional to these values and centred on the appropriate point in the scatterplot. Let’s begin by taking a look at the bubble plot of temp, wind, and SO2 that is given in Figure 2.7. The plot seems to suggest that cities with moderate annual temperatures and moderate annual wind speeds tend to suffer the greatest air
- 51. 2.3 The bubble and other glyph plots 35 q q q q q q q q q q q q q q q q q qq q qq q qq q qqq q q q q q q q q q q q q 050015002500 0500150025003500 Manufacturingenterpriseswith20ormoreworkers Populationsize(1970census)inthousands q q q qq q q q q q q q q q qq q q q q q q q q q q q q q q q q qq q q −1.0−0.50.00.51.0 −1.0−0.50.00.51.0 λ χ Fig.2.6.Chi-plotformanuandpopulshowingacleardeviationfromindependence. R> plot(blood_pcacor$sdev^2, xlab = "Component number", + ylab = "Component variance", type = "l", main = "Scree diagram")
- 52. 36 2 Looking at Multivariate Data: Visualisation pollution, but this is unlikely to be the whole story because none of the other variables in the data set are used in constructing Figure 2.7. We could try to include all variables on the basic temp and wind scatterplot by replacing the circles with five-sided “stars”, with the lengths of each side representing each of the remaining five variables. Such a plot is shown in Figure 2.8, but it fails to communicate much, if any, useful information about the data. R> ylim <- with(USairpollution, range(wind)) * c(0.95, 1) R> plot(wind ~ temp, data = USairpollution, + xlab = "Average annual temperature (Fahrenheit)", + ylab = "Average annual wind speed (m.p.h.)", pch = 10, + ylim = ylim) R> with(USairpollution, symbols(temp, wind, circles = SO2, + inches = 0.5, add = TRUE)) q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q 45 50 55 60 65 70 75 6789101112 Average annual temperature (Fahrenheit) Averageannualwindspeed(m.p.h.) q q q q q q q q q q q q qq q q q q q q q q q q q q q q Fig. 2.7. Bubble plot of temp, wind, and SO2.
- 53. 2.3 The bubble and other glyph plots 37 R> plot(wind ~ temp, data = USairpollution, + xlab = "Average annual temperature (Fahrenheit)", + ylab = "Average annual wind speed (m.p.h.)", pch = 10, + ylim = ylim) R> with(USairpollution, + stars(USairpollution[,-c(2,5)], locations = cbind(temp, wind), + labels = NULL, add = TRUE, cex = 0.5)) q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q 45 50 55 60 65 70 75 6789101112 Average annual temperature (Fahrenheit) Averageannualwindspeed(m.p.h.) Fig. 2.8. Scatterplot of temp and wind showing five-sided stars representing the other variables. In fact, both the bubble plot and “stars” plot are examples of symbol or glyph plots, in which data values control the symbol parameters. For example, a circle is a glyph where the values of one variable in a multivariate observation control the circle size. In Figure 2.8, the spatial positions of the cities in the scatterplot of temp and wind are combined with a star representation of the five other variables. An alternative is simply to represent the seven variables for each city by a seven-sided star and arrange the resulting stars in
- 54. 38 2 Looking at Multivariate Data: Visualisation a rectangular array; the result is shown in Figure 2.9. We see that some stars, for example those for New Orleans, Miami, Jacksonville, and Atlanta, have similar shapes, with their higher average annual temperature being distinctive, but telling a story about the data with this display is difficult. Stars, of course, are not the only symbols that could be used to represent data, and others have been suggested, with perhaps the most well known being the now infamous Chernoff’s faces (see Chernoff 1973). But, on the whole, such graphics for displaying multivariate data have not proved themselves to be effective for the task and are now largely confined to the past history of multivariate graphics. R> stars(USairpollution, cex = 0.55) Albany Albuquerque Atlanta Baltimore Buffalo Charleston Chicago Cincinnati Cleveland Columbus Dallas Denver Des Moines Detroit Hartford Houston Indianapolis Jacksonville Kansas City Little Rock Louisville Memphis Miami Milwaukee Minneapolis Nashville New Orleans Norfolk Omaha Philadelphia Phoenix Pittsburgh Providence Richmond Salt Lake City San Francisco Seattle St. Louis Washington Wichita Wilmington Fig. 2.9. Star plot of the air pollution data.
- 55. 2.4 The scatterplot matrix 39 2.4 The scatterplot matrix There are seven variables in the air pollution data, which between them gen- erate 21 possible scatterplots. But just making the graphs without any coor- dination will often result in a confusing collection of graphs that are hard to integrate visually. Consequently, it is very important that the separate plots be presented in the best way to aid overall comprehension of the data. The scatterplot matrix is intended to accomplish this objective. A scatterplot ma- trix is nothing more than a square, symmetric grid of bivariate scatterplots. The grid has q rows and columns, each one corresponding to a different vari- able. Each of the grid’s cells shows a scatterplot of two variables. Variable j is plotted against variable i in the ijth cell, and the same variables appear in cell ji, with the x- and y-axes of the scatterplots interchanged. The reason for including both the upper and lower triangles of the grid, despite the seeming redundancy, is that it enables a row and a column to be visually scanned to see one variable against all others, with the scales for the one variable lined up along the horizontal or the vertical. As a result, we can visually link features on one scatterplot with features on another, and this ability greatly increases the power of the graphic. The scatterplot matrix for the air pollution data is shown in Figure 2.10. The plot was produced using the function pairs(), here with slightly enlarged dot symbols, using the arguments pch = "." and cex = 1.5. The scatterplot matrix clearly shows the presence of possible outliers in many panels and the suggestion that the relationship between the two aspects of rainfall, namely precip, predays, and SO2 might be non-linear. Remem- bering that the multivariable aspect of these data, in which sulphur dioxide concentration is the response variable, with the remaining variables being ex- planatory, might be of interest, the scatterplot matrix may be made more helpful by including the linear fit of the two variables on each panel, and such a plot is shown in Figure 2.11. Here, the pairs() function was customised by a small function specified to the panel argument: in addition to plotting the x and y values, a regression line obtained via function lm() is added to each of the panels. Now the scatterplot matrix reveals that there is a strong linear relationship between SO2 and manu and between SO2 and popul, but the (3, 4) panel shows that manu and popul are themselves very highly related and thus predictive of SO2 in the same way. Figure 2.11 also underlines that assuming a linear relationship between SO2 and precip and SO2 and predays, as might be the case if a multiple linear regression model is fitted to the data with SO2 as the dependent variable, is unlikely to fully capture the relationship between each pair of variables. In the same way that the scatterplot should always be used alongside the numerical calculation of a correlation coefficient, so should the scatterplot matrix always be consulted when looking at the correlation matrix of a set of variables. The correlation matrix for the air pollution data is
- 56. 40 2 Looking at Multivariate Data: Visualisation R> pairs(USairpollution, pch = ".", cex = 1.5) SO2 45 65 0 2500 10 50 20100 4565 temp manu 02500 02500 popul wind 69 1050 precip 20 100 0 2500 6 9 40 140 40140 predays Fig. 2.10. Scatterplot matrix of the air pollution data. R> round(cor(USairpollution), 4) SO2 temp manu popul wind precip predays SO2 1.0000 -0.4336 0.6448 0.4938 0.0947 0.0543 0.3696 temp -0.4336 1.0000 -0.1900 -0.0627 -0.3497 0.3863 -0.4302 manu 0.6448 -0.1900 1.0000 0.9553 0.2379 -0.0324 0.1318 popul 0.4938 -0.0627 0.9553 1.0000 0.2126 -0.0261 0.0421 wind 0.0947 -0.3497 0.2379 0.2126 1.0000 -0.0130 0.1641 precip 0.0543 0.3863 -0.0324 -0.0261 -0.0130 1.0000 0.4961 predays 0.3696 -0.4302 0.1318 0.0421 0.1641 0.4961 1.0000 Focussing on the correlations between SO2 and the six other variables, we see that the correlation for SO2 and precip is very small and that for SO2 and predays is moderate. But relevant panels in the scatterplot indicate that the correlation coefficient that assesses only the linear relationship between
- 57. 2.4 The scatterplot matrix 41 R> pairs(USairpollution, + panel = function (x, y, ...) { + points(x, y, ...) + abline(lm(y ~ x), col = "grey") + }, pch = ".", cex = 1.5) SO2 45 65 0 2500 10 50 20100 4565 temp manu 02500 02500 popul wind 69 1050 precip 20 100 0 2500 6 9 40 140 40140 predays Fig. 2.11. Scatterplot matrix of the air pollution data showing the linear fit of each pair of variables. two variables may not be suitable here and that in a multiple linear regres- sion model for the data quadratic effects of predays and precip might be considered.
- 58. 42 2 Looking at Multivariate Data: Visualisation 2.5 Enhancing the scatterplot with estimated bivariate densities As we have seen above, scatterplots and scatterplot matrices are good at highlighting outliers in a multivariate data set. But in many situations an- other aim in examining scatterplots is to identify regions in the plot where there are high or low densities of observations that may indicate the pres- ence of distinct groups of observations; i.e., “clusters” (see Chapter 6). But humans are not particularly good at visually examining point density, and it is often a very helpful aid to add some type of bivariate density estimate to the scatterplot. A bivariate density estimate is simply an approximation to the bivariate probability density function of two variables obtained from a sample of bivariate observations of the variables. If, of course, we are willing to assume a particular form of the bivariate density of the two variables, for example the bivariate normal, then estimating the density is reduced to esti- mating the parameters of the assumed distribution. More commonly, however, we wish to allow the data to speak for themselves and so we need to look for a non-parametric estimation procedure. The simplest such estimator would be a two-dimensional histogram, but for small and moderately sized data sets that is not of any real use for estimating the bivariate density function simply because most of the “boxes” in the histogram will contain too few observa- tions; and if the number of boxes is reduced, the resulting histogram will be too coarse a representation of the density function. Other non-parametric density estimators attempt to overcome the deficien- cies of the simple two-dimensional histogram estimates by “smoothing” them in one way or another. A variety of non-parametric estimation procedures have been suggested, and they are described in detail in Silverman (1986) and Wand and Jones (1995). Here we give a brief description of just one popular class of estimators, namely kernel density estimators. 2.5.1 Kernel density estimators From the definition of a probability density, if the random variable X has a density f, f(x) = lim h→0 1 2h P(x − h < X < x + h). (2.1) For any given h, a na¨ıve estimator of P(x − h < X < x + h) is the proportion of the observations x1, x2, . . . , xn falling in the interval (x − h, x + h), ˆf(x) = 1 2hn n i=1 I(xi ∈ (x − h, x + h)); (2.2) i.e., the number of x1, . . . , xn falling in the interval (x − h, x + h) divided by 2hn. If we introduce a weight function W given by
- 59. 2.5 Enhancing the scatterplot with estimated bivariate densities 43 W(x) = 1 2 |x| < 1 0 else, then the na¨ıve estimator can be rewritten as ˆf(x) = 1 n n i=1 1 h W x − xi h . (2.3) Unfortunately, this estimator is not a continuous function and is not par- ticularly satisfactory for practical density estimation. It does, however, lead naturally to the kernel estimator defined by ˆf(x) = 1 hn n i=1 K x − xi h , (2.4) where K is known as the kernel function and h is the bandwidth or smoothing parameter. The kernel function must satisfy the condition ∞ −∞ K(x)dx = 1. Usually, but not always, the kernel function will be a symmetric density func- tion; for example, the normal. Three commonly used kernel functions are rectangular, K(x) = 1 2 |x| < 1 0 else. triangular, K(x) = 1 − |x| |x| < 1 0 else, Gaussian, K(x) = 1 √ 2π e− 1 2 x2 . The three kernel functions are implemented in R as shown in Figure 2.12. For some grid x, the kernel functions are plotted using the R statements in Figure 2.12. The kernel estimator ˆf is a sum of “bumps” placed at the observations. The kernel function determines the shape of the bumps, while the window width h determines their width. Figure 2.13 (redrawn from a similar plot in Silverman 1986) shows the individual bumps n−1 h−1 K((x − xi)/h) as well as the estimate ˆf obtained by adding them up for an artificial set of data points
- 60. 44 2 Looking at Multivariate Data: Visualisation R> rec <- function(x) (abs(x) < 1) * 0.5 R> tri <- function(x) (abs(x) < 1) * (1 - abs(x)) R> gauss <- function(x) 1/sqrt(2*pi) * exp(-(x^2)/2) R> x <- seq(from = -3, to = 3, by = 0.001) R> plot(x, rec(x), type = "l", ylim = c(0,1), lty = 1, + ylab = expression(K(x))) R> lines(x, tri(x), lty = 2) R> lines(x, gauss(x), lty = 3) R> legend("topleft", legend = c("Rectangular", "Triangular", + "Gaussian"), lty = 1:3, title = "kernel functions", + bty = "n") −3 −2 −1 0 1 2 3 0.00.20.40.60.81.0 x K(x) kernel functions Rectangular Triangular Gaussian Fig. 2.12. Three commonly used kernel functions. R> x <- c(0, 1, 1.1, 1.5, 1.9, 2.8, 2.9, 3.5) R> n <- length(x) For a grid R> xgrid <- seq(from = min(x) - 1, to = max(x) + 1, by = 0.01) on the real line, we can compute the contribution of each measurement in x, with h = 0.4, by the Gaussian kernel (defined in Figure 2.12, line 3) as follows:
- 61. 2.5 Enhancing the scatterplot with estimated bivariate densities 45 R> h <- 0.4 R> bumps <- sapply(x, function(a) gauss((xgrid - a)/h)/(n * h)) A plot of the individual bumps and their sum, the kernel density estimate ˆf, is shown in Figure 2.13. R> plot(xgrid, rowSums(bumps), ylab = expression(hat(f)(x)), + type = "l", xlab = "x", lwd = 2) R> rug(x, lwd = 2) R> out <- apply(bumps, 2, function(b) lines(xgrid, b)) −1 0 1 2 3 4 0.000.100.200.30 x f ^ (x) Fig. 2.13. Kernel estimate showing the contributions of Gaussian kernels evaluated for the individual observations with bandwidth h = 0.4. The kernel density estimator considered as a sum of “bumps” centred at the observations has a simple extension to two dimensions (and similarly for more than two dimensions). The bivariate estimator for data (x1, y1), (x2, y2), . . . , (xn, yn) is defined as
- 62. 46 2 Looking at Multivariate Data: Visualisation ˆf(x, y) = 1 nhxhy n i=1 K x − xi hx , y − yi hy . (2.5) In this estimator, each coordinate direction has its own smoothing parameter, hx or hy. An alternative is to scale the data equally for both dimensions and use a single smoothing parameter. For bivariate density estimation, a commonly used kernel function is the standard bivariate normal density K(x, y) = 1 2π e− 1 2 (x2 +y2 ) . Another possibility is the bivariate Epanechnikov kernel given by K(x, y) = 2 π (1 − x2 − y2 ) x2 + y2 < 1 0 else, which is implemented and depicted in Figure 2.14 by using the persp function for plotting in three dimensions. According to Venables and Ripley (2002), the bandwidth should be chosen to be proportional to n−1/5 ; unfortunately, the constant of proportionality depends on the unknown density. The tricky problem of bandwidth estimation is considered in detail in Silverman (1986). Our first illustration of enhancing a scatterplot with an estimated bivari- ate density will involve data from the Hertzsprung-Russell (H-R) diagram of the star cluster CYG OB1, calibrated according to Vanisma and De Greve (1972). The H-R diagram is the basis of the theory of stellar evolution and is essentially a plot of the energy output of stars as measured by the loga- rithm of their light intensity plotted against the logarithm of their surface temperature. Part of the data is shown in Table 2.1. A scatterplot of the data enhanced by the contours of the estimated bivariate density (Wand and Ripley 2010, obtained with the function bkde2D() from the package KernSmooth) is shown in Figure 2.15. The plot shows the presence of two distinct clusters of stars: the larger cluster consists of stars that have high surface temperatures and a range of light intensities, and the smaller cluster contains stars with low surface temperatures and high light intensities. The bivariate density estimate can also be displayed by means of a perspective plot rather than a contour plot, and this is shown in Figure 2.16. This again demonstrates that there are two groups of stars. Table 2.1: CYGOB1 data. Energy output and surface temperature of star cluster CYG OB1. logst logli logst logli logst logli 4.37 5.23 4.23 3.94 4.45 5.22
- 63. 2.6 Three-dimensional plots 47 Table 2.1: CYGOB1 data (continued). logst logli logst logli logst logli 4.56 5.74 4.42 4.18 3.49 6.29 4.26 4.93 4.23 4.18 4.23 4.34 4.56 5.74 3.49 5.89 4.62 5.62 4.30 5.19 4.29 4.38 4.53 5.10 4.46 5.46 4.29 4.22 4.45 5.22 3.84 4.65 4.42 4.42 4.53 5.18 4.57 5.27 4.49 4.85 4.43 5.57 4.26 5.57 4.38 5.02 4.38 4.62 4.37 5.12 4.42 4.66 4.45 5.06 3.49 5.73 4.29 4.66 4.50 5.34 4.43 5.45 4.38 4.90 4.45 5.34 4.48 5.42 4.22 4.39 4.55 5.54 4.01 4.05 3.48 6.05 4.45 4.98 4.29 4.26 4.38 4.42 4.42 4.50 4.42 4.58 4.56 5.10 For our next example of adding estimated bivariate densities to scatter- plots, we will use the body measurement data introduced in Chapter 1 (see Table 1.2), although there are rather too few observations on which to base the estimation. (The gender of each individual will not be used.) And in this case we will add the appropriate density estimate to each panel of the scat- terplot matrix of the chest, waist, and hips measurements. The resulting plot is shown in Figure 2.17. The waist/hips panel gives some evidence that there might be two groups in the data, which, of course, we know to be true, the groups being men and women. And the Waist histogram on the diagonal panel is also bimodal, underlining the two-group nature of the data. 2.6 Three-dimensional plots The scatterplot matrix allows us to display information about the univariate distributions of each variable (using histograms on the main diagonal, for example) and about the bivariate distribution of all pairs of variables in a set of multivariate data. But we should perhaps consider whether the use of three-dimensional plots offers any advantage over the series of two-dimensional scatterplots used in a scatterplot matrix. To begin, we can take a look at the three-dimensional plot of the body measurements data; a version of the plot that includes simply the points along with vertical lines dropped from each point to the x-y plane is shown in Figure 2.18. The plot, produced with the scatterplot3d package (Ligges 2010), suggests the presence of two relatively separate groups of points corresponding to the males and females in the data.
- 64. 48 2 Looking at Multivariate Data: Visualisation R> epa <- function(x, y) + ((x^2 + y^2) < 1) * 2/pi * (1 - x^2 - y^2) R> x <- seq(from = -1.1, to = 1.1, by = 0.05) R> epavals <- sapply(x, function(a) epa(a, x)) R> persp(x = x, y = x, z = epavals, xlab = "x", ylab = "y", + zlab = expression(K(x, y)), theta = -35, axes = TRUE, + box = TRUE) x y K(x,y) Fig. 2.14. Epanechnikov kernel for a grid between (−1.1, −1.1) and (1.1, 1.1). As a second example of using a three-dimensional plot, we can look at temp, wind, and SO2 from the air pollution data. The points and vertical lines versions of the required three-dimensional plot are shown in Figure 2.19. Two observations with high SO2 levels stand out, but the plot does not appear to add much to the bubble plot for the same three variables (Figure 2.7). Three-dimensional plots based on the original variables can be useful in some cases but may not add very much to, say, the bubble plot of the scat- terplot matrix of the data. When there are many variables in a multivariate
- 65. 2.6 Three-dimensional plots 49 R> library("KernSmooth") R> CYGOB1d <- bkde2D(CYGOB1, bandwidth = sapply(CYGOB1, dpik)) R> plot(CYGOB1, xlab = "log surface temperature", + ylab = "log light intensity") R> contour(x = CYGOB1d$x1, y = CYGOB1d$x2, + z = CYGOB1d$fhat, add = TRUE) q q q q q q q q q q q q q q q q q qq q q q q q q qq q q q q q q q q q q q q q q q qq q q q 3.6 3.8 4.0 4.2 4.4 4.6 4.04.55.05.56.0 log surface temperature loglightintensity 0.2 0.2 0.2 0.2 0.4 0.4 0.6 0.6 0.8 1 1.2 1.4 1.61.8 2 2.2 Fig. 2.15. Scatterplot of the log of light intensity and log of surface temperature for the stars in star cluster CYG OB1 showing the estimated bivariate density. data set, there will be many possible three-dimensional plots to look at and integrating and linking all the plots may be very difficult. But if the dimension- ality of the data could be reduced in some way with little loss of information, three-dimensional plots might become more useful, a point to which we will return in the next chapter.
- 66. 50 2 Looking at Multivariate Data: Visualisation R> persp(x = CYGOB1d$x1, y = CYGOB1d$x2, z = CYGOB1d$fhat, + xlab = "log surface temperature", + ylab = "log light intensity", + zlab = "density") log surface temperature loglightintensity density Fig. 2.16. Perspective plot of estimated bivariate density. 2.7 Trellis graphics Trellis graphics (see Becker, Cleveland, Shyu, and Kaluzny 1994) is an ap- proach to examining high-dimensional structure in data by means of one-, two-, and three-dimensional graphs. The problem addressed is how observa- tions of one or more variables depend on the observations of the other vari- ables. The essential feature of this approach is the multiple conditioning that allows some type of plot to be displayed for different values of a given variable (or variables). The aim is to help in understanding both the structure of the data and how well proposed models describe the structure. An example of the application of trellis graphics is given in Verbyla, Cullis, Kenward, and
- 67. 2.7 Trellis graphics 51 chest 22 26 30 0.0020.004 0.006 0.008 0.01 343842 0.005 0.01 0.015 222630 0.002 0.004 0.006 0.008 0.01 0.012 waist 0.002 0.004 0.006 0.008 0.01 34 38 42 0.005 0.01 0.015 0.002 0.004 0.006 0.008 0.01 32 36 40 323640 hips Fig. 2.17. Scatterplot matrix of body measurements data showing the estimated bivariate densities on each panel. Welham (1999). With the recent publication of Sarkar’s excellent book (see Sarkar 2008) and the development of the lattice (Sarkar 2010) package, trel- lis graphics are likely to become more popular, and in this section we will illustrate their use on multivariate data. For the first example, we return to the air pollution data and the temp, wind, and SO2 variables used previously to produce scatterplots of SO2 and temp conditioned on values of wind divided into two equal parts that we shall creatively label“Light”and“High”. The resulting plot is shown in Figure 2.20. The plot suggests that in cities with light winds, air pollution decreases with increasing temperature, but in cities with high winds, air pollution does not appear to be strongly related to temperature. A more complex example of trellis graphics is shown in Figure 2.21. Here three-dimensional plots of temp, wind, and precip are shown for four levels of SO2. The graphic looks pretty, but does it convey anything of interest about
- 68. 52 2 Looking at Multivariate Data: Visualisation R> library("scatterplot3d") R> with(measure, scatterplot3d(chest, waist, hips, + pch = (1:2)[gender], type = "h", angle = 55)) 32 34 36 38 40 42 44 323436384042 22 24 26 28 30 32 34 chest waist hips q q q q q q q q q q Fig. 2.18. A three-dimensional scatterplot for the body measurements data with points corresponding to male and triangles to female measurements. the data? Probably not, as there are few points in each of the three, three- dimensional displays. This is often a problem with multipanel plots when the sample size is not large. For the last example in this section, we will use a larger data set, namely data on earthquakes given in Sarkar (2008). The data consist of recordings of the location (latitude, longitude, and depth) and magnitude of 1000 seismic events around Fiji since 1964. In Figure 2.22, scatterplots of latitude and longitude are plotted for three ranges of depth. The distribution of locations in the latitude-longitude space is seen to be different in the three panels, particularly for very deep quakes. In
- 69. 2.8 Stalactite plots 53 R> with(USairpollution, + scatterplot3d(temp, wind, SO2, type = "h", + angle = 55)) 40 45 50 55 60 65 70 75 80 020406080100120 6 7 8 9 10 11 12 13 temp wind SO2 qqq q q q q q q qq q q q q q q q q q q q q q q q q q q q q qq q q qq q qq q Fig. 2.19. A three-dimensional scatterplot for the air pollution data. Figure 2.23 (a tour de force by Sarkar) the four panels are defined by ranges of magnitude and depth is encoded by different shading. Finally, in Figure 2.24, three-dimensional scatterplots of earthquake epi- centres (latitude, longitude, and depth) are plotted conditioned on earthquake magnitude. (Figures 2.22, 2.23, and 2.24 are reproduced with the kind per- mission of Dr. Deepayan Sarkar.) 2.8 Stalactite plots In this section, we will describe a multivariate graphic, the stalactite plot, specifically designed for the detection and identification of multivariate out-
- 70. 54 2 Looking at Multivariate Data: Visualisation R> plot(xyplot(SO2 ~ temp| cut(wind, 2), data = USairpollution)) temp SO2 20 40 60 80 100 50 60 70 q q q q q q q q qq q q q q q q q q q q q (5.99,9.35] 50 60 70 q q q q q q q q q q q q q q q q q q q q (9.35,12.7] Fig. 2.20. Scatterplot of SO2 and temp for light and high winds. liers. Like the chi-square plot for assessing multivariate normality, described in Chapter 1, the stalactite plot is based on the generalised distances of ob- servations from the multivariate mean of the data. But here these distances are calculated from the means and covariances estimated from increasing- sized subsets of the data. As mentioned previously when describing bivariate boxplots, the aim is to reduce the masking effects that can arise due to the influence of outliers on the estimates of means and covariances obtained from all the data. The central idea of this approach is that, given distances using, say, m observations for estimation of means and covariances, the m+ 1 obser- vations to be used for this estimation in the next stage are chosen to be those with the m+1 smallest distances. Thus an observation can be included in the subset used for estimation for some value of m but can later be excluded as m increases. Initially m is chosen to take the value q + 1, where q is the number of variables in the multivariate data set because this is the smallest number
- 71. 2.8 Stalactite plots 55 R> pollution <- with(USairpollution, equal.count(SO2,4)) R> plot(cloud(precip ~ temp * wind | pollution, panel.aspect = 0.9, + data = USairpollution)) tempwind precip pollution tempwind precip pollution tempwind precip pollution tempwind precip pollution Fig. 2.21. Three-dimensional plots of temp, wind, and precip conditioned on levels of SO2. allowing the calculation of the required generalised distances. The cutoff dis- tance generally employed to identify an outlier is the maximum expected value from a sample of n random variables each having a chi-squared distribution on q degrees of freedom. The stalactite plot graphically illustrates the evolution of the outliers as the size of the subset of observations used for estimation in- creases. We will now illustrate the application of the stalactite plot on the US cities air pollution data. The plot (produced via stalac(USairpollution)) is shown in Figure 2.25. Initially most cities are indicated as outliers (a“*”in the plot), but as the number of observations on which the generalised distances are calculated is increased, the number of outliers indicated by the plot de- creases. The plot clearly shows the outlying nature of a number of cities over
- 72. 56 2 Looking at Multivariate Data: Visualisation R> plot(xyplot(lat ~ long| cut(depth, 3), data = quakes, + layout = c(3, 1), xlab = "Longitude", + ylab = "Latitude")) Longitude Latitude −35 −30 −25 −20 −15 −10 165170175180185 q q q q q q q q q q q q q qq qq q q q qq q q qq q q q q q q qqq q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qqqqqqqq qqq q q q q q q q q q q q q q q qq qqqq q q q q q q q q q q q q q qq q q q q q qq q q q q q qq q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qqq q q q q q q q q q q q q q q qq q q qqqq q qq q q q q q q qqq q q q q q q q q q q q q q q q q q q q q q q q qq q qq q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q qq q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q qq q q q q q q q qq q q q q q q q q q q q q q qq q q q q q qq q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q qq q q q q q q q q q q q q q qq q q q q q qq qq q qq q q q q q q q (39.4,253] 165170175180185 q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q qq q qq q q q q q q q q q q q q qq q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q qq qq q q q q q q q q q q (253,467] 165170175180185 qq q q q q q q q q q q q q q q q q qq q q q q qq q q q q q q q q q q q q q q q q q q q q qqq q qq q q q q q q q q q q q q q q q qq q q q q q qq q q q qq q q q q q qq q q qq q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq qq q q q qq qqq q q q q q q q q qq q qq q q q q q q q q qq q q q q qq q qq q q q q qqqq q q q q qq q q q q q q q q q q q qq q q q q qq qq qq q q q q q qqq q q qq qqq q qq q q qq qq q q qqqq q qqq q qq q qq q q q q qq qq q q qqq q q q q q q q q qq q q q qq qq q q q q qqqq q q q q q q q qq q q q qq q q q q qqq q q q qq q q q q q qq q qq q q qq qq q q q q q q q q q qq q q q qqq q qq q q q q q q qqq q q q q q q q q q q (467,681] Fig. 2.22. Scatterplots of latitude and longitude conditioned on three ranges of depth. nearly all values of m. The effect of masking is also clear; when all 41 ob- servations are used to calculate the generalised distances, only observations Chicago, Phoenix, and Providence are indicated to be outliers. 2.9 Summary Plotting multivariate data is an essential first step in trying to understand the story they may have to tell. The methods covered in this chapter provide just some basic ideas for taking an initial look at the data, and with software such as R there are many other possibilities for graphing multivariate obser-
- 73. 2.9 Summary 57 Longitude Latitude −35 −30 −25 −20 −15 −10 165170175180185 q qqq qq qq q qqqq qqq q q qq q q qqq q qqq q q q q q qqq q qqq qq q qqqqq q qq q qq qq qq qq qq qq q q qq qq q q qq qq qqqq q q qqqqqq qqqqqqqqqq q qqq qqqqqqq qq q q q qqqqq q qq q qqq q q q qqq q qqq qqq q q qq qq qq qq q qqqqqqq qqqq q q qqq qqq q qq q qq q qqqq qqqq qqq qqq qq q q qq q qq q qq q q qq q q q q q qqqq q q q q qq q q q q q qqq q q q q q q qqq qq q q q q q qq q q q qq qq q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q qq q q q q q q qq q q q qq qq q q q q q q q q q q qq qqq q q q qqq q q q q q q qq q qq q q q q q q q q q q q q q q qq qq q q q q q q q q q q q q q qq q q q qq q q qq q q q q qq qq qqqqq q q q qq q q q q q q q qq q qq q q q q q q q q q q qqq q q q q q qq q q q q q q q q q q q q q q qq q q q q qq q q q q q q qqq q q q q q q q q q Magnitude q q q q qq q qqqq qqq q q q qq q q q q qq qq q q qq q q q qq q q qq q q q qqqqqqqq q q q q qqq qq qqqq q qqqq q q q q q qqq qqqqq qqqqqqqqqqqq q q q q qqqq qqqq qqq q qqqqqqq q q qq q q q q q qqq qqqq qqq qqq qqq q q qq q qq q q q qq q q q q qq qq q qqq qqq q q qq q qq q qq q qqq qqq q qqqqq q q q q q q q q q q qq q qqq q qq q q q q q q q q q q q q q q qq q q q q q q q qq q q q qqq q q q q q q q q qq qqq qq qq q q q q qqq q q qqq q q q q q q qq qq q qqq q q qq q qq q q q q qqqq q q q qq q q q qq q q q qq q q q q q q q q q q q q qq qq q q qq q q q q qq q q qq q q q qqqq q q q q q q q q q q qq qq q q qq q q q q qqq q q qq q q q q qq q q q q q q qq q q q qq q q q q q q q q q q q q qq q q q q q qq q q q q q q q q q q q qqqq q q q q q q q qq q q qq q qq qqq q q q q qq q q q q q q q q q qqq q qq q q q q q q q qq q q qq q q q Magnitude qq q q q qq q qqqq q q q q qqq qqq qq q q q q qqqq qq qqq qq qqq q q q q qqq q q q qq qq q q q q qq qqqqq qqqqqq q qqqq qq qqqqqqq q qqq qq qq q q q qq qqq q q q q q q qq q q q qq qq qqq q q q q qq qq q qqq q qq qqq q q q qq qq qqq qq q q q qqqq q qqq q qq qqq qq q q q q q qq qq q q q q q q q qqqq q q q qqq q q q q qqqq q q q qq q q qqq q q qqq qq qqq qq q qq q qqq q q q q qqq qq q q qq qq q q q q q q q q q q q q q q q qq q q q q qq q qq q qq qq q q q q q q qq q q qq q qq q q q qqqq q q q q q q q qq q q q q q q q q q qq q q qq q q q q q q q qq q qq qq q q q qq q q q q q q q q q q qq qq q q qqq qq q qq qq q q q q qq q q q q qq q q qq qqqq qq q q q q q q q q q q q q q q q qq qq q q q q q q q q q q q q q q Magnitude 165170175180185 −35 −30 −25 −20 −15 −10 qqq q q qq qqqq q qqq q q q q q q q q qq qq q q qqq q qq qq q q qq q qqqqq q qqqq q q qqqqqq q q qqqqqq qqqq q q qqq qqq q q q qq qqqqqq qq q q qq q q q qq q q q qqq q qqqqqq q qq q q qq q q qqqqq q qqq q qqqqqq qqq qq qq q qq qqq q q q q q q qqq q q qq q q q q q qq q qq q q q q q q q q q q q qqq q qq qqq qqq q qq q q q qqq q q q q q q q q q q q q qq q q q q q qq q q q q q q q q q q q q q q q q q q qqq q q qq q qq q q qq q qq q q q q qq q q qq q q q q q q q q q q q q q q qq qq qqq q qq q q q q q qq q qq q q q qqq q q qqq q q qq q qqq qq qq q q q q qq q q q q qq q q q q q q q q qq qqq qq q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q Magnitude 100 200 300 400 500 600 Fig. 2.23. Scatterplots of latitude and longitude conditioned on magnitude, with depth coded by shading. vations, and readers are encouraged to explore more fully what is available. But graphics can often flatter to deceive and it is important not to be seduced when looking at a graphic into responding “what a great graph” rather than “what interesting data”. A graph that calls attention to itself pictorially is al- most surely a failure (see Becker et al. 1994), and unless graphs are relatively simple, they are unlikely to survive the first glance. Three-dimensional plots and trellis plots provide great pictures, which may often also be very informa- tive (as the examples in Sarkar 2008, demonstrate), but for multivariate data with many variables, they may struggle. In many situations, the most useful graphic for a set of multivariate data may be the scatterplot matrix, perhaps with the panels enhanced in some way; for example, by the addition of bivari- ate density estimates or bivariate boxplots. And all the graphical approaches discussed in this chapter may become more helpful when applied to the data
- 74. 58 2 Looking at Multivariate Data: Visualisation R> plot(cloud(depth ~ lat * long | Magnitude, data = quakes, + zlim = rev(range(quakes$depth)), + screen = list(z = 105, x = -70), panel.aspect = 0.9, + xlab = "Longitude", ylab = "Latitude", zlab = "Depth")) Longitude Latitude Depth Magnitude Longitude Latitude Depth Magnitude Longitude Latitude Depth Magnitude Longitude Latitude Depth Magnitude Fig. 2.24. Scatterplots of latitude and longitude conditioned on magnitude. after their dimensionality has been reduced in some way, often by the method to be described in the next chapter.
- 75. 2.9 Summary 59 Number of observations used for estimation Albany Albuquerque Atlanta Baltimore Buffalo Charleston Chicago Cincinnati Cleveland Columbus Dallas Denver Des Moines Detroit Hartford Houston Indianapolis Jacksonville Kansas City Little Rock Louisville Memphis Miami Milwaukee Minneapolis Nashville New Orleans Norfolk Omaha Philadelphia Phoenix Pittsburgh Providence Richmond Salt Lake City San Francisco Seattle St. Louis Washington Wichita Wilmington 41 34 27 20 13 ********************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************** Fig. 2.25. Stalactite plot of US cities air pollution data.
- 76. 60 2 Looking at Multivariate Data: Visualisation 2.10 Exercises Ex. 2.1 Use the bivariate boxplot on the scatterplot of each pair of variables in the air pollution data to identify any outliers. Calculate the correlation between each pair of variables using all the data and the data with any identified outliers removed. Comment on the results. Ex. 2.2 Compare the chi-plots with the corresponding scatterplots for each pair of variables in the air pollution data. Do you think that there is any advantage in the former? Ex. 2.3 Construct a scatterplot matrix of the body measurements data that has the appropriate boxplot on the diagonal panels and bivariate boxplots on the other panels. Compare the plot with Figure 2.17, and say which diagram you find more informative about the data. Ex. 2.4 Construct a further scatterplot matrix of the body measurements data that labels each point in a panel with the gender of the individual, and plot on each scatterplot the separate estimated bivariate densities for men and women. Ex. 2.5 Construct a scatterplot matrix of the chemical composition of Romano-British pottery given in Chapter 1 (Table 1.3), identifying each unit by its kiln number and showing the estimated bivariate density on each panel. What does the resulting diagram tell you? Ex. 2.6 Construct a bubble plot of the earthquake data using latitude and longitude as the scatterplot and depth as the circles, with greater depths giving smaller circles. In addition, divide the magnitudes into three equal ranges and label the points in your bubble plot with a different symbol depending on the magnitude group into which the point falls.
- 77. 3 Principal Components Analysis 3.1 Introduction One of the problems with a lot of sets of multivariate data is that there are simply too many variables to make the application of the graphical techniques described in the previous chapters successful in providing an informative ini- tial assessment of the data. And having too many variables can also cause problems for other multivariate techniques that the researcher may want to apply to the data. The possible problem of too many variables is sometimes known as the curse of dimensionality (Bellman 1961). Clearly the scatterplots, scatterplot matrices, and other graphics included in Chapter 2 are likely to be more useful when the number of variables in the data, the dimensionality of the data, is relatively small rather than large. This brings us to principal com- ponents analysis, a multivariate technique with the central aim of reducing the dimensionality of a multivariate data set while accounting for as much of the original variation as possible present in the data set. This aim is achieved by transforming to a new set of variables, the principal components, that are linear combinations of the original variables, which are uncorrelated and are ordered so that the first few of them account for most of the variation in all the original variables. In the best of all possible worlds, the result of a principal components analysis would be the creation of a small number of new variables that can be used as surrogates for the originally large number of variables and consequently provide a simpler basis for, say, graphing or summarising the data, and also perhaps when undertaking further multivariate analyses of the data. 3.2 Principal components analysis (PCA) The basic goal of principal components analysis is to describe variation in a set of correlated variables, x = (x1, . . . , xq), in terms of a new set of uncor- related variables, y = (y1, . . . , yq), each of which is a linear combination of DOI 10.1007/978-1-4419-9650-3_3, © Springer Science+Business Media, LLC 2011 61B. Everitt and T. Hothorn, An Introduction to Applied Multivariate Analysis with R: Use R!,
- 78. 62 3 Principal Components Analysis the x variables. The new variables are derived in decreasing order of “impor- tance” in the sense that y1 accounts for as much as possible of the variation in the original data amongst all linear combinations of x. Then y2 is chosen to account for as much as possible of the remaining variation, subject to being uncorrelated with y1, and so on. The new variables defined by this process, y1, . . . , yq, are the principal components. The general hope of principal components analysis is that the first few components will account for a substantial proportion of the variation in the original variables, x1, . . . , xq, and can, consequently, be used to provide a con- venient lower-dimensional summary of these variables that might prove useful for a variety of reasons. Consider, for example, a set of data consisting of examination scores for several different subjects for each of a number of stu- dents. One question of interest might be how best to construct an informative index of overall examination performance. One obvious possibility would be the mean score for each student, although if the possible or observed range of examination scores varied from subject to subject, it might be more sensible to weight the scores in some way before calculating the average, or alterna- tively standardise the results for the separate examinations before attempting to combine them. In this way, it might be possible to spread the students out further and so obtain a better ranking. The same result could often be achieved by applying principal components to the observed examination results and us- ing the student’s scores on the first principal components to provide a measure of examination success that maximally discriminates between them. A further possible application for principal components analysis arises in the field of economics, where complex data are often summarised by some kind of index number; for example, indices of prices, wage rates, cost of living, and so on. When assessing changes in prices over time, the economist will wish to allow for the fact that prices of some commodities are more variable than others, or that the prices of some of the commodities are considered more important than others; in each case the index will need to be weighted accordingly. In such examples, the first principal component can often satisfy the investigator’s requirements. But it is not always the first principal component that is of most interest to a researcher. A taxonomist, for example, when investigating variation in morphological measurements on animals for which all the pairwise correlations are likely to be positive, will often be more concerned with the second and subsequent components since these might provide a convenient description of aspects of an animal’s “shape”. The latter will often be of more interest to the researcher than aspects of an animal’s “size” which here, because of the positive correlations, will be reflected in the first principal component. For essentially the same reasons, the first principal component derived from, say, clinical psychiatric scores on patients may only provide an index of the severity of symptoms, and it is the remaining components that will give the psychiatrist important information about the “pattern” of symptoms.
- 79. 3.3 Finding the sample principal components 63 The principal components are most commonly (and properly) used as a means of constructing an informative graphical representation of the data (see later in the chapter) or as input to some other analysis. One example of the latter is provided by regression analysis; principal components may be useful here when: ˆ There are too many explanatory variables relative to the number of obser- vations. ˆ The explanatory variables are highly correlated. Both situations lead to problems when applying regression techniques, problems that may be overcome by replacing the original explanatory vari- ables with the first few principal component variables derived from them. An example will be given later, and other applications of the technique are described in Rencher (2002). In some disciplines, particularly psychology and other behavioural sciences, the principal components may be considered an end in themselves and re- searchers may then try to interpret them in a similar fashion as for the factors in an exploratory factor analysis (see Chapter 5). We shall make some com- ments about this practise later in the chapter. 3.3 Finding the sample principal components Principal components analysis is overwhelmingly an exploratory technique for multivariate data. Although there are inferential methods for using the sample principal components derived from a random sample of individuals from some population to test hypotheses about population principal components (see Jolliffe 2002), they are very rarely seen in accounts of principal components analysis that appear in the literature. Quintessentially principal components analysis is an aid for helping to understand the observed data set whether or not this is actually a “sample” in any real sense. We use this observation as the rationale for describing only sample principal components in this chapter. The first principal component of the observations is that linear combina- tion of the original variables whose sample variance is greatest amongst all possible such linear combinations. The second principal component is defined as that linear combination of the original variables that accounts for a maxi- mal proportion of the remaining variance subject to being uncorrelated with the first principal component. Subsequent components are defined similarly. The question now arises as to how the coefficients specifying the linear com- binations of the original variables defining each component are found. A little technical material is needed to answer this question. The first principal component of the observations, y1, is the linear combi- nation y1 = a11x1 + a12x2 + · · · + a1qxq
- 80. 64 3 Principal Components Analysis whose sample variance is greatest among all such linear combinations. Be- cause the variance of y1 could be increased without limit simply by increasing the coefficients a1 = (a11, a12, . . . , a1q), a restriction must be placed on these coefficients. As we shall see later, a sensible constraint is to require that the sum of squares of the coefficients should take the value one, although other constraints are possible and any multiple of the vector a1 produces basically the same component. To find the coefficients defining the first principal com- ponent, we need to choose the elements of the vector a1 so as to maximise the variance of y1 subject to the sum of squares constraint, which can be written a1 a1 = 1. The sample variance of y1 that is a linear function of the x variables is given by (see Chapter 1) a1 Sa1, where S is the q × q sample covariance matrix of the x variables. To maximise a function of several vari- ables subject to one or more constraints, the method of Lagrange multipliers is used. Full details are given in Morrison (1990) and Jolliffe (2002), and we will not give them here. (The algebra of an example with q = 2 is, however, given in Section 3.5.) We simply state that the Lagrange multiplier approach leads to the solution that a1 is the eigenvector or characteristic vector of the sample covariance matrix, S, corresponding to this matrix’s largest eigenvalue or characteristic root. The eigenvalues λ and eigenvectors γ of a q × q matrix A are such that Aγ = λγ; for more details, see, for example, Mardia, Kent, and Bibby (1979). The second principal component, y2, is defined to be the linear combination y2 = a21x1 + a22x2 + · · · + a2qxq (i.e., y2 = a2 x, where a2 = (a21, a22, . . . , a2q) and x = (x1, x2, . . . , xq)) that has the greatest variance subject to the following two conditions: a2 a2 = 1, a2 a1 = 0. (The second condition ensures that y1 and y2 are uncorrelated; i.e., that the sample correlation is zero.) Similarly, the jth principal component is that linear combination yj = aj x that has the greatest sample variance subject to the conditions aj aj = 1, aj ai = 0 (i < j). Application of the Lagrange multiplier technique demonstrates that the vector of coefficients defining the jth principal component, aj, is the eigenvector of S associated with its jth largest eigenvalue. If the q eigenvalues of S are denoted by λ1, λ2, . . . , λq, then by requiring that ai ai = 1 it can be shown that the variance of the ith principal component is given by λi. The total variance of the q principal components will equal the total variance of the original variables so that
- 81. 3.4 Covariance or the correlation matrix? 65 q i=1 λi = s2 1 + s2 2 + · · · + s2 q, where s2 i is the sample variance of xi. We can write this more concisely as q i=1 λi = trace(S). Consequently, the jth principal component accounts for a proportion Pj of the total variation of the original data, where Pj = λj trace(S) . The first m principal components, where m < q account for a proportion P(m) of the total variation in the original data, where P(m) = m j=1 λj trace(S) . In geometrical terms, it is easy to show that the first principal component defines the line of best fit (in the sense of minimising residuals orthogonal to the line) to the q-dimensional observations in the sample. These observations may therefore be represented in one dimension by taking their projection onto this line; that is, finding their first principal component score. If the observations happen to be collinear in q dimensions, this representation would account completely for the variation in the data and the sample covariance matrix would have only one non-zero eigenvalue. In practise, of course, such collinearity is extremely unlikely, and an improved representation would be given by projecting the q-dimensional observations onto the space of the best fit, this being defined by the first two principal components. Similarly, the first m components give the best fit in m dimensions. If the observations fit exactly into a space of m dimensions, it would be indicated by the presence of q − m zero eigenvalues of the covariance matrix. This would imply the presence of q − m linear relationships between the variables. Such constraints are sometimes referred to as structural relationships. In practise, in the vast majority of applications of principal components analysis, all the eigenvalues of the covariance matrix will be non-zero. 3.4 Should principal components be extracted from the covariance or the correlation matrix? One problem with principal components analysis is that it is not scale- invariant. What this means can be explained using an example given in Mardia et al. (1979). Suppose the three variables in a multivariate data set are weight in pounds, height in feet, and age in years, but for some reason we would like our principal components expressed in ounces, inches, and decades. Intuitively two approaches seem feasible;
- 82. 66 3 Principal Components Analysis 1. Multiply the variables by 16, 12, and 1/10, respectively and then carry out a principal components analysis on the covariance matrix of the three variables. 2. Carry out a principal components analysis on the covariance matrix of the original variables and then multiply the elements of the relevant compo- nent by 16, 12, and 1/10. Unfortunately, these two procedures do not generally lead to the same result. So if we imagine a set of multivariate data where the variables are of com- pletely different types, for example length, temperature, blood pressure, or anxiety rating, then the structure of the principal components derived from the covariance matrix will depend upon the essentially arbitrary choice of units of measurement; for example, changing the length from centimetres to inches will alter the derived components. Additionally, if there are large differences between the variances of the original variables, then those whose variances are largest will tend to dominate the early components. Principal components should only be extracted from the sample covariance matrix when all the original variables have roughly the same scale. But this is rare in practise and consequently, in practise, principal components are extracted from the correla- tion matrix of the variables, R. Extracting the components as the eigenvectors of R is equivalent to calculating the principal components from the original variables after each has been standardised to have unit variance. It should be noted, however, that there is rarely any simple correspondence between the components derived from S and those derived from R. And choosing to work with R rather than with S involves a definite but possibly arbitrary decision to make variables “equally important”. To demonstrate how the principal components of the covariance matrix of a data set can differ from the components extracted from the data’s correlation matrix, we will use the example given in Jolliffe (2002). The data in this example consist of eight blood chemistry variables measured on 72 patients in a clinical trial. The correlation matrix of the data, together with the standard deviations of each of the eight variables, is R> blood_corr [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] [1,] 1.000 0.290 0.202 -0.055 -0.105 -0.252 -0.229 0.058 [2,] 0.290 1.000 0.415 0.285 -0.376 -0.349 -0.164 -0.129 [3,] 0.202 0.415 1.000 0.419 -0.521 -0.441 -0.145 -0.076 [4,] -0.055 0.285 0.419 1.000 -0.877 -0.076 0.023 -0.131 [5,] -0.105 -0.376 -0.521 -0.877 1.000 0.206 0.034 0.151 [6,] -0.252 -0.349 -0.441 -0.076 0.206 1.000 0.192 0.077 [7,] -0.229 -0.164 -0.145 0.023 0.034 0.192 1.000 0.423 [8,] 0.058 -0.129 -0.076 -0.131 0.151 0.077 0.423 1.000
- 83. 3.4 Covariance or the correlation matrix? 67 R> blood_sd rblood plate wblood neut lymph bilir sodium potass 0.371 41.253 1.935 0.077 0.071 4.037 2.732 0.297 There are considerable differences between these standard deviations. We can apply principal components analysis to both the covariance and correlation matrix of the data using the following R code: R> blood_pcacov <- princomp(covmat = blood_cov) R> summary(blood_pcacov, loadings = TRUE) Importance of components: Comp.1 Comp.2 Comp.3 Comp.4 Standard deviation 41.2877 3.880213 2.641973 1.624584 Proportion of Variance 0.9856 0.008705 0.004036 0.001526 Cumulative Proportion 0.9856 0.994323 0.998359 0.999885 Comp.5 Comp.6 Comp.7 Comp.8 Standard deviation 3.540e-01 2.562e-01 8.511e-02 2.373e-02 Proportion of Variance 7.244e-05 3.794e-05 4.188e-06 3.255e-07 Cumulative Proportion 1.000e+00 1.000e+00 1.000e+00 1.000e+00 Loadings: Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Comp.6 Comp.7 Comp.8 rblood 0.943 0.329 plate -0.999 wblood -0.192 -0.981 neut 0.758 -0.650 lymph -0.649 -0.760 bilir 0.961 0.195 -0.191 sodium 0.193 -0.979 potass 0.329 -0.942 R> blood_pcacor <- princomp(covmat = blood_corr) R> summary(blood_pcacor, loadings = TRUE) Importance of components: Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Standard deviation 1.671 1.2376 1.1177 0.8823 0.7884 Proportion of Variance 0.349 0.1915 0.1562 0.0973 0.0777 Cumulative Proportion 0.349 0.5405 0.6966 0.7939 0.8716 Comp.6 Comp.7 Comp.8 Standard deviation 0.6992 0.66002 0.31996 Proportion of Variance 0.0611 0.05445 0.01280 Cumulative Proportion 0.9327 0.98720 1.00000 Loadings:
- 84. 68 3 Principal Components Analysis Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Comp.6 Comp.7 Comp.8 [1,] -0.194 0.417 0.400 0.652 0.175 -0.363 0.176 0.102 [2,] -0.400 0.154 0.168 -0.848 0.230 -0.110 [3,] -0.459 0.168 -0.274 0.251 0.403 0.677 [4,] -0.430 -0.472 -0.171 0.169 0.118 -0.237 0.678 [5,] 0.494 0.360 -0.180 -0.139 0.136 0.157 0.724 [6,] 0.319 -0.320 -0.277 0.633 -0.162 0.384 0.377 [7,] 0.177 -0.535 0.410 -0.163 -0.299 -0.513 0.367 [8,] 0.171 -0.245 0.709 0.198 0.469 -0.376 (The “blanks” in this output represent very small values.) Examining the re- sults, we see that each of the principal components of the covariance matrix is largely dominated by a single variable, whereas those for the correlation matrix have moderate-sized coefficients on several of the variables. And the first component from the covariance matrix accounts for almost 99% of the total variance of the observed variables. The components of the covariance matrix are completely dominated by the fact that the variance of the plate variable is roughly 400 times larger than the variance of any of the seven other variables. Consequently, the principal components from the covariance matrix simply reflect the order of the sizes of the variances of the observed variables. The results from the correlation matrix tell us, in particular, that a weighted contrast of the first four and last four variables is the linear function with the largest variance. This example illustrates that when variables are on very different scales or have very different variances, a principal components analysis of the data should be performed on the correlation matrix, not on the covariance matrix. 3.5 Principal components of bivariate data with correlation coefficient r Before we move on to look at some practical examples of the application of principal components analysis, it will be helpful to look in a little more detail at the mathematics of the method in one very simple case. We will do this in this section for bivariate data where the two variables, x1 and x2, have correlation coefficient r. The sample correlation matrix in this case is simply R = 1.0 r r 0.1 . In order to find the principal components of the data we need to find the eigenvalues and eigenvectors of R. The eigenvalues are found as the roots of the equation det(R − λI) = 0. This leads to the quadratic equation in λ
- 85. 3.5 Principal components of bivariate data with correlation coefficient r 69 (1 − λ)2 − r2 = 0, and solving this equation leads to eigenvalues λ1 = 1 + r, λ2 = 1 − r. Note that the sum of the eigenvalues is two, equal to trace(R). The eigenvector corresponding to λ1 is obtained by solving the equation Ra1 = λ1a1. This leads to the equations a11 + ra12 = (1 + r)a11, ra11 + a12 = (1 + r)a12. The two equations are identical, and both reduce to requiring a11 = a12. If we now introduce the normalisation constraint a1 a1 = 1, we find that a11 = a12 = 1 √ 2 . Similarly, we find the second eigenvector is given by a21 = 1√ 2 and a22 = − 1√ 2 . The two principal components are then given by y1 = 1 √ 2 (x1 + x2), y2 = 1 √ 2 (x1 − x2). We can calculate the sample variance of the first principal component as Var(y1) = Var 1 √ 2 (x1 + x2) = 1 2 Var(x1 + x2) = 1 2 [Var(x1) + Var(x2) + 2Cov(x1, x2)] = 1 2 (1 + 1 + 2r) = 1 + r. Similarly, the variance of the second principal component is 1 − r. Notice that if r < 0, the order of the eigenvalues and hence of the principal components is reversed; if r = 0, the eigenvalues are both equal to 1 and any two solutions at right angles could be chosen to represent the two components. Two further points should be noted: 1. There is an arbitrary sign in the choice of the elements of ai. It is custom- ary (but not universal) to choose ai1 to be positive. 2. The coefficients that define the two components do not depend on r, al- though the proportion of variance explained by each does change with r. As r tends to 1, the proportion of variance accounted for by y1, namely (1+r)/2, also tends to one. When r = 1, the points all align on a straight line and the variation in the data is unidimensional.
- 86. 70 3 Principal Components Analysis 3.6 Rescaling the principal components The coefficients defining the principal components derived as described in the previous section are often rescaled so that they are correlations or covariances between the original variables and the derived components. The rescaled co- efficients are often useful in interpreting a principal components analysis. The covariance of variable i with component j is given by Cov(xi, yj) = λjaji. The correlation of variable xi with component yj is therefore rxi,yj = λjaji Var(xi)Var(yj) = λjaji si λj = aji λj si . If the components are extracted from the correlation matrix rather than the covariance matrix, the correlation between variable and component becomes rxi,yj = aji λj because in this case the standard deviation, si, is unity. (Although for con- venience we have used the same nomenclature for the eigenvalues and the eigenvectors extracted from the covariance matrix or the correlation matrix, they will, of course, not be equal.) The rescaled coefficients from a principal components analysis of a correlation matrix are analogous to factor loadings, as we shall see in Chapter 5. Often these rescaled coefficients are presented as the results of a principal components analysis and used in interpretation. 3.7 How the principal components predict the observed covariance matrix In this section, we will look at how the principal components reproduce the observed covariance or correlation matrix from which they were extracted. To begin, let the initial vectors a1, a2, . . . , aq, that define the principal compo- nents be used to form a q × q matrix, A = (a1, a2, . . . , aq); we assume that these are vectors extracted from the covariance matrix, S, and scaled so that ai ai = 1. Arrange the eigenvalues λ1, λ2, . . . , λq along the main diagonal of a diagonal matrix, Λ. Then it can be shown that the covariance matrix of the observed variables x1, x2, . . . , xq is given by S = AΛA . This is known as the spectral decomposition of S. Rescaling the vectors a1, a2, . . . , aq so that the sum of squares of their elements is equal to the corresponding eigenvalue (i.e., calculating a∗ i = λ 1 2 i ai) allows S to be written more simply as
- 87. 3.8 Choosing the number of components 71 S = A∗ A∗ , where A∗ = a∗ 1 . . . a∗ q . If the matrix A∗ m is formed from, say, the first m components rather than from all q, then A∗ mA∗ m gives the predicted value of S based on these m components. It is often useful to calculate such a predicted value based on the number of components considered adequate to describe the data to informally assess the “fit” of the principal components analysis. How this number of components might be chosen is considered in the next section. 3.8 Choosing the number of components As described earlier, principal components analysis is seen to be a technique for transforming a set of observed variables into a new set of variables that are uncorrelated with one another. The variation in the original q variables is only completely accounted for by all q principal components. The useful- ness of these transformed variables, however, stems from their property of accounting for the variance in decreasing proportions. The first component, for example, accounts for the maximum amount of variation possible for any linear combination of the original variables. But how useful is this artificial variate constructed from the observed variables? To answer this question we would first need to know the proportion of the total variance of the original variables for which it accounted. If, for example, 80% of the variation in a multivariate data set involving six variables could be accounted for by a sim- ple weighted average of the variable values, then almost all the variation can be expressed along a single continuum rather than in six-dimensional space. The principal components analysis would have provided a highly parsimonious summary (reducing the dimensionality of the data from six to one) that might be useful in later analysis. So the question we need to ask is how many components are needed to provide an adequate summary of a given data set. A number of informal and more formal techniques are available. Here we shall concentrate on the former; examples of the use of formal inferential methods are given in Jolliffe (2002) and Rencher (2002). The most common of the relatively ad hoc procedures that have been sug- gested for deciding upon the number of components to retain are the following: ˆ Retain just enough components to explain some specified large percentage of the total variation of the original variables. Values between 70% and 90% are usually suggested, although smaller values might be appropriate as q or n, the sample size, increases. ˆ Exclude those principal components whose eigenvalues are less than the average, q i=1 λi q . Since q i=1 λi = trace(S), the average eigenvalue is also the average variance of the original variables. This method then retains
- 88. 72 3 Principal Components Analysis those components that account for more variance than the average for the observed variables. ˆ When the components are extracted from the correlation matrix, trace(R) = q, and the average variance is therefore one, so applying the rule in the previous bullet point, components with eigenvalues less than one are excluded. This rule was originally suggested by Kaiser (1958), but Jolliffe (1972), on the basis of a number of simulation studies, proposed that a more appropriate procedure would be to exclude components ex- tracted from a correlation matrix whose associated eigenvalues are less than 0.7. ˆ Cattell (1966) suggests examination of the plot of the λi against i, the so- called scree diagram. The number of components selected is the value of i corresponding to an“elbow”in the curve, i.e., a change of slope from“steep” to “shallow”. In fact, Cattell was more specific than this, recommending to look for a point on the plot beyond which the scree diagram defines a more or less straight line, not necessarily horizontal. The first point on the straight line is then taken to be the last component to be retained. And it should also be remembered that Cattell suggested the scree diagram in the context of factor analysis rather than applied to principal components analysis. ˆ A modification of the scree digram described by Farmer (1971) is the log- eigenvalue diagram consisting of a plot of log(λi) against i. Returning to the results of the principal components analysis of the blood chemistry data given in Section 3.3, we find that the first four components account for nearly 80% of the total variance, but it takes a further two com- ponents to push this figure up to 90%. A cutoff of one for the eigenvalues leads to retaining three components, and with a cutoff of 0.7 four components are kept. Figure 3.1 shows the scree diagram and log-eigenvalue diagram for the data and the R code used to construct the two diagrams. The former plot may suggest four components, although this is fairly subjective, and the latter seems to be of little help here because it appears to indicate retaining seven components, hardly much of a dimensionality reduction. The example illustrates that the proposed methods for deciding how many components to keep can (and often do) lead to different conclusions. 3.9 Calculating principal components scores If we decide that we need, say, m principal components to adequately repre- sent our data (using one or another of the methods described in the previous section), then we will generally wish to calculate the scores on each of these components for each individual in our sample. If, for example, we have derived the components from the covariance matrix, S, then the m principal compo- nents scores for individual i with original q ×1 vector of variable values xi are obtained as
- 89. 3.9 Calculating principal components scores 73 R> plot(log(blood_pcacor$sdev^2), xlab = "Component number", + ylab = "log(Component variance)", type="l", + main = "Log(eigenvalue) diagram") 1 2 3 4 5 6 7 8 0.02.5 Scree diagram Component number Componentvariance 1 2 3 4 5 6 7 8 −2.01.0 Log(eigenvalue) diagram Component number log(Componentvariance) Fig. 3.1. Scree diagram and log-eigenvalue diagram for principal components of the correlation matrix of the blood chemistry data. yi1 = a1 xi yi2 = a2 xi ... yim = amxi If the components are derived from the correlation matrix, then xi would contain individual i’s standardised scores for each variable. R> plot(blood_pcacor$sdev^2, xlab = "Component number", + ylab = "Component variance", type = "l", main = "Scree diagram")
- 90. 74 3 Principal Components Analysis The principal components scores calculated as above have variances equal to λj for j = 1, . . . , m. Many investigators might prefer to have scores with mean zero and variance equal to unity. Such scores can be found as z = Λ−1 m Amx, where Λm is an m × m diagonal matrix with λ1, λ2, . . . , λm on the main diagonal, Am = (a1 . . . am), and x is the q × 1 vector of standardised scores. We should note here that the first m principal components scores are the same whether we retain all possible q components or just the first m. As we shall see in Chapter 5, this is not the case with the calculation of factor scores. 3.10 Some examples of the application of principal components analysis In this section, we will look at the application of PCA to a number of data sets, beginning with one involving only two variables, as this allows us to illustrate graphically an important point about this type of analysis. 3.10.1 Head lengths of first and second sons Table 3.1: headsize data. Head Size Data. head1 breadth1 head2 breadth2 head1 breadth1 head2 breadth2 191 155 179 145 190 159 195 157 195 149 201 152 188 151 187 158 181 148 185 149 163 137 161 130 183 153 188 149 195 155 183 158 176 144 171 142 186 153 173 148 208 157 192 152 181 145 182 146 189 150 190 149 175 140 165 137 197 159 189 152 192 154 185 152 188 152 197 159 174 143 178 147 192 150 187 151 176 139 176 143 179 158 186 148 197 167 200 158 183 147 174 147 190 163 187 150 174 150 185 152 The data in Table 3.1 give the head lengths and head breadths (in millimetres) for each of the first two adult sons in 25 families. Here we shall use only the head lengths; the head breadths will be used later in the chapter. The mean vector and covariance matrix of the head length measurements are found using
- 91. 3.10 Some examples of the application of principal components analysis 75 R> head_dat <- headsize[, c("head1", "head2")] R> colMeans(head_dat) head1 head2 185.7 183.8 R> cov(head_dat) head1 head2 head1 95.29 69.66 head2 69.66 100.81 The principal components of these data, extracted from their covariance ma- trix, can be found using R> head_pca <- princomp(x = head_dat) R> head_pca Call: princomp(x = head_dat) Standard deviations: Comp.1 Comp.2 12.691 5.215 2 variables and 25 observations. R> print(summary(head_pca), loadings = TRUE) Importance of components: Comp.1 Comp.2 Standard deviation 12.6908 5.2154 Proportion of Variance 0.8555 0.1445 Cumulative Proportion 0.8555 1.0000 Loadings: Comp.1 Comp.2 head1 0.693 -0.721 head2 0.721 0.693 and are y1 = 0.693x1 + 0.721x2 y2 = −0.721x1 + 0.693x2 with variances 167.77 and 28.33. The first principal component accounts for a proportion 167.77/(167.77+28.33) = 0.86 of the total variance in the original variables. Note that the total variance of the principal components is 196.10, which as expected is equal to the total variance of the original variables, found by adding the relevant terms in the covariance matrix given earlier; i.e., 95.29 + 100.81 = 196.10.
- 92. 76 3 Principal Components Analysis How should the two derived components be interpreted? The first compo- nent is essentially the sum of the head lengths of the two sons, and the second component is the difference in head lengths. Perhaps we can label the first component “size” and the second component “shape”, but later we will have some comments about trying to give principal components such labels. To calculate an individual’s score on a component, we simply multiply the variable values minus the appropriate mean by the loading for the variable and add these values over all variables. We can illustrate this calculation using the data for the first family, where the head length of the first son is 191 mm and for the second son 179 mm. The score for this family on the first principal component is calculated as 0.693 · (191 − 185.72) + 0.721 · (179 − 183.84) = 0.169, and on the second component the score is −0.721 · (191 − 185.72) + 0.693 · (179 − 183.84) = −7.61. The variance of the first principal components scores will be 167.77, and the variance of the second principal component scores will be 28.33. We can plot the data showing the axes corresponding to the principal components. The first axis passes through the mean of the data and has slope 0.721/0.693, and the second axis also passes through the mean and has slope −0.693/0.721. The plot is shown in Figure 3.2. This example illustrates that a principal components analysis is essentially simply a rotation of the axes of the multivariate data scatter. And we can also plot the principal components scores to give Figure 3.3. (Note that in this figure the range of the x-axis and the range for the y-axis have been made the same to account for the larger variance of the first principal component.) We can use the principal components analysis of the head size data to demonstrate how the principal components reproduce the observed covariance matrix. We first need to rescale the principal components we have at this point by multiplying them by the square roots of their respective variances to give the new components y1 = 12.952(0.693x1 + 0.721x2), i.e., y1 = 8.976x1 + 9.338x2 and y2 = 5.323(−0.721x1 + 0.693x2), i.e., y2 = −3.837x1 + 3.688x2, leading to the matrix A∗ 2 as defined in Section 1.5.1: A∗ 2 = 8.976 −3.837 9.338 3.688 . Multiplying this matrix by its transpose should recreate the covariance matrix of the head length data; doing the matrix multiplication shows that it does recreate S:
- 93. 3.10 Some examples of the application of principal components analysis 77 R> a1<-183.84-0.721*185.72/0.693 R> b1<-0.721/0.693 R> a2<-183.84-(-0.693*185.72/0.721) R> b2<--0.693/0.721 R> plot(head_dat, xlab = "First son's head length (mm)", + ylab = "Second son's head length") R> abline(a1, b1) R> abline(a2, b2, lty = 2) q q q q q q q q q qq q q q q q q q q q q q q q q 170 180 190 200 160170180190200 First son's head length (mm) Secondson'sheadlength Fig. 3.2. Head length of first and second sons, showing axes corresponding to the principal components of the sample covariance matrix of the data. A∗ 2(A∗ 2) = 95.29 69.66 69.66 100.81 . (As an exercise, readers might like to find the predicted covariance matrix using only the first component.) The head size example has been useful for discussing some aspects of prin- cipal components analysis but it is not, of course, typical of multivariate data sets encountered in practise, where many more than two variables will be recorded for each individual in a study. In the next two subsections, we con- sider some more interesting examples.
- 94. 78 3 Principal Components Analysis R> xlim <- range(head_pca$scores[,1]) R> plot(head_pca$scores, xlim = xlim, ylim = xlim) q qq q q q q q q q q q q q qq qq q q q q q q q −30 −20 −10 0 10 20 −30−20−1001020 Comp.1 Comp.2 Fig. 3.3. Plot of the first two principal component scores for the head size data. 3.10.2 Olympic heptathlon results The pentathlon for women was first held in Germany in 1928. Initially this consisted of the shot put, long jump, 100 m, high jump, and javelin events, held over two days. In the 1964 Olympic Games, the pentathlon became the first combined Olympic event for women, consisting now of the 80 m hurdles, shot, high jump, long jump, and 200 m. In 1977, the 200 m was replaced by the 800 m run, and from 1981 the IAAF brought in the seven-event heptathlon in place of the pentathlon, with day one containing the events 100 m hurdles, shot, high jump, and 200 m run, and day two the long jump, javelin, and 800 m run. A scoring system is used to assign points to the results from each event, and the winner is the woman who accumulates the most points over the two days. The event made its first Olympic appearance in 1984.
- 95. 3.10 Some examples of the application of principal components analysis 79 Table 3.2: heptathlon data. Results of Olympic heptathlon, Seoul, 1988. hurdles highjump shot run200m longjump javelin run800m score Joyner-Kersee (USA) 12.69 1.86 15.80 22.56 7.27 45.66 128.51 7291 John (GDR) 12.85 1.80 16.23 23.65 6.71 42.56 126.12 6897 Behmer (GDR) 13.20 1.83 14.20 23.10 6.68 44.54 124.20 6858 Sablovskaite (URS) 13.61 1.80 15.23 23.92 6.25 42.78 132.24 6540 Choubenkova (URS) 13.51 1.74 14.76 23.93 6.32 47.46 127.90 6540 Schulz (GDR) 13.75 1.83 13.50 24.65 6.33 42.82 125.79 6411 Fleming (AUS) 13.38 1.80 12.88 23.59 6.37 40.28 132.54 6351 Greiner (USA) 13.55 1.80 14.13 24.48 6.47 38.00 133.65 6297 Lajbnerova (CZE) 13.63 1.83 14.28 24.86 6.11 42.20 136.05 6252 Bouraga (URS) 13.25 1.77 12.62 23.59 6.28 39.06 134.74 6252 Wijnsma (HOL) 13.75 1.86 13.01 25.03 6.34 37.86 131.49 6205 Dimitrova (BUL) 13.24 1.80 12.88 23.59 6.37 40.28 132.54 6171 Scheider (SWI) 13.85 1.86 11.58 24.87 6.05 47.50 134.93 6137 Braun (FRG) 13.71 1.83 13.16 24.78 6.12 44.58 142.82 6109 Ruotsalainen (FIN) 13.79 1.80 12.32 24.61 6.08 45.44 137.06 6101 Yuping (CHN) 13.93 1.86 14.21 25.00 6.40 38.60 146.67 6087 Hagger (GB) 13.47 1.80 12.75 25.47 6.34 35.76 138.48 5975 Brown (USA) 14.07 1.83 12.69 24.83 6.13 44.34 146.43 5972 Mulliner (GB) 14.39 1.71 12.68 24.92 6.10 37.76 138.02 5746 Hautenauve (BEL) 14.04 1.77 11.81 25.61 5.99 35.68 133.90 5734 Kytola (FIN) 14.31 1.77 11.66 25.69 5.75 39.48 133.35 5686 Geremias (BRA) 14.23 1.71 12.95 25.50 5.50 39.64 144.02 5508 Hui-Ing (TAI) 14.85 1.68 10.00 25.23 5.47 39.14 137.30 5290 Jeong-Mi (KOR) 14.53 1.71 10.83 26.61 5.50 39.26 139.17 5289 Launa (PNG) 16.42 1.50 11.78 26.16 4.88 46.38 163.43 4566 In the 1988 Olympics held in Seoul, the heptathlon was won by one of the stars of women’s athletics in the USA, Jackie Joyner-Kersee. The results for all 25 competitors in all seven disciplines are given in Table 3.2 (from Hand, Daly, Lunn, McConway, and Ostrowski 1994). We shall analyse these data using principal components analysis with a view to exploring the structure of the data and assessing how the derived principal components scores (discussed later) relate to the scores assigned by the official scoring system. But before undertaking the principal components analysis, it is good data analysis practise to carry out an initial assessment of the data using one or another of the graphics described in Chapter 2. Some numerical summaries may also be helpful before we begin the main analysis. And before any of these, it will help to score all seven events in the same direction so that“large”values are indicative of a “better” performance. The R code for reversing the values for some events, then calculating the correlation coefficients between the ten events and finally constructing the scatterplot matrix of the data is R> heptathlon$hurdles <- with(heptathlon, max(hurdles)-hurdles) R> heptathlon$run200m <- with(heptathlon, max(run200m)-run200m) R> heptathlon$run800m <- with(heptathlon, max(run800m)-run800m)
- 96. 80 3 Principal Components Analysis R> score <- which(colnames(heptathlon) == "score") R> round(cor(heptathlon[,-score]), 2) hurdles highjump shot run200m longjump javelin run800m hurdles 1.00 0.81 0.65 0.77 0.91 0.01 0.78 highjump 0.81 1.00 0.44 0.49 0.78 0.00 0.59 shot 0.65 0.44 1.00 0.68 0.74 0.27 0.42 run200m 0.77 0.49 0.68 1.00 0.82 0.33 0.62 longjump 0.91 0.78 0.74 0.82 1.00 0.07 0.70 javelin 0.01 0.00 0.27 0.33 0.07 1.00 -0.02 run800m 0.78 0.59 0.42 0.62 0.70 -0.02 1.00 R> plot(heptathlon[,-score]) The scatterplot matrix appears in Figure 3.4. hurdles 1.50 1.85 0 2 4 36 44 02 1.501.85 highjump shot 1014 024 run200m longjump 5.07.0 3644 javelin 0 2 10 14 5.0 7.0 0 20 020 run800m Fig. 3.4. Scatterplot matrix of the seven heptathlon events after transforming some variables so that for all events large values are indicative of a better performance.
- 97. 3.10 Some examples of the application of principal components analysis 81 Examination of the correlation matrix shows that most pairs of events are positively correlated, some moderately (for example, high jump and shot) and others relatively highly (for example, high jump and hurdles). The exceptions to this general observation are the relationships between the javelin event and the others, where almost all the correlations are close to zero. One explanation might be that the javelin is a very “technical” event and perhaps the training for the other events does not help the competitors in the javelin. But before we speculate further, we should look at the scatterplot matrix of the seven events shown in Figure 3.4. One very clear observation in this plot is that for all events except the javelin there is an outlier who is very much poorer than the other athletes at these six events, and this is the competitor from Papua New Guinea (PNG), who finished last in the competition in terms of points scored. But surprisingly, in the scatterplots involving the javelin, it is this competitor who again stands out, but in this case she has the third highest value for the event. It might be sensible to look again at both the correlation matrix and the scatterplot matrix after removing the competitor from PNG; the relevant R code is R> heptathlon <- heptathlon[-grep("PNG", rownames(heptathlon)),] R> score <- which(colnames(heptathlon) == "score") R> round(cor(heptathlon[,-score]), 2) hurdles highjump shot run200m longjump javelin run800m hurdles 1.00 0.58 0.77 0.83 0.89 0.33 0.56 highjump 0.58 1.00 0.46 0.39 0.66 0.35 0.15 shot 0.77 0.46 1.00 0.67 0.78 0.34 0.41 run200m 0.83 0.39 0.67 1.00 0.81 0.47 0.57 longjump 0.89 0.66 0.78 0.81 1.00 0.29 0.52 javelin 0.33 0.35 0.34 0.47 0.29 1.00 0.26 run800m 0.56 0.15 0.41 0.57 0.52 0.26 1.00 The new scatterplot matrix is shown in Figure 3.5. Several of the cor- relations are changed to some degree from those shown before removal of the PNG competitor, particularly the correlations involving the javelin event, where the very small correlations between performances in this event and the others have increased considerably. Given the relatively large overall change in the correlation matrix produced by omitting the PNG competitor, we shall extract the principal components of the data from the correlation matrix after this omission. The principal components can now be found using R> heptathlon_pca <- prcomp(heptathlon[, -score], scale = TRUE) R> print(heptathlon_pca) Standard deviations: [1] 2.08 0.95 0.91 0.68 0.55 0.34 0.26 Rotation:
- 98. 82 3 Principal Components Analysis R> plot(heptathlon[,-score], pch = ".", cex = 1.5) hurdles 1.70 0 2 4 36 44 1.53.0 1.70 highjump shot 1014 024 run200m longjump 5.57.0 3644 javelin 1.5 3.0 10 14 5.5 7.0 20 35 2035 run800m Fig. 3.5. Scatterplot matrix for the heptathlon data after removing observations of the PNG competitor. PC1 PC2 PC3 PC4 PC5 PC6 PC7 hurdles -0.45 0.058 -0.17 0.048 -0.199 0.847 -0.070 highjump -0.31 -0.651 -0.21 -0.557 0.071 -0.090 0.332 shot -0.40 -0.022 -0.15 0.548 0.672 -0.099 0.229 run200m -0.43 0.185 0.13 0.231 -0.618 -0.333 0.470 longjump -0.45 -0.025 -0.27 -0.015 -0.122 -0.383 -0.749 javelin -0.24 -0.326 0.88 0.060 0.079 0.072 -0.211 run800m -0.30 0.657 0.19 -0.574 0.319 -0.052 0.077 The summary method can be used for further inspection of the details: R> summary(heptathlon_pca) Importance of components: PC1 PC2 PC3 PC4 PC5 PC6
- 99. 3.10 Some examples of the application of principal components analysis 83 Standard deviation 2.079 0.948 0.911 0.6832 0.5462 0.3375 Proportion of Variance 0.618 0.128 0.119 0.0667 0.0426 0.0163 Cumulative Proportion 0.618 0.746 0.865 0.9313 0.9739 0.9902 PC7 Standard deviation 0.26204 Proportion of Variance 0.00981 Cumulative Proportion 1.00000 The linear combination for the first principal component is 2 R> a1 <- heptathlon_pca$rotation[,1] R> a1 hurdles highjump shot run200m longjump javelin run800m -0.4504 -0.3145 -0.4025 -0.4271 -0.4510 -0.2423 -0.3029 We see that the hurdles and long jump events receive the highest weight but the javelin result is less important. For computing the first principal compo- nent, the data need to be rescaled appropriately. The center and the scaling used by prcomp internally can be extracted from the heptathlon_pca via R> center <- heptathlon_pca$center R> scale <- heptathlon_pca$scale Now, we can apply the scale function to the data and multiply it with the loadings matrix in order to compute the first principal component score for each competitor R> hm <- as.matrix(heptathlon[,-score]) R> drop(scale(hm, center = center, scale = scale) %*% + heptathlon_pca$rotation[,1]) Joyner-Kersee (USA) John (GDR) Behmer (GDR) -4.757530 -3.147943 -2.926185 Sablovskaite (URS) Choubenkova (URS) Schulz (GDR) -1.288136 -1.503451 -0.958467 Fleming (AUS) Greiner (USA) Lajbnerova (CZE) -0.953445 -0.633239 -0.381572 Bouraga (URS) Wijnsma (HOL) Dimitrova (BUL) -0.522322 -0.217701 -1.075984 Scheider (SWI) Braun (FRG) Ruotsalainen (FIN) 0.003015 0.109184 0.208868 Yuping (CHN) Hagger (GB) Brown (USA) 0.232507 0.659520 0.756855 Mulliner (GB) Hautenauve (BEL) Kytola (FIN) 1.880933 1.828170 2.118203 Geremias (BRA) Hui-Ing (TAI) Jeong-Mi (KOR) 2.770706 3.901167 3.896848
- 100. 84 3 Principal Components Analysis or, more conveniently, by extracting the first from all pre-computed principal components: R> predict(heptathlon_pca)[,1] Joyner-Kersee (USA) John (GDR) Behmer (GDR) -4.757530 -3.147943 -2.926185 Sablovskaite (URS) Choubenkova (URS) Schulz (GDR) -1.288136 -1.503451 -0.958467 Fleming (AUS) Greiner (USA) Lajbnerova (CZE) -0.953445 -0.633239 -0.381572 Bouraga (URS) Wijnsma (HOL) Dimitrova (BUL) -0.522322 -0.217701 -1.075984 Scheider (SWI) Braun (FRG) Ruotsalainen (FIN) 0.003015 0.109184 0.208868 Yuping (CHN) Hagger (GB) Brown (USA) 0.232507 0.659520 0.756855 Mulliner (GB) Hautenauve (BEL) Kytola (FIN) 1.880933 1.828170 2.118203 Geremias (BRA) Hui-Ing (TAI) Jeong-Mi (KOR) 2.770706 3.901167 3.896848 The first two components account for 75% of the variance. A barplot of each component’s variance (see Figure 3.6) shows how the first two compo- nents dominate. R> plot(heptathlon_pca) Variances 01234 Fig. 3.6. Barplot of the variances explained by the principal components (with observations for PNG removed).
- 101. 3.10 Some examples of the application of principal components analysis 85 The correlation between the score given to each athlete by the standard scoring system used for the heptathlon and the first principal component score can be found from R> cor(heptathlon$score, heptathlon_pca$x[,1]) [1] -0.9931 This implies that the first principal component is in good agreement with the score assigned to the athletes by official Olympic rules; a scatterplot of the official score and the first principal component is given in Figure 3.7. (The fact that the correlation is negative is unimportant here because of the arbi- trariness of the signs of the coefficients defining the first principal component; it is the magnitude of the correlation that is important.) R> plot(heptathlon$score, heptathlon_pca$x[,1]) q q q q q qq q qq q q qqqq qq qq q q qq 5500 6000 6500 7000 −4−2024 heptathlon$score heptathlon_pca$x[,1] Fig. 3.7. Scatterplot of the score assigned to each athlete in 1988 and the first principal component.
- 102. 86 3 Principal Components Analysis 3.10.3 Air pollution in US cities In this subsection, we will return to the air pollution data introduced in Chap- ter 1. The data were originally collected to investigate the determinants of pol- lution, presumably by regressing SO2 on the six other variables. Here, however, we shall examine how principal components analysis can be used to explore various aspects of the data, and will then look at how such an analysis can also be used to address the determinants of pollution question. To begin we shall ignore the SO2 variable and concentrate on the others, two of which relate to human ecology (popul, manu) and four to climate (temp, Wind, precip, predays). A case can be made to use negative temperature values in subsequent analyses since then all six variables are such that high values represent a less attractive environment. This is, of course, a personal view, but as we shall see later, the simple transformation of temp does aid interpretation. Prior to undertaking the principal components analysis on the air pollution data, we will again construct a scatterplot matrix of the six variables, but here we include the histograms for each variable on the main diagonal. The diagram that results is shown in Figure 3.8. A clear message from Figure 3.8 is that there is at least one city, and probably more than one, that should be considered an outlier. (This should come as no surprise given the investigation of the data in Chapter 2.) On the manu variable, for example, Chicago, with a value of 3344, has about twice as many manufacturing enterprises employing 20 or more workers as the city with the second highest number (Philadelphia). We shall return to this potential problem later in the chapter, but for the moment we shall carry on with a principal components analysis of the data for all 41 cities. For the data in Table 1.5, it seems necessary to extract the principal com- ponents from the correlation rather than the covariance matrix, since the six variables to be used are on very different scales. The correlation matrix and the principal components of the data can be obtained in R using the following command line code: R> cor(USairpollution[,-1]) manu popul wind precip predays negtemp manu 1.00000 0.95527 0.23795 -0.03242 0.13183 0.19004 popul 0.95527 1.00000 0.21264 -0.02612 0.04208 0.06268 wind 0.23795 0.21264 1.00000 -0.01299 0.16411 0.34974 precip -0.03242 -0.02612 -0.01299 1.00000 0.49610 -0.38625 predays 0.13183 0.04208 0.16411 0.49610 1.00000 0.43024 negtemp 0.19004 0.06268 0.34974 -0.38625 0.43024 1.00000 R> usair_pca <- princomp(USairpollution[,-1], cor = TRUE)
- 103. 3.10 Some examples of the application of principal components analysis 87 R> data("USairpollution", package = "HSAUR2") R> panel.hist <- function(x, ...) { + usr <- par("usr"); on.exit(par(usr)) + par(usr = c(usr[1:2], 0, 1.5) ) + h <- hist(x, plot = FALSE) + breaks <- h$breaks; nB <- length(breaks) + y <- h$counts; y <- y/max(y) + rect(breaks[-nB], 0, breaks[-1], y, col="grey", ...) + } R> USairpollution$negtemp <- USairpollution$temp * (-1) R> USairpollution$temp <- NULL R> pairs(USairpollution[,-1], diag.panel = panel.hist, + pch = ".", cex = 1.5) manu 0 2000 10 40 −75 −50 02000 02000 popul wind 6912 1040 precip predays 40120 0 2000 −75−50 6 9 12 40 120 negtemp Fig. 3.8. Scatterplot matrix of six variables in the air pollution data.
- 104. 88 3 Principal Components Analysis R> summary(usair_pca, loadings = TRUE) Importance of components: Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Standard deviation 1.482 1.225 1.1810 0.8719 0.33848 Proportion of Variance 0.366 0.250 0.2324 0.1267 0.01910 Cumulative Proportion 0.366 0.616 0.8485 0.9752 0.99426 Comp.6 Standard deviation 0.185600 Proportion of Variance 0.005741 Cumulative Proportion 1.000000 Loadings: Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Comp.6 manu -0.612 0.168 -0.273 -0.137 0.102 0.703 popul -0.578 0.222 -0.350 -0.695 wind -0.354 -0.131 0.297 0.869 -0.113 precip -0.623 -0.505 0.171 0.568 predays -0.238 -0.708 -0.311 -0.580 negtemp -0.330 -0.128 0.672 -0.306 0.558 -0.136 One thing to note about the correlations is the very high values for manu and popul, a finding we will return to later. We see that the first three components all have variances (eigenvalues) greater than one and together account for almost 85% of the variance of the original variables. Scores on these three components might be used to graph the data with little loss of information. We shall illustrate this possibility later. Many users of principal components analysis might be tempted to search for an interpretation of the derived components that allows them to be “la- belled” in some sense. This requires examining the coefficients defining each component (in the output shown above, these are scaled so that their sums of squares equal unity–“blanks” indicate near zero values). We see that the first component might be regarded as some index of “quality of life”, with high values indicating a relatively poor environment (in the authors’ opinion at least). The second component is largely concerned with a city’s rainfall having high coefficients for precip and predays and might be labelled as the “wet weather” component. Component three is essentially a contrast between precip and negtemp and will separate cities having high temperatures and high rainfall from those that are colder but drier. A suitable label might be simply “climate type”. Attempting to label components in this way is common, but perhaps it should be a little less common; the following quotation from Marriott (1974) should act as a salutary warning about the dangers of overinterpretation: It must be emphasised that no mathematical method is, or could be, designed to give physically meaningful results. If a mathematical ex-
- 105. 3.10 Some examples of the application of principal components analysis 89 pression of this sort has an obvious physical meaning, it must be at- tributed to a lucky change, or to the fact that the data have a strongly marked structure that shows up in analysis. Even in the latter case, quite small sampling fluctuations can upset the interpretation; for ex- ample, the first two principal components may appear in reverse order, or may become confused altogether. Reification then requires consid- erable skill and experience if it is to give a true picture of the physical meaning of the data. Even if we do not care to label the three components, they can still be used as the basis of various graphical displays of the cities. In fact, this is often the most useful aspect of a principal components analysis since regarding the principal components analysis as a means of providing an informative view of multivariate data has the advantage of making it less urgent or tempting to try to interpret and label the components. The first few component scores provide a low-dimensional “map” of the observations in which the Euclidean distances between the points representing the individuals best approximate in some sense the Euclidean distances between the individuals based on the original variables. We shall return to this point in Chapter 4. So we will begin by looking at the scatterplot matrix of the first three principal components and in each panel show the relevant bivariate boxplot; points are labelled by abbreviated city names. The plot and the code used to create it are shown in Figure 3.9. The plot again demonstrates clearly that Chicago is an outlier and suggests that Phoenix and Philadelphia may also be suspects in this respect. Phoenix appears to offer the best quality of life (on the limited basis of the six variables recorded), and Buffalo is a city to avoid if you prefer a drier environment. We leave further interpretation to the readers. We will now consider the main goal in the researcher’s mind when col- lecting the air pollution data, namely determining which of the climate and human ecology variables are the best predictors of the degree of air pollution in a city as measured by the sulphur dioxide content of the air. This ques- tion would normally be addressed by multiple linear regression, but there is a potential problem with applying this technique to the air pollution data, and that is the very high correlation between the manu and popul variables. We might, of course, deal with this problem by simply dropping either manu or popul, but here we will consider a possible alternative approach, and that is regressing the SO2 levels on the principal components derived from the six other variables in the data (see Figure 3.10). Using principal components in regression analysis is discussed in detail in Jolliffe (2002); here we simply give an example. The first question we need to ask is “how many principal compo- nents should be used as explanatory variables in the regression?” The obvious answer to this question is to use the number of principal components that were identified as important in the original analysis; for example, those with eigenvalues greater than one. But this is a case where the obvious answer is
- 106. 90 3 Principal Components Analysis R> pairs(usair_pca$scores[,1:3], ylim = c(-6, 4), xlim = c(-6, 4), + panel = function(x,y, ...) { + text(x, y, abbreviate(row.names(USairpollution)), + cex = 0.6) + bvbox(cbind(x,y), add = TRUE) + }) Comp.1 −6 −2 0 2 4 Albn Albq Atln Bltm Bffl Chrl Chcg Cncn Clvl Clmb DllsDnvrDsMn Dtrt Hrtf HstnIndn Jcks KnsC LttR LsvlMmph Miam MlwkMnnp Nshv NwOr Nrfl Omah Phld Phnx PttsPrvd Rchm SlLC SnFr Sttl St.LWshnWcht Wlmn q q q q q q q q q q q qq q q qq q q q qq q qq q q q q q q qq q q q q qqq q −6−2024 Albn Albq Atln Bltm Bffl Chrl Chcg Cncn Clvl ClmbDlls DnvrDsMn Dtrt Hrtf Hstn Indn Jcks KnsC LttR LsvlMmph Miam MlwkMnnp Nshv NwOr Nrfl Omah Phld Phnx PttsPrvd Rchm SlLC SnFr SttlSt.LWshnWcht Wlmn q q q q q q q q q qq qq q q q q q q q qq q qq q q q q q q qq q q q qqq q q −6−2024 Albn Albq Atln Bltm Bffl Chrl Chcg Cncn Clvl Clmb Dlls Dnvr DsMnDtrt Hrtf Hstn Indn Jcks KnsC LttRLsvlMmph Miam MlwkMnnp Nshv NwOrNrfl OmahPhld Phnx PttsPrvd Rchm SlLC SnFr Sttl St.LWshn Wcht Wlmnq q q q q q q q q q q q qq q q q q q qqq q qq q qq qq q qq q q q q qq q q Comp.2 Albn Albq Atln Bltm Bffl Chrl Chcg Cncn Clvl Clmb Dlls Dnvr DsMnDtrt Hrtf Hstn Indn Jcks KnsC LttRLsvlMmph Miam MlwkMnnp Nshv NwOr Nrfl OmahPhld Phnx PttsPrvd Rchm SlLC SnFr Sttl St.LWshn Wcht Wlmnq q q q q q q q q q q q qq q q q q q qqq q qqq q q qq q qq q q q q qq q q −6 −2 0 2 4 AlbnAlbq Atln Bltm Bffl Chrl Chcg Cncn Clvl Clmb Dlls Dnvr DsMn Dtrt Hrtf Hstn Indn Jcks KnsC LttR Lsvl Mmph Miam MlwkMnnp Nshv NwOr Nrfl Omah Phld Phnx PttsPrvd Rchm SlLC SnFr Sttl St.LWshn Wcht Wlmn q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q qq q q q q qq q q Albn Albq Atln Bltm Bffl Chrl Chcg Cncn ClvlClmb Dlls Dnvr DsMn Dtrt Hrtf Hstn Indn Jcks KnsC LttR Lsvl Mmph Miam MlwkMnnp Nshv NwOr Nrfl Omah Phld Phnx PttsPrvd Rchm SlLC SnFr Sttl St.LWshn Wcht Wlmn q q q q q q q q qq q q q q q q q q q q q q q qq q q q q q q qq q q q q qq q q −6 −2 0 2 4 −6−2024Comp.3 Fig. 3.9. Bivariate boxplots of the first three principal components. not necessarily correct and, Jolliffe (2002) gives an example where there are 13 original explanatory variables and therefore 13 principal components to consider, of which only the first four have variances greater than one. But using all the principal components as explanatory variables shows that, for example, component 12, with variance 0.04, is a significant predictor of the response. So, bearing this example in mind, we will regress the SO2 variables on all six principal components; the necessary R code is given on page 92.
- 107. 3.10 Some examples of the application of principal components analysis 91 R> out <- sapply(1:6, function(i) { + plot(USairpollution$SO2,usair_pca$scores[,i], + xlab = paste("PC", i, sep = ""), + ylab = "Sulphur dioxide concentration") + }) q q q q q q q q q qq qq q q q q q q q qq q q q q q q q q q q q qqq q qqq q 20 40 60 80 100 −6−22 PC1 Sulphurdioxideconcentration q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q 20 40 60 80 100 −2024 PC2 Sulphurdioxideconcentration qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q 20 40 60 80 100 −201 PC3 Sulphurdioxideconcentration q q q q q q q q q q q q q q q q q qq q q q q q q q q qq q q q q q q q q qq q q 20 40 60 80 100 −2012 PC4 Sulphurdioxideconcentration q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q 20 40 60 80 100 −0.60.00.6 PC5 Sulphurdioxideconcentration q q q q q q q q q q q q q q q q q q q q q q q q qq q qq q q q q q q q q q q q q 20 40 60 80 100 −0.40.00.4 PC6 Sulphurdioxideconcentration Fig. 3.10. Sulphur dioxide concentration depending on principal components.
- 108. 92 3 Principal Components Analysis R> usair_reg <- lm(SO2 ~ usair_pca$scores, + data = USairpollution) R> summary(usair_reg) Call: lm(formula = SO2 ~ usair_pca$scores, data = USairpollution) Residuals: Min 1Q Median 3Q Max -23.00 -8.54 -0.99 5.76 48.76 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 30.049 2.286 13.15 6.9e-15 usair_pca$scoresComp.1 -9.942 1.542 -6.45 2.3e-07 usair_pca$scoresComp.2 -2.240 1.866 -1.20 0.2384 usair_pca$scoresComp.3 -0.375 1.936 -0.19 0.8475 usair_pca$scoresComp.4 -8.549 2.622 -3.26 0.0025 usair_pca$scoresComp.5 15.176 6.753 2.25 0.0312 usair_pca$scoresComp.6 39.271 12.316 3.19 0.0031 Residual standard error: 14.6 on 34 degrees of freedom Multiple R-squared: 0.67, Adjusted R-squared: 0.611 F-statistic: 11.5 on 6 and 34 DF, p-value: 5.42e-07 Clearly, the first principal component score is the most predictive of sul- phur dioxide concentration, but it is also clear that components with small variance do not necessarily have small correlations with the response. We leave it as an exercise for the reader to investigate this example in more detail. 3.11 The biplot A biplot is a graphical representation of the information in an n × p data matrix. The “bi” reflects that the technique displays in a single diagram the variances and covariances of the variables and the distances (see Chapter 1) between units. The technique is based on the singular value decomposition of a matrix (see Gabriel 1971). A biplot is a two-dimensional representation of a data matrix obtained from eigenvalues and eigenvectors of the covariance matrix and obtained as X2 = (p1, p2) √ λ1 0 0 √ λ2 q1 q2 , where X2 is the“rank two”approximation of the data matrix X, λ1 and λ2 are the first two eigenvalues of the matrix nS, and q1 and q2 are the corresponding eigenvectors. The vectors p1 and p2 are obtained as
- 109. 3.12 Sample size for principal components analysis 93 pi = 1 √ λi Xqi; i = 1, 2. The biplot is the plot of the n rows of √ n(p1, p2) and the q rows of n−1/2 ( √ λ1q1, √ λ2q2) represented as vectors. The distance between the points representing the units reflects the generalised distance between the units (see Chapter 1), the length of the vector from the origin to the coordinates rep- resenting a particular variable reflects the variance of that variable, and the correlation of two variables is reflected by the angle between the two corre- sponding vectors for the two variables–the smaller the angle, the greater the correlation. Full technical details of the biplot are given in Gabriel (1981) and in Gower and Hand (1996). The biplot for the heptathlon data omitting the PNG competitor is shown in Figure 3.11. The plot in Figure 3.11 clearly shows that the winner of the gold medal, Jackie Joyner-Kersee, accumulates the majority of her points from the three events long jump, hurdles, and 200 m. We can also see from the biplot that the results of the 200 m, the hurdles and the long jump are highly correlated, as are the results of the javelin and the high jump; the 800 m time has relatively small correlation with all the other events and is almost uncorrelated with the high jump and javelin results. The first component largely separates the competitors by their overall score, with the second indicating which are their best events; for example, John, Choubenkova, and Behmer are placed near the end of the vector, representing the 800 m event because this is, relatively speaking, the event in which they give their best performance. Similarly Yuping, Scheider, and Braun can be seen to do well in the high jump. We shall have a little more to say about the biplot in the next chapter. 3.12 Sample size for principal components analysis There have been many suggestions about the number of units needed when applying principal components analysis. Intuitively, larger values of n should lead to more convincing results and make these results more generalisable. But unfortunately many of the suggestions made, for example that n should be greater than 100 or that n should be greater than five times the number of variables, are based on minimal empirical evidence. However, Guadagnoli and Velicer (1988) review several studies that reach the conclusion that it is the minimum value of n rather than the ratio of n to q that is most relevant, although the range of values suggested for the minimum value of n in these papers, from 50 to 400, sheds some doubt on their value. And indeed other authors, for example Gorsuch (1983) and Hatcher (1994), lean towards the ratio of the minimum value of n to q as being of greater importance and recommend at least 5:1. Perhaps the most detailed investigation of the problem is that reported in Osborne and Costello (2004), who found that the “best” results from principal
- 110. 94 3 Principal Components Analysis R> biplot(heptathlon_pca, col = c("gray", "black")) −0.4 0.0 0.2 0.4 0.6 −0.4−0.20.00.10.2 PC1 PC2 Jy−K John Bhmr Sblv Chbn Schl Flmn Grnr Ljbn Borg Wjns Dmtr Schd Bran Rtsl Ypng Hggr Brwn Mlln Htnv Kytl Grms H−In Jn−M −6 −4 −2 0 2 4 6 8 −4−202 hurdles highjump shot run200m longjump javelin run800m Fig. 3.11. Biplot of the (scaled) first two principal components (with observations for PNG removed). components analysis result when n and the ratio of n to q are both large. But the actual values needed depend largely on the separation of the eigenvalues defining the principal components structure. If these eigenvalues are “close together”, then a larger number of units will be needed to uncover the structure precisely than if they are far apart. 3.13 Canonical correlation analysis Principal components analysis considers interrelationships within a set of vari- ables. But there are situations where the researcher may be interested in assessing the relationships between two sets of variables. For example, in psy- chology, an investigator may measure a number of aptitude variables and a
- 111. 3.13 Canonical correlation analysis 95 number of achievement variables on a sample of students and wish to say something about the relationship between “aptitude” and “achievement”. And Krzanowski (1988) suggests an example in which an agronomist has taken, say, q1 measurements related to the yield of plants (e.g., height, dry weight, number of leaves) at each of n sites in a region and at the same time may have recorded q2 variables related to the weather conditions at these sites (e.g., average daily rainfall, humidity, hours of sunshine). The whole investigation thus consists of taking (q1 +q2) measurements on n units, and the question of interest is the measurement of the association between “yield” and “weather”. One technique for addressing such questions is canonical correlation analysis, although it has to be said at the outset that the technique is used less widely than other multivariate techniques, perhaps because the results from such an analysis are frequently difficult to interpret. For these reasons, the account given here is intentionally brief. One way to view canonical correlation analysis is as an extension of mul- tiple regression where a single variable (the response) is related to a number of explanatory variables and the regression solution involves finding the lin- ear combination of the explanatory variables that is most highly correlated with the response. In canonical correlation analysis where there is more than a single variable in each of the two sets, the objective is to find the linear functions of the variables in one set that maximally correlate with linear func- tions of variables in the other set. Extraction of the coefficients that define the required linear functions has similarities to the process of finding principal components. A relatively brief account of the technical aspects of canonical correlation analysis (CCA) follows; full details are given in Krzanowski (1988) and Mardia et al. (1979). The purpose of canonical correlation analysis is to characterise the inde- pendent statistical relationships that exist between two sets of variables, x = (x1, x2, . . . , xq1 ) and y = (y1, y2, . . . , yq2 ). The overall (q1 + q2) × (q1 + q2) correlation matrix contains all the information on associations between pairs of variables in the two sets, but attempting to extract from this matrix some idea of the association between the two sets of variables is not straightfor- ward. This is because the correlations between the two sets may not have a consistent pattern, and these between-set correlations need to be adjusted in some way for the within-set correlations. The question of interest is “how do we quantify the association between the two sets of variables x and y?” The approach adopted in CCA is to take the association between x and y to be the largest correlation between two single variables, u1 and v1, derived from x and y, with u1 being a linear combination of x1, x2, . . . , xq1 and v1 being a linear combination of y1, y2, . . . , yq2 . But often a single pair of variables (u1, v1) is not sufficient to quantify the association between the x and y variables, and we may need to consider some or all of s pairs (u1, v1), (u2, v2), . . . , (us, vs) to do this, where s = min(q1, q2). Each ui is a linear combination of the vari- ables in x, ui = ai x, and each vi is a linear combination of the variables y,
- 112. 96 3 Principal Components Analysis vi = bi y, with the coefficients (ai, bi) (i = 1 . . . s) being chosen so that the ui and vi satisfy the following: 1. The ui are mutually uncorrelated; i.e., Cov(ui, uj) = 0 for i = j. 2. The vi are mutually uncorrelated; i.e., Cov(vi, vj) = 0 for i = j. 3. The correlation between ui and vi is Ri for i = 1 . . . s, where R1 > R2 > · · · > Rs. The Ri are the canonical correlations. 4. The ui are uncorrelated with all vj except vi; i.e., Cov(ui, vj) = 0 for i = j. The vectors ai and bi i = 1, . . . , s, which define the required linear combi- nations of the x and y variables, are found as the eigenvectors of matrices E1(q1 × q1) (the ai) and E2(q2 × q2) (the bi), defined as E1 = R−1 11 R12R−1 22 R21, E2 = R−1 22 R21R−1 11 R12, where R11 is the correlation matrix of the variables in x, R22 is the correlation matrix of the variables in y, and R12 = R21 is the q1×q2 matrix of correlations across the two sets of variables. The canonical correlations R1, R2, . . . , Rs are obtained as the square roots of the non-zero eigenvalues of either E1 or E2. The s canonical correlations R1, R2, . . . , Rs express the association between the x and y variables after removal of the within-set correlation. Inspection of the coefficients of each original variable in each canonical variate can provide an interpretation of the canonical variate in much the same way as interpreting principal components. Such interpretation of the canonical variates may help to describe just how the two sets of original variables are related (see Krzanowski 2010). In practise, interpretation of canonical variates can be difficult because of the possibly very different variances and covariances among the original variables in the two sets, which affects the sizes of the coefficients in the canonical variates. Unfortunately, there is no convenient normalisation to place all coefficients on an equal footing (see Krzanowski 2010). In part, this problem can be dealt with by restricting interpretation to the standardised coefficients; i.e., the coefficients that are appropriate when the original variables have been standardised. We will now look at two relatively simple examples of the application of canonical correlation analysis. 3.13.1 Head measurements As our first example of CCA, we shall apply the technique to data on head length and head breadth for each of the first two adult sons in 25 families shown in Table 3.1. (Part of these data were used earlier in the chapter.) These data were collected by Frets (1921), and the question that was of interest to Frets was whether there is a relationship between the head measurements for pairs of sons. We shall address this question by using canonical correla- tion analysis. Here we shall develop the canonical correlation analysis from first principles as detailed above. Assuming the head measurements data are contained in the data frame headsize, the necessary R code is
- 113. 3.13 Canonical correlation analysis 97 R> headsize.std <- sweep(headsize, 2, + apply(headsize, 2, sd), FUN = "/") R> R <- cor(headsize.std) R> r11 <- R[1:2, 1:2] R> r22 <- R[-(1:2), -(1:2)] R> r12 <- R[1:2, -(1:2)] R> r21 <- R[-(1:2), 1:2] R> (E1 <- solve(r11) %*% r12 %*% solve(r22) %*%r21) head1 breadth1 head1 0.3225 0.3168 breadth1 0.3019 0.3021 R> (E2 <- solve(r22) %*% r21 %*% solve(r11) %*%r12) head2 breadth2 head2 0.3014 0.3002 breadth2 0.3185 0.3232 R> (e1 <- eigen(E1)) $values [1] 0.621745 0.002888 $vectors [,1] [,2] [1,] 0.7270 -0.7040 [2,] 0.6866 0.7102 R> (e2 <- eigen(E2)) $values [1] 0.621745 0.002888 $vectors [,1] [,2] [1,] -0.6838 -0.7091 [2,] -0.7297 0.7051 Here the four linear functions are found to be u1 = +0.73x1 + 0.69x2, u2 = −0.70x1 + 0.71x2, v1 = −0.68x3 − 0.73x4, v2 = −0.71x3 + 0.71x4. R> girth1 <- headsize.std[,1:2] %*% e1$vectors[,1] R> girth2 <- headsize.std[,3:4] %*% e2$vectors[,1]
- 114. 98 3 Principal Components Analysis R> shape1 <- headsize.std[,1:2] %*% e1$vectors[,2] R> shape2 <- headsize.std[,3:4] %*% e2$vectors[,2] R> (g <- cor(girth1, girth2)) [,1] [1,] -0.7885 R> (s <- cor(shape1, shape2)) [,1] [1,] 0.05374 R> plot(girth1, girth2) R> plot(shape1, shape2) q q q q q q q q q q q q q q q q q q q q q q q q q 25 27 29 −31−29−27−25 girth1 girth2 q q q q q q q q q q q q q q q q q q q q q q q q q 0.5 1.5 2.02.53.03.5 shape1 shape2 Fig. 3.12. Scatterplots of girth and shape for first and second sons. The first canonical variate for both first and second sons is simply a weighted sum of the two head measurements and might be labelled “girth”; these two variates have a correlation of −0.79. (The negative value arises be- cause of the arbitrariness of the sign in the first coefficient of an eigenvector– here both coefficients for girth in first sons are positive and for second sons
- 115. 3.13 Canonical correlation analysis 99 they are both negative. The correlation can also be found as the square root of the first eigenvalue of E1 (and E2), namely 0.6217.) Each second canonical variate is a weighted difference of the two head measurements and can be in- terpreted roughly as head “shape”; here the correlation is 0.05 (which can also be found as the square root of the second eigenvalue of E1, namely 0.0029). (Girth and shape are defined to be uncorrelated bioth within and between first and second sons.) In this example, it is clear that the association between the two head measurements of first and second sons is almost entirely expressed through the “girth” variables, with the two “shape” variables being almost uncorre- lated. The association between the two sets of measurements is essentially one-dimensional. A scatterplot of girth for first and second sons and a similar plot for shape reinforce this conclusion. Both plots are shown in Figure 3.12. 3.13.2 Health and personality We can now move on to a more substantial example taken from Affifi, Clark, and May (2004) and also discussed by Krzanowski (2010). The data for this example arise from a study of depression amongst 294 respondents in Los An- geles. The two sets of variables of interest were “health variables”, namely the CESD (the sum of 20 separate numerical scales measuring different aspects of depression) and a measure of general health and“personal” variables, of which there were four: gender, age, income, and educational level (numerically coded from the lowest “less than high school”, to the highest, “finished doctorate”). The sample correlation matrix between these variables is given in Table 3.3. Table 3.3: LAdepr data. Los Angeles Depression Data. CESD Health Gender Age Edu Income 1.000 0.212 0.124 -0.164 -0.101 -0.158 0.212 1.000 0.098 0.308 -0.207 -0.183 0.124 0.098 1.000 0.044 -0.106 -0.180 -0.164 0.308 0.044 1.000 -0.208 -0.192 -0.101 -0.207 -0.106 -0.208 1.000 0.492 -0.158 -0.183 -0.180 -0.192 0.492 1.000 Here the maximum number of canonical variate pairs is two, and they can be found using the following R code:
- 116. 100 3 Principal Components Analysis R> r11 <- LAdepr[1:2, 1:2] R> r22 <- LAdepr[-(1:2), -(1:2)] R> r12 <- LAdepr[1:2, -(1:2)] R> r21 <- LAdepr[-(1:2), 1:2] R> (E1 <- solve(r11) %*% r12 %*% solve(r22) %*%r21) CESD Health CESD 0.08356 -0.04312 Health -0.03256 0.13338 R> (E2 <- solve(r22) %*% r21 %*% solve(r11) %*%r12) Gender Age Edu Income Gender 0.015478 -0.001483 -0.01813 -0.02224 Age -0.002445 0.147069 -0.03981 -0.01599 Edu -0.014915 -0.026019 0.02544 0.02508 Income -0.021163 0.013027 0.02159 0.02895 R> (e1 <- eigen(E1)) $values [1] 0.15347 0.06347 $vectors [,1] [,2] [1,] 0.5250 -0.9065 [2,] -0.8511 -0.4222 R> (e2 <- eigen(E2)) $values [1] 1.535e-01 6.347e-02 -8.760e-18 6.633e-19 $vectors [,1] [,2] [,3] [,4] [1,] -0.002607 0.4904 -0.4617 0.79705 [2,] -0.980095 -0.3208 -0.1865 0.06251 [3,] 0.185801 -0.4270 -0.7974 -0.05817 [4,] -0.069886 -0.6887 0.3409 0.59784 (Note that the third and fourth eigenvalues of E2 are essentially zero, as we would expect in this case.) The first canonical correlation is 0.409, calculated as the square root of the first eigenvalue of E1, which is given above as 0.15347. If tested as outlined in Exercise 3.4, it has an associated p-value that is very small; there is strong evidence that the first canonical correlation is significant. The corresponding variates, in terms of standardised original variables, are u1 = +0.53 CESD − 0.85 Health v1 = −0.00 Gender − 0.98 Age + 0.19 Education − 0.07 Income .
- 117. 3.14 Summary 101 High coefficients correspond to CESD (positively) and health (negatively) for the perceived health variables, and to Age (negatively) and Education (positively) for the personal variables. It appears that relatively older and medicated people tend to have lower depression scores, but perceive their health as relatively poor, while relatively younger but educated people have the opposite health perception. (We are grateful to Krzanowski 2010, for this interpretation.) The second canonical correlation is 0.261, calculated as the square root of the second eigenvalue, found from the R output above to be 0.06347; this correlation is also significant (see Exercise 3.4). The corresponding canonical variates are u1 = −0.91 CESD − 0.42 Health v1 = +0.49 Gender − 0.32 Age − 0.43 Education − 0.69 Income . Since the higher value of the Gender variable is for females, the interpretation here is that relatively young, poor, and uneducated females are associated with higher depression scores and, to a lesser extent, with poor perceived health (again this interpretation is due to Krzanowski 2010). 3.14 Summary Principal components analysis is among the oldest of multivariate techniques, having been introduced originally by Pearson (1901) and independently by Hotelling (1933). It remains, however, one of the most widely employed meth- ods of multivariate analysis, useful both for providing a convenient method of displaying multivariate data in a lower dimensional space and for possibly simplifying other analyses of the data. The reduction in dimensionality that can often be achieved by a principal components analysis is possible only if the original variables are correlated; if the original variables are independent of one another a principal components analysis cannot lead to any simplification. The examples of principal components analysis given in this chapter have involved variables that are continuous. But when used as a descriptive tech- nique there is no reason for the variables in the analysis to be of any particular type. So variables might be a mixture of continuous, ordinal, and even binary variables. Linear functions of continuous variables are perhaps easier to inter- pret than corresponding functions of binary variables, but the basic objective of principal components analysis, namely to summarise most of the“variation” present in the original q variables, can be achieved regardless of the nature of the observed variables. The results of a canonical correlation analysis have the reputation of often being difficult to interpret, and in many respects this is a reputation that is well earned. Certainly one has to know the variables very well to have any hope of extracting a convincing explanation of the results. But in some
- 118. 102 3 Principal Components Analysis circumstances (the heads measurement data is an example) CCA can provide a useful description and informative insights into the association between two sets of variables. Finally, we should mention a technique known as independent component analysis (ICA), a potentially powerful technique that seeks to uncover hidden variables in high-dimensional data. In particular, ICA is often able to separate independent sources linearly mixed in several sensors; for example, in prenatal diagnostics a multichannel electrocardiogram (ECG) may be used to record a mixture of maternal and foetal electrical activity, including foetal heart rate and maternal heart rate. However, the maternal ECG will be much stronger than the foetal signal, and the signal recorded is also likely to be affected by other sources of electrical interference. Such data can be separated into their component signals by ICA, as demonstrated in Izenman (2008). Similar mixtures of signals arise in many other areas, for example monitor- ing of human brain-wave activity and functional magnetic resonance imaging experiments, and consequently ICA has become increasingly used in such ar- eas. Details of the methods are given in Hyv¨arinen (1999) and Hyv¨arinen, Karhunen, and Oja (2001). Interestingly, the first step in analysing multivari- ate data with ICA is often a principal components analysis. We have chosen not to describe ICA in this book because it is technically demanding and also rather specialised, but we should mention in passing that the method is available in R via the add-on package fastICA (Marchini, Heaton, and Ripley 2010). 3.15 Exercises Ex. 3.1 Construct the scatterplot of the heptathlon data showing the contours of the estimated bivariate density function on each panel. Is this graphic more useful than the unenhanced scatterplot matrix? Ex. 3.2 Construct a diagram that shows the SO2 variable in the air pollution data plotted against each of the six explanatory variables, and in each of the scatterplots show the fitted linear regression and a fitted locally weighted regression. Does this diagram help in deciding on the most ap- propriate model for determining the variables most predictive of sulphur dioxide levels? Ex. 3.3 Find the principal components of the following correlation matrix given by MacDonnell (1902) from measurements of seven physical char- acteristics in each of 3000 convicted criminals:
- 119. 3.15 Exercises 103 R = Head length Head breadth Face breadth Left finger length Left forearm length Left foot length Height 1.000 0.402 1.000 0.396 0.618 1.000 0.301 0.150 0.321 1.000 0.305 0.135 0.289 0.846 1.000 0.339 0.206 0.363 0.759 0.797 1.000 0.340 0.183 0.345 0.661 0.800 0.736 1.000 . How would you interpret the derived components? Ex. 3.4 Not all canonical correlations may be statistically significant. An approximate test proposed by Bartlett (1947) can be used to deter- mine how many significant relationships exist. The test statistic for testing that at least one canonical correlation is significant is Φ2 0 = − n − 1 2 (q1 + q2 + 1) s i=1 log(1 − λi), where the λi are the eigenvalues of E1 and E2. Under the null hypothesis that all correlations are zero, Φ2 0 has a chi-square distribution with q1 ×q2 degrees of freedom. Write R code to apply this test to the headsize data (Table 3.1) and the depression data (Table 3.3). Ex. 3.5 Repeat the regression analysis for the air pollution data described in the text after removing whatever cities you think should be regarded as outliers. For the results given in the text and the results from the outliers-removed data, produce scatterplots of sulphur dioxide concentra- tion against each of the principal component scores. Interpret your results.
- 121. 4 Multidimensional Scaling 4.1 Introduction In Chapter 3, we noted in passing that one of the most useful ways of using principal components analysis was to obtain a low-dimensional “map” of the data that preserved as far as possible the Euclidean distances between the ob- servations in the space of the original q variables. In this chapter, we will make this aspect of principal component analysis more explicit and also introduce a class of other methods, labelled multidimensional scaling, that aim to pro- duce similar maps of data but do not operate directly on the usual multivariate data matrix, X. Instead they are applied to distance matrices (see Chapter 1), which are derived from the matrix X (an example of a distance matrix de- rived from a small set of multivariate data is shown in Subsection 4.4.2), and also to so-called dissimilarity or similarity matrices that arise directly in a number of ways, in particular from judgements made by human raters about how alike pairs of objects, stimuli, etc., of interest are. An example of a di- rectly observed dissimilarity matrix is shown in Table 4.5, with judgements about political and war leaders that had major roles in World War II being given by a subject after receiving the simple instructions to rate each pair of politicians on a nine-point scale, with 1 indicating two politicians they regard as very similar and 9 indicating two they regard as very dissimilar. (If the nine point-scale had been 1 for very dissimilar and 9 for very similar, then the result would have been a rating of similarity, although similarities are often scaled to lie in a [0, 1] interval. The term proximity is often used to encompass both dissimilarity and similarity ratings.) 4.2 Models for proximity data Models are fitted to proximities in order to clarify, display, and help under- stand and possibly explain any structure or pattern amongst the observed or DOI 10.1007/978-1-4419-9650-3_4, © Springer Science+Business Media, LLC 2011 105B. Everitt and T. Hothorn, An Introduction to Applied Multivariate Analysis with R: Use R!,
- 122. 106 4 Multidimensional Scaling calculated proximities not readily apparent in the collection of numerical val- ues. In some areas, particularly psychology, the ultimate goal in the analysis of a set of proximities is more specific, namely the development of theories for ex- plaining similarity judgements; in other words, trying to answer the question, “what makes things seem alike or seem different?” According to Carroll and Arabie (1980) and Carroll, Clark, and DeSarbo (1984), models for the analysis of proximity data can be categorised into one of three major classes: spatial models, tree models, and hybrid models. In this chapter and in this book, we only deal with the first of these three classes. For details of tree models and hybrid models, see, for example, Everitt and Rabe-Hesketh (1997). 4.3 Spatial models for proximities: Multidimensional scaling (MDS) A spatial representation of a proximity matrix consists of a set of n m- dimensional coordinates, each one of which represents one of the n units in the data. The required coordinates are generally found by minimising some measure of “fit” between the distances implied by the coordinates and the ob- served proximities. In simple terms, a geometrical model is sought in which the larger the observed distance or dissimilarity between two units (or the smaller their similarity), the further apart should be the points representing them in the model. In general (but not exclusively), the distances between the points in the spatial model are assumed to be Euclidean. Finding the best-fitting set of coordinates and the appropriate value of m needed to adequately represent the observed proximities is the aim of the many methods of multidimensional scaling that have been proposed. The hope is that the number of dimensions, m, will be small, ideally two or three, so that the derived spatial configuration can be easily plotted. Multidimensional scaling is essentially a data reduction technique because the aim is to find a set of points in low dimension that approximate the possibly high-dimensional configuration represented by the original proximity matrix. The variety of methods that have been proposed largely differ in how agreement between fitted distances and observed prox- imities is assessed. In this chapter, we will consider two methods, classical multidimensional scaling and non-metric multidimensional scaling. 4.4 Classical multidimensional scaling First, like all MDS techniques, classical scaling seeks to represent a proximity matrix by a simple geometrical model or map. Such a model is characterised by a set of points x1, x2, . . . , xn, in m dimensions, each point representing one of the units of interest, and a measure of the distance between pairs of points. The objective of MDS is to determine both the dimensionality, m, of the model, and the n m-dimensional coordinates, x1, x2, . . . , xn, so that
- 123. 4.4 Classical multidimensional scaling 107 the model gives a “good” fit for the observed proximities. Fit will often be judged by some numerical index that measures how well the proximities and the distances in the geometrical model match. In essence, this simply means that the larger an observed dissimilarity between two stimuli (or the smaller their similarity), the further apart should be the points representing them in the final geometrical model. The question now arises as to how we estimate m, and the coordinate val- ues x1, x2, . . . , xn, from the observed proximity matrix. Classical scaling pro- vides an answer to this question based on the work of Young and Householder (1938). To begin, we must note that there is no unique set of coordinate values that give rise to a set of distances since the distances are unchanged by shifting the whole configuration of points from one place to another or by rotation or reflection of the configuration. In other words, we cannot uniquely determine either the location or the orientation of the configuration. The location prob- lem is usually overcome by placing the mean vector of the configuration at the origin. The orientation problem means that any configuration derived can be subjected to an arbitrary orthogonal transformation (see Chapter 5). Such transformations can often be used to facilitate the interpretation of solutions, as will be seen later (and again see Chapter 5). 4.4.1 Classical multidimensional scaling: Technical details To begin our account of the method, we shall assume that the proximity matrix we are dealing with is a matrix of Euclidean distances, D, derived from a raw n × q data matrix, X. In Chapter 1, we saw how to calculate Euclidean distances from X; classical multidimensional scaling is essentially concerned with the reverse problem: given the distances, how do we find X? First assume X is known and consider the n × n inner products matrix, B B = XX . (4.1) The elements of B are given by bij = q k=1 xikxjk. (4.2) It is easy to see that the squared Euclidean distances between the rows of X can be written in terms of the elements of B as d2 ij = bii + bjj − 2bij. (4.3) If the bs could be found in terms of the ds in the equation above, then the required coordinate values could be derived by factoring B as in (4.1). No unique solution exists unless a location constraint is introduced; usually the centre of the points x is set at the origin, so that n i=1 xik = 0 for all k = 1, 2 . . . m. These constraints and the relationship given in (4.2) imply that the
- 124. 108 4 Multidimensional Scaling sum of the terms in any row of B must be zero. Consequently, summing the relationship given in (4.2) over i, over j, and finally over both i and j leads to the series of equations n i=1 d2 ij = T + nbjj, n j=1 d2 ij = T + nbii, n i=1 n j=1 d2 ij = 2nT, where T = n i=1 bii is the trace of the matrix B. The elements of B can now be found in terms of squared Euclidean distances as bij = − 1 2 d2 ij − d2 i. − d2 .j + d2 .. , where d2 i. = 1 n n j=1 d2 ij, d2 .j = 1 n n i=1 d2 ij, d2 .. = 1 n2 n i=1 n j=1 d2 ij. Having now derived the elements of B in terms of Euclidean distances, it remains to factor it to give the coordinate values. In terms of its spectral decomposition (see Chapter 3), B can be written as B = VΛV , where Λ = diag (λ1, . . . , λn) is the diagonal matrix of eigenvalues of B and V = (V1, . . . , Vn) the corresponding matrix of eigenvectors, normalised so that the sum of squares of their elements is unity, that is, ViVi = 1. The eigenvalues are assumed to be labelled such that λ1 ≥ λ2 ≥ · · · ≥ λn. When D arises from an n × q matrix of full rank, then the rank of B is q, so that the last n − q of its eigenvalues will be zero. So B can be written as B = V1Λ1V1 , where V1 contains the first q eigenvectors and Λ1 the q non-zero eigenvalues. The required coordinate values are thus
- 125. 4.4 Classical multidimensional scaling 109 X = V1Λ 1 2 1 , where Λ 1 2 1 = diag λ 1 2 1 , . . . , λ 1 2 q . Using all q-dimensions will lead to complete recovery of the original Eu- clidean distance matrix. The best-fitting m-dimensional representation is given by the m eigenvectors of B corresponding to the m largest eigenval- ues. The adequacy of the m-dimensional representation can be judged by the size of the criterion Pm = m i=1 λi n i=1 λi . Values of Pm of the order of 0.8 suggest a reasonable fit. It should be mentioned here that where the proximity matrix contains Euclidean distances calculated from an n × q data matrix X, classical scal- ing can be shown to be equivalent to principal components analysis, with the required coordinate values corresponding to the scores on the principal com- ponent extracted from the covariance matrix of the data. One result of this duality is that classical multidimensional scaling is also referred to as principal coordinates–see Gower (1966). And the m-dimensional principal components solution (m < q) is “best” in the sense that it minimises the measure of fit S = n i=1 n i=1 d2 ij − d (m) ij 2 , where dij is the Euclidean distance between individuals i and j based on their original q variable values and d (m) ij is the corresponding distance calculated from the m principal component scores. When the observed proximity matrix is not Euclidean, the matrix B is not positive-definite. In such cases, some of the eigenvalues of B will be neg- ative; correspondingly, some coordinate values will be complex numbers. If, however, B has only a small number of small negative eigenvalues, a useful representation of the proximity matrix may still be possible using the eigen- vectors associated with the m largest positive eigenvalues. The adequacy of the resulting solution might be assessed using one of the following two criteria suggested by Mardia et al. (1979): P(1) m = m i=1 |λi| n i=1 |λi| , P(2) m = m i=1 λ2 i n i=1 λ2 i . Again we would look for values above 0.8 to claim a “good” fit. Alternatively, Sibson (1979) recommends one of the following two criteria for deciding on the number of dimensions for the spatial model to adequately represent the observed proximities:
- 126. 110 4 Multidimensional Scaling Trace criterion: Choose the number of coordinates so that the sum of the pos- itive eigenvalues is approximately equal to the sum of all the eigenvalues. Magnitude criterion: Accept as genuinely positive only those eigenvalues whose magnitude substantially exceeds that of the largest negative eigen- value. If, however, the matrix B has a considerable number of large negative eigenvalues, classical scaling of the proximity matrix may be inadvisable and some other methods of scaling, for example non-metric scaling (see the next section), might be better employed. 4.4.2 Examples of classical multidimensional scaling For our first example we will use the small set of multivariate data X [,1] [,2] [,3] [,4] [,5] [1,] 3 4 4 6 1 [2,] 5 1 1 7 3 [3,] 6 2 0 2 6 [4,] 1 1 1 0 3 [5,] 4 7 3 6 2 [6,] 2 2 5 1 0 [7,] 0 4 1 1 1 [8,] 0 6 4 3 5 [9,] 7 6 5 1 4 [10,] 2 1 4 3 1 and the associated matrix of Euclidean distances (computed via the dist() function) will be our proximity matrix R> (D <- dist(X)) 1 2 3 4 5 6 7 8 9 2 5.196 3 8.367 6.083 4 7.874 8.062 6.325 5 3.464 6.557 8.367 9.274 6 5.657 8.426 8.832 5.292 7.874 7 6.557 8.602 8.185 3.873 7.416 5.000 8 6.164 8.888 8.367 6.928 6.000 7.071 5.745 9 7.416 9.055 6.856 8.888 6.557 7.550 8.832 7.416 10 4.359 6.164 7.681 4.796 7.141 2.646 5.099 6.708 8.000 To apply classical scaling to this matrix in R, we can use the cmdscale() function to do the scaling: R> cmdscale(D, k = 9, eig = TRUE)
- 127. 4.4 Classical multidimensional scaling 111 $points [,1] [,2] [,3] [,4] [,5] [,6] [1,] -1.6038 -2.38061 -2.2301 -0.3657 0.11536 0.000e+00 [2,] -2.8246 2.30937 -3.9524 0.3419 0.33169 -2.797e-08 [3,] -1.6908 5.13970 1.2880 0.6503 -0.05134 -1.611e-09 [4,] 3.9528 2.43234 0.3834 0.6864 -0.03461 -7.393e-09 [5,] -3.5985 -2.75538 -0.2551 1.0784 -1.26125 -5.198e-09 [6,] 2.9520 -1.35475 -0.1899 -2.8211 0.12386 -2.329e-08 [7,] 3.4690 -0.76411 0.3017 1.6369 -1.94210 -1.452e-08 [8,] 0.3545 -2.31409 2.2162 2.9240 2.00450 -1.562e-08 [9,] -2.9362 0.01280 4.3117 -2.5123 -0.18912 -1.404e-08 [10,] 1.9257 -0.32527 -1.8734 -1.6189 0.90299 6.339e-09 [,7] [,8] [,9] [1,] 1.791e-08 NaN NaN [2,] -1.209e-09 NaN NaN [3,] 1.072e-09 NaN NaN [4,] 1.088e-08 NaN NaN [5,] -2.798e-09 NaN NaN [6,] -7.146e-09 NaN NaN [7,] 3.072e-09 NaN NaN [8,] 2.589e-10 NaN NaN [9,] 7.476e-09 NaN NaN [10,] 3.303e-09 NaN NaN $eig [1] 7.519e+01 5.881e+01 4.961e+01 3.043e+01 1.037e+01 [6] 2.101e-15 5.769e-16 -2.819e-15 -3.233e-15 -6.274e-15 $x NULL $ac [1] 0 $GOF [1] 1 1 Note that as q = 5 in this example, eigenvalues six to nine are essentially zero and only the first five columns of points represent the Euclidean distance ma- trix. First we should confirm that the five-dimensional solution achieves com- plete recovery of the observed distance matrix. We can do this simply by com- paring the original distances with those calculated from the five-dimensional scaling solution coordinates using the following R code: R> max(abs(dist(X) - dist(cmdscale(D, k = 5)))) [1] 1.243e-14
- 128. 112 4 Multidimensional Scaling This confirms that all the differences are essentially zero and that therefore the observed distance matrix is recovered by the five-dimensional classical scaling solution. We can also check the duality of classical scaling of Euclidean distances and principal components analysis mentioned previously in the chapter by com- paring the coordinates of the five-dimensional scaling solution given above with the first five principal component (up to signs) scores obtained by apply- ing PCA to the covariance matrix of the original data; the necessary R code is R> max(abs(prcomp(X)$x) - abs(cmdscale(D, k = 5))) [1] 3.035e-14 Now let us look at two examples involving distances that are not Eu- clidean. First, we will calculate the Manhattan distances between the rows of the small data matrix X. The Manhattan distance for units i and j is given by q k=1 |xik − xjk|, and these distances are not Euclidean. (Manhattan dis- tances will be familiar to those readers who have walked around New York.) The R code for calculating the Manhattan distances and then applying clas- sical multidimensional scaling to the resulting distance matrix is: R> X_m <- cmdscale(dist(X, method = "manhattan"), + k = nrow(X) - 1, eig = TRUE) The criteria P (1) m and P (2) m can be computed from the eigenvalues as follows: R> (X_eigen <- X_m$eig) [1] 2.807e+02 2.494e+02 2.289e+02 9.251e+01 4.251e+01 [6] 2.197e+01 -7.105e-15 -1.507e+01 -2.805e+01 -5.683e+01 Note that some of the eigenvalues are negative in this case. R> cumsum(abs(X_eigen)) / sum(abs(X_eigen)) [1] 0.2763 0.5218 0.7471 0.8382 0.8800 0.9016 0.9016 0.9165 [9] 0.9441 1.0000 R> cumsum(X_eigen^2) / sum(X_eigen^2) [1] 0.3779 0.6764 0.9276 0.9687 0.9773 0.9796 0.9796 0.9807 [9] 0.9845 1.0000 The values of both criteria suggest that a three-dimensional solution seems to fit well.
- 129. 4.4 Classical multidimensional scaling 113 Table 4.1: airdist data. Airline distances between ten US cities. ATL ORD DEN HOU LAX MIA JFK SFO SEA IAD ATL 0 ORD 587 0 DEN 1212 920 0 HOU 701 940 879 0 LAX 1936 1745 831 1374 0 MIA 604 1188 1726 968 2339 0 JFK 748 713 1631 1420 2451 1092 0 SFO 2139 1858 949 1645 347 2594 2571 0 SEA 218 1737 1021 1891 959 2734 2408 678 0 IAD 543 597 1494 1220 2300 923 205 2442 2329 0 For our second example of applying classical multidimensional scaling to non- Euclidean distances, we shall use the airline distances between ten US cities given in Table 4.1. These distances are not Euclidean since they relate essen- tially to journeys along the surface of a sphere. To apply classical scaling to these distances and to see the eigenvalues, we can use the following R code: R> airline_mds <- cmdscale(airdist, k = 9, eig = TRUE) R> airline_mds$points [,1] [,2] [,3] [,4] [,5] [,6] [,7] [,8] ATL -434.8 724.22 440.93 0.18579 -1.258e-02 NaN NaN NaN ORD -412.6 55.04 -370.93 4.39608 1.268e+01 NaN NaN NaN DEN 468.2 -180.66 -213.57 30.40857 -9.585e+00 NaN NaN NaN HOU -175.6 -515.22 362.84 9.48713 -4.860e+00 NaN NaN NaN LAX 1206.7 -465.64 56.53 1.34144 6.809e+00 NaN NaN NaN MIA -1161.7 -477.98 479.60 -13.79783 2.278e+00 NaN NaN NaN JFK -1115.6 199.79 -429.67 -29.39693 -7.137e+00 NaN NaN NaN SFO 1422.7 -308.66 -205.52 -26.06310 -1.983e+00 NaN NaN NaN SEA 1221.5 887.20 170.45 -0.06999 -8.943e-05 NaN NaN NaN IAD -1018.9 81.90 -290.65 23.50884 1.816e+00 NaN NaN NaN (The nineth column containing NaNs is omitted from the output.) The eigen- values are R> (lam <- airline_mds$eig) [1] 9.214e+06 2.200e+06 1.083e+06 3.322e+03 3.859e+02 [6] -5.204e-09 -9.323e+01 -2.169e+03 -9.091e+03 -1.723e+06 As expected (as the distances are not Euclidean), some of the eigenvalues are negative and so we will again use the criteria P (1) m and P (2) m to assess how many coordinates we need to adequately represent the observed distance matrix. The values of the two criteria calculated from the eigenvalues are
- 130. 114 4 Multidimensional Scaling R> cumsum(abs(lam)) / sum(abs(lam)) [1] 0.6473 0.8018 0.8779 0.8781 0.8782 0.8782 0.8782 0.8783 [9] 0.8790 1.0000 R> cumsum(lam^2) / sum(lam^2) [1] 0.9043 0.9559 0.9684 0.9684 0.9684 0.9684 0.9684 0.9684 [9] 0.9684 1.0000 These values suggest that the first two coordinates will give an adequate rep- resentation of the observed distances. The scatterplot of the two-dimensional coordinate values is shown in Figure 4.1. In this two-dimensional representa- tion, the geographical location of the cities has been very well recovered by the two-dimensional multidimensional scaling solution obtained from the airline distances. Our next example of the use of classical multidimensional scaling will in- volve the data shown in Table 4.2. These data show four measurements on male Egyptian skulls from five epochs. The measurements are: mb: maximum breadth of the skull; bh: basibregmatic height of the skull; bl: basialiveolar length of the skull; and nh: nasal height of the skull. Table 4.2: skulls data. Measurements of four variables taken from Egyptian skulls of five periods. epoch mb bh bl nh epoch mb bh bl nh epoch mb bh bl nh c4000BC 131 138 89 49 c3300BC 137 136 106 49 c200BC 132 133 90 53 c4000BC 125 131 92 48 c3300BC 126 131 100 48 c200BC 134 134 97 54 c4000BC 131 132 99 50 c3300BC 135 136 97 52 c200BC 135 135 99 50 c4000BC 119 132 96 44 c3300BC 129 126 91 50 c200BC 133 136 95 52 c4000BC 136 143 100 54 c3300BC 134 139 101 49 c200BC 136 130 99 55 c4000BC 138 137 89 56 c3300BC 131 134 90 53 c200BC 134 137 93 52 c4000BC 139 130 108 48 c3300BC 132 130 104 50 c200BC 131 141 99 55 c4000BC 125 136 93 48 c3300BC 130 132 93 52 c200BC 129 135 95 47 c4000BC 131 134 102 51 c3300BC 135 132 98 54 c200BC 136 128 93 54 c4000BC 134 134 99 51 c3300BC 130 128 101 51 c200BC 131 125 88 48 c4000BC 129 138 95 50 c1850BC 137 141 96 52 c200BC 139 130 94 53 c4000BC 134 121 95 53 c1850BC 129 133 93 47 c200BC 144 124 86 50 c4000BC 126 129 109 51 c1850BC 132 138 87 48 c200BC 141 131 97 53 c4000BC 132 136 100 50 c1850BC 130 134 106 50 c200BC 130 131 98 53 c4000BC 141 140 100 51 c1850BC 134 134 96 45 c200BC 133 128 92 51 c4000BC 131 134 97 54 c1850BC 140 133 98 50 c200BC 138 126 97 54 c4000BC 135 137 103 50 c1850BC 138 138 95 47 c200BC 131 142 95 53
- 131. 4.4 Classical multidimensional scaling 115 Table 4.2: skulls data (continued). epoch mb bh bl nh epoch mb bh bl nh epoch mb bh bl nh c4000BC 132 133 93 53 c1850BC 136 145 99 55 c200BC 136 138 94 55 c4000BC 139 136 96 50 c1850BC 136 131 92 46 c200BC 132 136 92 52 c4000BC 132 131 101 49 c1850BC 126 136 95 56 c200BC 135 130 100 51 c4000BC 126 133 102 51 c1850BC 137 129 100 53 cAD150 137 123 91 50 c4000BC 135 135 103 47 c1850BC 137 139 97 50 cAD150 136 131 95 49 c4000BC 134 124 93 53 c1850BC 136 126 101 50 cAD150 128 126 91 57 c4000BC 128 134 103 50 c1850BC 137 133 90 49 cAD150 130 134 92 52 c4000BC 130 130 104 49 c1850BC 129 142 104 47 cAD150 138 127 86 47 c4000BC 138 135 100 55 c1850BC 135 138 102 55 cAD150 126 138 101 52 c4000BC 128 132 93 53 c1850BC 129 135 92 50 cAD150 136 138 97 58 c4000BC 127 129 106 48 c1850BC 134 125 90 60 cAD150 126 126 92 45 c4000BC 131 136 114 54 c1850BC 138 134 96 51 cAD150 132 132 99 55 c4000BC 124 138 101 46 c1850BC 136 135 94 53 cAD150 139 135 92 54 c3300BC 124 138 101 48 c1850BC 132 130 91 52 cAD150 143 120 95 51 c3300BC 133 134 97 48 c1850BC 133 131 100 50 cAD150 141 136 101 54 c3300BC 138 134 98 45 c1850BC 138 137 94 51 cAD150 135 135 95 56 c3300BC 148 129 104 51 c1850BC 130 127 99 45 cAD150 137 134 93 53 c3300BC 126 124 95 45 c1850BC 136 133 91 49 cAD150 142 135 96 52 c3300BC 135 136 98 52 c1850BC 134 123 95 52 cAD150 139 134 95 47 c3300BC 132 145 100 54 c1850BC 136 137 101 54 cAD150 138 125 99 51 c3300BC 133 130 102 48 c1850BC 133 131 96 49 cAD150 137 135 96 54 c3300BC 131 134 96 50 c1850BC 138 133 100 55 cAD150 133 125 92 50 c3300BC 133 125 94 46 c1850BC 138 133 91 46 cAD150 145 129 89 47 c3300BC 133 136 103 53 c200BC 137 134 107 54 cAD150 138 136 92 46 c3300BC 131 139 98 51 c200BC 141 128 95 53 cAD150 131 129 97 44 c3300BC 131 136 99 56 c200BC 141 130 87 49 cAD150 143 126 88 54 c3300BC 138 134 98 49 c200BC 135 131 99 51 cAD150 134 124 91 55 c3300BC 130 136 104 53 c200BC 133 120 91 46 cAD150 132 127 97 52 c3300BC 131 128 98 45 c200BC 131 135 90 50 cAD150 137 125 85 57 c3300BC 138 129 107 53 c200BC 140 137 94 60 cAD150 129 128 81 52 c3300BC 123 131 101 51 c200BC 139 130 90 48 cAD150 140 135 103 48 c3300BC 130 129 105 47 c200BC 140 134 90 51 cAD150 147 129 87 48 c3300BC 134 130 93 54 c200BC 138 140 100 52 cAD150 136 133 97 51 We shall calculate Mahalanobis distances between each pair of epochs using the mahalanobis() function and apply classical scaling to the resulting dis- tance matrix. In this calculation, we shall use the estimate of the assumed common covariance matrix S S = 29S1 + 29S2 + 29S3 + 29S4 + 29S5 149 ,
- 132. 116 4 Multidimensional Scaling −1500 −500 0 500 1000 1500 −1500−50005001500 Coordinate 1 Coordinate2 ATL ORD DEN HOULAX MIA JFK SFO SEA IAD Fig. 4.1. Two-dimensional classical MDS solution for airline distances. The known spatial arrangement is clearly visible in the plot. where S1, S2, . . . , S5 are the covariance matrices of the data in each epoch. We shall then use the first two coordinate values to provide a map of the data showing the relationships between epochs. The necessary R code is: R> skulls_var <- tapply(1:nrow(skulls), skulls$epoch, + function(i) var(skulls[i,-1])) R> S <- 0 R> for (v in skulls_var) S <- S + 29 * v R> (S <- S / 149) mb bh bl nh mb 20.54407 0.03579 0.07696 1.955 bh 0.03579 22.85414 5.06040 2.769 bl 0.07696 5.06040 23.52998 1.103 nh 1.95503 2.76868 1.10291 9.880
- 133. 4.4 Classical multidimensional scaling 117 R> skulls_cen <- tapply(1:nrow(skulls), skulls$epoch, + function(i) apply(skulls[i,-1], 2, mean)) R> skulls_cen <- matrix(unlist(skulls_cen), + nrow = length(skulls_cen), byrow = TRUE) R> skulls_mah <- apply(skulls_cen, 1, + function(cen) mahalanobis(skulls_cen, cen, S)) R> skulls_mah [,1] [,2] [,3] [,4] [,5] [1,] 0.00000 0.09355 0.9280 1.9330 2.7712 [2,] 0.09355 0.00000 0.7490 1.6380 2.2357 [3,] 0.92799 0.74905 0.0000 0.4553 0.9360 [4,] 1.93302 1.63799 0.4553 0.0000 0.2253 [5,] 2.77121 2.23571 0.9360 0.2253 0.0000 R> cmdscale(skulls_mah, k = nrow(skulls_mah) - 1, + eig = TRUE)$eig [1] 5.113e+00 9.816e-02 -5.551e-16 -1.009e-01 -7.777e-01 R> skulls_mds <- cmdscale(skulls_mah) The resulting plot is shown in Figure 4.2 and shows that the scaling solu- tion for the skulls data is essentially unidimensional, with this single dimension time ordering the five epochs. There appears to be a change in the “shape” of the skulls over time, with maximum breadth increasing and basialiveolar length decreasing. Our final example of the application of classical multidimensional scal- ing involves an investigation by Corbet, Cummins, Hedges, and Krzanowski (1970), who report a study of water voles (genus Arvicola) in which the aim was to compare British populations of these animals with those in Europe to investigate whether more than one species might be present in Britain. The original data consisted of observations of the presence or absence of 13 charac- teristics in about 300 water vole skulls arising from six British populations and eight populations from the rest of Europe. Table 4.3 gives a distance matrix derived from the data as described in Corbet et al. (1970). The following code finds the classical scaling solution and computes the two criteria for assessing the required number of dimensions as described above. R> data("watervoles", package = "HSAUR2") R> voles_mds <- cmdscale(watervoles, k = 13, eig = TRUE) R> voles_mds$eig [1] 7.360e-01 2.626e-01 1.493e-01 6.990e-02 2.957e-02 [6] 1.931e-02 9.714e-17 -1.139e-02 -1.280e-02 -2.850e-02 [11] -4.252e-02 -5.255e-02 -7.406e-02 -1.098e-01 Note that some of the eigenvalues are negative. The criterion P (1) m can be computed by
- 134. 118 4 Multidimensional Scaling Table4.3:watervolesdata.Watervolesdata-dissimilaritymatrix. SrryShrpYrksPrthAbrdElnGAlpsYgslGrmnNrwyPyrIPyIINrtSSthS Surrey0.000 Shropshire0.0990.000 Yorkshire0.0330.0220.000 Perthshire0.1830.1140.0420.000 Aberdeen0.1480.2240.0590.0680.000 EleanGamhna0.1980.0390.0530.0850.0510.000 Alps0.4620.2660.3220.4350.2680.0250.000 Yugoslavia0.6280.4420.4440.4060.2400.1290.0140.000 Germany0.1130.0700.0460.0470.0340.0020.1060.1290.000 Norway0.1730.1190.1620.3310.1770.0390.0890.2370.0710.000 PyreneesI0.4340.4190.3390.5050.4690.3900.3150.3490.1510.4300.000 PyreneesII0.7620.6330.7810.7000.7580.6250.4690.6180.4400.5380.6070.000 NorthSpain0.5300.3890.4820.5790.5970.4980.3740.5620.2470.3830.3870.0840.000 SouthSpain0.5860.4350.5500.5300.5520.5090.3690.4710.2340.3460.4560.0900.0380.000
- 135. 4.4 Classical multidimensional scaling 119 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 −1.5−0.50.51.5 Coordinate 1 Coordinate2 c4000BC c3300BC c1850BC c200BC cAD150 Fig. 4.2. Two-dimensional solution from classical MDS applied to Mahalanobis distances between epochs for the skull data. R> cumsum(abs(voles_mds$eig))/sum(abs(voles_mds$eig)) [1] 0.4605 0.6248 0.7182 0.7619 0.7804 0.7925 0.7925 0.7996 [9] 0.8077 0.8255 0.8521 0.8850 0.9313 1.0000 and the criterion P (2) m is R> cumsum((voles_mds$eig)^2)/sum((voles_mds$eig)^2) [1] 0.8179 0.9220 0.9557 0.9631 0.9644 0.9649 0.9649 0.9651 [9] 0.9654 0.9666 0.9693 0.9735 0.9818 1.0000 Here the two criteria for judging the number of dimensions necessary to give an adequate fit to the data are quite different. The second criterion would suggest that two dimensions is adequate, but use of the first would suggest perhaps that three or even four dimensions might be required. Here we shall be guided by the second fit index and the two-dimensional solution that can be plotted by extracting the coordinates from the points element of the voles_mds object; the plot is shown in Figure 4.3. It appears that the six British populations are close to populations liv- ing in the Alps, Yugoslavia, Germany, Norway, and Pyrenees I (consisting
- 136. 120 4 Multidimensional Scaling R> x <- voles_mds$points[,1] R> y <- voles_mds$points[,2] R> plot(x, y, xlab = "Coordinate 1", ylab = "Coordinate 2", + xlim = range(x)*1.2, type = "n") R> text(x, y, labels = colnames(watervoles), cex = 0.7) −0.2 0.0 0.2 0.4 0.6 −0.3−0.10.1 Coordinate 1 Coordinate2 Surrey Shropshire YorkshirePerthshire AberdeenElean Gamhna Alps Yugoslavia GermanyNorway Pyrenees I Pyrenees II North Spain South Spain Fig. 4.3. Two-dimensional solution from classical multidimensional scaling of the distance matrix for water vole populations. of the species Arvicola terrestris) but rather distant from the populations in Pyrenees II, North Spain and South Spain (species Arvicola sapidus). This re- sult would seem to imply that Arvicola terrestris might be present in Britain but it is less likely that this is so for Arvicola sapidus. But here, as the two- dimensional fit may not be entirely what is needed to represent the observed distances, we shall investigate the solution in a little more detail using the minimum spanning tree. The minimum spanning tree is defined as follows. Suppose n points are given (possibly in many dimensions). Then a tree spanning these points (i.e., a spanning tree) is any set of straight line segments joining pairs of points such that ˆ o closed loops occur, ˆ every point is visited at least one time, and
- 137. 4.5 Non-metric multidimensional scaling 121 ˆ the tree is connected (i.e., it has paths between any pairs of points). The length of the tree is defined to be the sum of the length of its segments, and when a set of n points and the lengths of all n 2 segments are given, then the minimum spanning tree is defined as the spanning tree with minimum length. Algorithms to find the minimum spanning tree of a set of n points given the distances between them are given in Prim (1957) and Gower and Ross (1969). The links of the minimum spanning tree (of the spanning tree) of the proximity matrix of interest may be plotted onto the two-dimensional scaling representation in order to identify possible distortions produced by the scaling solutions. Such distortions are indicated when nearby points on the plot are not linked by an edge of the tree. To find the minimum spanning tree of the water vole proximity matrix, the function mst from the package ape (Paradis, Bolker, Claude, Cuong, Desper, Durand, Dutheil, Gascuel, Heibl, Lawson, Lefort, Legendre, Lemon, Noel, Nylander, Opgen-Rhein, Schliep, Strimmer, and de Vienne 2010) can be used, and we can plot the minimum spanning tree on the two-dimensional scaling solution as shown in Figure 4.4. The plot indicates, for example, that the apparent closeness of the popu- lations in Germany and Norway, suggested by the points representing them in the MDS solution, does not accurately reflect their calculated dissimilarity; the links of the minimum spanning tree show that the Aberdeen and Elean Gamhna populations are actually more similar to the German water voles than those from Norway. This suggests that the two-dimensional solution may not give an adequate representation of the whole distance matrix. 4.5 Non-metric multidimensional scaling In some psychological work and in market research, proximity matrices arise from asking human subjects to make judgements about the similarity or dis- similarity of objects or stimuli of interest. When collecting such data, the investigator may feel that realistically subjects are only able to give “ordinal” judgements; for example, when comparing a range of colours they might be able to specify with some confidence that one colour is brighter than another but would be far less confident if asked to put a value to how much brighter. Such considerations led, in the 1960s, to the search for a method of multidi- mensional scaling that uses only the rank order of the proximities to produce a spatial representation of them. In other words, a method was sought that would be invariant under monotonic transformations of the observed proxim- ity matrix; i.e., the derived coordinates will remain the same if the numerical values of the observed proximities are changed but their rank order is not. Such a method was proposed in landmark papers by Shepard (1962a,b) and by Kruskal (1964a). The quintessential component of the method proposed in
- 138. 122 4 Multidimensional Scaling R> library("ape") R> st <- mst(watervoles) R> plot(x, y, xlab = "Coordinate 1", ylab = "Coordinate 2", + xlim = range(x)*1.2, type = "n") R> for (i in 1:nrow(watervoles)) { + w1 <- which(st[i, ] == 1) + segments(x[i], y[i], x[w1], y[w1]) + } R> text(x, y, labels = colnames(watervoles), cex = 0.7) −0.2 0.0 0.2 0.4 0.6 −0.3−0.10.1 Coordinate 1 Coordinate2 Surrey Shropshire YorkshirePerthshire AberdeenElean Gamhna Alps Yugoslavia GermanyNorway Pyrenees I Pyrenees II North Spain South Spain Fig. 4.4. Minimum spanning tree for the watervoles data plotted onto the classical scaling two-dimensional solution. these papers is that the coordinates in the spatial representation of the ob- served proximities give rise to fitted distances, dij, and that these distances are related to a set of numbers which we will call disparities, ˆdij, by the formula dij = ˆdij + ij, where the ij are error terms representing errors of measure- ment plus distortion errors arising because the distances do not correspond to a configuration in the particular number of dimensions chosen. The disparities are monotonic with the observed proximities and, subject to this constraint, resemble the fitted distances as closely as possible. In general, only a weak monotonicity constraint is applied, so that if, say, the observed dissimilarities, δij are ranked from lowest to highest to give
- 139. 4.5 Non-metric multidimensional scaling 123 δi1j1 < δi2j2 < · · · < δiN jN , where N = n(n − 1)/2, then ˆdi1j1 ≤ ˆdi2j2 ≤ · · · ≤ ˆdiN jN . Monotonic regression (see Barlow, Bartholomew, Bremner, and Brunk 1972) is used to find the disparities, and then the required coordinates in the spatial representation of the observed dissimilarities, which we denote by ˆX(n × m), are found by minimising a criterion, S, known as Stress, which is a function of ˆX(n × m) and is defined as S( ˆX) = min i<j ( ˆdij − dij)2 / i<j d2 ij, where the minimum is taken over ˆdij such that ˆdij is monotonic with the observed dissimilarities. In essence, Stress represents the extent to which the rank order of the fitted distances disagrees with the rank order of the ob- served dissimilarities. The denominator in the formula for Stress is chosen to make the final spatial representation invariant under changes of scale; i.e., uniform stretching or shrinking. An algorithm to minimise Stress and so find the coordinates of the required spatial representation is described in a second paper by Kruskal (1964b). For each value of the number of dimensions, m, in the spatial configuration, the configuration that has the smallest Stress is called the best-fitting configuration in m dimensions, Sm, and a rule of thumb for judging the fit as given by Kruskal is Sm ≥ 20%, poor, Sm = 10%, fair, Sm ≤ 5%, good; and Sm = 0, perfect (this only occurs if the rank order of the fitted distances matches the rank order of the observed dissimilarities and event, which is, of course, very rare in practise). We will now look at some applications of non-metric multidimensional scaling. 4.5.1 House of Representatives voting Romesburg (1984) gives a set of data that shows the number of times 15 con- gressmen from New Jersey voted differently in the House of Representatives on 19 environmental bills. Abstentions are not recorded, but two congressmen abstained more frequently than the others, these being Sandman (nine absten- tions) and Thompson (six abstentions). The data are available in Table 4.4 and of interest is the question of whether party affiliations can be detected in the data. We shall now apply non-metric scaling to the voting behaviour shown in Table 4.4. Non-metric scaling is available with the function isoMDS from the package MASS (Venables and Ripley 2010, 2002) R> library("MASS") R> data("voting", package = "HSAUR2") R> voting_mds <- isoMDS(voting)
- 140. 124 4 Multidimensional Scaling and we again plot the two-dimensional solution (Figure 4.5). The figure sug- gests that voting behaviour is essentially along party lines, although there is more variation among Republicans. The voting behaviour of one of the Re- publicans (Rinaldo) seems to be closer to his Democratic colleagues rather than to the voting behaviour of other Republicans. Table 4.4: voting data. House of Representatives voting data; (R) is short for Republican, (D) for Democrat. Hnt Snd Hwr Thm Fry Frs Wdn Roe Hlt Rdn Mns Rnl Mrz Dnl Ptt Hunt(R) 0 Sandman(R) 8 0 Howard(D) 15 17 0 Thompson(D) 15 12 9 0 Freylinghuysen(R) 10 13 16 14 0 Forsythe(R) 9 13 12 12 8 0 Widnall(R) 7 12 15 13 9 7 0 Roe(D) 15 16 5 10 13 12 17 0 Heltoski(D) 16 17 5 8 14 11 16 4 0 Rodino(D) 14 15 6 8 12 10 15 5 3 0 Minish(D) 15 16 5 8 12 9 14 5 2 1 0 Rinaldo(R) 16 17 4 6 12 10 15 3 1 2 1 0 Maraziti(R) 7 13 11 15 10 6 10 12 13 11 12 12 0 Daniels(D) 11 12 10 10 11 6 11 7 7 4 5 6 9 0 Patten(D) 13 16 7 7 11 10 13 6 5 6 5 4 13 9 0 The quality of a multidimensional scaling can be assessed informally by plotting the original dissimilarities and the distances obtained from a mul- tidimensional scaling in a scatterplot, a so-called Shepard diagram. For the voting data, such a plot is shown in Figure 4.6. In an ideal situation, the points fall on the bisecting line; in our case, some deviations are observable. 4.5.2 Judgements of World War II leaders As the first example of the application of non-metric multidimensional scaling, we shall use the method to get a spatial representation of the judgements of the dissimilarities in ideology of a number of world leaders and politicians prominent at the time of the Second World War, shown in Table 4.5. The subject made judgements on a nine-point scale, with the anchor points of the scale, 1 and 9, being described as indicating “very similar” and “very dissimilar”, respectively; this was all the subject was told about the scale.
- 141. 4.5 Non-metric multidimensional scaling 125 R> x <- voting_mds$points[,1] R> y <- voting_mds$points[,2] R> plot(x, y, xlab = "Coordinate 1", ylab = "Coordinate 2", + xlim = range(voting_mds$points[,1])*1.2, type = "n") R> text(x, y, labels = colnames(voting), cex = 0.6) R> voting_sh <- Shepard(voting[lower.tri(voting)], + voting_mds$points) −10 −5 0 5 −6−202468 Coordinate 1 Coordinate2 Hunt(R) Sandman(R) Howard(D) Thompson(D) Freylinghuysen(R) Forsythe(R) Widnall(R) Roe(D) Heltoski(D) Rodino(D) Minish(D) Rinaldo(R) Maraziti(R) Daniels(D) Patten(D) Fig. 4.5. Two-dimensional solution from non-metric multidimensional scaling of distance matrix for voting matrix. Table 4.5: WWIIleaders data. Subjective distances between WWII leaders. Htl Mss Chr Esn Stl Att Frn DGl MT- Trm Chm Tit Hitler 0 Mussolini 3 0 Churchill 4 6 0 Eisenhower 7 8 4 0 Stalin 3 5 6 8 0 Attlee 8 9 3 9 8 0 Franco 3 2 5 7 6 7 0 De Gaulle 4 4 3 5 6 5 4 0
- 142. 126 4 Multidimensional Scaling Table 4.5: WWIIleaders data (continued). Htl Mss Chr Esn Stl Att Frn DGl MT- Trm Chm Tit Mao Tse-Tung 8 9 8 9 6 9 8 7 0 Truman 9 9 5 4 7 8 8 4 4 0 Chamberlin 4 5 5 4 7 2 2 5 9 5 0 Tito 7 8 2 4 7 8 3 2 4 5 7 0 The non-metric multidimensional scaling applied to these distances is R> (WWII_mds <- isoMDS(WWIIleaders)) initial value 20.504211 iter 5 value 15.216103 iter 5 value 15.207237 iter 5 value 15.207237 final value 15.207237 converged $points [,1] [,2] Hitler -2.5820 -1.75961 Mussolini -3.8807 -1.24756 Churchill 0.3110 1.46671 Eisenhower 2.9852 2.87822 Stalin -1.4274 -3.75699 Attlee -2.1067 5.07317 Franco -2.8590 0.07878 De Gaulle 0.6591 -0.20656 Mao Tse-Tung 4.1605 -4.57583 Truman 4.4962 0.29294 Chamberlin -2.1420 2.75877 Tito 2.3859 -1.00203 $stress [1] 15.21 The two-dimensional solution appears in Figure 4.7. Clearly, the three fascists group together as do the three British prime ministers. Stalin and Mao Tse- Tung are more isolated compared with the other leaders. Eisenhower seems more related to the British government than to his own President Truman. Interestingly, de Gaulle is placed in the center of the MDS solution.
- 143. 4.6 Correspondence analysis 127 R> plot(voting_sh, pch = ".", xlab = "Dissimilarity", + ylab = "Distance", xlim = range(voting_sh$x), + ylim = range(voting_sh$x)) R> lines(voting_sh$x, voting_sh$yf, type = "S") 5 10 15 51015 Dissimilarity Distance Fig. 4.6. The Shepard diagram for the voting data shows some discrepancies be- tween the original dissimilarities and the multidimensional scaling solution. 4.6 Correspondence analysis A form of multidimensional scaling known as correspondence analysis, which is essentially an approach to constructing a spatial model that displays the associations among a set of categorical variables, will be the subject of this section. Correspondence analysis has a relatively long history (see de Leeuw 1983) but for a long period was only routinely used in France, largely due to the almost evangelical efforts of Benz´ecri (1992). But nowadays the method is used rather more widely and is often applied to supplement, say, a stan- dard chi-squared test of independence for two categorical variables forming a contingency table. Mathematically, correspondence analysis can be regarded as either ˆ a method for decomposing the chi-squared statistic used to test for inde- pendence in a contingency table into components corresponding to differ- ent dimensions of the heterogeneity between its columns, or
- 144. 128 4 Multidimensional Scaling −4 −2 0 2 4 6 −4−20246 Coordinate 1 Coordinate2 Hitler Mussolini Churchill Eisenhower Stalin Attlee Franco De Gaulle Mao Tse−Tung Truman Chamberlin Tito Fig. 4.7. Non-metric multidimensional scaling of perceived distances of World War II leaders. ˆ a method for simultaneously assigning a scale to rows and a separate scale to columns so as to maximise the correlation between the two scales. Quintessentially, however, correspondence analysis is a technique for dis- playing multivariate (most often bivariate) categorical data graphically by deriving coordinates to represent the categories of both the row and column variables, which may then be plotted so as to display the pattern of association between the variables graphically. A detailed account of correspondence anal- ysis is given in Greenacre (2007), where its similarity to principal components and the biplot is stressed. Here we give only accounts of the method demon- strating the use of classical multidimensional scaling to get a two-dimensional map to represent a set of data in the form of a two-dimensional contingency table. The general two-dimensional contingency table in which there are r rows and c columns can be written as
- 145. 4.6 Correspondence analysis 129 y 1 . . . c 1 n11 . . . n1c n1· 2 n21 . . . n2c n2· x ... ... . . . ... ... r nr1 . . . nrc nr· n·1 . . . n·c n using an obvious dot notation for summing the counts in the contingency table over rows or over columns. From this table we can construct tables of column proportions and row proportions given by Column proportions pc ij = nij/n·j, Row proportions pr ij = nij/ni·. What is known as the chi-squared distance between columns i and j is defined as d (cols) ij = r k=1 1 pk· (pc ki − pc kj)2 , where pk· = nk·/n. The chi-square distance is seen to be a weighted Euclidean distance based on column proportions. It will be zero if the two columns have the same values for these proportions. It can also be seen from the weighting factors, 1/pk·, that rare categories of the column variable have a greater influence on the distance than common ones. A similar distance measure can be defined for rows i and j as d (rows) ij = c k=1 1 p·k (pr ik − pr jk)2 , where p·k = n·k/n. A correspondence analysis“map”of the data can be found by applying classical MDS to each distance matrix in turn and plotting usually the first two coordi- nates for column categories and those for row categories on the same diagram, suitably labelled to differentiate the points representing row categories from those representing column categories. The resulting diagram is interpreted by examining the positions of the points representing the row categories and the column categories. The relative values of the coordinates of these points reflect associations between the categories of the row variable and the categories of the column variable. Assuming that a two-dimensional solution provides an adequate fit for the data (see Greenacre 1992), row points that are close to- gether represent row categories that have similar profiles (conditional distribu- tions) across columns. Column points that are close together indicate columns
- 146. 130 4 Multidimensional Scaling with similar profiles (conditional distributions) down the rows. Finally, row points that lie close to column points represent a row/column combination that occurs more frequently in the table than would be expected if the row and column variables were independent. Conversely, row and column points that are distant from one another indicate a cell in the table where the count is lower than would be expected under independence. We will now look at a single simple example of the application of corre- spondence analysis. 4.6.1 Teenage relationships Consider the data shown in Table 4.6 concerned with the influence of a girl’s age on her relationship with her boyfriend. In this table, each of 139 girls has been classified into one of three groups: ˆ no boyfriend; ˆ boyfriend/no sexual intercourse; or ˆ boyfriend/sexual intercourse. In addition, the age of each girl was recorded and used to divide the girls into five age groups. Table 4.6: teensex data. The influence of age on relationships with boyfriends. Age Boyfriend <16 16-17 17-18 18-19 19-20 No boyfriend 21 21 14 13 8 Boyfriend no sex 8 9 6 8 2 Boyfriend sex 2 3 4 10 10 The calculation of the two-dimensional classical multidimensional scaling so- lution based on the row- and column-wise chi-squared distance measure can be computed via cmdscale(); however, we first have to compute the necessary row and column distance matrices, and we will do this by setting up a small convenience function as follows: R> D <- function(x) { + a <- t(t(x) / colSums(x)) + ret <- sqrt(colSums((a[,rep(1:ncol(x), ncol(x))] - + a[, rep(1:ncol(x), rep(ncol(x), ncol(x)))])^2 * + sum(x) / rowSums(x))) + matrix(ret, ncol = ncol(x)) + } R> (dcols <- D(teensex))
- 147. 4.7 Summary 131 [,1] [,2] [,3] [,4] [,5] [1,] 0.00000 0.08537 0.2574 0.6629 1.0739 [2,] 0.08537 0.00000 0.1864 0.5858 1.0142 [3,] 0.25743 0.18644 0.0000 0.4066 0.8295 [4,] 0.66289 0.58581 0.4066 0.0000 0.5067 [5,] 1.07385 1.01423 0.8295 0.5067 0.0000 R> (drows <- D(t(teensex))) [,1] [,2] [,3] [1,] 0.0000 0.2035 0.9275 [2,] 0.2035 0.0000 0.9355 [3,] 0.9275 0.9355 0.0000 Applying classical MDS to each of these distance matrices gives the re- quired two-dimensional coordinates with which to construct our “map” of the data. Plotting those with suitable labels and with the axes suitably scaled to reflect the greater variation on dimension one than on dimension two (see Greenacre 1992) is achieved using the R code presented with Figure 4.8. The points representing the age groups in Figure 4.8 give a two-dimensional representation in which the Euclidean distance between two points represents the chi-squared distance between the corresponding age groups (and similarly for the points representing the type of relationship). For a contingency table with r rows and c columns, it can be shown that the chi-squared distances can be represented exactly in min r − 1, c − 1 dimensions; here, since r = 3 and c = 5, this means that the Euclidean distances in Figure 4.8 will actually equal the corresponding chi-squared distances (readers might like to check that this is the case as an exercise). When both r and c are greater than three, an exact two-dimensional representation of the chi-squared distances is not possible. In such cases, the derived two-dimensional coordinates will give only an approximate representation, and so the question of the adequacy of the fit will need to be addressed. In some of these cases, more than two dimensions may be required to give an acceptable fit (again see Greenacre 1992, for details). Examining the plot in Figure 4.8, we see that it tells the age-old story of girls travelling through their teenage years, initially having no boyfriend, then acquiring a boyfriend, and then having sex with their boyfriend, a story that has broken the hearts of fathers everywhere, at least temporarily, until their wives suggest they reflect back to the time when they themselves were teenagers. 4.7 Summary Multidimensional scaling and correspondence analysis both aim to help in understanding particular types of data by displaying the data graphically.
- 148. 132 4 Multidimensional Scaling R> r1 <- cmdscale(dcols, eig = TRUE) R> c1 <- cmdscale(drows, eig = TRUE) R> plot(r1$points, xlim = range(r1$points[,1], c1$points[,1]) * 1.5, + ylim = range(r1$points[,1], c1$points[,1]) * 1.5, type = "n", + xlab = "Coordinate 1", ylab = "Coordinate 2", lwd = 2) R> text(r1$points, labels = colnames(teensex), cex = 0.7) R> text(c1$points, labels = rownames(teensex), cex = 0.7) R> abline(h = 0, lty = 2) R> abline(v = 0, lty = 2) −0.5 0.0 0.5 1.0 −0.50.00.51.0 Coordinate 1 Coordinate2 <16 16−17 17−18 18−19 19−20No boyfriend Boyfriend no sex Boyfriend sex Fig. 4.8. Correspondence analysis for teenage relationship data. Multidimensional scaling applied to proximity matrices is often useful in un- covering the dimensions on which similarity judgements are made, and corre- spondence analysis often allows more insight into the pattern of relationships in a contingency table than a simple chi-squared test. 4.8 Exercises Ex. 4.1 Consider 51 objects O1, . . . , O51 assumed to be arranged along a straight line with the jth object being located at a point with coordinate j. Define the similarity sij between object i and object j as
- 149. 4.8 Exercises 133 sij = 9 if i = j 8 if 1 ≤ |i − j| ≤ 3 7 if 4 ≤ |i − j| ≤ 6 · · · 1 if 22 ≤ |i − j| ≤ 24 0 if |i − j| ≥ 25. Convert these similarities into dissimilarities (δij) by using δij = sii + sjj − 2sij and then apply classical multidimensional scaling to the resulting dissim- ilarity matrix. Explain the shape of the derived two-dimensional solution. Ex. 4.2 Write an R function to calculate the chi-squared distance matrices for both rows and columns in a two-dimensional contingency table. Ex. 4.3 In Table 4.7 (from Kaufman and Rousseeuw 1990), the dissimilarity matrix of 18 species of garden flowers is shown. Use some form of multidi- mensional scaling to investigate which species share common properties.
- 150. 134 4 Multidimensional Scaling Table4.7:gardenflowersdata.Dissimilaritymatrixof18speciesofgarden- flowers. BgnBrmCmlDhlF-FchGrnGldHthHydIrsLlyL-PnyPncRdrScrTlp Begonia0.00 Broom0.910.00 Camellia0.490.670.00 Dahlia0.470.590.590.00 Forget-me-not0.430.900.570.610.00 Fuchsia0.230.790.290.520.440.00 Geranium0.310.700.540.440.540.240.00 Gladiolus0.490.570.710.260.490.680.490.00 Heather0.570.570.570.890.500.610.700.770.00 Hydrangea0.760.580.580.620.390.610.860.700.550.00 Iris0.320.770.630.750.460.520.600.630.460.470.00 Lily0.510.690.690.530.510.650.770.470.510.390.360.00 Lily-of-the-valley0.590.750.750.770.350.630.720.650.350.410.450.240.00 Peony0.370.680.680.380.520.480.630.490.520.390.370.170.390.00 Pinkcarnation0.740.540.700.580.540.740.500.490.360.520.600.480.390.490.00 Redrose0.840.410.750.370.820.710.610.640.810.430.840.620.670.470.450.00 Scotchrose0.940.200.700.480.770.830.740.450.770.380.800.580.620.570.400.210.00 Tulip0.440.500.790.480.590.680.470.220.590.920.590.670.720.670.610.850.670.00
- 151. 5 Exploratory Factor Analysis 5.1 Introduction In many areas of psychology, and other disciplines in the behavioural sciences, often it is not possible to measure directly the concepts of primary interest. Two obvious examples are intelligence and social class. In such cases, the re- searcher is forced to examine the concepts indirectly by collecting information on variables that can be measured or observed directly and can also realis- tically be assumed to be indicators, in some sense, of the concepts of real interest. The psychologist who is interested in an individual’s “intelligence”, for example, may record examination scores in a variety of different subjects in the expectation that these scores are dependent in some way on what is widely regarded as “intelligence” but are also subject to random errors. And a soci- ologist, say, concerned with people’s “social class” might pose questions about a person’s occupation, educational background, home ownership, etc., on the assumption that these do reflect the concept he or she is really interested in. Both “intelligence” and “social class” are what are generally referred to as latent variables–i.e., concepts that cannot be measured directly but can be as- sumed to relate to a number of measurable or manifest variables. The method of analysis most generally used to help uncover the relationships between the assumed latent variables and the manifest variables is factor analysis. The model on which the method is based is essentially that of multiple regression, except now the manifest variables are regressed on the unobservable latent variables (often referred to in this context as common factors), so that direct estimation of the corresponding regression coefficients (factor loadings) is not possible. A point to be made at the outset is that factor analysis comes in two dis- tinct varieties. The first is exploratory factor analysis, which is used to inves- tigate the relationship between manifest variables and factors without making any assumptions about which manifest variables are related to which factors. The second is confirmatory factor analysis which is used to test whether a specific factor model postulated a priori provides an adequate fit for the co- DOI 10.1007/978-1-4419-9650-3_5, © Springer Science+Business Media, LLC 2011 135B. Everitt and T. Hothorn, An Introduction to Applied Multivariate Analysis with R: Use R!,
- 152. 136 5 Exploratory Factor Analysis variances or correlations between the manifest variables. In this chapter, we shall consider only exploratory factor analysis. Confirmatory factor analysis will be the subject of Chapter 7. Exploratory factor analysis is often said to have been introduced by Spear- man (1904), but this is only partially true because Spearman proposed only the one-factor model as described in the next section. Fascinating accounts of the history of factor analysis are given in Thorndike (2005) and Bartholomew (2005). 5.2 A simple example of a factor analysis model To set the scene for the k-factor analysis model to be described in the next section, we shall in this section look at a very simple example in which there is only a single factor. Spearman considered a sample of children’s examination marks in three subjects, Classics (x1), French (x2), and English (x3), from which he calculated the following correlation matrix for a sample of children: R = Classics French English 1.00 0.83 1.00 0.78 0.67 1.00 . If we assume a single factor, then the single-factor model is specified as follows: x1 = λ1f + u1, x2 = λ2f + u2, x3 = λ3f + u3. We see that the model essentially involves the simple linear regression of each observed variable on the single common factor. In this example, the under- lying latent variable or common factor, f, might possibly be equated with intelligence or general intellectual ability. The terms λ1, λ2, and λ3 which are essentially regression coefficients are, in this context, known as factor load- ings, and the terms u1, u2, and u3 represent random disturbance terms and will have small variances if their associated observed variable is closely related to the underlying latent variable. The variation in ui actually consists of two parts, the extent to which an individual’s ability at Classics, say, differs from his or her general ability and the extent to which the examination in Classics is only an approximate measure of his or her ability in the subject. In practise no attempt is made to disentangle these two parts. We shall return to this simple example later when we consider how to estimate the parameters in the factor analysis model. Before this, however, we need to describe the factor analysis model itself in more detail. The description follows in the next section.
- 153. 5.3 The k-factor analysis model 137 5.3 The k-factor analysis model The basis of factor analysis is a regression model linking the manifest vari- ables to a set of unobserved (and unobservable) latent variables. In essence the model assumes that the observed relationships between the manifest variables (as measured by their covariances or correlations) are a result of the relation- ships of these variables to the latent variables. (Since it is the covariances or correlations of the manifest variables that are central to factor analysis, we can, in the description of the mathematics of the method given below, assume that the manifest variables all have zero mean.) To begin, we assume that we have a set of observed or manifest variables, x = (x1, x2, . . . , xq), assumed to be linked to k unobserved latent variables or common factors f1, f2, . . . , fk, where k < q, by a regression model of the form x1 = λ11f1 + λ12f2 + · · · + λ1kfk + u1, x2 = λ21f1 + λ22f2 + · · · + λ2kfk + u2, ... xq = λq1f1 + λq2f2 + · · · + λqkfk + uq. The λjs are essentially the regression coefficients of the x-variables on the common factors, but in the context of factor analysis these regression coeffi- cients are known as the factor loadings and show how each observed variable, xi, depends on the common factors. The factor loadings are used in the inter- pretation of the factors; i.e., larger values relate a factor to the corresponding observed variables and from these we can often, but not always, infer a mean- ingful description of each factor (we will give examples later). The regression equations above may be written more concisely as x = Λf + u, where Λ = λ11 . . . λ1k ... ... λq1 . . . λqk , f = f1 ... fq , u = u1 ... uq . We assume that the random disturbance terms u1, . . . , uq are uncorrelated with each other and with the factors f1, . . . , fk. (The elements of u are spe- cific to each xi and hence are generally better known in this context as specific variates.) The two assumptions imply that, given the values of the common factors, the manifest variables are independent; that is, the correlations of the observed variables arise from their relationships with the common factors. Because the factors are unobserved, we can fix their locations and scales arbi- trarily and we shall assume they occur in standardised form with mean zero and standard deviation one. We will also assume, initially at least, that the
- 154. 138 5 Exploratory Factor Analysis factors are uncorrelated with one another, in which case the factor loadings are the correlations of the manifest variables and the factors. With these ad- ditional assumptions about the factors, the factor analysis model implies that the variance of variable xi, σ2 i , is given by σ2 i = k j=1 λ2 ij + ψi, where ψi is the variance of ui. Consequently, we see that the factor analysis model implies that the variance of each observed variable can be split into two parts: the first, h2 i , given by h2 i = k j=1 λ2 ij, is known as the communality of the variable and represents the variance shared with the other variables via the common factors. The second part, ψi, is called the specific or unique variance and relates to the variability in xi not shared with other variables. In addition, the factor model leads to the following expression for the covariance of variables xi and xj: σij = k l=1 λilλjl. We see that the covariances are not dependent on the specific variates in any way; it is the common factors only that aim to account for the relationships between the manifest variables. The results above show that the k-factor analysis model implies that the population covariance matrix, Σ, of the observed variables has the form Σ = ΛΛ + Ψ, where Ψ = diag(Ψi). The converse also holds: if Σ can be decomposed into the form given above, then the k-factor model holds for x. In practise, Σ will be estimated by the sample covariance matrix S and we will need to obtain estimates of Λ and Ψ so that the observed covariance matrix takes the form required by the model (see later in the chapter for an account of estimation methods). We will also need to determine the value of k, the number of factors, so that the model provides an adequate fit for S. 5.4 Scale invariance of the k-factor model Before describing both estimation for the k-factor analysis model and how to determine the appropriate value of k, we will consider how rescaling the x variables affects the factor analysis model. Rescaling the x variables is equiv- alent to letting y = Cx, where C = diag(ci) and the ci, i = 1, . . . , q are the
- 155. 5.5 Estimating the parameters in the k-factor analysis model 139 scaling values. If the k-factor model holds for x with Λ = Λx and Ψ = Ψx, then y = CΨxf + Cu and the covariance matrix of y implied by the factor analysis model for x is Var(y) = CΣC = CΛxC + CΨxC. So we see that the k-factor model also holds for y with factor loading matrix Λy = CΛx and specific variances Ψy = CΨxC = c2 i ψi. So the factor loading matrix for the scaled variables y is found by scaling the factor loading matrix of the original variables by multiplying the ith row of Λx by ci and similarly for the specific variances. Thus factor analysis is essentially unaffected by the rescaling of the variables. In particular, if the rescaling factors are such that ci = 1/si, where si is the standard deviation of the xi, then the rescaling is equivalent to applying the factor analysis model to the correlation matrix of the x variables and the factor loadings and specific variances that result can be found simply by scaling the corresponding loadings and variances obtained from the covariance matrix. Consequently, the factor analysis model can be applied to either the covariance matrix or the correlation matrix because the results are essentially equivalent. (Note that this is not the same as when using principal components analysis, as pointed out in Chapter 3, and we will return to this point later in the chapter.) 5.5 Estimating the parameters in the k-factor analysis model To apply the factor analysis model outlined in the previous section to a sample of multivariate observations, we need to estimate the parameters of the model in some way. These parameters are the factor loadings and specific variances, and so the estimation problem in factor analysis is essentially that of finding ˆΛ (the estimated factor loading matrix) and ˆΨ (the diagonal matrix containing the estimated specific variances), which, assuming the factor model outlined in Section 5.3, reproduce as accurately as possible the sample covariance matrix, S. This implies S ≈ ˆΛˆΛ + ˆΨ. Given an estimate of the factor loading matrix, ˆΛ, it is clearly sensible to estimate the specific variances as ˆψi = s2 i − k j=1 ˆλ2 ij, i = 1, . . . , q so that the diagonal terms in S are estimated exactly.
- 156. 140 5 Exploratory Factor Analysis Before looking at methods of estimation used in practise, we shall for the moment return to the simple single-factor model considered in Section 5.2 because in this case estimation of the factor loadings and specific variances is very simple, the reason being that in this case the number of parameters in the model, 6 (three factor loadings and three specific variances), is equal to the number of independent elements in R (the three correlations and the three diagonal standardised variances), and so by equating elements of the observed correlation matrix to the corresponding values predicted by the single-factor model, we will be able to find estimates of λ1, λ2, λ3, ψ1, ψ2, and ψ3 such that the model fits exactly. The six equations derived from the matrix equality implied by the factor analysis model, R = λ1 λ2 λ3 λ1 λ2λ3 + ψ1 0 0 0 ψ2 0 0 0 ψ3 , are ˆλ1λ2 = 0.83, ˆλ1λ3 = 0.78, ˆλ1λ4 = 0.67, ψ1 = 1.0 − ˆλ2 1, ψ2 = 1.0 − ˆλ2 2, ψ3 = 1.0 − ˆλ2 3. The solutions of these equations are ˆλ1 = 0.99, ˆλ2 = 0.84, ˆλ3 = 0.79, ˆψ1 = 0.02, ˆψ2 = 0.30, ˆψ3 = 0.38. Suppose now that the observed correlations had been R = Classics French English 1.00 0.84 1.00 0.60 0.35 1.00 . In this case, the solution for the parameters of a single-factor model is ˆλ1 = 1.2, ˆλ2 = 0.7, ˆλ3 = 0.5, ˆψ1 = −0.44, ˆψ2 = 0.51, ˆψ3 = 0.75. Clearly this solution is unacceptable because of the negative estimate for the first specific variance. In the simple example considered above, the factor analysis model does not give a useful description of the data because the number of parameters in
- 157. 5.5 Estimating the parameters in the k-factor analysis model 141 the model equals the number of independent elements in the correlation ma- trix. In practise, where the k-factor model has fewer parameters than there are independent elements of the covariance or correlation matrix (see Sec- tion 5.6), the fitted model represents a genuinely parsimonious description of the data and methods of estimation are needed that try to make the covari- ance matrix predicted by the factor model as close as possible in some sense to the observed covariance matrix of the manifest variables. There are two main methods of estimation leading to what are known as principal factor anal- ysis and maximum likelihood factor analysis, both of which are now briefly described. 5.5.1 Principal factor analysis Principal factor analysis is an eigenvalue and eigenvector technique similar in many respects to principal components analysis (see Chapter 3) but operating not directly on S (or R) but on what is known as the reduced covariance matrix, S∗ , defined as S∗ = S − ˆΨ, where ˆΨ is a diagonal matrix containing estimates of the ψi. The “ones” on the diagonal of S have in S∗ been replaced by the estimated communalities, k j=1 ˆλ2 ij, the parts of the variance of each observed variable that can be explained by the common factors. Unlike principal components analysis, factor analysis does not try to account for all the observed variance, only that shared through the common factors. Of more concern in factor analysis is accounting for the covariances or correlations between the manifest variables. To calculate S∗ (or with R replacing S, R∗ ) we need values for the com- munalities. Clearly we cannot calculate them on the basis of factor loadings because these loadings still have to be estimated. To get around this seem- ingly “chicken and egg” situation, we need to find a sensible way of finding initial values for the communalities that does not depend on knowing the fac- tor loadings. When the factor analysis is based on the correlation matrix of the manifest variables, two frequently used methods are: ˆ Take the communality of a variable xi as the square of the multiple corre- lation coefficient of xi with the other observed variables. ˆ Take the communality of xi as the largest of the absolute values of the correlation coefficients between xi and one of the other variables. Each of these possibilities will lead to higher values for the initial communality when xi is highly correlated with at least some of the other manifest variables, which is essentially what is required. Given the initial communality values, a principal components analysis is performed on S∗ and the first k eigenvectors used to provide the estimates of the loadings in the k-factor model. The estimation process can stop here or the loadings obtained at this stage can provide revised communality estimates
- 158. 142 5 Exploratory Factor Analysis calculated as k j=1 ˆλ2 ij, where the ˆλ2 ijs are the loadings estimated in the pre- vious step. The procedure is then repeated until some convergence criterion is satisfied. Difficulties can sometimes arise with this iterative approach if at any time a communality estimate exceeds the variance of the corresponding manifest variable, resulting in a negative estimate of the variable’s specific variance. Such a result is known as a Heywood case (see Heywood 1931) and is clearly unacceptable since we cannot have a negative specific variance. 5.5.2 Maximum likelihood factor analysis Maximum likelihood is regarded, by statisticians at least, as perhaps the most respectable method of estimating the parameters in the factor analysis. The essence of this approach is to assume that the data being analysed have a multivariate normal distribution (see Chapter 1). Under this assumption and assuming the factor analysis model holds, the likelihood function L can be shown to be −1 2 nF plus a function of the observations where F is given by F = ln |ΛΛ + Ψ| + trace(S|ΛΛ + Ψ|−1 ) − ln |S| − q. The function F takes the value zero if ΛΛ +Ψ is equal to S and values greater than zero otherwise. Estimates of the loadings and the specific variances are found by minimising F with respect to these parameters. A number of iterative numerical algorithms have been suggested; for details see Lawley and Maxwell (1963), Mardia et al. (1979), Everitt (1984, 1987), and Rubin and Thayer (1982). Initial values of the factor loadings and specific variances can be found in a number of ways, including that described above in Section 5.5.1. As with iterated principal factor analysis, the maximum likelihood approach can also experience difficulties with Heywood cases. 5.6 Estimating the number of factors The decision over how many factors, k, are needed to give an adequate rep- resentation of the observed covariances or correlations is generally critical when fitting an exploratory factor analysis model. Solutions with k = m and k = m + 1 will often produce quite different factor loadings for all factors, unlike a principal components analysis, in which the first m components will be identical in each solution. And, as pointed out by Jolliffe (2002), with too few factors there will be too many high loadings, and with too many factors, factors may be fragmented and difficult to interpret convincingly. Choosing k might be done by examining solutions corresponding to dif- ferent values of k and deciding subjectively which can be given the most convincing interpretation. Another possibility is to use the scree diagram ap- proach described in Chapter 3, although the usefulness of this method is not
- 159. 5.7 Factor rotation 143 so clear in factor analysis since the eigenvalues represent variances of principal components, not factors. An advantage of the maximum likelihood approach is that it has an associ- ated formal hypothesis testing procedure that provides a test of the hypothesis Hk that k common factors are sufficient to describe the data against the alter- native that the population covariance matrix of the data has no constraints. The test statistic is U = N min(F), where N = n + 1 − 1 6 (2q + 5) − 2 3 k. If k common factors are adequate to account for the observed covariances or correlations of the manifest variables (i.e., Hk is true), then U has, asymptotically, a chi-squared distribution with ν degrees of freedom, where ν = 1 2 (q − k)2 − 1 2 (q + k). In most exploratory studies, k cannot be specified in advance and so a sequen- tial procedure is used. Starting with some small value for k (usually k = 1), the parameters in the corresponding factor analysis model are estimated using maximum likelihood. If U is not significant, the current value of k is accepted; otherwise k is increased by one and the process is repeated. If at any stage the degrees of freedom of the test become zero, then either no non-trivial solution is appropriate or alternatively the factor model itself, with its assumption of linearity between observed and latent variables, is questionable. (This proce- dure is open to criticism because the critical values of the test criterion have not been adjusted to allow for the fact that a set of hypotheses are being tested in sequence.) 5.7 Factor rotation Up until now, we have conveniently ignored one problematic feature of the factor analysis model, namely that, as formulated in Section 5.3, there is no unique solution for the factor loading matrix. We can see that this is so by introducing an orthogonal matrix M of order k × k and rewriting the basic regression equation linking the observed and latent variables as x = (ΛM)(M f) + u. This “new” model satisfies all the requirements of a k-factor model as previ- ously outlined with new factors f∗ = Mf and the new factor loadings ΛM. This model implies that the covariance matrix of the observed variables is Σ = (ΛM)(ΛM) + Ψ,
- 160. 144 5 Exploratory Factor Analysis which, since MM = I, reduces to Σ = ΛΛ +Ψ as before. Consequently, factors f with loadings Λ and factors f∗ with loadings ΛM are, for any or- thogonal matrix M, equivalent for explaining the covariance matrix of the observed variables. Essentially then there are an infinite number of solutions to the factor analysis model as previously formulated. The problem is generally solved by introducing some constraints in the original model. One possibility is to require the matrix G given by G = ΛΨ−1 Λ to be diagonal, with its elements arranged in descending order of magnitude. Such a requirement sets the first factor to have maximal contribution to the common variance of the observed variables, and the second has maximal con- tribution to this variance subject to being uncorrelated with the first and so on (cf. principal components analysis in Chapter 3). The constraint above ensures that Λ is uniquely determined, except for a possible change of sign of the columns. (When k = 1, the constraint is irrelevant.) The constraints on the factor loadings imposed by a condition such as that given above need to be introduced to make the parameter estimates in the factor analysis model unique, and they lead to orthogonal factors that are arranged in descending order of importance. These properties are not, however, inherent in the factor model, and merely considering such a solution may lead to difficulties of interpretation. For example, two consequences of a factor solution found when applying the constraint above are: ˆ The factorial complexity of variables is likely to be greater than one re- gardless of the underlying true model; consequently variables may have substantial loadings on more than one factor. ˆ Except for the first factor, the remaining factors are often bipolar; i.e., they have a mixture of positive and negative loadings. It may be that a more interpretable orthogonal solution can be achieved using the equivalent model with loadings Λ∗ = ΛM for some particular orthogonal matrix, M. Such a process is generally known as factor rotation, but before we consider how to choose M (i.e., how to “rotate” the factors), we need to address the question “is factor rotation an acceptable process?” Certainly factor analysis has in the past been the subject of severe criticism because of the possibility of rotating factors. Critics have suggested that this apparently allows investigators to impose on the data whatever type of solu- tion they are looking for; some have even gone so far as to suggest that factor analysis has become popular in some areas precisely because it does enable users to impose their preconceived ideas of the structure behind the observed correlations (Blackith and Reyment 1971). But, on the whole, such suspicions are not justified and factor rotation can be a useful procedure for simplifying an exploratory factor analysis. Factor rotation merely allows the fitted factor analysis model to be described as simply as possible; rotation does not alter
- 161. 5.7 Factor rotation 145 the overall structure of a solution but only how the solution is described. Ro- tation is a process by which a solution is made more interpretable without changing its underlying mathematical properties. Initial factor solutions with variables loading on several factors and with bipolar factors can be difficult to interpret. Interpretation is more straightforward if each variable is highly loaded on at most one factor and if all factor loadings are either large and positive or near zero, with few intermediate values. The variables are thus split into disjoint sets, each of which is associated with a single factor. This aim is essentially what Thurstone (1931) referred to as simple structure. In more detail, such structure has the following properties: ˆ Each row or the factor loading matrix should contain at least one zero. ˆ Each column of the loading matrix should contain at least k zeros. ˆ Every pair of columns of the loading matrix should contain several vari- ables whose loadings vanish in one column but not in the other. ˆ If the number of factors is four or more, every pair of columns should contain a large number of variables with zero loadings in both columns. ˆ Conversely, for every pair of columns of the loading matrix only a small number of variables should have non-zero loadings in both columns. When simple structure is achieved, the observed variables will fall into mu- tually exclusive groups whose loadings are high on single factors, perhaps moderate to low on a few factors, and of negligible size on the remaining factors. Medium-sized, equivocal loadings are to be avoided. The search for simple structure or something close to it begins after an initial factoring has determined the number of common factors necessary and the communalities of each observed variable. The factor loadings are then transformed by post-multiplication by a suitably chosen orthogonal matrix. Such a transformation is equivalent to a rigid rotation of the axes of the origi- nally identified factor space. And during the rotation phase of the analysis, we might choose to abandon one of the assumptions made previously, namely that factors are orthogonal, i.e., independent (the condition was assumed initially simply for convenience in describing the factor analysis model). Consequently, two types of rotation are possible: ˆ orthogonal rotation, in which methods restrict the rotated factors to being uncorrelated, or ˆ oblique rotation, where methods allow correlated factors. As we have seen above, orthogonal rotation is achieved by post-multiplying the original matrix of loadings by an orthogonal matrix. For oblique rotation, the original loadings matrix is post-multiplied by a matrix that is no longer constrained to be orthogonal. With an orthogonal rotation, the matrix of correlations between factors after rotation is the identity matrix. With an oblique rotation, the corresponding matrix of correlations is restricted to have unit elements on its diagonal, but there are no restrictions on the off-diagonal elements.
- 162. 146 5 Exploratory Factor Analysis So the first question that needs to be considered when rotating factors is whether we should use an orthogonal or an oblique rotation. As for many questions posed in data analysis, there is no universal answer to this ques- tion. There are advantages and disadvantages to using either type of rotation procedure. As a general rule, if a researcher is primarily concerned with get- ting results that “best fit” his or her data, then the factors should be rotated obliquely. If, on the other hand, the researcher is more interested in the gen- eralisability of his or her results, then orthogonal rotation is probably to be preferred. One major advantage of an orthogonal rotation is simplicity since the loadings represent correlations between factors and manifest variables. This is not the case with an oblique rotation because of the correlations between the factors. Here there are two parts of the solution to consider; ˆ factor pattern coefficients, which are regression coefficients that multiply with factors to produce measured variables according to the common factor model, and ˆ factor structure coefficients, correlation coefficients between manifest vari- ables and the factors. Additionally there is a matrix of factor correlations to consider. In many cases where these correlations are relatively small, researchers may prefer to return to an orthogonal solution. There are a variety of rotation techniques, although only relatively few are in general use. For orthogonal rotation, the two most commonly used techniques are known as varimax and quartimax. ˆ Varimax rotation, originally proposed by Kaiser (1958), has as its rationale the aim of factors with a few large loadings and as many near-zero load- ings as possible. This is achieved by iterative maximisation of a quadratic function of the loadings–details are given in Mardia et al. (1979). It pro- duces factors that have high correlations with one small set of variables and little or no correlation with other sets. There is a tendency for any general factor to disappear because the factor variance is redistributed. ˆ Quartimax rotation, originally suggested by Carroll (1953), forces a given variable to correlate highly on one factor and either not at all or very low on other factors. It is far less popular than varimax. For oblique rotation, the two methods most often used are oblimin and pro- max. ˆ Oblimin rotation, invented by Jennrich and Sampson (1966), attempts to find simple structure with regard to the factor pattern matrix through a parameter that is used to control the degree of correlation between the factors. Fixing a value for this parameter is not straightforward, but Pett, Lackey, and Sullivan (2003) suggest that values between about −0.5 and 0.5 are sensible for many applications.
- 163. 5.8 Estimating factor scores 147 ˆ Promax rotation, a method due to Hendrickson and White (1964), operates by raising the loadings in an orthogonal solution (generally a varimax rotation) to some power. The goal is to obtain a solution that provides the best structure using the lowest possible power loadings and the lowest correlation between the factors. Factor rotation is often regarded as controversial since it apparently allows the investigator to impose on the data whatever type of solution is required. But this is clearly not the case since although the axes may be rotated about their origin or may be allowed to become oblique, the distribution of the points will remain invariant. Rotation is simply a procedure that allows new axes to be chosen so that the positions of the points can be described as simply as possible. (It should be noted that rotation techniques are also often applied to the results from a principal components analysis in the hope that they will aid in their interpretability. Although in some cases this may be acceptable, it does have several disadvantages, which are listed by Jolliffe (1989). The main problem is that the defining property of principal components, namely that of accounting for maximal proportions of the total variation in the observed variables, is lost after rotation. 5.8 Estimating factor scores The first stage of an exploratory factor analysis consists of the estimation of the parameters in the model and the rotation of the factors, followed by an (often heroic) attempt to interpret the fitted model. The second stage is concerned with estimating latent variable scores for each individual in the data set; such factor scores are often useful for a number of reasons: 1. They represent a parsimonious summary of the original data possibly use- ful in subsequent analyses (cf. principal component scores in Chapter 3). 2. They are likely to be more reliable than the observed variable values. 3. The factor score is a“pure”measure of a latent variable, while an observed value may be ambiguous because we do not know what combination of latent variables may be represented by that observed value. But the calculation of factor scores is not as straightforward as the calcula- tion of principal component scores. In the original equation defining the factor analysis model, the variables are expressed in terms of the factors, whereas to calculate scores we require the relationship to be in the opposite direction. Bartholomew and Knott (1987) make the point that to talk about “estimat- ing” factor scores is essentially misleading since they are random variables and the issue is really one of prediction. But if we make the assumption of normality, the conditional distribution of f given x can be found. It is N(Λ Σ−1 x, (Λ Ψ−1 Λ + I)−1 ).
- 164. 148 5 Exploratory Factor Analysis Consequently, one plausible way of calculating factor scores would be to use the sample version of the mean of this distribution, namely ˆf = ˆΛ S−1 x, where the vector of scores for an individual, x, is assumed to have mean zero; i.e., sample means for each variable have already been subtracted. Other possible methods for deriving factor scores are described in Rencher (1995), and helpful detailed calculations of several types of factor scores are given in Hershberger (2005). In many respects, the most damaging problem with factor analysis is not the rotational indeterminacy of the loadings but the indeterminacy of the factor scores. 5.9 Two examples of exploratory factor analysis 5.9.1 Expectations of life The data in Table 5.1 show life expectancy in years by country, age, and sex. The data come from Keyfitz and Flieger (1971) and relate to life expectancies in the 1960s. Table 5.1: life data. Life expectancies for different countries by age and gender. m0 m25 m50 m75 w0 w25 w50 w75 Algeria 63 51 30 13 67 54 34 15 Cameroon 34 29 13 5 38 32 17 6 Madagascar 38 30 17 7 38 34 20 7 Mauritius 59 42 20 6 64 46 25 8 Reunion 56 38 18 7 62 46 25 10 Seychelles 62 44 24 7 69 50 28 14 South Africa (C) 50 39 20 7 55 43 23 8 South Africa (W) 65 44 22 7 72 50 27 9 Tunisia 56 46 24 11 63 54 33 19 Canada 69 47 24 8 75 53 29 10 Costa Rica 65 48 26 9 68 50 27 10 Dominican Rep. 64 50 28 11 66 51 29 11 El Salvador 56 44 25 10 61 48 27 12 Greenland 60 44 22 6 65 45 25 9 Grenada 61 45 22 8 65 49 27 10 Guatemala 49 40 22 9 51 41 23 8 Honduras 59 42 22 6 61 43 22 7 Jamaica 63 44 23 8 67 48 26 9 Mexico 59 44 24 8 63 46 25 8
- 165. 5.9 Two examples of exploratory factor analysis 149 Table 5.1: life data (continued). m0 m25 m50 m75 w0 w25 w50 w75 Nicaragua 65 48 28 14 68 51 29 13 Panama 65 48 26 9 67 49 27 10 Trinidad (62) 64 63 21 7 68 47 25 9 Trinidad (67) 64 43 21 6 68 47 24 8 United States (66) 67 45 23 8 74 51 28 10 United States (NW66) 61 40 21 10 67 46 25 11 United States (W66) 68 46 23 8 75 52 29 10 United States (67) 67 45 23 8 74 51 28 10 Argentina 65 46 24 9 71 51 28 10 Chile 59 43 23 10 66 49 27 12 Colombia 58 44 24 9 62 47 25 10 Ecuador 57 46 28 9 60 49 28 11 To begin, we will use the formal test for the number of factors incorporated into the maximum likelihood approach. We can apply this test to the data, assumed to be contained in the data frame life with the country names labelling the rows and variable names as given in Table 5.1, using the following R code: R> sapply(1:3, function(f) + factanal(life, factors = f, method ="mle")$PVAL) objective objective objective 1.880e-24 1.912e-05 4.578e-01 These results suggest that a three-factor solution might be adequate to account for the observed covariances in the data, although it has to be remembered that, with only 31 countries, use of an asymptotic test result may be rather suspect. The three-factor solution is as follows (note that the solution is that resulting from a varimax solution. the default for the factanal() function): R> factanal(life, factors = 3, method ="mle") Call: factanal(x = life, factors = 3, method = "mle") Uniquenesses: m0 m25 m50 m75 w0 w25 w50 w75 0.005 0.362 0.066 0.288 0.005 0.011 0.020 0.146 Loadings: Factor1 Factor2 Factor3
- 166. 150 5 Exploratory Factor Analysis m0 0.964 0.122 0.226 m25 0.646 0.169 0.438 m50 0.430 0.354 0.790 m75 0.525 0.656 w0 0.970 0.217 w25 0.764 0.556 0.310 w50 0.536 0.729 0.401 w75 0.156 0.867 0.280 Factor1 Factor2 Factor3 SS loadings 3.375 2.082 1.640 Proportion Var 0.422 0.260 0.205 Cumulative Var 0.422 0.682 0.887 Test of the hypothesis that 3 factors are sufficient. The chi square statistic is 6.73 on 7 degrees of freedom. The p-value is 0.458 (“Blanks” replace negligible loadings.) Examining the estimated factor load- ings, we see that the first factor is dominated by life expectancy at birth for both males and females; perhaps this factor could be labelled “life force at it “life force amongst the elderly”. The third factor from the varimax rotation has its highest loadings for the life expectancies of men aged 50 and 75 and in the same vein might be labelled “life force for elderly men”. (When labelling factors in this way, factor analysts can often be extremely creative!) The estimated factor scores are found as follows; R> (scores <- factanal(life, factors = 3, method = "mle", + scores = "regression")$scores) Factor1 Factor2 Factor3 Algeria -0.258063 1.90096 1.91582 Cameroon -2.782496 -0.72340 -1.84772 Madagascar -2.806428 -0.81159 -0.01210 Mauritius 0.141005 -0.29028 -0.85862 Reunion -0.196352 0.47430 -1.55046 Seychelles 0.367371 0.82902 -0.55214 South Africa (C) -1.028568 -0.08066 -0.65422 South Africa (W) 0.946194 0.06400 -0.91995 Tunisia -0.862494 3.59177 -0.36442 Canada 1.245304 0.29564 -0.27343 Costa Rica 0.508736 -0.50500 1.01329 Dominican Rep. 0.106044 0.01111 1.83872 El Salvador -0.608156 0.65101 0.48836 Greenland 0.235114 -0.69124 -0.38559 birth”. The second reflects life expectancies at older ages, and we might label
- 167. 5.9 Two examples of exploratory factor analysis 151 Grenada 0.132008 0.25241 -0.15221 Guatemala -1.450336 -0.67766 0.65912 Honduras 0.043253 -1.85176 0.30633 Jamaica 0.462125 -0.51918 0.08033 Mexico -0.052333 -0.72020 0.44418 Nicaragua 0.268974 0.08407 1.70568 Panama 0.442333 -0.73778 1.25219 Trinidad (62) 0.711367 -0.95989 -0.21545 Trinidad (67) 0.787286 -1.10729 -0.51958 United States (66) 1.128331 0.16390 -0.68177 United States (NW66) 0.400059 -0.36230 -0.74299 United States (W66) 1.214345 0.40877 -0.69225 United States (67) 1.128331 0.16390 -0.68177 Argentina 0.731345 0.24812 -0.12818 Chile 0.009752 0.75223 -0.49199 Colombia -0.240603 -0.29544 0.42920 Ecuador -0.723452 0.44246 1.59165 We can use the scores to provide the plot of the data shown in Figure 5.1. Ordering along the first axis reflects life force at birth ranging from Cameroon and Madagascar to countries such as the USA. And on the third axis Algeria is prominent because it has high life expectancy amongst men at higher ages, with Cameroon at the lower end of the scale with a low life expectancy for men over 50. 5.9.2 Drug use by American college students The majority of adult and adolescent Americans regularly use psychoactive substances during an increasing proportion of their lifetimes. Various forms of licit and illicit psychoactive substance use are prevalent, suggesting that patterns of psychoactive substance taking are a major part of the individual’s behavioural repertory and have pervasive implications for the performance of other behaviours. In an investigation of these phenomena, Huba, Wingard, and Bentler (1981) collected data on drug usage rates for 1634 students in the seventh to ninth grades in 11 schools in the greater metropolitan area of Los Angeles. Each participant completed a questionnaire about the number of times a particular substance had ever been used. The substances asked about were as follows: ˆ cigarettes; ˆ beer; ˆ wine; ˆ liquor; ˆ cocaine; ˆ tranquillizers; ˆ drug store medications used to get high;
- 168. 152 5 Exploratory Factor Analysis −2 −1 0 1 −2−10123 Factor 1 Factor2 Alger CamrnMdgsc Marts Reunn Sychl SA(C) SA(W) Tunis Canad CstRc DmRp. ElSlv Grnln Grend Gutml Hndrs Jamac Mexic Nicrg Panam T(62)T(67) US(66 US(NW US(W6 US(67Argnt Chile Colmb Ecudr −2 −1 0 1 −1012 Factor 1 Factor3 Alger Camrn Mdgsc Marts Reunn SychlSA(C) SA(W) Tunis Canad CstRc DmRp. ElSlv Grnln Grend Gutml Hndrs Jamac Mexic Nicrg Panam T(62) T(67) US(66US(NW US(W6US(67 Argnt Chile Colmb Ecudr −2 −1 0 1 2 3 −1012 Factor 2 Factor3 Alger Camrn Mdgsc Marts Reunn SychlSA(C) SA(W) TunisCanad CstRc DmRp. ElSlv Grnln Grend Gutml Hndrs Jamac Mexic Nicrg Panam T(62) T(67) US(66US(NW US(W6US(67 Argnt Chile Colmb Ecudr Fig. 5.1. Individual scatterplots of three factor scores for life expectancy data, with points labelled by abbreviated country names.
- 169. 5.9 Two examples of exploratory factor analysis 153 ˆ heroin and other opiates; ˆ marijuana; ˆ hashish; ˆ inhalants (glue, gasoline, etc.); ˆ hallucinogenics (LSD, mescaline, etc.); ˆ amphetamine stimulants. Responses were recorded on a five-point scale: never tried, only once, a few times, many times, and regularly. The correlations between the usage rates of the 13 substances are shown in Figure 5.2. The plot was produced using the levelplot() function from the package lattice with a somewhat lengthy panel function, so we refer the interested reader to the R code contained in the demo for this chapter (see the Preface for how to access this document). The figure depicts each correlation by an ellipse whose shape tends towards a line with slope 1 for correlations near 1, to a circle for correlations near zero, and to a line with negative slope −1 for negative correlations near −1. In addition, 100 times the correlation coefficient is printed inside the ellipse, and a colourcoding indicates strong negative (dark) to strong positive (light) correlations. We first try to determine the number of factors using the maximum likeli- hood test. The R code for finding the results of the test for number of factors here is: R> sapply(1:6, function(nf) + factanal(covmat = druguse, factors = nf, + method = "mle", n.obs = 1634)$PVAL) objective objective objective objective objective objective 0.000e+00 9.786e-70 7.364e-28 1.795e-11 3.892e-06 9.753e-02 These values suggest that only the six-factor solution provides an adequate fit. The results from the six-factor varimax solution are obtained from R> (factanal(covmat = druguse, factors = 6, + method = "mle", n.obs = 1634)) Call: factanal(factors = 6, covmat = druguse, n.obs = 1634) Uniquenesses: cigarettes beer 0.563 0.368 wine liquor 0.374 0.412 cocaine tranquillizers 0.681 0.522 drug store medication heroin 0.785 0.669
- 170. 154 5 Exploratory Factor Analysis drug store medication cocaine heroin inhalants hallucinogenics tranquillizers amphetamine marijuana hashish cigarettes liquor beer wine drugstoremedication cocaine heroin inhalants hallucinogenics tranquillizers amphetamine marijuana hashish cigarettes liquor beer wine 1002120312322231516 9 121011 21100322728352819301112 7 5 2032100293236311522 8 10 6 7 312729100343239303024262018 23283234100375120371014 9 7 223536323710055323820261514 232831395155100394724292018 151915302032391005351484436 163022303738475310030373224 9 11 8 241020245130100444542 121210261426294837441006058 10 7 6 20 9 15204432456010062 11 5 7 18 7 14183624425862100 −1.0 −0.5 0.0 0.5 1.0 Fig. 5.2. Visualisation of the correlation matrix of drug use. The numbers in the cells correspond to 100 times the correlation coefficient. The color and the shape of the plotting symbols also correspond to the correlation in this cell. marijuana hashish 0.318 0.005 inhalants hallucinogenics 0.541 0.620 amphetamine 0.005 Loadings: Factor1 Factor2 Factor3 Factor4 Factor5 cigarettes 0.494 0.407 beer 0.776 0.112 wine 0.786 liquor 0.720 0.121 0.103 0.115 0.160
- 171. 5.9 Two examples of exploratory factor analysis 155 cocaine 0.519 0.132 tranquillizers 0.130 0.564 0.321 0.105 0.143 drug store medication 0.255 heroin 0.532 0.101 marijuana 0.429 0.158 0.152 0.259 0.609 hashish 0.244 0.276 0.186 0.881 0.194 inhalants 0.166 0.308 0.150 0.140 hallucinogenics 0.387 0.335 0.186 amphetamine 0.151 0.336 0.886 0.145 0.137 Factor6 cigarettes 0.110 beer wine liquor cocaine 0.158 tranquillizers drug store medication 0.372 heroin 0.190 marijuana 0.110 hashish 0.100 inhalants 0.537 hallucinogenics 0.288 amphetamine 0.187 Factor1 Factor2 Factor3 Factor4 Factor5 Factor6 SS loadings 2.301 1.415 1.116 0.964 0.676 0.666 Proportion Var 0.177 0.109 0.086 0.074 0.052 0.051 Cumulative Var 0.177 0.286 0.372 0.446 0.498 0.549 Test of the hypothesis that 6 factors are sufficient. The chi square statistic is 22.41 on 15 degrees of freedom. The p-value is 0.0975 Substances that load highly on the first factor are cigarettes, beer, wine, liquor, and marijuana and we might label it “social/soft drug use”. Cocaine, tranquil- lizers, and heroin load highly on the second factor–the obvious label for the factor is “hard drug use”. Factor three is essentially simply amphetamine use, and factor four hashish use. We will not try to interpret the last two factors, even though the formal test for number of factors indicated that a six-factor solution was necessary. It may be that we should not take the results of the formal test too literally; rather, it may be a better strategy to consider the value of k indicated by the test to be an upper bound on the number of factors with practical importance. Certainly a six-factor solution for a data set with only 13 manifest variables might be regarded as not entirely satisfactory, and clearly we would have some difficulties interpreting all the factors.
- 172. 156 5 Exploratory Factor Analysis One of the problems is that with the large sample size in this example, even small discrepancies between the correlation matrix predicted by a proposed model and the observed correlation matrix may lead to rejection of the model. One way to investigate this possibility is simply to look at the differences between the observed and predicted correlations. We shall do this first for the six-factor model using the following R code: R> pfun <- function(nf) { + fa <- factanal(covmat = druguse, factors = nf, + method = "mle", n.obs = 1634) + est <- tcrossprod(fa$loadings) + diag(fa$uniquenesses) + ret <- round(druguse - est, 3) + colnames(ret) <- rownames(ret) <- + abbreviate(rownames(ret), 3) + ret + } R> pfun(6) cgr ber win lqr ccn trn dsm hrn cgr 0.000 -0.001 0.014 -0.018 0.010 0.001 -0.020 -0.004 ber -0.001 0.000 -0.002 0.004 0.004 -0.011 -0.001 0.007 win 0.014 -0.002 0.000 -0.001 -0.001 -0.005 0.008 0.008 lqr -0.018 0.004 -0.001 0.000 -0.008 0.021 -0.006 -0.018 ccn 0.010 0.004 -0.001 -0.008 0.000 0.000 0.008 0.004 trn 0.001 -0.011 -0.005 0.021 0.000 0.000 0.006 -0.004 dsm -0.020 -0.001 0.008 -0.006 0.008 0.006 0.000 -0.015 hrn -0.004 0.007 0.008 -0.018 0.004 -0.004 -0.015 0.000 mrj 0.001 0.002 -0.004 0.003 -0.004 -0.004 0.008 0.006 hsh 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 inh 0.010 -0.004 -0.007 0.012 -0.003 0.002 0.004 -0.002 hll -0.005 0.005 -0.001 -0.005 -0.008 -0.008 -0.002 0.020 amp 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 mrj hsh inh hll amp cgr 0.001 0 0.010 -0.005 0 ber 0.002 0 -0.004 0.005 0 win -0.004 0 -0.007 -0.001 0 lqr 0.003 0 0.012 -0.005 0 ccn -0.004 0 -0.003 -0.008 0 trn -0.004 0 0.002 -0.008 0 dsm 0.008 0 0.004 -0.002 0 hrn 0.006 0 -0.002 0.020 0 mrj 0.000 0 -0.006 0.003 0 hsh 0.000 0 0.000 0.000 0 inh -0.006 0 0.000 -0.002 0 hll 0.003 0 -0.002 0.000 0 amp 0.000 0 0.000 0.000 0
- 173. 5.10 Factor analysis and principal components analysis compared 157 The differences are all very small, underlining that the six-factor model does describe the data very well. Now let us look at the corresponding matrices for the three- and four-factor solutions found in a similar way in Figure 5.3. Again, in both cases the residuals are all relatively small, suggesting perhaps that use of the formal test for number of factors leads, in this case, to overfitting. The three-factor model appears to provide a perfectly adequate fit for these data. 5.10 Factor analysis and principal components analysis compared Factor analysis, like principal components analysis, is an attempt to explain a set of multivariate data using a smaller number of dimensions than one begins with, but the procedures used to achieve this goal are essentially quite different in the two approaches. Some differences between the two are as follows: ˆ Factor analysis tries to explain the covariances or correlations of the ob- served variables by means of a few common factors. Principal components analysis is primarily concerned with explaining the variance of the observed variables. ˆ If the number of retained components is increased, say from m to m+1, the first m components are unchanged. This is not the case in factor analysis, where there can be substantial changes in all factors if the number of factors is changed. ˆ The calculation of principal component scores is straightforward, but the calculation of factor scores is more complex, and a variety of methods have been suggested. ˆ There is usually no relationship between the principal components of the sample correlation matrix and the sample covariance matrix. For maximum likelihood factor analysis, however, the results of analysing either matrix are essentially equivalent (which is not true of principal factor analysis). Despite these differences, the results from both types of analyses are fre- quently very similar. Certainly, if the specific variances are small, we would expect both forms of analyses to give similar results. However, if the specific variances are large, they will be absorbed into all the principal components, both retained and rejected, whereas factor analysis makes special provision for them. Lastly, it should be remembered that both principal components analy- sis and factor analysis are similar in one important respect–they are both pointless if the observed variables are almost uncorrelated. In this case, factor analysis has nothing to explain and principal components analysis will simply lead to components that are similar to the original variables.
- 174. 158 5 Exploratory Factor Analysis R>pfun(3) cgrberwinlqrccntrndsmhrnmrjhshinhhllamp cgr0.000-0.0010.009-0.0130.0110.009-0.011-0.0040.003-0.0270.039-0.0170.002 ber-0.0010.000-0.0020.0020.002-0.0140.0000.005-0.0010.019-0.0020.009-0.007 win0.009-0.0020.0000.000-0.002-0.0040.0120.0130.001-0.017-0.0070.0040.002 lqr-0.0130.0020.0000.000-0.0080.024-0.017-0.020-0.0010.014-0.002-0.0150.006 ccn0.0110.002-0.002-0.0080.0000.0310.0380.082-0.0020.0410.023-0.030-0.075 trn0.009-0.014-0.0040.0240.0310.000-0.0210.026-0.002-0.016-0.038-0.0580.044 dsm-0.0110.0000.012-0.0170.038-0.0210.0000.0210.007-0.0400.1130.000-0.038 hrn-0.0040.0050.013-0.0200.0820.0260.0210.0000.006-0.0350.031-0.005-0.049 mrj0.003-0.0010.001-0.001-0.002-0.0020.0070.0060.0000.0010.003-0.002-0.002 hsh-0.0270.019-0.0170.0140.041-0.016-0.040-0.0350.0010.000-0.0350.0340.010 inh0.039-0.002-0.007-0.0020.023-0.0380.1130.0310.003-0.0350.0000.007-0.015 hll-0.0170.0090.004-0.015-0.030-0.0580.000-0.005-0.0020.0340.0070.0000.041 amp0.002-0.0070.0020.006-0.0750.044-0.038-0.049-0.0020.010-0.0150.0410.000 R>pfun(4) cgrberwinlqrccntrndsmhrnmrjhshinhhllamp cgr0.000-0.0010.008-0.0120.0090.008-0.015-0.0070.001-0.0230.037-0.0200.000 ber-0.0010.000-0.0010.0010.000-0.016-0.0020.003-0.0010.018-0.0050.0060.000 win0.008-0.0010.0000.000-0.001-0.0050.0120.0140.001-0.020-0.0080.0010.000 lqr-0.0120.0010.0000.000-0.0040.029-0.015-0.015-0.0010.0180.001-0.010-0.001 ccn0.0090.000-0.001-0.0040.0000.024-0.0140.007-0.0030.035-0.022-0.0280.000 trn0.008-0.016-0.0050.0290.0240.000-0.0200.027-0.0010.001-0.032-0.0280.001 dsm-0.015-0.0020.012-0.015-0.014-0.0200.000-0.0180.003-0.0420.0900.0080.000 hrn-0.0070.0030.014-0.0150.0070.027-0.0180.0000.003-0.037-0.0010.0050.000 mrj0.001-0.0010.001-0.001-0.003-0.0010.0030.0030.0000.0000.001-0.0020.000 hsh-0.0230.018-0.0200.0180.0350.001-0.042-0.0370.0000.000-0.0310.055-0.001 inh0.037-0.005-0.0080.001-0.022-0.0320.090-0.0010.001-0.0310.0000.0210.000 hll-0.0200.0060.001-0.010-0.028-0.0280.0080.005-0.0020.0550.0210.0000.000 amp0.0000.0000.000-0.0010.0000.0010.0000.0000.000-0.0010.0000.0000.000 Fig.5.3.Differencesbetweenthree-andfour-factorsolutionsandactualcorrelationmatrixforthedrugusedata.
- 175. 5.12 Exercises 159 5.11 Summary Factor analysis has probably attracted more critical comments than any other statistical technique. Hills (1977), for example, has gone so far as to suggest that factor analysis is not worth the time necessary to understand it and carry it out. And Chatfield and Collins (1980) recommend that factor analysis should not be used in most practical situations. The reasons for such an openly sceptical view about factor analysis arise first from the central role of latent variables in the factor analysis model and second from the lack of uniqueness of the factor loadings in the model, which gives rise to the possibility of rotating factors. It certainly is the case that, since the common factors cannot be measured or observed, the existence of these hypothetical variables is open to question. A factor is a construct operationally defined by its factor loadings, and overly enthusiastic reification is not recommended. And it is the case that, given one factor loading matrix, there are an infinite number of factor loading matrices that could equally well (or equally badly) account for the variances and covariances of the manifest variables. Rotation methods are designed to find an easily interpretable solution from among this infinitely large set of alternatives by finding a solution that exhibits the best simple structure. Factor analysis can be a useful tool for investigating particular features of the structure of multivariate data. Of course, like many models used in data analysis, the one used in factor analysis may be only a very idealised approx- imation to the truth. Such an approximation may, however, prove a valuable starting point for further investigations, particularly for the confirmatory fac- tor analysis models that are the subject of Chapter 7. For exploratory factor analysis, similar comments apply about the size of n and q needed to get convincing results, such as those given in Chapter 3 for principal components analysis. And the maximum likelihood method for the estimation of factor loading and specific variances used in this chapter is only suitable for data having a multivariate normal distribution (or at least a reasonable approximation to such a distribution). Consequently, for the factor analysis of, in particular, binary variables, special methods are needed; see, for example, Muthen (1978). 5.12 Exercises Ex. 5.1 Show how the result Σ = ΛΛ + Ψ arises from the assumptions of uncorrelated factors, independence of the specific variates, and indepen- dence of common factors and specific variances. What form does Σ take if the factors are allowed to be correlated? Ex. 5.2 Show that the communalities in a factor analysis model are unaffected by the transformation Λ∗ = ΛM. Ex. 5.3 Give a formula for the proportion of variance explained by the jth factor estimated by the principal factor approach.
- 176. 160 5 Exploratory Factor Analysis Ex. 5.4 Apply the factor analysis model separately to the life expectancies of men and women and compare the results. Ex. 5.5 The correlation matrix given below arises from the scores of 220 boys in six school subjects: (1) French, (2) English, (3) History, (4) Arithmetic, (5) Algebra, and (6) Geometry. Find the two-factor solution from a max- imum likelihood factor analysis. By plotting the derived loadings, find an orthogonal rotation that allows easier interpretation of the results. R = French English History Arithmetic Algebra Geometry 1.00 0.44 1.00 0.41 0.35 1.00 0.29 0.35 0.16 1.00 0.33 0.32 0.19 0.59 1.00 0.25 0.33 0.18 0.47 0.46 1.00 . Ex. 5.6 The matrix below shows the correlations between ratings on nine statements about pain made by 123 people suffering from extreme pain. Each statement was scored on a scale from 1 to 6, ranging from agreement to disagreement. The nine pain statements were as follows: 1. Whether or not I am in pain in the future depends on the skills of the doctors. 2. Whenever I am in pain, it is usually because of something I have done or not done, 3. Whether or not I am in pain depends on what the doctors do for me. 4. I cannot get any help for my pain unless I go to seek medical advice. 5. When I am in pain I know that it is because I have not been taking proper exercise or eating the right food. 6. People’s pain results from their own carelessness. 7. I am directly responsible for my pain, 8. relief from pain is chiefly controlled by the doctors. 9. People who are never in pain are just plain lucky. 1.00 −0.04 1.00 0.61 −0.07 1.00 0.45 −0.12 0.59 1.00 0.03 0.49 0.03 −0.08 1.00 −0.29 0.43 −0.13 −0.21 0.47 1.00 −0.30 0.30 −0.24 −0.19 0.41 0.63 1.00 0.45 −0.31 0.59 0.63 −0.14 −0.13 −0.26 1.00 0.30 −0.17 0.32 0.37 −0.24 −0.15 −0.29 0.40 1.00 . (a) Perform a principal components analysis on these data, and exam- ine the associated scree plot to decide on the appropriate number of components. (b) Apply maximum likelihood factor analysis, and use the test described in the chapter to select the necessary number of common factors.
- 177. 5.12 Exercises 161 (c) Rotate the factor solution selected using both an orthogonal and an oblique procedure, and interpret the results.
- 179. 6 Cluster Analysis 6.1 Introduction An intelligent being cannot treat every object it sees as a unique entity unlike anything else in the universe. It has to put objects in categories so that it may apply its hard-won knowledge about similar objects encountered in the past to the object at hand (Pinker 1997). One of the most basic abilities of living creatures involves the grouping of similar objects to produce a classification. The idea of sorting similar things into categories is clearly a primitive one because early humans, for example, must have been able to realise that many individual objects shared certain properties such as being edible, or poisonous, or ferocious, and so on. And classification in its widest sense is needed for the development of language, which consists of words that help us to recognise and discuss the different types of events, objects, and people we encounter. Each noun in a language, for example, is essentially a label used to describe a class of things that have striking features in common; thus animals are called cats, dogs, horses, etc., and each name collects individuals into groups. Naming and classifying are essentially synonymous. As well as being a basic human conceptual activity, classification of the phenomena being studied is an important component of virtually all scientific research. In the behavioural sciences, for example, these “phenomena” may be individuals or societies, or even patterns of behaviour or perception. The investigator is usually interested in finding a classification in which the items of interest are sorted into a small number of homogeneous groups or clusters, the terms being synonymous. Most commonly the required classification is one in which the groups are mutually exclusive (an item belongs to a single group) rather than overlapping (items can be members of more than one group). At the very least, any derived classification scheme should provide a convenient method of organizing a large, complex set of multivariate data, with the class labels providing a parsimonious way of describing the patterns of similarities DOI 10.1007/978-1-4419-9650-3_6, © Springer Science+Business Media, LLC 2011 163B. Everitt and T. Hothorn, An Introduction to Applied Multivariate Analysis with R: Use R!,
- 180. 164 6 Cluster Analysis and differences in the data. In market research, for example, it might be useful to group a large number of potential customers according to their needs in a particular product area. Advertising campaigns might then be tailored for the different types of consumers as represented by the different groups. But often a classification may seek to serve a more fundamental purpose. In psychiatry, for example, the classification of psychiatric patients with different symptom profiles into clusters might help in the search for the causes of mental illnesses and perhaps even lead to improved therapeutic methods. And these twin aims of prediction (separating diseases that require different treatments) and aetiology (searching for the causes of disease) for classifications will be the same in other branches of medicine. Clearly, a variety of classifications will always be possible for whatever is being classified. Human beings could, for example, be classified with respect to economic status into groups labelled lower class, middle class, and upper class or they might be classified by annual consumption of alcohol into low, medium, and high. Clearly, different classifications may not collect the same set of individuals into groups, but some classifications will be more useful than others, a point made clearly by the following extract from Needham (1965) in which he considers the classification of human beings into men and women: The usefulness of this classification does not begin and end with all that can, in one sense, be strictly inferred from it–namely a statement about sexual organs. It is a very useful classification because classify- ing a person as man or woman conveys a great deal more information, about probable relative size, strength, certain types of dexterity and so on. When we say that persons in class man are more suitable than persons in class woman for certain tasks and conversely, we are only incidentally making a remark about sex, our primary concern being with strength, endurance, etc. The point is that we have been able to use a classification of persons which conveys information on many properties. On the contrary a classification of persons into those with hairs on their forearms between 3/16 and 1 4 inch long and those with- out, though it may serve some particular use, is certainly of no general use, for imputing membership in the former class to a person conveys information on this property alone. Put another way, there are no known properties which divide up a set of people in a similar way. In a similar vein, a classification of books based on subject matter into classes such as dictionaries, novels, biographies, and so on is likely to be far more useful than one based on, say, the colour of the book’s binding. Such examples illustrate that any classification of a set of multivariate data is likely to be judged on its usefulness.
- 181. 6.2 Cluster analysis 165 6.2 Cluster analysis Cluster analysis is a generic term for a wide range of numerical methods with the common goal of uncovering or discovering groups or clusters of obser- vations that are homogeneous and separated from other groups. Clustering techniques essentially try to formalise what human observers do so well in two or three dimensions. Consider, for example, the scatterplot shown in Fig- ure 6.1. The conclusion that there are three natural groups or clusters of dots is reached with no conscious effort or thought. Clusters are identified by the assessment of the relative distances between points, and in this example the relative homogeneity of each cluster and the degree of their separation makes the task very simple. The examination of scatterplots based either on the orig- inal data or perhaps on the first few principal component scores of the data is often a very helpful initial phase when intending to apply some form of cluster analysis to a set of multivariate data. Cluster analysis techniques are described in detail in Gordon (1987, 1999) in and Everitt, Landau, Leese, and Stahl (2011). In this chapter, we give a relatively brief account of three types of clustering methods: agglomerative hierarchical techniques, k-means clustering, and model-based clustering. q q q q q q q q q qq q q q q q q q q q q q q q q q q q qq q q q q q q q q qq q q q q q q q q qq q q q q q 5 10 15 20 0246810 x1 x2 Fig. 6.1. Bivariate data showing the presence of three clusters.
- 182. 166 6 Cluster Analysis 6.3 Agglomerative hierarchical clustering This class of clustering methods produces a hierarchical classification of data. In a hierarchical classification, the data are not partitioned into a particular number of classes or groups at a single step. Instead the classification consists of a series of partitions that may run from a single “cluster” containing all individuals to n clusters, each containing a single individual. Agglomerative hierarchical clustering techniques produce partitions by a series of successive fusions of the n individuals into groups. With such methods, fusions, once made, are irreversible, so that when an agglomerative algorithm has placed two individuals in the same group they cannot subsequently appear in different groups. Since all agglomerative hierarchical techniques ultimately reduce the data to a single cluster containing all the individuals, the investigator seeking the solution with the best-fitting number of clusters will need to decide which division to choose. The problem of deciding on the“correct”number of clusters will be taken up later. An agglomerative hierarchical clustering procedure produces a series of partitions of the data, Pn, Pn−1, . . . , P1. The first, Pn, consists of n single- member clusters, and the last, P1, consists of a single group containing all n individuals. The basic operation of all methods is similar: (START) Clusters C1, C2, . . . , Cn each containing a single individual. (1) Find the nearest pair of distinct clusters, say Ci and Cj, merge Ci and Cj, delete Cj, and decrease the number of clusters by one. (2) If the number of clusters equals one, then stop; otherwise return to 1. But before the process can begin, an inter-individual distance matrix or similarity matrix needs to be calculated. There are many ways to calculate distances or similarities between pairs of individuals, but here we only deal with a commonly used distance measure, Euclidean distance, which was de- fined in Chapter 1 but as a reminder is calculated as dij = q k=1 (xik − xjk)2, where dij is the Euclidean distance between individual i with variable values xi1, xi2, . . . , xiq and individual j with variable values xj1, xj2, . . . , xjq. (De- tails of other possible distance measures and similarity measures are given in Everitt et al. 2011). The Euclidean distances between each pair of individuals can be arranged in a matrix that is symmetric because dij = dji and has zeros on the main diagonal. Such a matrix is the starting point of many clustering examples, although the calculation of Euclidean distances from the raw data may not be sensible when the variables are on very different scales. In such cases, the variables can be standardised in the usual way before calculating the distance matrix, although this can be unsatisfactory in some cases (see Everitt et al. 2011, for details).
- 183. 6.3 Agglomerative hierarchical clustering 167 Given an inter-individual distance matrix, the hierarchical clustering can begin, and at each stage in the process the methods fuse individuals or groups of individuals formed earlier that are closest (or most similar). So as groups are formed, the distance between an individual and a group containing several individuals and the distance between two groups of individuals will need to be calculated. How such distances are defined leads to a variety of different techniques. Two simple inter-group measures are dAB = min i∈A i∈B (dij), dAB = max i∈A i∈B (dij), where dAB is the distance between two clusters A and B, and dij is the distance between individuals i and j found from the initial inter-individual distance matrix. The first inter-group distance measure above is the basis of single linkage clustering, the second that of complete linkage clustering. Both these tech- niques have the desirable property that they are invariant under monotone transformations of the original inter-individual distances; i.e., they only de- pend on the ranking on these distances, not their actual values. A further possibility for measuring inter-cluster distance or dissimilarity is dAB = 1 nAnB i∈Ai∈B dij, where nA and nB are the numbers of individuals in clusters A and B. This measure is the basis of a commonly used procedure known as group average clustering. All three inter-group measures described above are illustrated in Figure 6.2. Hierarchical classifications may be represented by a two-dimensional di- agram known as a dendrogram, which illustrates the fusions made at each stage of the analysis. An example of such a diagram is given in Figure 6.3. The structure of Figure 6.3 resembles an evolutionary tree, a concept intro- duced by Darwin under the term “Tree of Life” in his book On the Origin of Species by Natural Selection in 1859, and it is in biological applications that hierarchical classifications are most relevant and most justified (although this type of clustering has also been used in many other areas). As a first example of the application of the three clustering methods, single linkage, complete linkage, and group average, each will be applied to the chest, waist, and hip measurements of 20 individuals given in Chapter 1, Table 1.2. First Euclidean distances are calculated on the unstandardised measurements using the following R code:
- 184. 168 6 Cluster Analysis Fig. 6.2. Inter-cluster distance measures. Fig. 6.3. Darwin’s Tree of Life.
- 185. 6.3 Agglomerative hierarchical clustering 169 R> (dm <- dist(measure[, c("chest", "waist", "hips")])) 1 2 3 4 5 6 7 8 9 10 2 6.16 3 5.66 2.45 4 7.87 2.45 4.69 5 4.24 5.10 3.16 7.48 6 11.00 6.08 5.74 7.14 7.68 7 12.04 5.92 7.00 5.00 10.05 5.10 8 8.94 3.74 4.00 3.74 7.07 5.74 4.12 9 7.81 3.61 2.24 5.39 4.58 3.74 5.83 3.61 10 10.10 4.47 4.69 5.10 7.35 2.24 3.32 3.74 3.00 11 7.00 8.31 6.40 9.85 5.74 11.05 12.08 8.06 7.48 10.25 12 7.35 7.07 5.48 8.25 6.00 9.95 10.25 6.16 6.40 8.83 13 7.81 8.54 7.28 9.43 7.55 12.08 11.92 7.81 8.49 10.82 14 8.31 11.18 9.64 12.45 8.66 14.70 15.30 11.18 11.05 13.75 15 7.48 6.16 4.90 7.07 6.16 9.22 9.00 4.90 5.74 7.87 16 7.07 6.00 4.24 7.35 5.10 8.54 9.11 5.10 5.00 7.48 17 7.81 7.68 6.71 8.31 7.55 11.40 10.77 6.71 7.87 9.95 18 6.71 6.08 4.58 7.28 5.39 9.27 9.49 5.39 5.66 8.06 19 9.17 5.10 4.47 5.48 7.07 6.71 5.74 2.00 4.12 5.10 20 7.68 9.43 7.68 10.82 7.00 12.41 13.19 9.11 8.83 11.53 11 12 13 14 15 16 17 18 19 2 3 4 5 6 7 8 9 10 11 12 2.24 13 2.83 2.24 14 3.74 5.20 3.74 15 3.61 1.41 3.00 6.40 16 3.00 1.41 3.61 6.40 1.41 17 3.74 2.24 1.41 5.10 2.24 3.32 18 2.83 1.00 2.83 5.83 1.00 1.00 2.45 19 6.71 4.69 6.40 9.85 3.46 3.74 5.39 4.12 20 1.41 3.00 2.45 2.45 4.36 4.12 3.74 3.74 7.68 Application of each of the three clustering methods described earlier to the distance matrix and a plot of the corresponding dendrogram are achieved using the hclust() function:
- 186. 170 6 Cluster Analysis R> plot(cs <- hclust(dm, method = "single")) R> plot(cc <- hclust(dm, method = "complete")) R> plot(ca <- hclust(dm, method = "average")) The resulting plots (for single, complete, and average linkage) are given in the upper part of Figure 6.4. 1756104239819141317161512181120 1.02.03.04.0 Single hclust (*, "single") dm Height −4 0 4 −42 PC1 PC2 1 2 2 2 2 22 22 2 2 2 2 2 22 2 22 2 7610248193915161512181317141120 051015 Complete hclust (*, "complete") dm Height −4 0 4 −42 PC1 PC2 1 2 2 2 1 22 22 2 1 1 1 1 11 1 12 1 7610539819241161512181317141120 02468 Average hclust (*, "average") dm Height −4 0 4 −42 PC1 PC2 1 2 2 2 2 22 22 2 1 1 1 1 11 1 12 1 Fig. 6.4. Cluster solutions for measure data. The top row gives the cluster dendro- grams along with the cutoff used to derive the classes presented (in the space of the first two principal components) in the bottom row. We now need to consider how we select specific partitions of the data (i.e., a solution with a particular number of groups) from these dendrograms. The answer is that we “cut” the dendrogram at some height and this will give a partition with a particular number of groups. How do we choose where to cut or, in other words, how do we decide on a particular number of groups that is, in some sense, optimal for the data? This is a more difficult question to answer.
- 187. 6.3 Agglomerative hierarchical clustering 171 One informal approach is to examine the sizes of the changes in height in the dendrogram and take a “large” change to indicate the appropriate number of clusters for the data. (More formal approaches are described in Everitt et al. 2011) Even using this informal approach on the dendrograms in Figure 6.4, it is not easy to decide where to “cut”. So instead, because we know that these data consist of measurements on ten men and ten women, we will look at the two-group solutions from each method that are obtained by cutting the dendrograms at suitable heights. We can display and compare the three solutions graphically by plotting the first two principal component scores of the data, labelling the points to identify the cluster solution of one of the methods by using the following code: R> body_pc <- princomp(dm, cor = TRUE) R> xlim <- range(body_pc$scores[,1]) R> plot(body_pc$scores[,1:2], type = "n", + xlim = xlim, ylim = xlim) R> lab <- cutree(cs, h = 3.8) R> text(body_pc$scores[,1:2], labels = lab, cex = 0.6) The resulting plots are shown in the lower part of Figure 6.4. The plots of dendrograms and principal components scatterplots are combined into a single diagram using the layout() function (see the chapter demo for the complete R code). The plot associated with the single linkage solution immediately demonstrates one of the problems with using this method in practise, and that is a phenomenon known as chaining, which refers to the tendency to incorporate intermediate points between clusters into an existing cluster rather than initiating a new one. As a result, single linkage solutions often contain long “straggly” clusters that do not give a useful description of the data. The two-group solutions from complete linkage and average linkage, also shown in Figure 6.4, are similar and in essence place the men (observations 1 to 10) together in one cluster and the women (observations 11 to 20) in the other. 6.3.1 Clustering jet fighters The data shown in Table 6.1 as originally given in Stanley and Miller (1979) and also in Hand et al. (1994) are the values of six variables for 22 US fighter aircraft. The variables are as follows: FFD: first flight date, in months after January 1940; SPR: specific power, proportional to power per unit weight; RGF: flight range factor; PLF: payload as a fraction of gross weight of aircraft; SLF: sustained load factor; CAR: a binary variable that takes the value 1 if the aircraft can land on a carrier and 0 otherwise.
- 188. 172 6 Cluster Analysis Table 6.1: jet data. Jet fighters data. FFD SPR RGF PLF SLF CAR 82 1.468 3.30 0.166 0.10 no 89 1.605 3.64 0.154 0.10 no 101 2.168 4.87 0.177 2.90 yes 107 2.054 4.72 0.275 1.10 no 115 2.467 4.11 0.298 1.00 yes 122 1.294 3.75 0.150 0.90 no 127 2.183 3.97 0.000 2.40 yes 137 2.426 4.65 0.117 1.80 no 147 2.607 3.84 0.155 2.30 no 166 4.567 4.92 0.138 3.20 yes 174 4.588 3.82 0.249 3.50 no 175 3.618 4.32 0.143 2.80 no 177 5.855 4.53 0.172 2.50 yes 184 2.898 4.48 0.178 3.00 no 187 3.880 5.39 0.101 3.00 yes 189 0.455 4.99 0.008 2.64 no 194 8.088 4.50 0.251 2.70 yes 197 6.502 5.20 0.366 2.90 yes 201 6.081 5.65 0.106 2.90 yes 204 7.105 5.40 0.089 3.20 yes 255 8.548 4.20 0.222 2.90 no 328 6.321 6.45 0.187 2.00 yes We shall apply complete linkage to the data but using only variables two to five. And given that the variables are on very different scales, we will standardise them to unit variance before clustering. The required R code for standardisation and clustering is as follows: R> X <- scale(jet[, c("SPR", "RGF", "PLF", "SLF")], + center = FALSE, scale = TRUE) R> dj <- dist(X) R> plot(cc <- hclust(dj), main = "Jets clustering") R> cc Call: hclust(d = dj) Cluster method : complete Distance : euclidean Number of objects: 22
- 189. 6.3 Agglomerative hierarchical clustering 173 The resulting dendrogram in Figure 6.5 strongly suggests the presence of two groups of fighters. In Figure 6.6, the data are plotted in the space of the first two principal components of the correlation matrix of the relevant variables (SPR to SLF). And in Figure 6.6 the points are labelled by cluster number for the two-group solution and the colours used are the values of the CAR variable. The two-group solution largely corresponds to planes that can and cannot land on a carrier. Call: hclust(d = dj) Cluster method : complete Distance : euclidean Number of objects: 22 YF−107A F−106A F−100A F−102A F−104B F−4B F−101A F−111A F4D−1 F−105B F9F−2 F−94A F3D−1 FH−1 FJ−1 F−89A F−8A XF10F−1 F−86A F3H−2 F9F−6 F11F−1 0.00.51.01.52.0 Jets clustering hclust (*, "complete") dj Height Fig. 6.5. Hierarchical clustering (complete linkage) of jet data.
- 190. 174 6 Cluster Analysis Again, we cut the dendrogram in such a way that two clusters remain and plot the corresponding classes in the space of the first two principal compo- nents; see Figure 6.6. R> pr <- prcomp(dj)$x[, 1:2] R> plot(pr, pch = (1:2)[cutree(cc, k = 2)], + col = c("black", "darkgrey")[jet$CAR], + xlim = range(pr) * c(1, 1.5)) R> legend("topright", col = c("black", "black", + "darkgrey", "darkgrey"), + legend = c("1 / no", "2 / no", "1 / yes", "2 / yes"), + pch = c(1:2, 1:2), title = "Cluster / CAR", bty = "n") q q q q q q q q q q q q −2 −1 0 1 2 3 −2−101 PC1 PC2 q q Cluster / CAR 1 / no 2 / no 1 / yes 2 / yes Fig. 6.6. Hierarchical clustering (complete linkage) of jet data plotted in PCA space.
- 191. 6.4 K-means clustering 175 6.4 K-means clustering The k-means clustering technique seeks to partition the n individuals in a set of multivariate data into k groups or clusters, (G1, G2, . . . , Gk), where Gi denotes the set of ni individuals in the ith group, and k is given (or a possible range is specified by the researcher–the problem of choosing the“true”value of k will be taken up later) by minimising some numerical criterion, low values of which are considered indicative of a“good”solution. The most commonly used implementation of k-means clustering is one that tries to find the partition of the n individuals into k groups that minimises the within-group sum of squares (WGSS) over all variables; explicitly, this criterion is WGSS = q j=1 k l=1 i∈Gl (xij − x (l) j )2 , where x (l) j = 1 ni i∈Gl xij is the mean of the individuals in group Gl on variable j. The problem then appears relatively simple; namely, consider every pos- sible partition of the n individuals into k groups, and select the one with the lowest within-group sum of squares. Unfortunately, the problem in practise is not so straightforward. The numbers involved are so vast that complete enu- meration of every possible partition remains impossible even with the fastest computer. The scale of the problem immediately becomes clear by looking at the numbers in Table 6.2. Table 6.2: Number of possible partitions depending on the sample size n and number of clusters k. n k Number of possible partitions 15 3 2, 375, 101 20 4 45, 232, 115, 901 25 8 690, 223, 721, 118, 368, 580 100 5 1068 The impracticability of examining every possible partition has led to the development of algorithms designed to search for the minimum values of the clustering criterion by rearranging existing partitions and keeping the new one only if it provides an improvement. Such algorithms do not, of course, guarantee finding the global minimum of the criterion. The essential steps in these algorithms are as follows:
- 192. 176 6 Cluster Analysis 1. Find some initial partition of the individuals into the required number of groups. (Such an initial partition could be provided by a solution from one of the hierarchical clustering techniques described in the previous section.) 2. Calculate the change in the clustering criterion produced by“moving”each individual from its own cluster to another cluster. 3. Make the change that leads to the greatest improvement in the value of the clustering criterion. 4. Repeat steps (2) and (3) until no move of an individual causes the clus- tering criterion to improve. (For a more detailed account of the typical k-means algorithm see Steinley 2008) The k-means approach to clustering using the minimisation of the within- group sum of squares over all the variables is widely used but suffers from the two problems of (1) not being scale-invariant (i.e., different solutions may result from clustering the raw data and the data standardised in some way) and (2) of imposing a “spherical” structure on the data; i.e., it will find clus- ters shaped like hyper-footballs even if the “true” clusters in the data are of some other shape (see Everitt et al. 2011, for some examples of the latter phenomenon). Nevertheless, the k-means method remains very popular. With k-means clustering, the investigator can choose to partition the data into a specified number of groups. In practice, solutions for a range of values for the number of groups are found and in some way the optimal or “true” number of groups for the data must be chosen. Several suggestions have been made as to how to answer the number of groups question, but none is completely satisfactory. The method we shall use in the forthcoming example is to plot the within-groups sum of squares associated with the k-means solution for each number of groups. As the number of groups increases the sum of squares will necessarily decrease, but an obvious “elbow” in the plot may be indicative of the most useful solution for the investigator to look at in detail. (Compare this with the scree plot described in Chapter 3.) 6.4.1 Clustering the states of the USA on the basis of their crime rate profiles The Statistical Abstract of the USA (Anonymous 1988, Table 265) gives rates of different types of crime per 100,000 residents of the 50 states of the USA plus the District of Columbia for the year 1986. The data are given in Table 6.3. Table 6.3: crime data. Crime data. Murder Rape Robbery Assault Burglary Theft Vehicle ME 2.0 14.8 28 102 803 2347 164 NH 2.2 21.5 24 92 755 2208 228 VT 2.0 21.8 22 103 949 2697 181 MA 3.6 29.7 193 331 1071 2189 906
- 193. 6.4 K-means clustering 177 Table 6.3: crime data (continued). Murder Rape Robbery Assault Burglary Theft Vehicle RI 3.5 21.4 119 192 1294 2568 705 CT 4.6 23.8 192 205 1198 2758 447 NY 10.7 30.5 514 431 1221 2924 637 NJ 5.2 33.2 269 265 1071 2822 776 PA 5.5 25.1 152 176 735 1654 354 OH 5.5 38.6 142 235 988 2574 376 IN 6.0 25.9 90 186 887 2333 328 IL 8.9 32.4 325 434 1180 2938 628 MI 11.3 67.4 301 424 1509 3378 800 WI 3.1 20.1 73 162 783 2802 254 MN 2.5 31.8 102 148 1004 2785 288 IA 1.8 12.5 42 179 956 2801 158 MO 9.2 29.2 170 370 1136 2500 439 ND 1.0 11.6 7 32 385 2049 120 SD 4.0 17.7 16 87 554 1939 99 NE 3.1 24.6 51 184 748 2677 168 KS 4.4 32.9 80 252 1188 3008 258 DE 4.9 56.9 124 241 1042 3090 272 MD 9.0 43.6 304 476 1296 2978 545 DC 31.0 52.4 754 668 1728 4131 975 VA 7.1 26.5 106 167 813 2522 219 WV 5.9 18.9 41 99 625 1358 169 NC 8.1 26.4 88 354 1225 2423 208 SC 8.6 41.3 99 525 1340 2846 277 GA 11.2 43.9 214 319 1453 2984 430 FL 11.7 52.7 367 605 2221 4373 598 KY 6.7 23.1 83 222 824 1740 193 TN 10.4 47.0 208 274 1325 2126 544 AL 10.1 28.4 112 408 1159 2304 267 MS 11.2 25.8 65 172 1076 1845 150 AR 8.1 28.9 80 278 1030 2305 195 LA 12.8 40.1 224 482 1461 3417 442 OK 8.1 36.4 107 285 1787 3142 649 TX 13.5 51.6 240 354 2049 3987 714 MT 2.9 17.3 20 118 783 3314 215 ID 3.2 20.0 21 178 1003 2800 181 WY 5.3 21.9 22 243 817 3078 169 CO 7.0 42.3 145 329 1792 4231 486 NM 11.5 46.9 130 538 1845 3712 343 AZ 9.3 43.0 169 437 1908 4337 419 UT 3.2 25.3 59 180 915 4074 223 NV 12.6 64.9 287 354 1604 3489 478
- 194. 178 6 Cluster Analysis Table 6.3: crime data (continued). Murder Rape Robbery Assault Burglary Theft Vehicle WA 5.0 53.4 135 244 1861 4267 315 OR 6.6 51.1 206 286 1967 4163 402 CA 11.3 44.9 343 521 1696 3384 762 AK 8.6 72.7 88 401 1162 3910 604 HI 4.8 31.0 106 103 1339 3759 328 Murder 10 50 100 600 1500 020 1050 Rape Robbery 0600 100600 Assault Burglary 500 1500 Theft 0 20 0 600 500 200 1000 2001000 Vehicle Fig. 6.7. Scatterplot matrix of crime data. To begin, let’s look at the scatterplot matrix of the data shown in Fig- ure 6.7. The plot suggests that at least one of the cities is considerably dif- ferent from the others in its murder rate at least. The city is easily identified using
- 195. 6.4 K-means clustering 179 R> subset(crime, Murder > 15) Murder Rape Robbery Assault Burglary Theft Vehicle DC 31 52.4 754 668 1728 4131 975 i.e., the murder rate is very high in the District of Columbia. In order to check if the other crime rates are also higher in DC, we label the corresponding points in the scatterplot matrix in Figure 6.8. Clearly, DC is rather extreme in most crimes (the clear message is don’t live in DC). Murder 10 50 + + 100 600 + + 1500 + 020 + 1050 + Rape + + + + + + + Robbery + + + 0600 + 100600 + + + Assault + + + + + + + Burglary + 500 + 1500 + + + + + Theft + 0 20 + + 0 600 + + 500 + + 200 1000 2001000 Vehicle Fig. 6.8. Scatterplot matrix of crime data with DC observation labelled using a plus sign. We will now apply k-means clustering to the crime rate data after removing the outlier, DC. If we first calculate the variances of the crime rates for the different types of crimes we find the following: R> sapply(crime, var)
- 196. 180 6 Cluster Analysis Murder Rape Robbery Assault Burglary Theft Vehicle 23.2 212.3 18993.4 22004.3 177912.8 582812.8 50007.4 The variances are very different, and using k-means on the raw data would not be sensible; we must standardise the data in some way, and here we standardise each variable by its range. After such standardisation, the variances become R> rge <- sapply(crime, function(x) diff(range(x))) R> crime_s <- sweep(crime, 2, rge, FUN = "/") R> sapply(crime_s, var) Murder Rape Robbery Assault Burglary Theft Vehicle 0.02578 0.05687 0.03404 0.05440 0.05278 0.06411 0.06517 The variances of the standardised data are very similar, and we can now progress with clustering the data. First we plot the within-groups sum of squares for one- to six-group solutions to see if we can get any indication of the number of groups. The plot is shown in Figure 6.9. The only“elbow”in the plot occurs for two groups, and so we will now look at the two-group solution. The group means for two groups are computed by R> kmeans(crime_s, centers = 2)$centers * rge Murder Rape Robbery Assault Burglary Theft Vehicle 1 4.893 305.1 189.6 259.70 31.0 540.5 873.0 2 21.098 483.3 1031.4 19.26 638.9 2096.1 578.6 A plot of the two-group solution in the space of the first two principal compo- nents of the correlation matrix of the data is shown in Figure 6.10. The two groups are created essentially on the basis of the first principal component score, which is a weighted average of the crime rates. Perhaps all the cluster analysis is doing here is dividing into two parts a homogenous set of data. This is always a possibility, as is discussed in some detail in Everitt et al. (2011). 6.4.2 Clustering Romano-British pottery The second application of k-means clustering will be to the data on Romano- British pottery given in Chapter 1. We begin by computing the Euclidean distance matrix for the standardised measurements of the 45 pots. The result- ing 45 × 45 matrix can be inspected graphically by using an image plot, here obtained with the function levelplot available in the package lattice (Sarkar 2010, 2008). Such a plot associates each cell of the dissimilarity matrix with a colour or a grey value. We choose a very dark grey for cells with distance zero (i.e., the diagonal elements of the dissimilarity matrix) and pale values for cells with greater Euclidean distance. Figure 6.11 leads to the impression that there are at least three distinct groups with small inter-cluster differences (the dark rectangles), whereas much larger distances can be observed for all other cells.
- 197. 6.4 K-means clustering 181 R> n <- nrow(crime_s) R> wss <- rep(0, 6) R> wss[1] <- (n - 1) * sum(sapply(crime_s, var)) R> for (i in 2:6) + wss[i] <- sum(kmeans(crime_s, + centers = i)$withinss) R> plot(1:6, wss, type = "b", xlab = "Number of groups", + ylab = "Within groups sum of squares") q q q q q q 1 2 3 4 5 6 681012141618 Number of groups Withingroupssumofsquares Fig. 6.9. Plot of within-groups sum of squares against number of clusters. We plot the within-groups sum of squares for one to six group k-means solutions to see if we can get any indication of the number of groups (see Figure 6.12). Again, the plot leads to the relatively clear conclusion that the data contain three clusters. Our interest is now in a comparison of the kiln sites at which the pottery was found.
- 198. 182 6 Cluster Analysis q q q q q q q q q q q qq q q q q q q q q q −0.5 0.0 0.5 1.0 −0.4−0.20.00.20.4 PC1 PC2 Fig. 6.10. Plot of k-means two-group solution for the standardised crime rate data. R> set.seed(29) R> pottery_cluster <- kmeans(pots, centers = 3)$cluster R> xtabs(~ pottery_cluster + kiln, data = pottery) kiln pottery_cluster 1 2 3 4 5 1 21 0 0 0 0 2 0 12 2 0 0 3 0 0 0 5 5 The contingency table shows that cluster 1 contains all pots found at kiln site number one, cluster 2 contains all pots from kiln sites numbers two and three, and cluster three collects the ten pots from kiln sites four and five. In fact, the five kiln sites are from three different regions: region 1 contains just kiln one, region 2 contains kilns two and three, and region 3 contains kilns four
- 199. 6.5 Model-based clustering 183 R> library("lattice") R> levelplot(as.matrix(pottery_dist), xlab = "Pot Number", + ylab = "Pot Number") Pot Number PotNumber 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 Fig. 6.11. Image plot of the dissimilarity matrix of the pottery data. and five. So the clusters found actually correspond to pots from three different regions. 6.5 Model-based clustering The agglomerative hierarchical and k-means clustering methods described in the previous two sections are based largely on heuristic but intuitively reason- able procedures. But they are not based on formal models for cluster structure in the data, making problems such as deciding between methods, estimating R> pottery_dist <- dist(pots <- scale(pottery[, colnames(pottery) != "kiln"], + center = FALSE))
- 200. 184 6 Cluster Analysis R> n <- nrow(pots) R> wss <- rep(0, 6) R> wss[1] <- (n - 1) * sum(sapply(pots, var)) R> for (i in 2:6) + wss[i] <- sum(kmeans(pots, + centers = i)$withinss) R> plot(1:6, wss, type = "b", xlab = "Number of groups", + ylab = "Within groups sum of squares") q q q q q 1 2 3 4 5 6 20304050 Number of groups Withingroupssumofsquares Fig. 6.12. Plot of within-groups sum of squares against number of clusters. the number of clusters, etc, particularly difficult. And, of course, without a reasonable model, formal inference is precluded. In practise, these may not be insurmountable objections to the use of either the agglomerative meth- ods or k-means clustering because cluster analysis is most often used as an “exploratory” tool for data analysis. But if an acceptable model for cluster structure could be found, then the cluster analysis based on the model might give more persuasive solutions (more persuasive to statisticians at least). In
- 201. 6.5 Model-based clustering 185 q q q q q q q q qq q q q qq q q q qq q −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 −1.5−0.50.00.51.01.5 PC1 PC2 Fig. 6.13. Plot of the k-means three-group solution for the pottery data displayed in the space of the first two principal components of the correlation matrix of the data. this section, we describe an approach to clustering that postulates a formal statistical model for the population from which the data are sampled, a model that assumes that this population consists of a number of subpopulations (the “clusters”), each having variables with a different multivariate probability den- sity function, resulting in what is known as a finite mixture density for the population as a whole. By using finite mixture densities as models for cluster analysis, the clustering problem becomes that of estimating the parameters of the assumed mixture and then using the estimated parameters to calculate the posterior probabilities of cluster membership. And determining the number of clusters reduces to a model selection problem for which objective procedures exist. Finite mixture densities often provide a sensible statistical model for the clustering process, and cluster analyses based on finite mixture models are also
- 202. 186 6 Cluster Analysis known as model-based clustering methods; see Banfield and Raftery (1993). Finite mixture models have been increasingly used in recent years to clus- ter data in a variety of disciplines, including behavioural, medical, genetic, computer, environmental sciences, and robotics and engineering; see, for ex- ample, Everitt and Bullmore (1999), Bouguila and Amayri (2009), Bran- chaud, Cham, Nenadic, Andersen, and Burdick (2010), Dai, Erkkila, Yli- Harja, and Lahdesmaki (2009), Dunson (2009), Ganesalingam, Stahl, Wi- jesekera, Galtrey, Shaw, Leigh, and Al-Chalabi (2009), Marin, Mengersen, and Roberts (2005), Meghani, Lee, Hanlon, and Bruner (2009), Pledger and Phillpot (2008), and van Hattum and Hoijtink (2009). Finite mixture modelling can be seen as a form of latent variable analysis (see, for example, Skrondal and Rabe-Hesketh 2004), with “subpopulation” being a latent categorical variable and the latent classes being described by the different components of the mixture density; consequently, cluster analysis based on such models is also often referred to as latent class cluster analysis. 6.5.1 Finite mixture densities Finite mixture densities are described in detail in Everitt and Hand (1981), Titterington, Smith, and Makov (1985), McLachlan and Basford (1988), McLachlan and Peel (2000), and Fr¨uhwirth-Schnatter (2006); they are a fam- ily of probability density functions of the form f(x; p, θ) = c j=1 pjgj(x; θj), (6.1) where x is a p-dimensional random variable, p = (p1, p2, . . . , pc−1), and θ = (θ1 , θ2 , . . . , θc ), with the pj being known as mixing proportions and the gj, j = 1, . . . , c, being the component densities, with density gj being parameterised by θj. The mixing proportions are non-negative and are such that c j=1 pj = 1. The number of components forming the mixture (i.e., the postulated number of clusters) is c. Finite mixtures provide suitable models for cluster analysis if we assume that each group of observations in a data set suspected to contain clusters comes from a population with a different probability distribution. The latter may belong to the same family but differ in the values they have for the parameters of the distribution; it is such an example that we consider in the next section, where the components of the mixture are multivariate normal with different mean vectors and possibly different covariance matrices. Having estimated the parameters of the assumed mixture density, observa- tions can be associated with particular clusters on the basis of the maximum value of the estimated posterior probability ˆP(cluster j|xi) = ˆpjgj(xi; ˆθj) f(xi; ˆp, ˆθ) , j = 1, . . . , c. (6.2)
- 203. 6.5 Model-based clustering 187 6.5.2 Maximum likelihood estimation in a finite mixture density with multivariate normal components Given a sample of observations x1, x2, . . . , xn, from the mixture density given in Equation (6.1) the log-likelihood function, l, is l(p, θ) = n i=1 ln f(xi; p, θ). (6.3) Estimates of the parameters in the density would usually be obtained as a solution of the likelihood equations ∂l(ϕ) ∂(ϕ) = 0, (6.4) where ϕ = (p , θ ). In the case of finite mixture densities, the likelihood function is too complicated to employ the usual methods for its maximisation; for example, an iterative Newton–Raphson method that approximates the gradient vector of the log-likelihood function l(ϕ) by a linear Taylor series expansion (see Everitt (1984)). Consequently, the required maximum likelihood estimates of the parame- ters in a finite mixture model have to be computed in some other way. In the case of a mixture in which the jth component density is multivariate normal with mean vector µj and covariance matrix Σj, it can be shown (see Everitt and Hand 1981, for details) that the application of maximum likelihood results in the series of equations ˆpj = 1 n n i=1 ˆP(j|xi), (6.5) ˆµj = 1 nˆpj n i=1 xi ˆP(j|xi), (6.6) ˆΣj = 1 n n i=1 (xi − µj)(xi − µj) ˆP(j|xi), (6.7) where the ˆP(j|xi)s are the estimated posterior probabilities given in equation (6.2). Hasselblad (1966, 1969), Wolfe (1970), and Day (1969) all suggest an it- erative scheme for solving the likelihood equations given above that involves finding initial estimates of the posterior probabilities given initial estimates of the parameters of the mixture and then evaluating the right-hand sides of Equations 6.5 to 6.7 to give revised values for the parameters. From these, new estimates of the posterior probabilities are derived, and the procedure is repeated until some suitable convergence criterion is satisfied. There are po- tential problems with this process unless the component covariance matrices
- 204. 188 6 Cluster Analysis are constrained in some way; for example, it they are all assumed to be the same–again see Everitt and Hand (1981) for details. This procedure is a particular example of the iterative expectation max- imisation (EM) algorithm described by Dempster, Laird, and Rubin (1977) in the context of likelihood estimation for incomplete data problems. In estimat- ing parameters in a mixture, it is the “labels” of the component density from which an observation arises that are missing. As an alternative to the EM algo- rithm, Bayesian estimation methods using the Gibbs sampler or other Monte Carlo Markov Chain (MCMC) methods are becoming increasingly popular– see Marin et al. (2005) and McLachlan and Peel (2000). Fraley and Raftery (2002, 2007) developed a series of finite mixture den- sity models with multivariate normal component densities in which they allow some, but not all, of the features of the covariance matrix (orientation, size, and shape–discussed later) to vary between clusters while constraining others to be the same. These new criteria arise from considering the reparameterisa- tion of the covariance matrix Σj in terms of its eigenvalue description Σj = DjΛjDj , (6.8) where Dj is the matrix of eigenvectors and Λj is a diagonal matrix with the eigenvalues of Σj on the diagonal (this is simply the usual principal compo- nents transformation–see Chapter 3). The orientation of the principal compo- nents of Σj is determined by Dj, whilst Λj specifies the size and shape of the density contours. Specifically, we can write Λj = λjAj, where λj is the largest eigenvalue of Σj and Aj = diag(1, α2, . . . , αp) contains the eigenvalue ratios after division by λj. Hence λj controls the size of the jth cluster and Aj its shape. (Note that the term “size” here refers to the volume occupied in space, not the number of objects in the cluster.) In two dimensions, the parame- ters would reflect, for each cluster, the correlation between the two variables, and the magnitudes of their standard deviations. More details are given in Banfield and Raftery (1993) and Celeux and Govaert (1995), but Table 6.4 gives a series of models corresponding to various constraints imposed on the covariance matrix. The models make up what Fraley and Raftery (2003, 2007) term the “MCLUST” family of mixture models. The mixture likelihood ap- proach based on the EM algorithm for parameter estimation is implemented in the Mclust() function in the R package mclust and fits the models in the MCLUST family described in Table 6.4. Model selection is a combination of choosing the appropriate clustering model for the population from which the n observations have been taken (i.e., are all clusters spherical, all elliptical, all different shapes or somewhere in between?) and the optimal number of clusters. A Bayesian approach is used (see Fraley and Raftery 2002), applying what is known as the Bayesian Information Criterion (BIC). The result is a cluster solution that “fits” the observed data as well as possible, and this can include a solution that has only one “cluster” implying that cluster analysis is not really a useful technique for the data.
- 205. 6.5 Model-based clustering 189 Table 6.4: mclust family of mixture models. Model names describe model restrictions of volume λj, shape Aj, and orientation Dj, V = variable, parameter unconstrained, E= equal, parameter con- strained, I = matrix constrained to identity matrix. Abbreviation Model EII spherical, equal volume VII spherical, unequal volume EEI diagonal, equal volume and shape VEI diagonal, varying volume, equal shape EVI diagonal, equal volume, varying shape VVI diagonal, varying volume and shape EEE ellipsoidal, equal volume, shape, and orientation EEV ellipsoidal, equal volume and equal shape VEV ellipsoidal, equal shape VVV ellipsoidal, varying volume, shape, and orientation To illustrate the use of the finite mixture approach to cluster analysis, we will apply it to data that arise from a study of what gastroenterologists in Europe tell their cancer patients (Thomsen, Wulff, Martin, and Singer 1993). A questionnaire was sent to about 600 gastroenterologists in 27 European countries (the study took place before the recent changes in the political map of the continent) asking what they would tell a patient with newly diagnosed cancer of the colon, and his or her spouse, about the diagnosis. The respondent gastroenterologists were asked to read a brief case history and then to answer six questions with a yes/no answer. The questions were as follows: Q1: Would you tell this patient that he/she has cancer, if he/she asks no questions? Q2: Would you tell the wife/husband that the patient has cancer (In the patient’s absence)? Q3: Would you tell the patient that he or she has a cancer, if he or she directly asks you to disclose the diagnosis. (During surgery the surgeon notices several small metastases in the liver.) Q4: Would you tell the patient about the metastases (supposing the patient asks to be told the results of the operation)? Q5: Would you tell the patient that the condition is incurable? Q6: Would you tell the wife or husband that the operation revealed metas- tases? The data are shown in a graphical form in Figure 6.14 (we are aware that using finite mixture clustering on this type of data is open to criticism–it may even be a statistical sin–but we hope that even critics will agree it provides an interesting example).
- 206. 190 6 Cluster Analysis qqqqq qqqqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqqq qqqqqqqqqqqqqqqqqqq q qqqqqqqqqqqqqq qqqqqqqq qq qqqqq qqqqqqqqqq qqqqqqqqqqqq qqqqqq qqqqqqqqqqqqqqq qqqqqqqq qqqqqqqqqqqqqqq qqqqq qqq qqqqqqq qqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqq qqqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqqq qqqqq qqqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqqq qqqqqqqqqqqqqqqqqq qqq qqqqqqqqqqqqq qqqqqqqq qq qqqqq qqqqqqqq qqqqqqqqqqqq qqqqqq qqqqqqqqqqqqqqq qqqqqqqq qqqqqqqqqqqqqqq qqqqq qqq qqqqqq qqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqqq qqqqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqq qqqq qqqqqq qqqqqqq qqqqqq qqqqqqqqqqqq qqqqqqqqqqqqqqqqqqqq qqq qqqqqqqqqqqqq qqqqqqqq qq qqqq qqqqqqqq qqqqqqqqqq qqqqqq qqqqqqqqqqqqqq qqqqqqqq qqqqqqqqqqqqqqq qqqqq qqq qqqqqqq qqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqq qqqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqq qqqqq qqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqq qqqqqqqqqqqqqqqqqqqq qq qqqqqqqqqqqqqq qqqqqqqq qq qqqq qqqqqqqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqqqqq qqqqqqqq qqqqqqqqqqqqqqq qqqqq qqq qqqqqqq qqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqqqq qqqqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqq qqqqq qqqq qqqqqqqqqq qqqqqq qqqqqqqqqqq qqqqqqqqqqqqqqqqqqq qq qqqqqqqqqqqqq qqqqqqq qq qqqqq qqqqqqqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqqqqq qqqqqqqq qqqqqqqqqqqqqqq qqqqqq qqq qqqqqqq qqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqq qqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqq qqqqq qqqqq qqqqqqqqq qqqqqq qqqqqqqqqqqqq qqqqqqqqqqqqqqqqq qqq qqqqqqqqqqqqqq qqqqqqqq qq qqqq qqqqqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqqqqq qqqqqqq qqqqqqqqqqqqqq qqqqq qqq qqqqqqq qqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqqq qqqqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqqEstonia Latvia Lithuania CIS Poland Slovakia Czechia Hungary Romania Bulgaria Albania Yugoslavia Greece Italy Portugal Spain France Switzerland Belgium Netherlands Germany Eire UK Denmark Finland Sweden Norway Iceland Question1Question2Question3Question4Question5Question6 Fig.6.14.Gastroenterologistsquestionnairedata.Darkcirclesindicatea‘yes’,opencirclesa‘no’.
- 207. 6.6 Displaying clustering solutions graphically 191 Applying the finite mixture approach to the proportions of ‘yes’ answers for each question for each country computed from these data using the R code utilizing functionality offered by package mclust (Fraley and Raftery 2010) R> library("mclust") by using mclust, invoked on its own or through another package, you accept the license agreement in the mclust LICENSE file and at http://www.stat.washington.edu/mclust/license.txt R> (mc <- Mclust(thomsonprop)) best model: ellipsoidal, equal shape with 3 components where thomsonprob is the matrix of proportions of “yes” answers to ques- tions Q1–Q6 in the different countries (i.e., the proportion of filled dots in Figure 6.14) available from the MVA add-on package. We can first examine the resulting plot of BIC values shown in Figure 6.15. In this diagram, the plot symbols and abbreviations refer to different model assumptions about the shapes of clusters as given in Table 6.4. The BIC criterion selects model VEV (ellipsoidal, equal shape) and three clusters as the optimal solution. The three-cluster solution is illustrated graph- ically in Figure 6.16. The first cluster consists of countries in which the large majority of respondents gave “yes” answers to questions 1, 2, 3, 4, and 6 and about half also gave a “yes” answer to question 5. This cluster includes all the Scandinavian countries the UK, Iceland, Germany, the Netherlands, and Switzerland. In the second cluster, the majority of respondents answer “no” to questions 1, 4, and 5 and “yes” to questions 2, 3 and 6; in these countries it appears that the clinicians do not mind giving bad news to the spouses of patients but not to the patients themselves unless they are directly asked by the patient about hispr her condition. This cluster contains Catholic countries such as Spain, Portugal, and Italy. In cluster three, the large majority of re- spondents answer “no” to questions 1, 3, 4, and 5 and again a large majority answer “yes” to questions 2 and 6. In these countries, very few clinicians ap- pear to be willing to give the patient bad news even if asked directly by the patient about his or her condition. 6.6 Displaying clustering solutions graphically Plotting cluster solutions in the space of the first few principal components as illustrated earlier in this chapter is often a useful way to display clustering solutions, but other methods of displaying clustering solutions graphically are also available. Leisch (2010), for example, describes several graphical displays that can be used to visualise cluster analysis solutions. The basis of a number of these graphics is the shadow value, s(x), of each multivariate observation, x, defined as
- 208. 192 6 Cluster Analysis 2 4 6 8 −100−500 number of components BIC q q q q q q q q q q q q q q q q qq q q q qEII VII EEI VEI EVI VVI EEE EEV VEV VVV Fig. 6.15. BIC values for gastroenterologists questionnaire. s(x) = 2d(x, c(x)) d(x, c(x)) + d(x, ˜c(x)) , where d(x, c(x)) is the distance of the observation x from the centroid of its own cluster and d(x, ˜c(x)) is the distance of x from the second closest cluster centroid. If s(x) is close to zero, then the observation is close to its cluster centroid; if s(x) is close to one, then the observation is almost equidistant from the two centroids (a similar approach is used in defining silhouette plots, see Chapter 5). The average shadow value of all observations where cluster i is closest and cluster j is second closest can be used as a simple measure of cluster similarity, sij = 1 ni x∈Aij s(x), where ni is the number of observations that are closest to the centroid of cluster i and Aij is the set of observations for which the centroid of cluster i is
- 209. 6.6 Displaying clustering solutions graphically 193 qqqqq qqqqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqqq qqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqq qqqqqqqq qqqqq q qq qqqqqqqqqq qqqqqqqqqqqq qqqqqq qqqqqqqqqqqqqqq qqqqqqqq qqqqqqqqqqqqqqq qqqqq qqqqq qqqqqqqqqqqqq qqqq qqq qqqqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqqqq qqqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqqq qqqqqqqqqqqqqqqqqq qqqqqqqqqqqqq qqqqqqqq qqqqq qqq qq qqqqqqqq qqqqqqqqqqqq qqqqqq qqqqqqqqqqqqqqq qqqqqqqq qqqqqqqqqqqqqqq qqqqq qqqqq qqqqqqqqqqqqqqq qqq qqq qqqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqqq qqqqqq qqqqqqq qqqqqq qqqqqqqqqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqq qqqqqqqq qqqq qqq qq qqqqqqqq qqqqqqqqqq qqqqqq qqqqqqqqqqqqqq qqqqqqqq qqqqqqqqqqqqqqq qqqqq qqqqq qqqqqqqqqqqqqq qqq qqq qqqqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqqqq qqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqq qqqqqqqq qqqq qq qq qqqqqqqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqqqqq qqqqqqqq qqqqqqqqqqqqqqq qqqqq qqqqq qqqqqqqqqqqqqqqq qqq qqq qqqqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqqqq qqqq qqqqqqqqqq qqqqqq qqqqqqqqqqq qqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqq qqqqqqq qqqqq qq qq qqqqqqqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqqqqq qqqqqqqq qqqqqqqqqqqqqqq qqqqqq qqqqq qqqqqqqqqqqqq qqq qqq qqqqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqq qqqqq qqqqq qqqqqqqqq qqqqqq qqqqqqqqqqqqq qqqqqqqqqqqqqqqqq qqqqqqqqqqqqqq qqqqqqqq qqqq qqq qq qqqqqqqq qqqqqqqqqqq qqqqqq qqqqqqqqqqqqqqq qqqqqqq qqqqqqqqqqqqqq qqqqq qqqqq qqqqqqqqqqqqqqq qqq qqq qqqqqqq qqqqqqqqqqqqqqqqqqqq qqqqqqqqqqqqqqqqqqq qqqqqqq qqqqqqqq qqqqqLatvia Lithuania CIS Poland Czechia Romania Bulgaria Estonia Slovakia Hungary Albania Yugoslavia Greece Italy Portugal Spain France Belgium Eire Switzerland Netherlands Germany UK Denmark Finland Sweden Norway Iceland Question1Question2Question3Question4Question5Question6 Cluster1 Cluster2 Cluster3 Fig.6.16.Model-basedclusteringintothreeclustersforgastroenterologistsquestionnairedata.Darkcirclesindicatea‘yes’,open circlesa‘no’.
- 210. 194 6 Cluster Analysis closest and the centroid of cluster j is second closest. The denominator of sij is taken to be ni rather than nij, the number of observations in the set Aij, to prevent inducing large cluster similarity when nij is small and the set of observations consists of poorly clustered points with large shadow values. For a cluster solution derived from bivariate data, a neighbourhood graph can be constructed using the scatterplot of the two variables, and where two cluster centroids are joined if there exists at least one observation for which these two are closest, and second closest with the thickness of the joining lines being made proportional to the average value of the corresponding sij. When there are more than two variables in the data set, the neighbourhood graph can be constructed on some suitable projection of the data into two dimensions; for example, the first two principal components of the data could be used. Such plots may help to establish which clusters are “real” and which are not, as we will try to illustrate with two examples. The first example uses some two-dimensional data generated to contain three clusters. The neighbourhood graph for the k-means five-cluster solution from the application of k-means clustering is shown in Figure 6.17. The thicker lines joining the centroids of clusters 1 and 5 and clusters 2 and 4 strongly suggest that both pairs of clusters overlap to a considerable extent and are probably each divisions of a single cluster. For the second example we return to the pottery data previously met in the chapter. From the k-means analysis, it is clear that these data contain three clusters; Figure 6.18 shows the neighbourhood plot for the k-means three- cluster solution in the space of the first two principal components of the data. The three clusters are clearly visible in this plot. A further graphic for displaying clustering solutions is known as a stripes plot. This graphic is a simple but often effective way of visualising the distance of each point from its closest and second closest cluster centroids. For each cluster, k = 1, . . . , K, a stripes plot has a rectangular area that is vertically divided into K smaller rectangles, with each smaller rectangle, i, containing information about distances of the observations in cluster i from the centroid of that cluster along with the corresponding information about observations that have cluster i as their second closest cluster. The explanation of how the plot is constructed becomes more transparent if we look at an actual ex- ample. Figure 6.19 shows a stripes plot produced with that package flexclust (Leisch and Dimitriadou 2019) for a five-cluster solution on a set of data gen- erated to contain five relatively distinct clusters. Looking first at the rectangle for cluster one, we see that observations in clusters two and three have the cluster one centroid as their second closest. These observations form the two other stripes within the rectangle. Observations in cluster three are further away from cluster one, but a number of observations in cluster three have a distance to the centroid of cluster one similar to those observations that belong to cluster one. Overall though, the stripes plot in Figure 6.19 suggests that the five-cluster solution matches quite well the actual structure in the data. The situation is quite different in Figure 6.20, where the stripes plot
- 211. 6.6 Displaying clustering solutions graphically 195 R> library("flexclust") R> library("mvtnorm") R> set.seed(290875) R> x <- rbind(rmvnorm(n = 20, mean = c(0, 0), + sigma = diag(2)), + rmvnorm(n = 20, mean = c(3, 3), + sigma = 0.5 * diag(2)), + rmvnorm(n = 20, mean = c(7, 6), + sigma = 0.5 * (diag(2) + 0.25))) R> k <- cclust(x, k = 5, save.data = TRUE) R> plot(k, hull = FALSE, col = rep("black", 5), xlab = "x", ylab = "y") q q q q q q q q q q q q q −2 0 2 4 6 8 −202468 x y 1 2 3 4 5 Fig. 6.17. Neighbourhood plot of k-means five-cluster solution for bivariate data containing three clusters.
- 212. 196 6 Cluster Analysis R> k <- cclust(pots, k = 3, save.data = TRUE) R> plot(k, project = prcomp(pots), hull = FALSE, col = rep("black", 3), + xlab = "PC1", ylab = "PC2") qq q q q q q q q q −5 0 5 −4−202 PC1 PC2 1 2 3 Fig. 6.18. Neighbourhood plot of k-means three-cluster solution for pottery data. for the k-means five-group solution suggests that the clusters in this solution are not well separated, implying perhaps that the five-group solution is not appropriate for the data in this case. Lastly, the stripes plot for the k-means three-group solution on the pottery data is shown in Figure 6.21. The graphic confirms the three-group structure of the data. All the information in a stripes plot is also available from a neighbour- hood plot, but the former is dimension independent and may work well even for high-dimensional data where projections to two dimensions lose a lot of information about the structure in the data. Neither neighbourhood graphs nor stripes plots are infallible, but both offer some help in the often difficult task of evaluating and validating the solutions from a cluster analysis of a set of data.
- 213. 6.7 Summary 197 R> set.seed(912345654) R> x <- rbind(matrix(rnorm(100, sd = 0.5), ncol= 2 ), + matrix(rnorm(100, mean =4, sd = 0.5), ncol = 2), + matrix(rnorm(100, mean =7, sd = 0.5), ncol = 2), + matrix(rnorm(100, mean =-1.0, sd = 0.7), ncol = 2), + matrix(rnorm(100, mean =-4.0, sd = 1.0), ncol = 2)) R> c5 <- cclust(x, 5, save.data = TRUE) R> stripes(c5, type = "second", col = 1) 0 1 2 3 4 5 6 distancefromcentroid 1 2 3 4 5 Fig. 6.19. Stripes plot of k-means solution for artificial data. 6.7 Summary Cluster analysis techniques are used to search for clusters or groups in a priori unclassified multivariate data. Although clustering techniques are potentially very useful for the exploration of multivariate data, they require care in their application if misleading solutions are to be avoided. Many methods of clus- ter analysis have been developed, and most studies have shown that no one method is best for all types of data. But the more statistical techniques covered
- 214. 198 6 Cluster Analysis R> set.seed(912345654) R> x <- rbind(matrix(rnorm(100, sd = 2.5), ncol = 2), + matrix(rnorm(100, mean = 3, sd = 0.5), ncol = 2), + matrix(rnorm(100, mean = 5, sd = 0.5), ncol = 2), + matrix(rnorm(100, mean = -1.0, sd = 1.5), ncol = 2), + matrix(rnorm(100, mean = -4.0, sd = 2.0), ncol = 2)) R> c5 <- cclust(x, 5, save.data = TRUE) R> stripes(c5, type = "second", col = 1) 0 2 4 6 8 10 distancefromcentroid 1 2 3 4 5 Fig. 6.20. Stripes plot of k-means solution for artificial data. briefly in Section 6.5 and in more detail in Everitt et al. (2011) have definite statistical advantages because the clustering is based on sensible models for the data. Cluster analysis is a large area and has been covered only briefly in this chapter. The many problems that need to be considered when using clustering in practice have barely been touched upon. For a detailed discussion of these problems, again see Everitt et al. (2011).
- 215. 6.7 Summary 199 R> set.seed(15) R> c5 <- cclust(pots, k = 3, save.data = TRUE) R> stripes(c5, type = "second", col = "black") 0 0.5 1 1.5 2 2.5 3 distancefromcentroid 1 2 3 Fig. 6.21. Stripes plot of three-group k-means solution for pottery data. Finally, we should mention in passing a technique known as projection pursuit. In essence, and like principal components analysis, projection pursuit seeks a low-dimensional projection of a multivariate data set but one that may be more likely to be successful in uncovering any cluster (or more ex- otic) structure in the data than principal component plots using the first few principal component scores. The technique is described in detail in Jones and Sibson (1987) and more recently in Cook and Swayne (2007).
- 216. 200 6 Cluster Analysis 6.8 Exercises Ex. 6.1 Apply k-means to the crime rate data after standardising each vari- able by its standard deviation. Compare the results with those given in the text found by standardising by a variable’s range. Ex. 6.2 Calculate the first five principal components scores for the Romano- British pottery data, and then construct the scatterplot matrix of the scores, displaying the contours of the estimated bivariate density for each panel of the plot and a boxplot of each score in the appropriate place on the diagonal. Label the points in the scatterplot matrix with their kiln numbers. Ex. 6.3 Return to the air pollution data given in Chapter 1 and use finite mixtures to cluster the data on the basis of the six climate and ecology variables (i.e., excluding the sulphur dioxide concentration). Investigate how sulphur dioxide concentration varies in the clusters you find both graphically and by formal significance testing.
- 217. 7 Confirmatory Factor Analysis and Structural Equation Models 7.1 Introduction An exploratory factor analysis as described in Chapter 5 is used in the early investigation of a set of multivariate data to determine whether the factor analysis model is useful in providing a parsimonious way of describing and accounting for the relationships between the observed variables. The analysis will determine which observed variables are most highly correlated with the common factors and how many common factors are needed to give an ade- quate description of the data. In an exploratory factor analysis, no constraints are placed on which manifest variables load on which factors. In this chapter, we will consider confirmatory factor analysis models in which particular man- ifest variables are allowed to relate to particular factors whilst other manifest variables are constrained to have zero loadings on some of the factors. A con- firmatory factor analysis model may arise from theoretical considerations or be based on the results of an exploratory factor analysis where the investigator might wish to postulate a specific model for a new set of similar data, one in which the loadings of some variables on some factors are fixed at zero because they were “small” in the exploratory analysis and perhaps to allow some pairs of factors but not others to be correlated. It is important to emphasise that whilst it is perfectly appropriate to arrive at a factor model to submit to a confirmatory analysis from an exploratory factor analysis, the model must be tested on a fresh set of data. Models must not be generated and tested on the same data. Confirmatory factor analysis models are a subset of a more general ap- proach to modelling latent variables known as structural equation modelling or covariance structure modelling. Such models allow both response and ex- planatory latent variables linked by a series of linear equations. Although more complex than confirmatory factor analysis models, the aim of structural equation models is essentially the same, namely to explain the correlations or covariances of the observed variables in terms of the relationships of these variables to the assumed underlying latent variables and the relationships pos- DOI 10.1007/978-1-4419-9650-3_7, © Springer Science+Business Media, LLC 2011 201B. Everitt and T. Hothorn, An Introduction to Applied Multivariate Analysis with R: Use R!,
- 218. 202 7 Confirmatory Factor Analysis and Structural Equation Models tulated between the latent variables themselves. Structural equation models represent the convergence of relatively independent research traditions in psy- chiatry, psychology, econometrics, and biometrics. The idea of latent variables in psychometrics arises from Spearman’s early work on general intelligence. The concept of simultaneous directional influences of some variables on others has been part of economics for several decades, and the resulting simultaneous equation models have been used extensively by economists but essentially only with observed variables. Path analysis was introduced by Wright (1934) in a biometrics context as a method for studying the direct and indirect effects of variables. The quintessential feature of path analysis is a diagram show- ing how a set of explanatory variables influence a dependent variable under consideration. How the paths are drawn determines whether the explanatory variables are correlated causes, mediated causes, or independent causes. Some examples of path diagrams appear later in the chapter. (For more details of path analysis, see Schumaker and Lomax 1996). Later, path analysis was taken up by sociologists such as Blalock (1961), Blalock (1963) and then by Duncan (1969), who demonstrated the value of combining path-analytic representation with simultaneous equation models. And, finally, in the 1970s, several workers most prominent of whom were J¨oreskog (1973), Bentler (1980), and Browne (1974), combined all these var- ious approaches into a general method that could in principle deal with ex- tremely complex models in a routine manner. 7.2 Estimation, identification, and assessing fit for confirmatory factor and structural equation models 7.2.1 Estimation Structural equation models will contain a number of parameters that need to be estimated from the covariance or correlation matrix of the manifest variables. Estimation involves finding values for the model parameters that minimise a discrepancy function indicating the magnitude of the differences between the elements of S, the observed covariance matrix of the manifest variables and those of Σ(θ), the covariance matrix implied by the fitted model (i.e., a matrix the elements of which are functions of the parameters of the model), contained in the vector θ = (θ1, . . . , θt) . There are a number of possibilities for discrepancy functions; for example, the ordinary least squares discrepancy function, FLS, is FLS(S, Σ(θ)) = i<j j (sij − σij(θ))2 , where sij and σij(θ) are the elements of S and Σ(θ). But this criterion has several problems that make it unsuitable for estimation; for example, it is not
- 219. 7.2 Estimation, identification, and assessing fit 203 independent of the scale of the manifest variables, and so different estimates of the model parameters would be produced using the sample covariance matrix and the sample correlation matrix. Other problems with the least squares criterion are detailed in Everitt (1984). The most commonly used method of estimating the parameters in confir- matory factor and structural equation models is maximum likelihood under the assumption that the observed data have a multivariate normal distribu- tion. It is easy to show that maximising the likelihood is now equivalent to minimising the discrepancy function, FML, given by FML(S, Σ(θ)) = log |Σ(θ)| − log |S| + tr(SΣ(θ)−1 ) − q (cf. maximum likelihood factor analysis in Chapter 5). We see that by varying the parameters θ1, . . . , θt so that Σ(θ) becomes more like S, FML becomes smaller. Iterative numerical algorithms are needed to minimise the function FML with respect to the parameters, but for details see Everitt (1984) and Everitt and Dunn (2001). 7.2.2 Identification Consider the following simple example of a model in which there are three manifest variables, x, x , and y, and two latent variables, u and v, with the relationships between the manifest and latent variables being x = u + δ, y = v + , x = u + δ . If we assume that δ, δ , and have expected values of zero, that δ and δ are uncorrelated with each other and with u, and that is uncorrelated with v, then the covariance matrix of the three manifest variables may be expressed in terms of parameters representing the variances and covariances of the residuals and the latent variables as Σ(θ) = θ1 + θ2 θ3 θ4 + θ5 θ3 θ4 θ4 + θ6 , where θ = (θ1, θ2, θ3, θ4, θ5, θ6) and θ1 = Var(v), θ2 = Var( ), θ3 = Cov(v, u), θ4 = Var(u), θ5 = Var(δ), and θ6 = Var(δ ). It is immediately apparent that estimation of the parameters in this model poses a problem. The two param- eters θ1 and θ2 are not uniquely determined because one can be, for example, increased by some amount and the other decreased by the same amount with- out altering the covariance matrix predicted by the model. In other words, in this example, different sets of parameter values (i.e., different θs) will lead to the same predicted covariance matrix, Σ(θ). The model is said to be uniden- tifiable. Formally, a model is identified if and only if Σ(θ1) = Σ(θ2) implies
- 220. 204 7 Confirmatory Factor Analysis and Structural Equation Models that θ1 = θ2. In Chapter 5, it was pointed out that the parameters in the ex- ploratory factor analysis model are not identifiable unless some constraints are introduced because different sets of factor loadings can give rise to the same predicted covariance matrix. In confirmatory factor analysis models and more general covariance structure models, identifiability depends on the choice of model and on the specification of fixed, constrained (for example, two param- eters constrained to equal one another), and free parameters. If a parameter is not identified, it is not possible to find a consistent estimate of it. Establishing model identification in confirmatory factor analysis models (and in structural equation models) can be difficult because there are no simple, practicable, and universally applicable rules for evaluating whether a model is identified, although there is a simple necessary but not sufficient condition for identifi- cation, namely that the number of free parameters in a model, t, be less than q(q + 1)/2. For a more detailed discussion of the identifiability problem, see Bollen and Long (1993). 7.2.3 Assessing the fit of a model Once a model has been pronounced identified and its parameters estimated, the next step becomes that of assessing how well the model-predicted co- variance matrix fits the covariance matrix of the manifest variables. A global measure of fit of a model is provided by the likelihood ratio statistic given by X2 = (N − 1)FMLmin, where N is the sample size and FMLmin is the minimised value of the maximum likelihood discrepancy function given in Subsection 7.2.1. If the sample size is sufficiently large, the X2 statistic pro- vides a test that the population covariance matrix of the manifest variables is equal to the covariance implied by the fitted model against the alternative hypothesis that the population matrix is unconstrained. Under the equality hypothesis, X2 has a chi-squared distribution with degrees of freedom ν given by 1 2 q(q + 1) − t, where t is the number of free parameters in the model. The likelihood ratio statistic is often the only measure of fit quoted for a fitted model, but on its own it has limited practical use because in large samples even relatively trivial departures from the equality null hypothesis will lead to its rejection. Consequently, in large samples most models may be rejected as statistically untenable. A more satisfactory way to use the test is for a comparison of a series of nested models where a large difference in the statistic for two models compared with the difference in the degrees of freedom of the models indicates that the additional parameters in one of the models provide a genuine improvement in fit. Further problems with the likelihood ratio statistic arise when the observa- tions come from a population where the manifest variables have a non-normal distribution. Browne (1982) demonstrates that in the case of a distribution with substantial kurtosis, the chi-squared distribution may be a poor approx- imation for the null distribution of X2 . Browne suggests that before using the test it is advisable to assess the degree of kurtosis of the data by using
- 221. 7.2 Estimation, identification, and assessing fit 205 Mardia’s coefficient of multivariate kurtosis (see Mardia et al. 1979). Browne’s suggestion appears to be little used in practise. Perhaps the best way to assess the fit of a model is to use the X2 statistic alongside one or more of the following procedures: ˆ Visual inspection of the residual covariances (i.e., the differences between the covariances of the manifest variables and those predicted by the fitted model). These residuals should be small when compared with the values of the observed covariances or correlations. ˆ Examination of the standard errors of the parameters and the correlations between these estimates. If the correlations are large, it may indicate that the model being fitted is almost unidentified. ˆ Estimated parameter values outside their possible range; i.e., negative vari- ances or absolute values of correlations greater than unity are often an indication that the fitted model is fundamentally wrong for the data. In addition, a number of fit indices have been suggested that can sometimes be useful. For example, the goodness-of-fit index (GFI) is based on the ratio of the sum of squared distances between the matrices observed and those reproduced by the model covariance, thus allowing for scale. The GFI measures the amount of variance and covariance in S that is ac- counted for by the covariance matrix predicted by the putative model, namely Σ(θ), which for simplicity we shall write as Σ. For maximum likelihood esti- mation, the GFI is given explicitly by GFI = 1 − tr S ˆΣ−1 − I S ˆΣ−1 − I tr S ˆΣ−1S ˆΣ−1 . The GFI can take values between zero (no fit) and one (perfect fit); in practise, only values above about 0.9 or even 0.95 suggest an acceptable level of fit. The adjusted goodness of fit index (AGFI) adjusts the GFI index for the degrees of freedom of a model relative to the number of variables. The AGFI is calculated as follow; AGFI = 1 − (k/df)(1 − GFI), where k is the number of unique values in S and df is the number of degrees of freedom in the model (discussed later). The GFI and AGFI can be used to compare the fit of two different models with the same data or compare the fit of models with different data, for example male and female data sets. A further fit index is the root-mean-square residual (RMSR), which is the square root of the mean squared differences between the elements in S and ˆΣ. It can be used to compare the fit of two different models with the same data. A value of RMSR < 0.05 is generally considered to indicate a reasonable fit. A variety of other fit indices have been proposed, including the Tucker- Lewis index and the normed fit index; for details, see Bollen and Long (1993).
- 222. 206 7 Confirmatory Factor Analysis and Structural Equation Models 7.3 Confirmatory factor analysis models In a confirmatory factor model the loadings for some observed variables on some of the postulated common factors will be set a priori to zero. Addi- tionally, some correlations between factors might also be fixed at zero. Such a model is fitted to a set of data by estimating its free parameters; i.e., those not fixed at zero by the investigator. Estimation is usually by maximum likelihood using the FML discrepancy function. We will now illustrate the application of confirmatory factor analysis with two examples. 7.3.1 Ability and aspiration Calsyn and Kenny (1977) recorded the values of the following six variables for 556 white eighth-grade students: SCA: self-concept of ability; PPE: perceived parental evaluation; PTE: perceived teacher evaluation; PFE: perceived friend’s evaluation; EA: educational aspiration; CP: college plans. Calsyn and Kenny (1977) postulated that two underlying latent variables, ability and aspiration, generated the relationships between the observed vari- ables. The first four of the manifest variables were assumed to be indicators of ability and the last two indicators of aspiration; the latent variables, ability and aspiration, are assumed to be correlated. The regression-like equations that specify the postulated model are SCA = λ1f1 + 0f2 + u1, PPE = λ2f1 + 0f2 + u2, PTE = λ3f1 + 0f2 + u3, PFE = λ4f1 + 0f2 + u4, AE = 0f1 + λ5f2 + u5, CP = 0f1 + λ6f2 + u6, where f1 represents the ability latent variable and f2 represents the aspiration latent variable. Note that, unlike in exploratory factor analysis, a number of factor loadings are fixed at zero and play no part in the estimation process. The model has a total of 13 parameters to estimate, six factor loadings (λ1 to λ6), six specific variances (ψ1 to ψ6), and one correlation between ability and aspiration (ρ). (To be consistent with the nomenclature used in Sub- section 7.2.1, all parameters should be suffixed thetas; this could, however, become confusing, so we have changed the nomenclature and use lambdas, etc., in a manner similar to how they are used in Chapter 5.) The observed
- 223. 7.3 Confirmatory factor analysis models 207 correlation matrix given in Figure 7.1 has six variances and 15 correlations, a total of 21 terms. Consequently, the postulated model has 21 − 13 = 8 de- grees of freedom. The figure depicts each correlation by an ellipse whose shape tends towards a line with slope 1 for correlations near 1, to a circle for correla- tions near zero, and to a line with negative slope −1 for negative correlations near −1. In addition, 100 times the correlation coefficient is printed inside the ellipse and colour-coding indicates strong negative (dark) to strong positive (light) correlations. EA CP PFE PTE SCA PPE EA CP PFE PTE SCA PPE 100 72 37 40 46 43 72 100 41 48 56 52 37 41 100 57 58 61 40 48 57 100 70 68 46 56 58 70 100 73 43 52 61 68 73 100 −1.0 −0.5 0.0 0.5 1.0 Fig. 7.1. Correlation matrix of ability and aspiration data; values given are corre- lation coefficients ×100. The R code, contained in the package sem (Fox, Kramer, and Friendly 2010), for fitting the model is R> ability_model <- specify.model(file = "ability_model.txt") R> ability_sem <- sem(ability_model, ability, 556)
- 224. 208 7 Confirmatory Factor Analysis and Structural Equation Models Here, the model is specified in a text file (called ability_model.txt in our case) with the following content: Ability -> SCA, lambda1, NA Ability -> PPE, lambda2, NA Ability -> PTE, lambda3, NA Ability -> PFE, lambda4, NA Aspiration -> EA, lambda5, NA Aspiration -> CP, lambda6, NA Ability <-> Aspiration, rho, NA SCA <-> SCA, theta1, NA PPE <-> PPE, theta2, NA PTE <-> PTE, theta3, NA PFE <-> PFE, theta4, NA EA <-> EA, theta5, NA CP <-> CP, theta6, NA Ability <-> Ability, NA, 1 Aspiration <-> Aspiration, NA, 1 The model is specified via arrows in the so-called reticular action model (RAM) notation. The text consists of three columns. The first one corresponds to an arrow specification where single-headed or directional arrows correspond to regression coefficients and double-headed or bidirectional arrows correspond to variance parameters. The second column denotes parameter names, and the third one assigns values to fixed parameters. Further details are available from the corresponding pages of the manual for the sem package. The results from fitting the ability and aspiration model to the observed correlations are available via R> summary(ability_sem) Model Chisquare = 9.2557 Df = 8 Pr(>Chisq) = 0.32118 Chisquare (null model) = 1832.0 Df = 15 Goodness-of-fit index = 0.99443 Adjusted goodness-of-fit index = 0.98537 RMSEA index = 0.016817 90% CI: (NA, 0.05432) Bentler-Bonnett NFI = 0.99495 Tucker-Lewis NNFI = 0.9987 Bentler CFI = 0.9993 SRMR = 0.012011 BIC = -41.310 Normalized Residuals Min. 1st Qu. Median Mean 3rd Qu. Max. -0.4410 -0.1870 0.0000 -0.0131 0.2110 0.5330 Parameter Estimates
- 225. 7.3 Confirmatory factor analysis models 209 Estimate Std Error z value Pr(>|z|) lambda1 0.86320 0.035182 24.5355 0.0000000 lambda2 0.84932 0.035489 23.9323 0.0000000 lambda3 0.80509 0.036409 22.1123 0.0000000 lambda4 0.69527 0.038678 17.9757 0.0000000 lambda5 0.77508 0.040365 19.2020 0.0000000 lambda6 0.92893 0.039417 23.5665 0.0000000 rho 0.66637 0.030955 21.5273 0.0000000 theta1 0.25488 0.023502 10.8450 0.0000000 theta2 0.27865 0.024263 11.4847 0.0000000 theta3 0.35184 0.026916 13.0715 0.0000000 theta4 0.51660 0.034820 14.8365 0.0000000 theta5 0.39924 0.038214 10.4475 0.0000000 theta6 0.13709 0.043530 3.1493 0.0016366 lambda1 SCA <--- Ability lambda2 PPE <--- Ability lambda3 PTE <--- Ability lambda4 PFE <--- Ability lambda5 EA <--- Aspiration lambda6 CP <--- Aspiration rho Aspiration <--> Ability theta1 SCA <--> SCA theta2 PPE <--> PPE theta3 PTE <--> PTE theta4 PFE <--> PFE theta5 EA <--> EA theta6 CP <--> CP Iterations = 28 (Note that the two latent variables have their variances fixed at one, although it is the fixing that is important, not the value at which they are fixed; these variances cannot be free parameters to be estimated.) The z values test whether parameters are significantly different from zero. All have very small associated p-values. Of particular note amongst the parameter estimates is the correlation between “true” ability and “true” aspiration; this is known as a disattenuated correlation and is uncontaminated by measurement error in the observed indicators of the two latent variables. In this case the estimate is 0.666, with a standard error of 0.031. An approximate 95% confidence interval for the disattenuated correlation is [0.606, 0.727]. A path diagram (see Everitt and Dunn 2001) for the correlated, two-factor model is shown in Figure 7.2. Note that the R function path.diagram()“only” produces a textual representation of the graph, here in file ability_sem.dot.
- 226. 210 7 Confirmatory Factor Analysis and Structural Equation Models The graphviz graph visualisation software (Gansner and North 2000) needs to be installed in order to compile the corresponding PDF file. R> path.diagram(ability_sem) Fig. 7.2. Ability path diagram. The fit of the model can be partially judged using the chi-square statistic described in Subsection 7.2.3, which in this case takes the value 9.256 with eight degrees of freedom and an associated p-value of 0.321, suggesting that the postulated model fits the data very well. (The chi-square test for the null model is simply a test that the population covariance matrix of the observed variables is diagonal; i.e., that the observed variables are independent. In most cases, this null model will be rejected; if it is not, the model-fitting exercise is a waste of time.) The various fit indices also indicate that the
- 227. 7.3 Confirmatory factor analysis models 211 model is a good fit for the data. Also helpful in assessing the fit of the model are the summary statistics for the normed residuals, which are essentially the differences between corresponding elements of S and ˆΣ(θ) but scaled so that they are unaffected by the differences in the variances of the observed variables. The normed residuals, r∗ ij, are defined as r∗ ij = sij − ˆσij (ˆσij ˆσj 2 + ˆσ2 ij)/n)1/2 . Generally the absolute values of the normed residuals should all be less than 2 to claim that the current model fits the data well. In the ability and aspiration example, this seems to be the case. (Note that with confirmatory factor models the standard errors of parameters become of importance because they allow the investigator to assess whether parameters might be dropped from the model to find a more parsimonious model that still provides an adequate fit to the data. In exploratory factor analysis, standard errors of factor loadings can be calculated, but they are hardly ever used; instead an informal interpretation of factors is made.) 7.3.2 A confirmatory factor analysis model for drug use For our second example of fitting a confirmatory factor analysis model, we return to the drug use among students data introduced in Chapter 5. In the original investigation of these data reported by Huba et al. (1981), a confirmatory factor analysis model was postulated, the model arising from consideration of previously reported research on student drug use. The model involved the following three latent variables: f1: Alcohol use, with non-zero loadings on beer, wine, spirits, and cigarette use. f2: Cannabis use, with non-zero loadings on marijuana, hashish, cigarette, and wine use. The cigarette variable is assumed to load on both the first and second latent variables because it sometimes occurs with both alcohol and marijuana use and at other times does not. The non-zero loading on wine was allowed because of reports that wine is frequently used with marijuana and that consequently some of the use of wine might be an indicator of tendencies toward cannabis use. f3: Hard drug use, with non-zero loadings on amphetamines, tranquillizers, hallucinogenics, hashish, cocaine, heroin, drug store medication, inhalants, and spirits. The use of each of these substances was considered to suggest a strong commitment to the notion of psychoactive drug use. Each pair of latent variables is assumed to be correlated so that these correlations are allowed to be free parameters that need to be estimated. The variance of each latent variance must, however, be fixed–they are not free parameters that can be estimated–and here as usual we will specify that each of these variances takes the value one. So the proposed model can be specified by the following series of equations:
- 228. 212 7 Confirmatory Factor Analysis and Structural Equation Models cigarettes = λ1f1 + λ2f2 + 0f3 + u1, beer = λ3f1 + 0f2 + 0f3 + u2, wine = λ4f1 + λ5f2 + 0f3 + u3, spirits = λ6f1 + 0f2 + λ7f3 + u4, cocaine = 0f1 + 0f2 + λ8f3 + u5, tranquillizers = 0f1 + 0f2 + λ9f3 + u6, drug store medication = 0f1 + 0f2 + λ10f3 + u7, heroin = 0f1 + 0f2 + λ11f3 + u8, marijuana = 0f1 + λ12f2 + 0f3 + u9, inhalants = 0f1 + 0f2 + λ15f3 + u11, hallucinogenics = 0f1 + 0f2 + λ16f3 + u12, amphetamines = 0f1 + 0f2 + λ17f3 + u13. The proposed model also allows for non-zero correlations between each pair of latent variables and so has a total of 33 parameters to estimate–17 loadings (λ1 to λ17), 13 specific variances (ψ1 to ψ13), and three correlations between latent variables (ρ1 to ρ3). Consequently, the model has 91−33 = 58 degrees of freedom. We first abbreviate the names of the variables via R> rownames(druguse) <- colnames(druguse) <- c("Cigs", + "Beer", "Wine", "Liqr", "Cocn", "Tran", "Drug", + "Hern", "Marj", "Hash", "Inhl", "Hall", "Amph") To fit the model, we can use the R code R> druguse_model <- specify.model(file = "druguse_model.txt") R> druguse_sem <- sem(druguse_model, druguse, 1634) where the model (stored in the text file druguse_model.txt) reads Alcohol -> Cigs, lambda1, NA Alcohol -> Beer, lambda3, NA Alcohol -> Wine, lambda4, NA Alcohol -> Liqr, lambda6, NA Cannabis -> Cigs, lambda2, NA Cannabis -> Wine, lambda5, NA Cannabis -> Marj, lambda12, NA Cannabis -> Hash, lambda13, NA Hard -> Liqr, lambda7, NA Hard -> Cocn, lambda8, NA Hard -> Tran, lambda9, NA Hard -> Drug, lambda10, NA Hard -> Hern, lambda11, NA
- 229. 7.3 Confirmatory factor analysis models 213 Hard -> Hash, lambda14, NA Hard -> Inhl, lambda15, NA Hard -> Hall, lambda16, NA Hard -> Amph, lambda17, NA Cigs <-> Cigs, theta1, NA Beer <-> Beer, theta2, NA Wine <-> Wine, theta3, NA Liqr <-> Liqr, theta4, NA Cocn <-> Cocn, theta5, NA Tran <-> Tran, theta6, NA Drug <-> Drug, theta7, NA Hern <-> Hern, theta8, NA Marj <-> Marj, theta9, NA Hash <-> Hash, theta10, NA Inhl <-> Inhl, theta11, NA Hall <-> Hall, theta12, NA Amph <-> Amph, theta13, NA Alcohol <-> Alcohol, NA, 1 Cannabis <-> Cannabis, NA, 1 Hard <-> Hard, NA, 1 Alcohol <-> Cannabis, rho1, NA Alcohol <-> Hard, rho2, NA Cannabis <-> Hard, rho3, NA The results of fitting the proposed model are R> summary(druguse_sem) Model Chisquare = 324.09 Df = 58 Pr(>Chisq) = 0 Chisquare (null model) = 6613.7 Df = 78 Goodness-of-fit index = 0.9703 Adjusted goodness-of-fit index = 0.9534 RMSEA index = 0.053004 90% CI: (0.047455, 0.058705) Bentler-Bonnett NFI = 0.951 Tucker-Lewis NNFI = 0.94525 Bentler CFI = 0.95929 SRMR = 0.039013 BIC = -105.04 Normalized Residuals Min. 1st Qu. Median Mean 3rd Qu. Max. -3.0500 -0.8800 0.0000 -0.0217 0.9990 4.5800 Parameter Estimates Estimate Std Error z value Pr(>|z|) lambda1 0.35758 0.034332 10.4153 0.0000e+00
- 230. 214 7 Confirmatory Factor Analysis and Structural Equation Models lambda3 0.79159 0.022684 34.8962 0.0000e+00 lambda4 0.87588 0.037963 23.0716 0.0000e+00 lambda6 0.72176 0.023575 30.6150 0.0000e+00 lambda2 0.33203 0.034661 9.5793 0.0000e+00 lambda5 -0.15202 0.037155 -4.0914 4.2871e-05 lambda12 0.91237 0.030833 29.5907 0.0000e+00 lambda13 0.39549 0.030061 13.1559 0.0000e+00 lambda7 0.12347 0.022878 5.3971 6.7710e-08 lambda8 0.46467 0.025954 17.9038 0.0000e+00 lambda9 0.67554 0.024001 28.1468 0.0000e+00 lambda10 0.35842 0.026488 13.5312 0.0000e+00 lambda11 0.47591 0.025813 18.4367 0.0000e+00 lambda14 0.38199 0.029533 12.9343 0.0000e+00 lambda15 0.54297 0.025262 21.4940 0.0000e+00 lambda16 0.61825 0.024566 25.1667 0.0000e+00 lambda17 0.76336 0.023224 32.8695 0.0000e+00 theta1 0.61155 0.023495 26.0284 0.0000e+00 theta2 0.37338 0.020160 18.5210 0.0000e+00 theta3 0.37834 0.023706 15.9597 0.0000e+00 theta4 0.40799 0.019119 21.3398 0.0000e+00 theta5 0.78408 0.029381 26.6863 0.0000e+00 theta6 0.54364 0.023469 23.1644 0.0000e+00 theta7 0.87154 0.031572 27.6051 0.0000e+00 theta8 0.77351 0.029066 26.6126 0.0000e+00 theta9 0.16758 0.044839 3.7374 1.8592e-04 theta10 0.54692 0.022352 24.4691 0.0000e+00 theta11 0.70518 0.027316 25.8159 0.0000e+00 theta12 0.61777 0.025158 24.5551 0.0000e+00 theta13 0.41729 0.021422 19.4797 0.0000e+00 rho1 0.63317 0.028006 22.6079 0.0000e+00 rho2 0.31320 0.029574 10.5905 0.0000e+00 rho3 0.49893 0.027212 18.3349 0.0000e+00 lambda1 Cigs <--- Alcohol lambda3 Beer <--- Alcohol lambda4 Wine <--- Alcohol lambda6 Liqr <--- Alcohol lambda2 Cigs <--- Cannabis lambda5 Wine <--- Cannabis lambda12 Marj <--- Cannabis lambda13 Hash <--- Cannabis lambda7 Liqr <--- Hard lambda8 Cocn <--- Hard lambda9 Tran <--- Hard lambda10 Drug <--- Hard
- 231. 7.3 Confirmatory factor analysis models 215 lambda11 Hern <--- Hard lambda14 Hash <--- Hard lambda15 Inhl <--- Hard lambda16 Hall <--- Hard lambda17 Amph <--- Hard theta1 Cigs <--> Cigs theta2 Beer <--> Beer theta3 Wine <--> Wine theta4 Liqr <--> Liqr theta5 Cocn <--> Cocn theta6 Tran <--> Tran theta7 Drug <--> Drug theta8 Hern <--> Hern theta9 Marj <--> Marj theta10 Hash <--> Hash theta11 Inhl <--> Inhl theta12 Hall <--> Hall theta13 Amph <--> Amph rho1 Cannabis <--> Alcohol rho2 Hard <--> Alcohol rho3 Hard <--> Cannabis Iterations = 31 Here the chi-square test for goodness of fit takes the value 324.092, which with 58 degrees of freedom has an associated p-value that is very small; the model does not appear to fit very well. But before we finally decide that the fitted model is unsuitable for the data, we should perhaps investigate its fit in other ways. Here we will look at the differences of the elements of the observed covariance matrix and the covariance matrix of the fitted model. We can find these differences using the following R code: R> round(druguse_sem$S - druguse_sem$C, 3) Cigs Beer Wine Liqr Cocn Tran Drug Hern Cigs 0.000 -0.002 0.010 -0.009 -0.015 0.015 -0.009 -0.050 Beer -0.002 0.000 0.002 0.002 -0.047 -0.021 0.014 -0.055 Wine 0.010 0.002 0.000 -0.004 -0.039 0.005 0.039 -0.028 Liqr -0.009 0.002 -0.004 0.000 -0.047 0.022 -0.003 -0.069 Cocn -0.015 -0.047 -0.039 -0.047 0.000 0.035 0.042 0.100 Tran 0.015 -0.021 0.005 0.022 0.035 0.000 -0.021 0.034 Drug -0.009 0.014 0.039 -0.003 0.042 -0.021 0.000 0.030 Hern -0.050 -0.055 -0.028 -0.069 0.100 0.034 0.030 0.000 Marj 0.003 -0.012 -0.002 0.009 -0.026 0.007 -0.013 -0.063 Hash -0.023 0.025 0.005 0.029 0.034 -0.014 -0.045 -0.057 Inhl 0.094 0.068 0.075 0.065 0.020 -0.044 0.115 0.030
- 232. 216 7 Confirmatory Factor Analysis and Structural Equation Models Hall -0.071 -0.065 -0.049 -0.077 -0.008 -0.051 0.010 0.026 Amph 0.033 0.010 0.032 0.026 -0.077 0.029 -0.042 -0.049 Marj Hash Inhl Hall Amph Cigs 0.003 -0.023 0.094 -0.071 0.033 Beer -0.012 0.025 0.068 -0.065 0.010 Wine -0.002 0.005 0.075 -0.049 0.032 Liqr 0.009 0.029 0.065 -0.077 0.026 Cocn -0.026 0.034 0.020 -0.008 -0.077 Tran 0.007 -0.014 -0.044 -0.051 0.029 Drug -0.013 -0.045 0.115 0.010 -0.042 Hern -0.063 -0.057 0.030 0.026 -0.049 Marj 0.000 -0.001 0.054 -0.077 0.047 Hash -0.001 0.000 -0.013 0.010 0.025 Inhl 0.054 -0.013 0.000 0.004 -0.022 Hall -0.077 0.010 0.004 0.000 0.039 Amph 0.047 0.025 -0.022 0.039 0.000 Some of these “raw” residuals look quite large in terms of a correlational scale; for example, that corresponding to drug store medication and inhalants. And the summary statistics for the normalised residuals show that the largest is far greater than the acceptable value of 2 and the smallest is rather less than the acceptable value of −2. Perhaps the overall message for the goodness-of-fit measures is that the fitted model does not provide an entirely adequate fit for the relationships between the observed variables. Readers are referred to the original paper by Huba et al. (1981) for details of how the model was changed to try to achieve a better fit. 7.4 Structural equation models Confirmatory factor analysis models are relatively simple examples of a more general framework for modelling latent variables and are known as either structural equation models or covariance structure models. In such models, observed variables are again assumed to be indicators of underlying latent variables, but now regression equations linking the latent variables are incor- porated. Such models are fitted as described in Subsection 7.2.1. We shall illustrate these more complex models by way of a single example. 7.4.1 Stability of alienation To illustrate the fitting of a structural equation model, we shall look at a study reported by Wheaton, Muthen, Alwin, and Summers (1977) concerned with the stability over time of attitudes such as alienation and the relationship of such attitudes to background variables such as education and occupation. For this purpose, data on attitude scales were collected from 932 people in two rural regions in Illinois at three time points, 1966, 1967, and 1971. Here we
- 233. 7.4 Structural equation models 217 Fig. 7.3. Drug use path diagram. shall only consider the data from 1967 and 1971. Scores on the anomia scale and powerlessness scale were taken to be indicators of the assumed latent variable, alienation. A respondent’s years of schooling (education) and Dun- can’s socioeconomic index were assumed to be indicators of a respondent’s socioeconomic status. The correlation matrix for the six observed variables is shown in Figure 7.4. The path diagram for the model to be fitted is shown in Figure 7.5. The latent variable socioeconomic status is considered to affect alienation at both time points, and alienation in 1967 also affects alienation in 1971.
- 234. 218 7 Confirmatory Factor Analysis and Structural Equation Models Educ SEI Anomia71 Powles71 Anomia67 Powles67 Educ SEI Anomia71 Powles71 Anomia67 Powles67 100 54 −35 −37 −36 −41 54 100 −29 −28 −30 −29 −35 −29 100 67 56 47 −37 −28 67 100 44 52 −36 −30 56 44 100 66 −41 −29 47 52 66 100 −1.0 −0.5 0.0 0.5 1.0 Fig. 7.4. Correlation matrix of alienation data; values given are correlation coeffi- cients ×100. The scale of the three latent variables, SES, alienation 67, and alienation 71, are arbitrary and have to be fixed in some way to make the model identi- fiable. Here each is set to the scale of one of its indicator variables by fixing the corresponding regression coefficient to one. Consequently, the equations defining the model to be fitted are as follows: Education = SES + u1, SEI = λ1SES + u2, Anomia67 = Alienation67 + u3, Powerlessness67 = λ2Alienation67 + u4,
- 235. 7.4 Structural equation models 219 Anomia71 = Alienation71 + u5, Powerlessness71 = λ3Alienation71 + u6, Alienation67 = β1SES + u7, Alienation71 = β2SES + β3Alienation67 + u8. In addition to the six regression coefficients in these equations, the model also has to estimate the variances of the eight error terms, u1, . . . , u8, and the variance of the error term for the latent variables, SES. The necessary R code for fitting the model is R> alienation_model <- specify.model( + file = "alienation_model.txt") R> alienation_sem <- sem(alienation_model, alienation, 932) where the model reads SES -> Educ, NA, 1 SES -> SEI, lambda1, NA Alienation67 -> Anomia67, NA, 1 Alienation67 -> Powles67, lambda2, NA Alienation71 -> Anomia71, NA, 1 Alienation71 -> Powles71, lambda3, NA SES -> Alienation67, beta1, NA SES -> Alienation71, beta2, NA Alienation67 -> Alienation71, beta3, NA Educ <-> Educ, theta1, NA SEI <-> SEI, theta2, NA SES <-> SES, delta0, NA Anomia67 <-> Anomia67, theta3, NA Powles67 <-> Powles67, theta4, NA Anomia71 <-> Anomia71, theta5, NA Powles71 <-> Powles71, theta6, NA Alienation67 <-> Alienation67, delta1, NA Alienation71 <-> Alienation71, delta2, NA The parameter estimates are R> summary(alienation_sem) Model Chisquare = 71.532 Df = 6 Pr(>Chisq) = 1.9829e-13 Chisquare (null model) = 2131.5 Df = 15 Goodness-of-fit index = 0.97514 Adjusted goodness-of-fit index = 0.913 RMSEA index = 0.10831 90% CI: (0.086636, 0.13150) Bentler-Bonnett NFI = 0.96644 Tucker-Lewis NNFI = 0.9226 Bentler CFI = 0.96904
- 236. 220 7 Confirmatory Factor Analysis and Structural Equation Models Fig.7.5.Alienationpathdiagram.
- 237. 7.4 Structural equation models 221 SRMR = 0.021256 BIC = 30.508 Normalized Residuals Min. 1st Qu. Median Mean 3rd Qu. Max. -1.2600 -0.2090 -0.0001 -0.0151 0.2440 1.3300 Parameter Estimates Estimate Std Error z value Pr(>|z|) lambda1 5.33054 0.430948 12.3693 0.0000e+00 lambda2 0.88883 0.043229 20.5609 0.0000e+00 lambda3 0.84892 0.041567 20.4229 0.0000e+00 beta1 -0.61361 0.056262 -10.9063 0.0000e+00 beta2 -0.17447 0.054221 -3.2178 1.2920e-03 beta3 0.70463 0.053387 13.1984 0.0000e+00 theta1 2.93614 0.500979 5.8608 4.6064e-09 theta2 260.93220 18.275902 14.2774 0.0000e+00 delta0 6.66392 0.641907 10.3814 0.0000e+00 theta3 4.02305 0.359231 11.1990 0.0000e+00 theta4 3.18933 0.284033 11.2288 0.0000e+00 theta5 3.70315 0.391837 9.4508 0.0000e+00 theta6 3.62334 0.304359 11.9048 0.0000e+00 delta1 5.30685 0.484186 10.9603 0.0000e+00 delta2 3.73998 0.388683 9.6222 0.0000e+00 lambda1 SEI <--- SES lambda2 Powles67 <--- Alienation67 lambda3 Powles71 <--- Alienation71 beta1 Alienation67 <--- SES beta2 Alienation71 <--- SES beta3 Alienation71 <--- Alienation67 theta1 Educ <--> Educ theta2 SEI <--> SEI delta0 SES <--> SES theta3 Anomia67 <--> Anomia67 theta4 Powles67 <--> Powles67 theta5 Anomia71 <--> Anomia71 theta6 Powles71 <--> Powles71 delta1 Alienation67 <--> Alienation67 delta2 Alienation71 <--> Alienation71 Iterations = 84 The value of the chi-square fit statistic is 71.532, which with 6 degrees of freedom suggests that the model does not fit well. J¨oreskog and S¨orbom (1981)
- 238. 222 7 Confirmatory Factor Analysis and Structural Equation Models suggest that the model can be improved by allowing the measurement errors for anomia in 1967 and in 1971 to be correlated. Fitting such a model in R requires the addition of the following line to the code above: Anomia67 <-> Anomia71, psi, NA The chi-square fit statistic is now 6.359 with 5 degrees of freedom. Clearly the introduction of correlated measurement errors for the two measurements of anomia has greatly improved the fit of the model. However, Bentler (1982), in a discussion of this example, suggests that the importance of the structure remaining to be explained after fitting the original model is in practical terms very small, and Browne (1982) criticises the tendency to allow error terms to become correlated simply to obtain an improvement in fit unless there are sound theoretical reasons why particular error terms should be related. 7.5 Summary The possibility of making causal inferences about latent variables is one that has great appeal, particularly for social and behavioural scientists, simply because the concepts in which they are most interested are rarely measurable directly. And because such models can nowadays be relatively easily fitted, researchers can routinely investigate quite complex models. But perhaps a caveat issued more than 20 years ago still has some relevance–the following is from Cliff (1983): Correlational data are still correlational and no computer program can take account of variables that are not in the analysis. Causal relations can only be established through patient, painstaking attention to all the relevant variables, and should involve active manipulation as a final confirmation. The maximum likelihood estimation approach described in this chapter is based on the assumption of multivariate normality for the data. When a multivariate normality assumption is clearly untenable, for example with cate- gorical variables, applying the maximum likelihood methods can lead to biased results, although when there are five or more response categories (and the dis- tribution of the data is normal) the problems from disregarding the categorical nature of variables are likely to be minimised (see Marcoulides 2005). Muthen (1984) describes a very general approach for structural equation modelling that can be used when the data consist of a mixture of continuous, ordinal, and dichotomous variables. Sample size issues for structural equation modelling have been consid- ered by MacCullum, Browne, and Sugawara (1996), and Muthen and Muthen (2002) illustrate how to undertake a Monte Carlo study to help decide on sample size and determine power.
- 239. 7.6 Exercises 223 7.6 Exercises Ex. 7.1 The matrix below shows the correlations between ratings on nine statements about pain made by 123 people suffering from extreme pain. Each statement was scored on a scale from 1 to 6, ranging from agreement to disagreement. The nine pain statements were as follows: 1. Whether or not I am in pain in the future depends on the skills of the doctors. 2. Whenever I am in pain, it is usually because of something I have done or not done. 3. Whether or not I am in pain depends on what the doctors do for me. 4. I cannot get any help for my pain unless I go to seek medical advice. 5. When I am in pain, I know that it is because I have not been taking proper exercise or eating the right food. 6. People’s pain results from their own carelessness. 7. I am directly responsible for my pain, 8. relief from pain is chiefly controlled by the doctors. 9. People who are never in pain are just plain lucky. 1.00 −0.04 1.00 0.61 −0.07 1.00 0.45 −0.12 0.59 1.00 0.03 0.49 0.03 −0.08 1.00 −0.29 0.43 −0.13 −0.21 0.47 1.00 −0.30 0.30 −0.24 −0.19 0.41 0.63 1.00 0.45 −0.31 0.59 0.63 −0.14 −0.13 −0.26 1.00 0.30 −0.17 0.32 0.37 −0.24 −0.15 −0.29 0.40 1.00 . Fit a correlated two-factor model in which questions 1, 3, 4, and 8 are assumed to be indicators of the latent variable Doctor’s Responsibility and questions 2, 5, 6, and 7 are assumed to be indicators of the latent variable Patient’s Responsibility. Find a 95% confidence interval for the correlation between the two latent variables. Ex. 7.2 For the stability of alienation example, fit the model in which the measurement errors for anomia in 1967 and anomia in 1971 are allowed to be correlated. Ex. 7.3 Meyer and Bendig (1961) administered the five Thurstone Primary Mental Ability tests, verbal meaning (V), space (S), reasoning (R), nu- merical (N), and word fluency (W), to 49 boys and 61 girls in grade 8 and again three and a half years later in grade 11. The observed correlation matrix is shown below. Fit a single-factor model to the correlations that allows the factor at time one to be correlated with the factor at time two.
- 240. 224 7 Confirmatory Factor Analysis and Structural Equation Models V S R N W V S R N W V 1.00 S 0.37 1.00 R 0.42 0.33 1.00 N 0.53 0.14 0.38 1.00 W 0.38 0.10 0.20 0.24 1.00 V 0.81 0.34 0.49 0.58 0.32 1.00 S 0.35 0.65 0.20 −0.04 0.11 0.34 1.00 R 0.42 0.32 0.75 0.46 0.26 0.46 0.18 1.00 N 0.40 0.14 0.39 0.73 0.19 0.55 0.06 0.54 1.00 W 0.24 0.15 0.17 0.15 0.43 0.24 0.15 0.20 0.16 1.00
- 241. 8 The Analysis of Repeated Measures Data 8.1 Introduction The multivariate data sets considered in previous chapters have involved mea- surements or observations on a number of different variables for each object or individual in the study. In this chapter, however, we will consider multivariate data of a different nature, namely data resulting from the repeated measure- ments of the same variable on each unit in the data set. Examples of such data are common in many disciplines. But before we introduce some actual repeated measures data sets, we need to make a small digression in order to introduce the two different “formats”, the wide and the long forms, in which such data are commonly stored and dealt with. The simplest way to do this is by way of a small example data set: R> ex_wide ID Group Day.1 Day.2 Day.5 Day.7 1 1 1 15 15 10 7 2 2 1 10 9 11 12 3 3 1 8 7 6 9 4 4 2 11 8 13 7 5 5 2 11 12 11 11 6 6 2 12 12 6 10 We can pretend that these data come from a clinical trial in which individuals have been assigned to two treatment groups and have a response variable of interest recorded on days 1, 2, 5 and 7. As given in this table, the data are in their wide form (as are most of the data sets met in earlier chapters); each row of data corresponds to an individual and contains all the repeated measures for the individual as well as other variables that might have been recorded–in this case the treatment group of the individual. These data can be put into their long form, in which each row of data now corresponds to one of the repeated measurements along with the values of other variables associated with this DOI 10.1007/978-1-4419-9650-3_8, © Springer Science+Business Media, LLC 2011 225B. Everitt and T. Hothorn, An Introduction to Applied Multivariate Analysis with R: Use R!,
- 242. 226 8 The Analysis of Repeated Measures Data particular measurement (for example, the time the measurement was taken) by using the reshape() function in R. The code needed to rearrange the data ex_wide into their long form is as follows: R> reshape(ex_wide, direction = "long", idvar = "ID", + varying = colnames(ex_wide)[-(1:2)]) ID Group time Day 1.1 1 1 1 15 2.1 2 1 1 10 3.1 3 1 1 8 4.1 4 2 1 11 5.1 5 2 1 11 6.1 6 2 1 12 1.2 1 1 2 15 2.2 2 1 2 9 3.2 3 1 2 7 4.2 4 2 2 8 5.2 5 2 2 12 6.2 6 2 2 12 1.5 1 1 5 10 2.5 2 1 5 11 3.5 3 1 5 6 4.5 4 2 5 13 5.5 5 2 5 11 6.5 6 2 5 6 1.7 1 1 7 7 2.7 2 1 7 12 3.7 3 1 7 9 4.7 4 2 7 7 5.7 5 2 7 11 6.7 6 2 7 10 Here, varying contains the names of the variables containing the repeated measurements. Note that the name of the variable consists of the name itself and the time point, separated by a dot. This long form of repeated measures data is used when applying the models to be described in Section 8.2, although the wide format is often more convenient for plotting the data and computing summary statistics. So now let us take a look at two repeated measurement data sets that we shall be concerned with in this chapter. The first, shown in its long form in Table 8.1, is taken from Crowder (1998) and gives the loads required to produce slippage x of a timber specimen in a clamp. There are eight speci- mens each with 15 repeated measurements. The second data set, in Table 8.2 reported in Zerbe (1979) and also given in Davis (2003), consists of plasma inorganic phosphate measurements obtained from 13 control and 20 obese pa- tients 0, 0.5, 1, 1.5, 2, and 3 hours after an oral glucose challenge. These two
- 243. 8.1 Introduction 227 data sets illustrate that although repeated measurements often arise from the passing of time (longitudinal data), this is not always the case. Table 8.1: timber data. Data giving loads needed for a given slip- page in eight specimens of timber, with data in “long” form. specimen slippage loads specimen slippage loads spec1 0.0 0.00 spec5 0.7 12.25 spec2 0.0 0.00 spec6 0.7 12.85 spec3 0.0 0.00 spec7 0.7 12.13 spec4 0.0 0.00 spec8 0.7 12.21 spec5 0.0 0.00 spec1 0.8 12.98 spec6 0.0 0.00 spec2 0.8 14.08 spec7 0.0 0.00 spec3 0.8 13.18 spec8 0.0 0.00 spec4 0.8 12.23 spec1 0.1 2.38 spec5 0.8 13.35 spec2 0.1 2.69 spec6 0.8 13.83 spec3 0.1 2.85 spec7 0.8 13.15 spec4 0.1 2.46 spec8 0.8 13.16 spec5 0.1 2.97 spec1 0.9 13.94 spec6 0.1 3.96 spec2 0.9 14.66 spec7 0.1 3.17 spec3 0.9 14.12 spec8 0.1 3.36 spec4 0.9 13.24 spec1 0.2 4.34 spec5 0.9 14.54 spec2 0.2 4.75 spec6 0.9 14.85 spec3 0.2 4.89 spec7 0.9 14.09 spec4 0.2 4.28 spec8 0.9 14.05 spec5 0.2 4.68 spec1 1.0 14.74 spec6 0.2 6.46 spec2 1.0 15.37 spec7 0.2 5.33 spec3 1.0 15.09 spec8 0.2 5.45 spec4 1.0 14.19 spec1 0.3 6.64 spec5 1.0 15.53 spec2 0.3 7.04 spec6 1.0 15.79 spec3 0.3 6.61 spec7 1.0 15.11 spec4 0.3 5.88 spec8 1.0 14.96 spec5 0.3 6.66 spec1 1.2 16.13 spec6 0.3 8.14 spec2 1.2 16.89 spec7 0.3 7.14 spec3 1.2 16.68 spec8 0.3 7.08 spec4 1.2 16.07 spec1 0.4 8.05 spec5 1.2 17.38 spec2 0.4 9.20 spec6 1.2 17.39 spec3 0.4 8.09 spec7 1.2 16.69 spec4 0.4 7.43 spec8 1.2 16.24 spec5 0.4 8.11 spec1 1.4 17.98 spec6 0.4 9.35 spec2 1.4 17.78
- 244. 228 8 The Analysis of Repeated Measures Data Table 8.1: timber data (continued). specimen slippage loads specimen slippage loads spec7 0.4 8.29 spec3 1.4 17.94 spec8 0.4 8.32 spec4 1.4 17.43 spec1 0.5 9.78 spec5 1.4 18.76 spec2 0.5 10.94 spec6 1.4 18.44 spec3 0.5 9.72 spec7 1.4 17.69 spec4 0.5 8.32 spec8 1.4 17.34 spec5 0.5 9.64 spec1 1.6 19.52 spec6 0.5 10.72 spec2 1.6 18.41 spec7 0.5 9.86 spec3 1.6 18.22 spec8 0.5 9.91 spec4 1.6 18.36 spec1 0.6 10.97 spec5 1.6 19.81 spec2 0.6 12.23 spec6 1.6 19.46 spec3 0.6 11.03 spec7 1.6 18.71 spec4 0.6 9.92 spec8 1.6 18.23 spec5 0.6 11.06 spec1 1.8 19.97 spec6 0.6 11.84 spec2 1.8 18.97 spec7 0.6 11.07 spec3 1.8 19.40 spec8 0.6 11.06 spec4 1.8 18.93 spec1 0.7 12.05 spec5 1.8 20.62 spec2 0.7 13.19 spec6 1.8 20.05 spec3 0.7 12.14 spec7 1.8 19.54 spec4 0.7 11.10 spec8 1.8 18.87 Table 8.2: plasma data. Plasma inorganic phosphate levels from 33 subjects, with data in “long” form. Subject group time plasma Subject group time plasma id01 control 1 4.3 id01 control 5 2.2 id02 control 1 3.7 id02 control 5 2.9 id03 control 1 4.0 id03 control 5 2.9 id04 control 1 3.6 id04 control 5 2.9 id05 control 1 4.1 id05 control 5 3.6 id06 control 1 3.8 id06 control 5 3.8 id07 control 1 3.8 id07 control 5 3.1 id08 control 1 4.4 id08 control 5 3.6 id09 control 1 5.0 id09 control 5 3.3 id10 control 1 3.7 id10 control 5 1.5 id11 control 1 3.7 id11 control 5 2.9 id12 control 1 4.4 id12 control 5 3.7
- 245. 8.1 Introduction 229 Table 8.2: plasma data (continued). Subject group time plasma Subject group time plasma id13 control 1 4.7 id13 control 5 3.2 id14 control 1 4.3 id14 control 5 2.2 id15 control 1 5.0 id15 control 5 3.7 id16 control 1 4.6 id16 control 5 3.7 id17 control 1 4.3 id17 control 5 3.1 id18 control 1 3.1 id18 control 5 2.6 id19 control 1 4.8 id19 control 5 2.2 id20 control 1 3.7 id20 control 5 2.9 id21 obese 1 5.4 id21 obese 5 2.8 id22 obese 1 3.0 id22 obese 5 2.1 id23 obese 1 4.9 id23 obese 5 3.7 id24 obese 1 4.8 id24 obese 5 4.7 id25 obese 1 4.4 id25 obese 5 3.5 id26 obese 1 4.9 id26 obese 5 3.3 id27 obese 1 5.1 id27 obese 5 3.4 id28 obese 1 4.8 id28 obese 5 4.1 id29 obese 1 4.2 id29 obese 5 3.3 id30 obese 1 6.6 id30 obese 5 4.3 id31 obese 1 3.6 id31 obese 5 2.1 id32 obese 1 4.5 id32 obese 5 2.4 id33 obese 1 4.6 id33 obese 5 3.8 id01 control 2 3.3 id01 control 6 2.5 id02 control 2 2.6 id02 control 6 3.2 id03 control 2 4.1 id03 control 6 3.1 id04 control 2 3.0 id04 control 6 3.9 id05 control 2 3.8 id05 control 6 3.4 id06 control 2 2.2 id06 control 6 3.6 id07 control 2 3.0 id07 control 6 3.4 id08 control 2 3.9 id08 control 6 3.8 id09 control 2 4.0 id09 control 6 3.6 id10 control 2 3.1 id10 control 6 2.3 id11 control 2 2.6 id11 control 6 2.2 id12 control 2 3.7 id12 control 6 4.3 id13 control 2 3.1 id13 control 6 4.2 id14 control 2 3.3 id14 control 6 2.5 id15 control 2 4.9 id15 control 6 4.1 id16 control 2 4.4 id16 control 6 4.2 id17 control 2 3.9 id17 control 6 3.1 id18 control 2 3.1 id18 control 6 1.9 id19 control 2 5.0 id19 control 6 3.1 id20 control 2 3.1 id20 control 6 3.6 id21 obese 2 4.7 id21 obese 6 3.7
- 246. 230 8 The Analysis of Repeated Measures Data Table 8.2: plasma data (continued). Subject group time plasma Subject group time plasma id22 obese 2 2.5 id22 obese 6 2.6 id23 obese 2 5.0 id23 obese 6 4.1 id24 obese 2 4.3 id24 obese 6 3.7 id25 obese 2 4.2 id25 obese 6 3.4 id26 obese 2 4.3 id26 obese 6 4.1 id27 obese 2 4.1 id27 obese 6 4.2 id28 obese 2 4.6 id28 obese 6 4.0 id29 obese 2 3.5 id29 obese 6 3.1 id30 obese 2 6.1 id30 obese 6 3.8 id31 obese 2 3.4 id31 obese 6 2.4 id32 obese 2 4.0 id32 obese 6 2.3 id33 obese 2 4.4 id33 obese 6 3.6 id01 control 3 3.0 id01 control 7 3.4 id02 control 3 2.6 id02 control 7 3.1 id03 control 3 3.1 id03 control 7 3.9 id04 control 3 2.2 id04 control 7 3.8 id05 control 3 2.1 id05 control 7 3.6 id06 control 3 2.0 id06 control 7 3.0 id07 control 3 2.4 id07 control 7 3.5 id08 control 3 2.8 id08 control 7 4.0 id09 control 3 3.4 id09 control 7 4.0 id10 control 3 2.9 id10 control 7 2.7 id11 control 3 2.6 id11 control 7 3.1 id12 control 3 3.1 id12 control 7 3.9 id13 control 3 3.2 id13 control 7 3.7 id14 control 3 3.0 id14 control 7 2.4 id15 control 3 4.1 id15 control 7 4.7 id16 control 3 3.9 id16 control 7 4.8 id17 control 3 3.1 id17 control 7 3.6 id18 control 3 3.3 id18 control 7 2.3 id19 control 3 2.9 id19 control 7 3.5 id20 control 3 3.3 id20 control 7 4.3 id21 obese 3 3.9 id21 obese 7 3.5 id22 obese 3 2.3 id22 obese 7 3.2 id23 obese 3 4.1 id23 obese 7 4.7 id24 obese 3 4.7 id24 obese 7 3.6 id25 obese 3 4.2 id25 obese 7 3.8 id26 obese 3 4.0 id26 obese 7 4.2 id27 obese 3 4.6 id27 obese 7 4.4 id28 obese 3 4.6 id28 obese 7 3.8 id29 obese 3 3.8 id29 obese 7 3.5 id30 obese 3 5.2 id30 obese 7 4.2
- 247. 8.1 Introduction 231 Table 8.2: plasma data (continued). Subject group time plasma Subject group time plasma id31 obese 3 3.1 id31 obese 7 2.5 id32 obese 3 3.7 id32 obese 7 3.1 id33 obese 3 3.8 id33 obese 7 3.8 id01 control 4 2.6 id01 control 8 4.4 id02 control 4 1.9 id02 control 8 3.9 id03 control 4 2.3 id03 control 8 4.0 id04 control 4 2.8 id04 control 8 4.0 id05 control 4 3.0 id05 control 8 3.7 id06 control 4 2.6 id06 control 8 3.5 id07 control 4 2.5 id07 control 8 3.7 id08 control 4 2.1 id08 control 8 3.9 id09 control 4 3.4 id09 control 8 4.3 id10 control 4 2.2 id10 control 8 2.8 id11 control 4 2.3 id11 control 8 3.9 id12 control 4 3.2 id12 control 8 4.8 id13 control 4 3.3 id13 control 8 4.3 id14 control 4 2.6 id14 control 8 3.4 id15 control 4 3.7 id15 control 8 4.9 id16 control 4 3.9 id16 control 8 5.0 id17 control 4 3.1 id17 control 8 4.0 id18 control 4 2.6 id18 control 8 2.7 id19 control 4 2.8 id19 control 8 3.6 id20 control 4 2.8 id20 control 8 4.4 id21 obese 4 4.1 id21 obese 8 3.7 id22 obese 4 2.2 id22 obese 8 3.5 id23 obese 4 3.7 id23 obese 8 4.9 id24 obese 4 4.6 id24 obese 8 3.9 id25 obese 4 3.4 id25 obese 8 4.0 id26 obese 4 4.0 id26 obese 8 4.3 id27 obese 4 4.1 id27 obese 8 4.9 id28 obese 4 4.4 id28 obese 8 3.8 id29 obese 4 3.6 id29 obese 8 3.9 id30 obese 4 4.1 id30 obese 8 4.8 id31 obese 4 2.8 id31 obese 8 3.5 id32 obese 4 3.3 id32 obese 8 3.3 id33 obese 4 3.8 id33 obese 8 3.8 The distinguishing feature of a repeated measures study is that the re- sponse variable of interest has been recorded several times on each unit in the data set. In addition, a set of explanatory variables (covariates is an alterna-
- 248. 232 8 The Analysis of Repeated Measures Data tive term that is often used in this context) are available for each; some of the explanatory variables may have been recorded only once for each unit and so take the same value for each of the repeated response values for that unit– an example would be treatment group in a clinical trial. Other explanatory variables may take different values for each of the different response variable values–an example would be age; these are sometimes called time-varying co- variates. The main objective in such a study is to characterise change in the repeated values of the response variable and to determine the explanatory variables most associated with any change. Because several observations of the response variable are made on the same individual, it is likely that the measurements will be correlated rather than independent, even after conditioning on the explanatory variables. Consequently, repeated measures data require special methods of analysis, and models for such data need to include parameters linking the explanatory variables to the repeated measurements, parameters analogous to those in the usual multiple regression model, and, in addition, parameters that account for the correlational structure of the repeated mea- surements. It is the former parameters that are generally of most interest, with the latter often being regarded as nuisance parameters. But providing an adequate model for the correlational structure of the repeated measures is necessary to avoid misleading inferences about the parameters that are of most importance to the researcher. Over the last decade, methodology for the analysis of repeated measures data has been the subject of much research and development, and there are now a variety of powerful techniques available. Comprehensive accounts of these methods are given in Diggle, Heagerty, Liang, and Zeger (2003), Davis (2003), and Skrondal and Rabe-Hesketh (2004). Here we will concentrate on a single class of methods, linear mixed-effects models. 8.2 Linear mixed-effects models for repeated measures data Linear mixed-effects models for repeated measures data formalise the sensible idea that an individual’s pattern of responses is likely to depend on many characteristics of that individual, including some that are unobserved. These unobserved variables are then included in the model as random variables, that is, random effects. The essential feature of the model is that correlation amongst the repeated measurements on the same unit arises from shared, unobserved variables. Conditional on the values of the random effects, the repeated measurements are assumed to be independent, the so-called local independence assumption. Linear mixed-effects models are introduced in the next subsection by de- scribing two commonly used models, the random intercept and random inter- cept and slope models, in the context of the timber slippage data in Table 8.1.
- 249. 8.2 Linear mixed-effects models for repeated measures data 233 8.2.1 Random intercept and random intercept and slope models for the timber slippage data Let yij represent the load in specimen i needed to produce a slippage of xj, with i = 1, . . . , 8 and j = 1, . . . , 15. If we choose to ignore the repeated measures structure of the, data we could fit a simple linear regression model of the form yij = β0 + β1xj + ij. (8.1) The model in Equation (8.1) can be fitted using the long form of the data in association with the lm() function in R as follows: R> summary(lm(loads ~ slippage, data = timber)) Call: lm(formula = loads ~ slippage, data = timber) Residuals: Min 1Q Median 3Q Max -3.516 -0.981 0.408 1.298 2.491 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 3.516 0.264 13.3 <2e-16 slippage 10.373 0.283 36.6 <2e-16 Residual standard error: 1.65 on 118 degrees of freedom Multiple R-squared: 0.919, Adjusted R-squared: 0.918 F-statistic: 1.34e+03 on 1 and 118 DF, p-value: <2e-16 We see that the slippage effect is large and highly significant. But such a model assumes that the repeated observations are independent of one another which is totally unrealistic for most repeated measures data sets. A more realistic model is the random intercept model, where by partition- ing the total residual that is present in the usual linear regression model (8.1) into a subject-specific random component ui that is constant over slippage plus an error term ij that varies randomly over slippage, some correlational struc- ture for the repeated measures is introduced. The random intercept model is yij = (β0 + ui) + β1xj + ij. (8.2) (The error terms and parameters in Equations (8.1) and (8.2) are, of course, not the same.) The ui are assumed to be normally distributed with zero mean and variance σ2 u, and the ij are assumed normally distributed with zero mean and variance σ2 . The ui and the ij are assumed to be independent of each other and the xj. The subject-specific random effects allow for differences in the intercepts of each individual’s regression fit of load on slippage. The repeated measurements for each timber specimen vary about that specimen’s
- 250. 234 8 The Analysis of Repeated Measures Data own regression line, and this can differ in intercept but not in slope from the regression lines of other specimens. In the random-effects model, there is possible heterogeneity in the intercepts of the individuals. In this model, slippage has a fixed effect. How does the random intercept model introduce a correlational structure for the repeated measurements? First, the random intercept model implies that the total variance of each repeated measurement is Var(ui + ij) = σ2 u + σ2 . (Due to this decomposition of the total residual variance into a between- subject component, σ2 u, and a within-subject component, σ2 , the model is sometimes referred to as a variance component model.) The covariance be- tween the total residuals at two slippage levels j and j in the same specimen i is Cov(ui + ij, ui + ij ) = σ2 u. The covariance will be non-zero if the variance of the subject-specific random effects is non-zero. (Note that this covariance is induced by the shared random intercept; for specimens with ui > 0, the total residuals will tend to be greater than the mean, and for specimens with ui < 0 they will tend to be less than the mean.) It follows from the two relations above that the residual correlation (i.e., the correlation between pairs of repeated measurements) is given by Cor(ui + ij, ui + ij ) = σ2 u σ2 u + σ2 . This is an intra-class correlation that is interpreted as the proportion of the total residual variance that is due to residual variability between subjects. So a random intercept model constrains the variance of each repeated measure to be the same and the covariance between any pair of repeated measurements to be equal. This is usually called the compound symmetry structure. These constraints often are not realistic for repeated measures data. For example, for longitudinal data, it is more common for measures taken closer to each other in time to be more highly correlated than those taken further apart. In addition, the variances of the later repeated measures are often greater than those taken earlier. Consequently, for many such data sets, the random intercept model will not do justice to the observed pattern of covariances between the repeated measures. A model that allows a more realistic structure for the covariances is the random slope and intercept model that provides for heterogeneity in both slopes and intercepts. In this model, there are two types of random effects, the first modelling heterogeneity in intercepts, ui1, and the second modelling heterogeneity in slopes, ui2. Explicitly, the model is yij = (β0 + ui1) + (β1 + ui2)xj + ij, (8.3)
- 251. 8.2 Linear mixed-effects models for repeated measures data 235 where the parameters are not, of course, the same as in (8.1). The two random effects are assumed to have a bivariate normal distribution with zero means for both variables, variances σ2 u1 and σ2 u2 , and covariance σu1u2 . With this model, the total residual is ui1 + ui2xj + ij with variance Var(ui1 + ui2xj + ij) = σ2 u1 + 2σu1u2 xj + σ2 u2 x2 j + σ2 , which is no longer constant for different values of xj. Similarly, the covariance between two total residuals of the same individual, Cov(ui1 + xjui2 + ij, ui1 + xj ui2 + ij ) = σ2 u1 + σu1u2 (xj + xj ) + σ2 u2 xjxj , is not constrained to be the same for all pairs j and j . The random intercept and slope model allows for the repeated measurements to have different vari- ances and for pairs of repeated measurements to have different correlations. Linear mixed-effects models can be estimated by maximum likelihood. However, this method tends to underestimate the variance components. A modified version of maximum likelihood, known as restricted maximum like- lihood, is therefore often recommended; this provides consistent estimates of the variance components. Details are given in Diggle et al. (2003), Longford (1993), and Skrondal and Rabe-Hesketh (2004). Competing linear mixed-effects models can be compared using a likelihood ratio test. If, however, the models have been estimated by restricted maximum likelihood, this test can only be used if both models have the same set of fixed effects (Longford 1993). 8.2.2 Applying the random intercept and the random intercept and slope models to the timber slippage data Before beginning any formal model-fitting exercise, it is good data analysis practise to look at some informative graphic (or graphics) of the data. Here we first produce a plot of the trajectories of each timber specimen over the slippage levels; see Figure 8.1. The figure shows that loads required to achieve a given slippage level in- crease with slippage value, with the increase gradually leveling off; this ex- plains the large slippage effect found earlier when applying simple linear re- gression to the data. We can now fit the two linear mixed-effects models (8.2) and (8.3) as described in the previous subsection using the lme() function from the package nlme (Pinheiro, Bates, DebRoy, Sarkar, and R Development Core Team 2010): R> timber.lme <- lme(loads ~ slippage, + random = ~1 | specimen, + data = timber, method = "ML") R> timber.lme1 <- lme(loads ~ slippage, + random = ~slippage | specimen, + data = timber, method = "ML")
- 252. 236 8 The Analysis of Repeated Measures Data R> xyplot(loads ~ slippage | specimen, data = timber, + layout = c(4, 2)) slippage loads 0 5 10 15 20 0.0 0.5 1.0 1.5 q q q q q q q q q q q q q qq spec1 q q q q q q q q qq q q qqq spec2 0.0 0.5 1.0 1.5 q q q q q q q q q q q q qq q spec3 q q q q q q q q q q q q q qq spec4 q q q q q q q q q q q q q q q spec5 0.0 0.5 1.0 1.5 q q q q q q q q q q q q q qq spec6 q q q q q q q q q q q q q q q spec7 0.0 0.5 1.0 1.5 0 5 10 15 20 q q q q q q q q q q q q q qq spec8 Fig. 8.1. Plot of the eight specimens in timber data. Similar to a linear model fitted using lm(), the lme() function takes a model formula describing the response and exploratory variables, here associated with fixed effects. In addition, a second formula is needed (specified via the random argument) that specifies the random effects. The first model assigns a random intercept to each level of specimen, while the second model fits a random intercept and slope parameter in variable slippage for each level of specimen. First we can test the random intercept model against the simple linear regression model via a likelihood ratio test for zero variances of the random- effect parameters, which is applied by the exactLRT() function from RLRsim (Scheipl 2010, Scheipl, Greven, and K¨uchenhoff 2008):
- 253. 8.2 Linear mixed-effects models for repeated measures data 237 R> library("RLRsim") R> exactRLRT(timber.lme) Using restricted likelihood evaluated at ML estimators. Refit with method="REML" for exact results. simulated finite sample distribution of RLRT. (p-value based on 10000 simulated values) data: RLRT = 0, p-value = 0.4419 We see that the random intercept model does not improve upon the simple linear model. Nevertheless, we continue to work with the mixed-effects models simply because the “observations”, due to their repeated measurement struc- ture, cannot be assumed to be independent. Now we can do the same for the two random-effect models R> anova(timber.lme, timber.lme1) Model df AIC BIC logLik Test L.Ratio timber.lme 1 4 466.7 477.8 -229.3 timber.lme1 2 6 470.7 487.4 -229.3 1 vs 2 1.497e-08 p-value timber.lme timber.lme1 1 The p-value associated with the likelihood ratio test is very small, indicating that the random intercept and slope model is to be preferred over the simpler random intercept model for these data. The results from this model are: R> summary(timber.lme1) Linear mixed-effects model fit by maximum likelihood Data: timber AIC BIC logLik 470.7 487.4 -229.3 Random effects: Formula: ~slippage | specimen Structure: General positive-definite, Log-Cholesky param. StdDev Corr (Intercept) 3.106e-04 (Intr) slippage 1.028e-05 0 Residual 1.636e+00 Fixed effects: loads ~ slippage Value Std.Error DF t-value p-value
- 254. 238 8 The Analysis of Repeated Measures Data (Intercept) 3.516 0.2644 111 13.30 0 slippage 10.373 0.2835 111 36.59 0 Correlation: (Intr) slippage -0.822 Standardized Within-Group Residuals: Min Q1 Med Q3 Max -2.1492 -0.5997 0.2491 0.7936 1.5226 Number of Observations: 120 Number of Groups: 8 (Determining the degrees of freedom for the t-value given in this output is not always easy except in special cases where the data are balanced and the model for the mean has a relatively simple form; consequently, an approximation is described in Kenward and Roger 1997) The regression coefficient for slippage is highly significant. We can find the predicted values under this model and then plot them alongside a plot of the raw data using the following R code: R> timber$pred1 <- predict(timber.lme1) The resulting plot is shown in Figure 8.2. Clearly the fit is not good. In fact, under the random intercept and slope model, the predicted values for each specimen are almost identical, reflecting the fact that the estimated variances of both random effects are essentially zero. The plot of the observed values in Figure 8.2 shows a leveling-off of the increase in load needed to achieve a given slippage level as slippage increases; i.e., it suggests that a quadratic term in slippage is essential in any model for these data. Including this as a fixed effect, the required model is yij = (β0 + ui1) + (β1 + ui2)xj + β2x2 j + ij. (8.4) The necessary R code to fit this model and test it against the previous random intercept and slope model is R> timber.lme2 <- lme(loads ~ slippage + I(slippage^2), + random = ~slippage | specimen, + data = timber, method = "ML") R> anova(timber.lme1, timber.lme2) Model df AIC BIC logLik Test L.Ratio p-value timber.lme1 1 6 470.7 487.4 -229.33 timber.lme2 2 7 213.5 233.1 -99.77 1 vs 2 259.1 <.0001 The p-value from the likelihood ratio test is less than 0.0001, indicating that the model that includes a quadratic term does provide a much improved fit. Both the linear and quadratic effects of slippage are highly significant.
- 255. 8.2 Linear mixed-effects models for repeated measures data 239 R> pfun <- function(x, y) { + panel.xyplot(x, y[1:length(x)]) + panel.lines(x, y[1:length(x) + length(x)], lty = 1) + } R> plot(xyplot(cbind(loads, pred1) ~ slippage | specimen, + data = timber, panel = pfun, layout = c(4, 2), + ylab = "loads")) slippage loads 0 5 10 15 20 0.0 0.5 1.0 1.5 q q q q q q q q q q q q q qq spec1 q q q q q q q q qqq q qqq spec2 0.0 0.5 1.0 1.5 q q q q q q q q q q q q qq q spec3 q q q q q q q q q q q q q qq spec4 q q q q q q q q q q q q q q q spec5 0.0 0.5 1.0 1.5 q q q q q q q q q q q q q qq spec6 q q q q q q q q q q q q q q q spec7 0.0 0.5 1.0 1.5 0 5 10 15 20 q q q q q q q q q q q q q qq spec8 Fig. 8.2. Plot of the predicted values of the random intercept model for the timber data. We can now produce a plot similar to that in Figure 8.2 but showing the predicted values from the model in (8.4). The code is similar to that given above and so is not repeated again here. The resulting plot is shown in Figure 8.3. Clearly the model describes the data more satisfactorily, although there remains an obvious problem, which is taken up in Exercise 8.1.
- 256. 240 8 The Analysis of Repeated Measures Data slippage loads 0 5 10 15 20 0.0 0.5 1.0 1.5 q q q q q q q q q q q q q qq spec1 q q q q q q q q qq q q qqq spec2 0.0 0.5 1.0 1.5 q q q q q q q q q q q q qq q spec3 q q q q q q q q q q q q q qq spec4 q q q q q q q q q q q q q q q spec5 0.0 0.5 1.0 1.5 q q q q q q q q q q q q q qq spec6 q q q q q q q q q q q q q q q spec7 0.0 0.5 1.0 1.5 0 5 10 15 20 q q q q q q q q q q q q q qq spec8 Fig. 8.3. Plot of the predicted values of the random intercept model including a quadratic term for the timber data. 8.2.3 Fitting random-effect models to the glucose challenge data Now we can move on to consider the glucose challenge data given in its long form in Table 8.2. Again we will begin by plotting the data so that we get some ideas as to what form of linear mixed-effect model might be appropriate. First we plot the raw data separately for the control and the obese groups in Figure 8.4. First, we transform the plasma data from long into wide form and apply the parallel() function from package lattice to set up a parallel- coordinates plot. The profiles in both groups show some curvature, suggesting that a quadratic effect of time may be needed in any model. There also appears to be some difference in the shapes of the curves in the two groups, suggesting perhaps the need to consider a group × time interaction. Next we plot the scatterplot matrices of the repeated measurements for the two groups using the code in Figure 8.5.
- 257. 8.2 Linear mixed-effects models for repeated measures data 241 R> x <- reshape(plasma, direction = "wide", timevar = "time", + idvar = "Subject", v.names = "plasma") R> parallel(~ x[,-(1:2)] | group, data = x, horizontal = FALSE, + col = "black", scales = list(x = list(labels = 1:8)), + ylab = "Plasma inorganic phosphate", + xlab = "Time (hours after oral glucose challenge)") Time (hours after oral glucose challenge) Plasmainorganicphosphate Min Max 1 2 3 4 5 6 7 8 control 1 2 3 4 5 6 7 8 obese Fig. 8.4. Glucose challenge data for control and obese groups over time. Each line represents the trajectory of one individual. The plot indicates that the correlations of pairs of measurements made at different times differ so that the compound symmetry structure for these correlations is unlikely to be appropriate. On the basis of these plots, we will begin by fitting the model in (8.4) with the addition, in this case, of an extra covariate, namely a dummy variable coding the group, control or obese, to which a subject belongs. We can fit the required model using R> plasma.lme1 <- lme(plasma ~ time + I(time^2) + group, + random = ~ time | Subject, + data = plasma, method = "ML") R> summary(plasma.lme1)
- 258. 242 8 The Analysis of Repeated Measures Data R> plot(splom(~ x[, grep("plasma", colnames(x))] | group, data = x, + cex = 1.5, pch = ".", pscales = NULL, varnames = 1:8)) Scatter Plot Matrix 1 2 3 4 5 6 7 8 control 1 2 3 4 5 6 7 8 obese Fig. 8.5. Scatterplot matrix for glucose challenge data. Linear mixed-effects model fit by maximum likelihood Data: plasma AIC BIC logLik 390.5 419.1 -187.2 Random effects: Formula: ~time | Subject Structure: General positive-definite, Log-Cholesky param. StdDev Corr (Intercept) 0.69772 (Intr) time 0.09383 -0.7 Residual 0.38480 Fixed effects: plasma ~ time + I(time^2) + group Value Std.Error DF t-value p-value (Intercept) 4.880 0.17091 229 28.552 0.0000 time -0.803 0.05075 229 -15.827 0.0000 I(time^2) 0.085 0.00521 229 16.258 0.0000 groupobese 0.437 0.18589 31 2.351 0.0253 Correlation:
- 259. 8.2 Linear mixed-effects models for repeated measures data 243 (Intr) time I(t^2) time -0.641 I(time^2) 0.457 -0.923 groupobese -0.428 0.000 0.000 Standardized Within-Group Residuals: Min Q1 Med Q3 Max -2.771508 -0.548688 -0.002765 0.564435 2.889633 Number of Observations: 264 Number of Groups: 33 The regression coefficients for linear and quadratic time are both highly significant. The group effect is also significant, and an asymptotic 95% confi- dence interval for the group effect is obtained from 0.437±1.96×0.186, giving [−3.209, 4.083]. Here, to demonstrate what happens if we make a very misleading assump- tion about the correlational structure of the repeated measurements, we will compare the results with those obtained if we assume that the repeated mea- surements are independent. The independence model can be fitted in the usual way with the lm() function R> summary(lm(plasma ~ time + I(time^2) + group, data = plasma)) Call: lm(formula = plasma ~ time + I(time^2) + group, data = plasma) Residuals: Min 1Q Median 3Q Max -1.6323 -0.4401 0.0347 0.4750 2.0170 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 4.85761 0.16686 29.11 < 2e-16 time -0.80328 0.08335 -9.64 < 2e-16 I(time^2) 0.08467 0.00904 9.37 < 2e-16 groupobese 0.49332 0.08479 5.82 1.7e-08 Residual standard error: 0.673 on 260 degrees of freedom Multiple R-squared: 0.328, Adjusted R-squared: 0.32 F-statistic: 42.3 on 3 and 260 DF, p-value: <2e-16 We see that, under the independence assumption, the standard error for the group effect is about one-half of that given for model plasma.lme1 and if used would lead to a much stronger claim of evidence of a difference between control and obese patients.
- 260. 244 8 The Analysis of Repeated Measures Data We will now plot the predicted values from the fitted linear mixed-effects model for each group using the code presented with Figure 8.5. Time (hours after oral glucose challenge) Plasmainorganicphosphate 2 4 6 2 4 6 8 q qqqqq q q id01 q qqq qqqq id02 2 4 6 8 qq q qqq qq id03 qq qqq qqq id04 2 4 6 8 qq q qqqqq id05 q qqq qqqq id06 q qqqqqqq id07 qq qq qqqq id08 q qqqqqqq id09 qqqqq qqq id10 q qqqqq q q id11 2 4 6 qqqqqqq q id12 2 4 6 q qqqq qqq id13 q qqqqqq q id14 qq qqqqqq id15 qqqqqqqq id16 qq qqqqqq id17 qqqqqqqq id18 qq qqq qqq id19 qqqqq qqq id20 qq qq q qqq id21 qqqqqqqq id22 qq qqqqqq id23 2 4 6qqqqq qqq id24 2 4 6 qqq qqqqq id25 qqqqq qqq id26 q qqqq qqq id27 qqqqqqqq id28 qqqqqqqq id29 qq q qqqqq id30 qqqqqqq q id31 2 4 6 8 qqqq qqqq id32 2 4 6 qqqqqqqq id33 Fig. 8.6. Predictions for glucose challenge data. We can see that the model has captured the profiles of the control group relatively well but not those of the obese group. We need to consider a further model that contains a group × time interaction. R> plasma.lme2 <- lme(plasma ~ time*group +I(time^2), + random = ~time | Subject, + data = plasma, method = "ML") The required model can be fitted and tested against the previous model using R> anova(plasma.lme1, plasma.lme2)
- 261. 8.2 Linear mixed-effects models for repeated measures data 245 Model df AIC BIC logLik Test L.Ratio p-value plasma.lme1 1 8 390.5 419.1 -187.2 plasma.lme2 2 9 383.3 415.5 -182.7 1 vs 2 9.157 0.0025 The p-value associated with the likelihood ratio test is 0.0011, indicating that the model containing the interaction term is to be preferred. The results for this model are R> summary(plasma.lme2) Linear mixed-effects model fit by maximum likelihood Data: plasma AIC BIC logLik 383.3 415.5 -182.7 Random effects: Formula: ~time | Subject Structure: General positive-definite, Log-Cholesky param. StdDev Corr (Intercept) 0.64190 (Intr) time 0.07626 -0.631 Residual 0.38480 Fixed effects: plasma ~ time * group + I(time^2) Value Std.Error DF t-value p-value (Intercept) 4.659 0.17806 228 26.167 0.0000 time -0.759 0.05178 228 -14.662 0.0000 groupobese 0.997 0.25483 31 3.911 0.0005 I(time^2) 0.085 0.00522 228 16.227 0.0000 time:groupobese -0.112 0.03476 228 -3.218 0.0015 Correlation: (Intr) time gropbs I(t^2) time -0.657 groupobese -0.564 0.181 I(time^2) 0.440 -0.907 0.000 time:groupobese 0.385 -0.264 -0.683 0.000 Standardized Within-Group Residuals: Min Q1 Med Q3 Max -2.72436 -0.53605 -0.01071 0.58568 2.95029 Number of Observations: 264 Number of Groups: 33 The interaction effect is highly significant. The fitted values from this model are shown in Figure 8.7 (the code is very similar to that given for producing Figure 8.6). The plot shows that the new model has produced predicted values
- 262. 246 8 The Analysis of Repeated Measures Data that more accurately reflect the raw data plotted in Figure 8.4. The predicted profiles for the obese group are “flatter” as required. Time (hours after oral glucose challenge) Plasmainorganicphosphate 2 4 6 2 4 6 8 q qqqqq q q id01 q qqq qqqq id02 2 4 6 8 qq q qqq qq id03 qq qqq qqq id04 2 4 6 8 qq q qqqqq id05 q qqq qqqq id06 q qqqqqqq id07 qq qq qqqq id08 q qqqqqqq id09 qqqqq qqq id10 q qqqqq q q id11 2 4 6 qqqqqqq q id12 2 4 6 q qqqq qqq id13 q qqqqqq q id14 qq qqqqqq id15 qqqqqqqq id16 qq qqqqqq id17 qqqqqqqq id18 qq qqq qqq id19 qqqqq qqq id20 qq qq q qqq id21 qqqqqqqq id22 qq qqqqqq id23 2 4 6qqqqq qqq id24 2 4 6 qqq qqqqq id25 qqqqq qqq id26 q qqqq qqq id27 qqqqqqqq id28 qqqqqqqq id29 qq q qqqqq id30 qqqqqqq q id31 2 4 6 8 qqqq qqqq id32 2 4 6 qqqqqqqq id33 Fig. 8.7. Predictions for glucose challenge data. We can check the assumptions of the final model fitted to the glucose chal- lenge data (i.e., the normality of the random-effect terms and the residuals) by first using the random.effects() function to predict the former and the resid() function to calculate the differences between the observed data val- ues and the fitted values and then using normal probability plots on each. How the random effects are predicted is explained briefly in Section 8.3. The necessary R code to obtain the effects, residuals, and plots is as follows: R> res.int <- random.effects(plasma.lme2)[,1] R> res.slope <- random.effects(plasma.lme2)[,2]
- 263. 8.3 Prediction of random effects 247 The resulting plot is shown in Figure 8.8. The plot of the residuals is linear as required, but there is some slight deviation from linearity for each of the predicted random effects. q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q −2 0 1 2 −1.5−0.50.51.01.5 Random intercepts Theoretical Quantiles Estimatedrandomintercepts q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q −2 0 1 2 −0.100.000.050.10 Random slopes Theoretical Quantiles Estimatedrandomslopes q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qqq q qq q q q qq q q q q q q q q q q q q q q q qq q q q q q q q q q qq qqq q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q qq q q q q q q q q q q q q q q q q q q q q qq qq qq q qq q q q q qq q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q qq q q q q q q q q qq q q q qq q q −3 −1 1 3 −1.0−0.50.00.51.0 Residuals Theoretical Quantiles Estimatedresiduals Fig. 8.8. Probability plots of predicted random intercepts, random slopes, and residuals for the final model fitted to glucose challenge data. 8.3 Prediction of random effects The random effects are not estimated as part of the model. However, having estimated the model, we can predict the values of the random effects. Accord- ing to Bayes’ Theorem, the posterior probability of the random effects is given by Pr(u|y, x) = f(y|u, x)g(u), where f(y|u, x) is the conditional density of the responses given the random effects and covariates (a product of normal densities) and g(u) is the prior density of the random effects (multivariate normal). The means of this poste- rior distribution can be used as estimates of the random effects and are known as empirical Bayes estimates. The empirical Bayes estimator is also known as a shrinkage estimator because the predicted random effects are smaller in absolute value than their fixed-effect counterparts. Best linear unbiased pre- dictions (BLUPs) are linear combinations of the responses that are unbiased estimators of the random effects and minimise the mean square error.
- 264. 248 8 The Analysis of Repeated Measures Data 8.4 Dropouts in longitudinal data A problem that frequently occurs when collecting longitudinal data is that some of the intended measurements are, for one reason or another, not made. In clinical trials, for example, some patients may miss one or more protocol scheduled visits after treatment has begun and so fail to have the required outcome measure taken. There will be other patients who do not complete the intended follow-up for some reason and drop out of the study before the end date specified in the protocol. Both situations result in missing values of the outcome measure. In the first case, these are intermittent, but dropping out of the study implies that once an observation at a particular time point is missing, so are all the remaining planned observations. Many studies will contain missing values of both types, although in practise it is dropouts that cause the most problems when analysing the resulting data set. An example of a set of longitudinal data in which a number of patients have dropped out is given in Table 8.3. These data are essentially a subset of those collected in a clinical trial that is described in detail in Proudfoot, Goldberg, Mann, Everitt, Marks, and Gray (2003). The trial was designed to assess the effectiveness of an interactive program using multimedia techniques for the delivery of cognitive behavioural therapy for depressed patients and known as Beat the Blues (BtB). In a randomised controlled trial of the program, patients with depression recruited in primary care were randomised to either the BtB program or to Treatment as Usual (TAU). The outcome measure used in the trial was the Beck Depression Inventory II (see Beck, Steer, and Brown 1996), with higher values indicating more depression. Measurements of this variable were made on five occasions, one prior to the start of treatment and at two monthly intervals after treatment began. In addition, whether or not a participant in the trial was already taking anti-depressant medication was noted along with the length of time they had been depressed. Table 8.3: BtheB data. Data of a randomised trial evaluating the effects of Beat the Blues. drug length treatment bdi.pre bdi.2m bdi.3m bdi.5m bdi.8m No >6m TAU 29 2 2 NA NA Yes >6m BtheB 32 16 24 17 20 Yes <6m TAU 25 20 NA NA NA No >6m BtheB 21 17 16 10 9 Yes >6m BtheB 26 23 NA NA NA Yes <6m BtheB 7 0 0 0 0 Yes <6m TAU 17 7 7 3 7 No >6m TAU 20 20 21 19 13 Yes <6m BtheB 18 13 14 20 11 Yes >6m BtheB 20 5 5 8 12 No >6m TAU 30 32 24 12 2
- 265. 8.4 Dropouts in longitudinal data 249 Table 8.3: BtheB data (continued). drug length treatment bdi.pre bdi.2m bdi.3m bdi.5m bdi.8m Yes <6m BtheB 49 35 NA NA NA No >6m TAU 26 27 23 NA NA Yes >6m TAU 30 26 36 27 22 Yes >6m BtheB 23 13 13 12 23 No <6m TAU 16 13 3 2 0 No >6m BtheB 30 30 29 NA NA No <6m BtheB 13 8 8 7 6 No >6m TAU 37 30 33 31 22 Yes <6m BtheB 35 12 10 8 10 No >6m BtheB 21 6 NA NA NA No <6m TAU 26 17 17 20 12 No >6m TAU 29 22 10 NA NA No >6m TAU 20 21 NA NA NA No >6m TAU 33 23 NA NA NA No >6m BtheB 19 12 13 NA NA Yes <6m TAU 12 15 NA NA NA Yes >6m TAU 47 36 49 34 NA Yes >6m BtheB 36 6 0 0 2 No <6m BtheB 10 8 6 3 3 No <6m TAU 27 7 15 16 0 No <6m BtheB 18 10 10 6 8 Yes <6m BtheB 11 8 3 2 15 Yes <6m BtheB 6 7 NA NA NA Yes >6m BtheB 44 24 20 29 14 No <6m TAU 38 38 NA NA NA No <6m TAU 21 14 20 1 8 Yes >6m TAU 34 17 8 9 13 Yes <6m BtheB 9 7 1 NA NA Yes >6m TAU 38 27 19 20 30 Yes <6m BtheB 46 40 NA NA NA No <6m TAU 20 19 18 19 18 Yes >6m TAU 17 29 2 0 0 No >6m BtheB 18 20 NA NA NA Yes >6m BtheB 42 1 8 10 6 No <6m BtheB 30 30 NA NA NA Yes <6m BtheB 33 27 16 30 15 No <6m BtheB 12 1 0 0 NA Yes <6m BtheB 2 5 NA NA NA No >6m TAU 36 42 49 47 40 No <6m TAU 35 30 NA NA NA No <6m BtheB 23 20 NA NA NA No >6m TAU 31 48 38 38 37
- 266. 250 8 The Analysis of Repeated Measures Data Table 8.3: BtheB data (continued). drug length treatment bdi.pre bdi.2m bdi.3m bdi.5m bdi.8m Yes <6m BtheB 8 5 7 NA NA Yes <6m TAU 23 21 26 NA NA Yes <6m BtheB 7 7 5 4 0 No <6m TAU 14 13 14 NA NA No <6m TAU 40 36 33 NA NA Yes <6m BtheB 23 30 NA NA NA No >6m BtheB 14 3 NA NA NA No >6m TAU 22 20 16 24 16 No >6m TAU 23 23 15 25 17 No <6m TAU 15 7 13 13 NA No >6m TAU 8 12 11 26 NA No >6m BtheB 12 18 NA NA NA No >6m TAU 7 6 2 1 NA Yes <6m TAU 17 9 3 1 0 Yes <6m BtheB 33 18 16 NA NA No <6m TAU 27 20 NA NA NA No <6m BtheB 27 30 NA NA NA No <6m BtheB 9 6 10 1 0 No >6m BtheB 40 30 12 NA NA No >6m TAU 11 8 7 NA NA No <6m TAU 9 8 NA NA NA No >6m TAU 14 22 21 24 19 Yes >6m BtheB 28 9 20 18 13 No >6m BtheB 15 9 13 14 10 Yes >6m BtheB 22 10 5 5 12 No <6m TAU 23 9 NA NA NA No >6m TAU 21 22 24 23 22 No >6m TAU 27 31 28 22 14 Yes >6m BtheB 14 15 NA NA NA No >6m TAU 10 13 12 8 20 Yes <6m TAU 21 9 6 7 1 Yes >6m BtheB 46 36 53 NA NA No >6m BtheB 36 14 7 15 15 Yes >6m BtheB 23 17 NA NA NA Yes >6m TAU 35 0 6 0 1 Yes <6m BtheB 33 13 13 10 8 No <6m BtheB 19 4 27 1 2 No <6m TAU 16 NA NA NA NA Yes <6m BtheB 30 26 28 NA NA Yes <6m BtheB 17 8 7 12 NA No >6m BtheB 19 4 3 3 3 No >6m BtheB 16 11 4 2 3
- 267. 8.4 Dropouts in longitudinal data 251 Table 8.3: BtheB data (continued). drug length treatment bdi.pre bdi.2m bdi.3m bdi.5m bdi.8m Yes >6m BtheB 16 16 10 10 8 Yes <6m TAU 28 NA NA NA NA No >6m BtheB 11 22 9 11 11 No <6m TAU 13 5 5 0 6 Yes <6m TAU 43 NA NA NA NA To begin, we shall graph the data here by plotting the boxplots of each of the five repeated measures separately for each treatment group. Assuming the data are available as the data frame BtheB, the necessary code is given with Figure 8.9. Figure 8.9 shows that there is a decline in BDI values in both groups, with perhaps the values in the BtheB group being lower at each post-randomisation visit. We shall fit both random intercept and random intercept and slope models to the data including the pre-BDI values, treatment group, drugs, and length as fixed-effect covariates. First we need to rearrange the data into the long form using the following code: R> BtheB$subject <- factor(rownames(BtheB)) R> nobs <- nrow(BtheB) R> BtheB_long <- reshape(BtheB, idvar = "subject", + varying = c("bdi.2m", "bdi.3m", "bdi.5m", "bdi.8m"), + direction = "long") R> BtheB_long$time <- rep(c(2, 3, 5, 8), rep(nobs, 4)) The resulting data frame BtheB_long contains a number of missing values, and in applying the lme() function these will need to be dropped. But notice it is only the missing values that are removed, not participants that have at least one missing value. All the available data are used in the model-fitting process. We can fit the two models and test which is most appropriate using R> BtheB_lme1 <- lme(bdi ~ bdi.pre + time + treatment + drug + + length, random = ~ 1 | subject, data = BtheB_long, + na.action = na.omit) R> BtheB_lme2 <- lme(bdi ~ bdi.pre + time + treatment + drug + + length, random = ~ time | subject, data = BtheB_long, + na.action = na.omit) This results in R> anova(BtheB_lme1, BtheB_lme2) Model df AIC BIC logLik Test L.Ratio p-value BtheB_lme1 1 8 1883 1912 -933.5 BtheB_lme2 2 10 1886 1922 -933.2 1 vs 2 0.5665 0.7533
- 268. 252 8 The Analysis of Repeated Measures Data R> ylim <- range(BtheB[,grep("bdi", names(BtheB))], + na.rm = TRUE) R> tau <- subset(BtheB, treatment == "TAU")[, + grep("bdi", names(BtheB))] R> boxplot(tau, main = "Treated as Usual", ylab = "BDI", + xlab = "Time (in months)", names = c(0, 2, 3, 5, 8), + ylim = ylim) R> btheb <- subset(BtheB, treatment == "BtheB")[, + grep("bdi", names(BtheB))] R> boxplot(btheb, main = "Beat the Blues", ylab = "BDI", + xlab = "Time (in months)", names = c(0, 2, 3, 5, 8), + ylim = ylim) 0 3 8 01020304050 Treated as Usual Time (in months) BDI q qq 0 3 8 01020304050 Beat the Blues Time (in months) BDI Fig. 8.9. Boxplots for the repeated measures by treatment group for the BtheB data. Clearly, the simpler random intercept model is adequate for these data. The results from this model can be found using R> summary(BtheB_lme1) Linear mixed-effects model fit by REML Data: BtheB_long AIC BIC logLik 1883 1912 -933.5
- 269. 8.4 Dropouts in longitudinal data 253 Random effects: Formula: ~1 | subject (Intercept) Residual StdDev: 7.206 5.029 Fixed effects: bdi ~ bdi.pre + time + treatment + drug + length Value Std.Error DF t-value p-value (Intercept) 5.574 2.2995 182 2.424 0.0163 bdi.pre 0.640 0.0799 92 8.013 0.0000 time -0.702 0.1469 182 -4.775 0.0000 treatmentBtheB -2.315 1.7152 92 -1.350 0.1804 drugYes -2.816 1.7729 92 -1.588 0.1156 length>6m 0.179 1.6816 92 0.106 0.9155 Correlation: (Intr) bdi.pr time trtmBB drugYs bdi.pre -0.683 time -0.232 0.019 treatmentBtheB -0.390 0.121 0.017 drugYes -0.074 -0.236 -0.022 -0.323 length>6m -0.244 -0.241 -0.036 0.002 0.157 Standardized Within-Group Residuals: Min Q1 Med Q3 Max -2.68699 -0.50847 -0.06085 0.42067 3.81414 Number of Observations: 280 Number of Groups: 97 The effect of most interest in this study is, of course, the treatment effect, and our analysis shows that this is not significant at the 5% level. The only effect that is significant is time, the negative value of its regression coefficient showing that the BDI values decline over the eight months of the study. We now need to consider briefly how the dropouts may affect the analyses reported above. To understand the problems that patients dropping out can cause for the analysis of data from a longitudinal trial, we need to consider a classification of dropout mechanisms first introduced by Rubin (1976). The type of mechanism involved has implications for which approaches to analysis are suitable and which are not. Rubin’s suggested classification involves three types of dropout mechanisms: ˆ Dropout completely at random (DCAR). Here the probability that a pa- tient drops out does not depend on either the observed or missing values of the response. Consequently, the observed (non-missing) values effectively constitute a simple random sample of the values for all subjects. Possible examples include missing laboratory measurements because of a dropped
- 270. 254 8 The Analysis of Repeated Measures Data test tube (if it was not dropped because of the knowledge of any measure- ment), the accidental death of a participant in a study, or a participant moving to another area. Intermittent missing values in a longitudinal data set, whereby a patient misses a clinic visit for transitory reasons (“went shopping instead” or the like) can reasonably be assumed to be DCAR. Completely random dropout causes the least problems for data analysis, but it is a strong assumption. ˆ Dropout at random (DAR). The dropout at random mechanism occurs when the probability of dropping out depends on the outcome measures that have been observed in the past but given this information is condi- tionally independent of all the future (unrecorded) values of the outcome variable following dropout. Here “missingness” depends only on the ob- served data, with the distribution of future values for a subject who drops out at a particular time being the same as the distribution of the future values of a subject who remains in at that time, if they have the same covariates and the same past history of outcome up to and including the specific time point. Murray and Findlay (1988) provide an example of this type of missing value from a study of hypertensive drugs in which the outcome measure was diastolic blood pressure. The protocol of the study specified that the participant was to be removed from the study when his or her blood pressure got too large. Here blood pressure at the time of dropout was observed before the participant dropped out, so although the dropout mechanism is not DCAR since it depends on the values of blood pressure, it is DAR, because dropout depends only on the observed part of the data. A further example of a DAR mechanism is provided by Heitjan (1997) and involves a study in which the response measure is body mass index (BMI). Suppose that the measure is missing because subjects who had high body mass index values at earlier visits avoided being measured at later visits out of embarrassment, regardless of whether they had gained or lost weight in the intervening period. The missing values here are DAR but not DCAR; consequently methods applied to the data that assumed the latter might give misleading results (see later discussion). ˆ Non-ignorable (sometimes referred to as informative). The final type of dropout mechanism is one where the probability of dropping out depends on the unrecorded missing values–observations are likely to be missing when the outcome values that would have been observed had the pa- tient not dropped out are systematically higher or lower than usual (cor- responding perhaps to their condition becoming worse or improving). A non-medical example is when individuals with lower income levels or very high incomes are less likely to provide their personal income in an inter- view. In a medical setting, possible examples are a participant dropping out of a longitudinal study when his or her blood pressure became too high and this value was not observed, or when their pain became intolera- ble and we did not record the associated pain value. For the BDI example introduced above, if subjects were more likely to avoid being measured if
- 271. 8.4 Dropouts in longitudinal data 255 they had put on extra weight since the last visit, then the data are non- ignorably missing. Dealing with data containing missing values that result from this type of dropout mechanism is difficult. The correct analyses for such data must estimate the dependence of the missingness probability on the missing values. Models and software that attempt this are available (see for example, Diggle and Kenward 1994), but their use is not routine and, in addition, it must be remembered that the associated parameter estimates can be unreliable. Under what type of dropout mechanism are the mixed-effects models con- sidered in this chapter valid? The good news is that such models can be shown to give valid results under the relatively weak assumption that the dropout mechanism is DAR (Carpenter, Pocock, and Lamm 2002). When the missing values are thought to be informative, any analysis is potentially problematical. But Diggle and Kenward (1994) have developed a modelling framework for longitudinal data with informative dropouts, in which random or completely random dropout mechanisms are also included as explicit models. The essen- tial feature of the procedure is a logistic regression model for the probability of dropping out, in which the explanatory variables can include previous values of the response variable and, in addition, the unobserved value at dropout as a latent variable (i.e., an unobserved variable). In other words, the dropout probability is allowed to depend on both the observed measurement history and the unobserved value at dropout. This allows both a formal assessment of the type of dropout mechanism in the data and the estimation of effects of interest, for example, treatment effects under different assumptions about the dropout mechanism. A full technical account of the model is given in Dig- gle and Kenward (1994), and a detailed example that uses the approach is described in Carpenter et al. (2002). One of the problems for an investigator struggling to identify the dropout mechanism in a data set is that there are no routine methods to help, although a number of largely ad hoc graphical procedures can be used as described in Diggle (1998), Everitt (2002), and Carpenter et al. (2002). We shall now illustrate one very simple graphical procedure for assessing the dropout mechanism suggested in Carpenter et al. (2002). That involves plotting the observations for each treatment group, at each time point, dif- ferentiating between two categories of patients: those who do and those who do not attend their next scheduled visit. Any clear difference between the distributions of values for these two categories indicates that dropout is not completely at random. For the Beat the Blues data, such a plot is shown in Figure 8.10. When comparing the distribution of BDI values for patients who do (circles) and do not (bullets) attend the next scheduled visit, there is no apparent difference, and so it is reasonable to assume dropout completely at random.
- 272. 256 8 The Analysis of Repeated Measures Data R> bdi <- BtheB[, grep("bdi", names(BtheB))] R> plot(1:4, rep(-0.5, 4), type = "n", axes = FALSE, + ylim = c(0, 50), xlab = "Months", ylab = "BDI") R> axis(1, at = 1:4, labels = c(0, 2, 3, 5)) R> axis(2) R> for (i in 1:4) { + dropout <- is.na(bdi[,i + 1]) + points(rep(i, nrow(bdi)) + ifelse(dropout, 0.05, -0.05), + jitter(bdi[,i]), pch = ifelse(dropout, 20, 1)) + } Months BDI 0 2 3 5 01020304050 q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qq q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q q qqq q q q q q q q q q q q q q q q q q q Fig. 8.10. Distribution of BDI values for patients who do (circles) and do not (bullets) attend the next scheduled visit.
- 273. 8.6 Exercises 257 8.5 Summary Linear mixed-effects models are extremely useful for modelling longitudinal data in particular and repeated measures data more generally. The models allow the correlations between the repeated measurements to be accounted for so that correct inferences can be drawn about the effects of covariates of interest on the repeated response values. In this chapter, we have concentrated on responses that are continuous and conditional on the explanatory variables, and random effects have a normal distribution. But random-effects models can also be applied to non-normal responses, for example binary variables–see, for example Everitt (2002) and Skrondal and Rabe-Hesketh (2004). The lack of independence of repeated measures data is what makes the modelling of such data a challenge. But even when only a single measure- ment of a response is involved, correlation can, in some circumstances, occur between the response values of different individuals and cause similar prob- lems. As an example, consider a randomised clinical trial in which subjects are recruited at multiple study centres. The multicentre design can help to provide adequate sample sizes and enhance the generalisability of the results. However, factors that vary by centre, including patient characteristics and medical practise patterns, may exert a sufficiently powerful effect to make inferences that ignore the “clustering” seriously misleading. Consequently, it may be necessary to incorporate random effects for centres into the analysis. 8.6 Exercises Ex. 8.1 The final model fitted to the timber data did not constrain the fitted curves to go through the origin, although this is clearly necessary. Fit an amended model where this constraint is satisfied, and plot the new predicted values. Ex. 8.2 Investigate a further model for the glucose challenge data that allows a random quadratic effect. Ex. 8.3 Fit an independence model to the Beat the Blues data, and com- pare the estimated treatment effect confidence interval with that from the random intercept model described in the text. Ex. 8.4 Construct a plot of the mean profiles of the two treatment groups in the Beat the Blues study showing also the predicted mean profiles under the model used in the chapter. Repeat the exercise with a model that includes only a time effect. Ex. 8.5 Investigate whether there is any evidence of an interaction between treatment and time for the Beat the Blues data.
- 275. References Affifi, A. A., Clark, V. A., and May, S. (2004), Computer-Aided Multivariate Analysis, London, UK: Chapman & Hall, 4th edition. Cited on p. 99. Anonymous (1988), “Statistical abstract of the USA,” Technical Report 159, U.S. Department of Commerce, Washington, DC. Cited on p. 176. Banfield, J. D. and Raftery, A. E. (1993), “Model-based Gaussian and non- Gaussian clustering,” Biometrics, 49, 803–821. Cited on p. 186, 188. Barlow, R. E., Bartholomew, D. J., Bremner, J. M., and Brunk, H. D. (1972), Statistical Inference under Order Restrictions, New York: John Wiley & Sons. Cited on p. 123. Bartholomew, D. and Knott, M. (1987), Latent Variable Models and Factor Analysis, London, UK: Griffin. Cited on p. 147. Bartholomew, D. J. (2005), “History of factor analysis: A statistical perspec- tive,” in Encyclopedia of Statistics in Behavioral Science, eds. B. S. Everitt and D. Howell, Chichester, UK: John Wiley & Sons. Cited on p. 136. Bartlett, M. S. (1947),“Multivariate analysis,”Journal of the Royal Statistical Society, Series B, 9, 176–197. Cited on p. 103. Beck, A., Steer, R., and Brown, G. (1996), BDI-II Manual, The Psychological Corporation, San Antonio, TX, 2nd edition. Cited on p. 248. Becker, C. and Gather, U. (2001),“The largest nonidentifiable outlier: A com- parison of multivariate simultaneous outlier identification rules,” Computa- tional Statistics & Data Analysis, 36, 119–127. Cited on p. 23. Becker, R., Cleveland, W., Shyu, M., and Kaluzny, S. (1994), “Trellis dis- plays: User’s guide,”Technical report, AT&T Research, URL http://www2. research.att.com/areas/stat/doc/94.10.ps. Cited on p. 50, 57. Bellman, R. (1961), Adaptive Control Processes, Princeton, NJ: Princeton University Press. Cited on p. 61. Bentler, P. (1980), “Multivariate analysis with latent variables: Causal mod- eling,” Annual Review of Psychology, 31, 419–456. Cited on p. 202. Bentler, P. (1982), “Linear systems with multiple levels and types of latent variables,” in Systems under Indirect Observation, eds. K. G. J¨oreskog and H. Wold, Amsterdam: North-Holland, pp. 101–130. Cited on p. 222. 259 DOI 10.1007/978-1-4419-9650-3, © Springer Science+Business Media, LLC 2011 B. Everitt and T. Hothorn, An Introduction to Applied Multivariate Analysis with R: Use R!,
- 276. 260 References Benz´ecri, J. (1992), Correspondence Analysis Handbook, New York: Marcel Dekker Inc. Cited on p. 127. Blackith, R. E. and Reyment, R. A. (1971), Multivariate Morphometrics, Lon- don, UK: Academic Press. Cited on p. 144. Blalock, H. M. J. (1961), “Correlation and causality: The multivariate case,” Social Forces, 39, 246–251. Cited on p. 202. Blalock, H. M. J. (1963), “Making causal inferences for unmeasured variables from correlations among indicators,” The American Journal of Sociology, 69, 53–62. Cited on p. 202. Bollen, K. and Long, J. (1993), Testing Structural Equation Models, Newbury Park, CA: Sage Publications. Cited on p. 204, 205. Bouguila, N. and Amayri, O. (2009), “A discrete mixture-based kernel for SVMs: Application to spam and image categorization,” Information Pro- cessing and Management, 45, 631–642. Cited on p. 186. Branchaud, E. A., Cham, J. G., Nenadic, Z., Andersen, R. A., and Burdick, J. (2010), “A miniature robot for autonomous single neuron recordings,” Proceedings of the 2005 IEEE International Conference on Robotics and Automation, pp. 1920–1926. Cited on p. 186. Browne, M. (1982), “Covariance structures,”in Topics in Applied Multivariate Analysis, ed. D. M. Hawkins, London, UK: Cambridge University Press. Cited on p. 204, 222. Browne, M. W. (1974), “Generalized least squares estimates in the analysis of covariance structures,” South African Statistical Journal, 8, 1–24. Cited on p. 202. Calsyn, J. R. and Kenny, D. A. (1977), “Self-concept of ability and perceived evaluation of others. Cause or effect of academic achievement?” Journal of Educational Psychology, 69, 136–145. Cited on p. 206. Carpenter, J., Pocock, S., and Lamm, C. J. (2002), “Coping with missing data in clinical trials: A model-based approach applied to asthma trials,” Statistics in Medicine, 21, 1043–1066. Cited on p. 255. Carroll, J., Clark, L., and DeSarbo, W. (1984), “The representation of three- way proximity data by single and multiple tree structure models,” Journal of Classification, 1, 25–74. Cited on p. 106. Carroll, J. B. (1953), “An analytical solution for approximating simple struc- ture in factor analysis,” Psychometrika, 18, 23–38. Cited on p. 146. Carroll, J. D. and Arabie, P. (1980), “Multidimensional scaling,” Annual Re- view of Psychology, 31, 607–649. Cited on p. 106. Cattell, R. B. (1966), “The scree test for the number of factors,” Multivariate Behavioural Research, 1, 245–276. Cited on p. 72. Celeux, G. and Govaert, G. (1995),“A classification EM algorithm for cluster- ing and two stochastic versions,”Computational Statistics & Data Analysis, 14, 315–332. Cited on p. 188. Chambers, J. M., Cleveland, W. S., Kleiner, B., and Tukey, P. A. (1983), Graphical Methods for Data Analysis, London, UK: Chapman & Hall/CRC. Cited on p. 25.
- 277. References 261 Chatfield, C. and Collins, A. (1980), Introduction to Multivariate Analysis, London, UK: Chapman & Hall. Cited on p. 159. Chernoff, H. (1973), “The use of faces to represent points in k-dimensional space graphically,”Journal of the American Statistical Association, 68, 361– 368. Cited on p. 38. Cliff, N. (1983), “Some cautions concerning the application of causal modeling methods,” Multivariate Behavioral Research, 18, 115–126. Cited on p. 222. Cook, D. and Swayne, D. F. (2007), Interactive and Dynamic Graphics for Data Analysis, New York: Springer-Verlag. Cited on p. 26, 199. Corbet, G. B., Cummins, J., Hedges, S. R., and Krzanowski, W. J. (1970), “The taxonomic structure of British water voles, genus Arvicola,” Journal of Zoology, 61, 301–316. Cited on p. 117. Crowder, M. J. (1998),“Nonlinear growth curve,”in Encyclopedia of Biostatis- tics, eds. P. Armitage and T. Colton, Chichester, UK: John Wiley & Sons. Cited on p. 226. Dai, X. F., Erkkila, T., Yli-Harja, O., and Lahdesmaki, H. (2009), “A joint finite mixture model for clustering genes from independent Gaussian and beta distributed data,” BMC Bioinformatics, 10. Cited on p. 186. Dalgaard, P. (2002), Introductory Statistics with R, New York: Springer- Verlag. Cited on p. viii. Davis, C. S. (2003), Statistical Methods in the Analysis of Repeated Measure- ments, New York: Springer-Verlag. Cited on p. 226, 232. Day, N. (1969), “Estimating the components of a mixture of normal distribu- tions,” Biometrika, 56, 463–474. Cited on p. 187. de Leeuw, J. (1983),“On the prehistory of correspondence analysis,”Statistica Neerlandica, 37, 161–164. Cited on p. 127. Dempster, A., Laird, N., and Rubin, D. (1977), “Maximum likelihood from incomplete data via the EM algorithm,” Journal of the Royal Statistical Society, Series B, 39, 1–38. Cited on p. 188. Diggle, P. J. (1998), “Dealing with missing values in longitudinal studies,” in Statistical Analysis of Medical Data, eds. B. S. Everitt and G. Dunn, London, UK: Arnold. Cited on p. 255. Diggle, P. J., Heagerty, P. J., Liang, K. Y., and Zeger, S. L. (2003), Analysis of Longitudinal Data, Oxford, UK: Oxford University Press. Cited on p. 232, 235. Diggle, P. J. and Kenward, M. G. (1994),“Informative dropout in longitudinal data analysis,” Journal of the Royal Statistical Society, Series C, 43, 49–93. Cited on p. 255. Duncan, O. (1969),“Some linear models for two-wave, two-variable panel anal- ysis,” Psychological Bulletin, 72, 177–182. Cited on p. 202. Dunson, D. B. (2009), “Bayesian nonparametric hierarchical modeling,” Bio- metrical Journal, 51, 273–284. Cited on p. 186. Everitt, B. (1984), An Introduction to Latent Variable Models, London, UK: Chapman & Hall. Cited on p. 142, 187, 203.
- 278. 262 References Everitt, B. (1987), Introduction to Optimization Methods and Their Applica- tion in Statistics, London, UK: Chapman & Hall. Cited on p. 142. Everitt, B. and Bullmore, E. (1999), “Mixture model mapping of brain acti- vation in functional magnetic resonance images,” Human Brain Mapping, 7, 1–14. Cited on p. 186. Everitt, B. and Hand, D. (1981), Finite Mixture Distributions, Chapman & Hall/CRC. Cited on p. 186, 187, 188. Everitt, B. S. (2002), Modern Medical Statistics, London, UK: Arnold. Cited on p. 255, 257. Everitt, B. S. and Dunn, G. (2001), Applied Multivariate Data Analysis, Lon- don, UK: Arnold, 2nd edition. Cited on p. 203, 209. Everitt, B. S., Landau, S., Leese, M., and Stahl, D. (2011), Cluster Analysis, Chichester, UK: John Wiley & Sons, 5th edition. Cited on p. 165, 166, 171, 176, 180, 198. Everitt, B. S. and Rabe-Hesketh, S. (1997), The Analysis of Proximity Data, London, UK: Arnold. Cited on p. 106. Farmer, S. A. (1971), “An investigation into the results of principal com- ponents analysis of data derived from random numbers,” Statistician, 20, 63–72. Cited on p. 72. Fayyad, U. M., Piatetsky-Shapiro, G., Smyth, P., and Uthurusamy, R., eds. (1996), Advances in Knowledge Discovery and Data Mining, Menlo Park, CA: AAAI Press/The MIT Press. Cited on p. 4. Fisher, N. I. and Switzer, P. (1985), “Chi-plots for assessing dependence,” Biometrika, 72, 253–265. Cited on p. 34. Fisher, N. I. and Switzer, P. (2001), “Graphical assessment of dependence: Is a picture worth a 100 tests?”The American Statistician, 55, 233–239. Cited on p. 34. Fox, J., Kramer, A., and Friendly, M. (2010), sem: Structural Equation Mod- els, URL http://CRAN.R-project.org/package=sem, R package version 0.9-21. Cited on p. 207. Fraley, C. and Raftery, A. (2010), mclust: Model-Based Clustering / Normal Mixture Modeling, URL http://CRAN.R-project.org/package=mclust, R package version 3.4.8. Cited on p. 191. Fraley, C. and Raftery, A. E. (2002), “Model-based clustering, discriminant analysis, and density estimation,” Journal of the American Statistical As- sociation, 97, 611–631. Cited on p. 188. Fraley, C. and Raftery, A. E. (2003), “Enhanced model-based clustering, den- sity estimation, and discriminant analysis software: MCLUST,” Journal of Classification, 20, 263–286. Cited on p. 188. Fraley, C. and Raftery, A. E. (2007), “Bayesian regularization for normal mix- ture estimation and model-based clustering,” Journal of Classification, 24, 155–181. Cited on p. 188. Frets, G. P. (1921), “Heredity of head form in man,” Genetica, 3, 193–384. Cited on p. 96.
- 279. References 263 Fr¨uhwirth-Schnatter, S. (2006), Finite Mixture and Markov Switching Models, New York: Springer-Verlag. Cited on p. 186. Gabriel, K. R. (1971),“The biplot graphic display of matrices with application to principal component analysis,” Biometrika, 58, 453–467. Cited on p. 92. Gabriel, K. R. (1981), “Biplot display of multivariate matrices for inspection of data and diagnosis,” in Interpreting Multivariate Data, ed. V. Barnett, Chichester, UK: John Wiley & Sons. Cited on p. 93. Ganesalingam, J., Stahl, D., Wijesekera, L., Galtrey, C., Shaw, C. E., Leigh, P. N., and Al-Chalabi, A. (2009),“Latent cluster analysis of ALS phenotypes identifies prognostically differing groups,” PLoS One, 4. Cited on p. 186. Gansner, E. and North, S. (2000), “An open graph visualization system and its applications to software engineering,” Software-Practice & Experience, 30, 1203–1233. Cited on p. 210. Goldberg, K. and Iglewicz, B. (1992), “Bivariate extensions of the boxplot,” Technometrics, 34, 307–320. Cited on p. 28, 31. Gordon, A. (1999), Classification, Boca Raton, FL: Chapman & Hall/CRC, 2nd edition. Cited on p. 165. Gordon, A. D. (1987), “A review of hierarchical classification,” Journal of the Royal Statistical Society, Series A, 150, 119–137. Cited on p. 165. Gorsuch, R. L. (1983), Factor Analysis, Hillsdale, NJ: Lawrence Erlbaum Associates, 2nd edition. Cited on p. 93. Gower, J. (1966),“Some distance properties of latent root and vector methods used in multivariate analysis,” Biometrika, 53, 325–338. Cited on p. 109. Gower, J. C. and Hand, D. J. (1996), Biplots, London, UK: Chapman & Hall/CRC. Cited on p. 93. Gower, J. C. and Ross, G. J. S. (1969), “Minimum spanning trees and single linkage cluster analysis,” Applied Statistics, 18, 54–64. Cited on p. 121. Greenacre, M. (1992), “Correspondence analysis in medical research,” Statis- tical Methods in Medical Research, 1, 97. Cited on p. 129, 131. Greenacre, M. (2007), Correspondence Analysis in Practice, Boca Raton, FL: Chapman & Hall/CRC, 2nd edition. Cited on p. 128. Guadagnoli, E. and Velicer, W. F. (1988), “Relation of sample size to the sta- bility of component patterns,” Psychological Bulletin, 103, 265–275. Cited on p. 93. Hand, D. J., Daly, F., Lunn, A. D., McConway, K. J., and Ostrowski, E. (1994), A Handbook of Small Datasets, London, UK: Chapman & Hall/CRC. Cited on p. 79, 171. Hand, D. J., Mannila, H., and Smyth, P. (2001), Principles of Data Mining, Cambridge, MA: The MIT Press. Cited on p. 4. Hasselblad, V. (1966), “Estimation of parameters for a mixture of normal distributions,” Technometrics, 8, 431–444. Cited on p. 187. Hasselblad, V. (1969), “Estimation of finite mixtures of distributions from the exponential family,” Journal of the American Statistical Association, 64, 1459–1471. Cited on p. 187.
- 280. 264 References Hatcher, L. (1994), A Step-by-Step Approach to Using the SAS System for Factor Analysis and Structural Equation Modeling, Cary, NC: SAS Insti- tute. Cited on p. 93. Heitjan, D. F. (1997), “Annotation: What can be done about missing data? Approaches to imputation,” American Journal of Public Health, 87, 548– 550. Cited on p. 254. Hendrickson, A. and White, P. (1964), “Promax: A quick method for rotation to oblique simple structure,”British Journal of Mathematical and Statistical Psychology, 17, 65–70. Cited on p. 147. Hershberger, S. L. (2005), “Factor score estimation,” in Encyclopedia of Be- havioral Statistics, eds. B. S. Everitt and D. C. Howell, Chichester, UK: John Wiley & Sons. Cited on p. 148. Heywood, H. (1931), “On finite sequences of real numbers,” Proceedings of the Royal Society of London, Series A, Containing Papers of a Mathematical and Physical Character, 134, 486–501. Cited on p. 142. Hills, M. (1977), “Book review,” Applied Statistics, 26, 339–340. Cited on p. 159. Hotelling, H. (1933), “Analysis of a complex of statistical variables into prin- cipal components,” Journal of Educational Psychology, 24, 417–441. Cited on p. 101. Huba, G. J., Wingard, J., and Bentler, P. (1981), “A comparison of two latent variable causal models for adolescent drug use,” Journal of Personality and Social Psychology, 40, 180–193. Cited on p. 151, 211, 216. Hyv¨arinen, A. (1999), “The fixed-plot algorithm and maximum likelihood es- timation for independent component analysis,” Neural Processing Letters, 10, 1–5. Cited on p. 102. Hyv¨arinen, A., Karhunen, L., and Oja, E. (2001), Independent Component Analysis, New York: John Wiley & Sons. Cited on p. 102. Izenman, A. J. (2008), Modern Multivariate Techniques, New York: Springer- Verlag. Cited on p. 102. Jennrich, R. and Sampson, P. (1966), “Rotation for simple loadings,” Psy- chometrika, 31, 313–323. Cited on p. 146. Jolliffe, I. (1972), “Discarding variables in a principal component analysis. I: Artificial data,” Journal of the Royal Statistical Society, Series C, 21, 160–173. Cited on p. 72. Jolliffe, I. (1989), “Rotation of ill-defined principal components,” Journal of the Royal Statistical Society, Series C, 38, 139–148. Cited on p. 147. Jolliffe, I. (2002), Principal Component Analysis, New York: Springer-Verlag. Cited on p. 63, 64, 66, 71, 89, 90, 142. Jones, M. C. and Sibson, R. (1987), “What is projection pursuit?” Journal of the Royal Statistical Society, Series A, 150, 1–38. Cited on p. 199. J¨oreskog, K. G. (1973), “Analysis of covariance structures,” in Multivariate Analysis, Volume II , ed. P. R. Krishnaiah, New York: Academic Press. Cited on p. 202.
- 281. References 265 J¨oreskog, K. G. and S¨orbom, D. (1981), “Analysis of linear structural rela- tionships by maximum likelihood and least squares,”Technical Report 81-8, University of Uppsala, Uppsala, Sweden. Cited on p. 221. Kaiser, H. (1958), “The varimax criterion for analytic rotation in factor anal- ysis,” Psychometrika, 23, 187–200. Cited on p. 72, 146. Kaufman, L. and Rousseeuw, P. J. (1990), Finding Groups in Data: An In- troduction to Cluster Analysis, New York: John Wiley & Sons. Cited on p. 133. Kenward, M. G. and Roger, J. H. (1997), “Small sample inference for fixed effects from restricted maximum likelihood,”Biometrics, 53, 983–997. Cited on p. 238. Keyfitz, N. and Flieger, W. (1971), Population: Facts and Methods of Demog- raphy, San Francisco, CA: W.H. Freeman. Cited on p. 148. Kruskal, J. B. (1964a), “Multidimensional scaling by optimizing goodness-of- fit to a nonmetric hypothesis,” Psychometrika, 29, 1–27. Cited on p. 121. Kruskal, J. B. (1964b), “Nonmetric multidimensional scaling: A numerical method,” Psychometrika, 29, 115–129. Cited on p. 123. Krzanowski, W. J. (1988), Principles of Multivariate Analysis, Oxford, UK: Oxford University Press. Cited on p. 95. Krzanowski, W. J. (2010), “Canonical correlation,” in Encyclopaedic Compan- ion to Medical Statistics, eds. B. Everitt and C. Palmer, Chichester, UK: John Wiley & Sons, 2nd edition. Cited on p. 96, 99, 101. Lawley, D. and Maxwell, A. (1963), Factor Analysis as a Statistical Method, London, UK: Butterworths. Cited on p. 142. Leisch, F. (2010), “Neighborhood graphs, stripes and shadow plots for cluster visualization,” Statistics and Computing, 20, 457–469. Cited on p. 191. Leisch, F. and Dimitriadou, E. (2019), flexclust: Flexible Cluster Algorithms, URL http://CRAN.R-project.org/package=flexclust, R package ver- sion 1.3-1. Cited on p. 194. Ligges, U. (2010), scatterplot3d: 3D Scatter Plot, URL http://CRAN. R-project.org/package=scatterplot3d, R package version 0.3-31. Cited on p. 47. Little, R. (2005), “Missing data,” in Encyclopedia of Statistics in Behavioral Science, eds. B. S. Everitt and D. Howell, Chichester, UK: John Wiley & Sons, pp. 1234–1238. Cited on p. 6. Little, R. and Rubin, D. (1987), Statistical Analysis with Missing Data, New York: John Wiley & Sons. Cited on p. 6. Longford, N. T. (1993), Random Coefficient Models, Oxford, UK: Oxford Uni- versity Press. Cited on p. 235. MacCullum, R. C., Browne, M. W., and Sugawara, H. M. (1996), “Power analysis and determination of ssample size for covariance structure models,” Psychological Methods, 1, 130–149. Cited on p. 222. MacDonnell, W. R. (1902),“On criminal anthropometry and the identification of criminals,” Biometrika, 1, 177–227. Cited on p. 102.
- 282. 266 References Marchini, J. L., Heaton, C., and Ripley, B. D. (2010), fastICA: FastICA Algorithms to perform ICA and Projection Pursuit, URL http://CRAN. R-project.org/package=fastICA, R package version 1.1-13. Cited on p. 102. Marcoulides, G. A. (2005), “Structural equation modeling: Nontraditional al- ternatives,” in Encyclopedia of Statistics in Behavioral Science, eds. B. S. Everitt and D. Howell, Chichester, UK: John Wiley & Sons. Cited on p. 222. Mardia, K. V., Kent, J. T., and Bibby, J. M. (1979), Multivariate Analysis, London, UK: Academic Press. Cited on p. 64, 65, 95, 109, 142, 146, 205. Marin, J. M., Mengersen, K., and Roberts, C. (2005),“Bayesian modelling and inferences on mixtures of distributions,” in Bayesian Thinking, Modeling and Computation, eds. D. Dey and C. R. Rao, Amsterdam: Elsevier, pp. 15840–15845. Cited on p. 186, 188. Marriott, F. (1974), The Interpretation of Multiple Observations, London, UK: Academic Press. Cited on p. 88. McLachlan, G. and Basford, K. (1988), Mixture Models: Inference and Appli- cations to Clustering, New York: Marcel Dekker Inc. Cited on p. 186. McLachlan, G. and Peel, D. (2000), Finite Mixture Models, New York: John Wiley & Sons. Cited on p. 186, 188. Meghani, S. H., Lee, C. S., Hanlon, A. L., and Bruner, D. W. (2009), “Latent class cluster analysis to understand heterogeneity in prostate cancer treat- ment utilities,” BMC Medical Informatics and Decision Making, 9. Cited on p. 186. Meyer, W. and Bendig, A. (1961), “A longitudinal study of the primary men- tal abilities test,” Journal of Educational Psychology, 52, 50–60. Cited on p. 223. Morrison, D. (1990), Multivariate Statistical Methods, New York: McGraw- Hill, 3rd edition. Cited on p. 64. Murray, G. D. and Findlay, J. G. (1988), “Correcting for bias caused by dropouts in hypertension trials,” Statistics in Medicine, 7, 941–946. Cited on p. 254. Muthen, B. O. (1978), “Contributions to factor analysis of dichotomous vari- ables,” Psychometrika, 43, 551–560. Cited on p. 159. Muthen, B. O. (1984),“A general structural equation model with dichotomous, ordered categorical, and continuous latent variable indicators,” Psychome- trika, 49, 115–132. Cited on p. 222. Muthen, L. K. and Muthen, B. O. (2002),“How to use a Monte Carlo study to decide on sample size and determine power,” Structural Equation Modeling, 9, 599–620. Cited on p. 222. Needham, R. M. (1965), “Computer methods for classification and grouping,” in The Use of Computers in Anthropology, ed. J. Hymes, The Hague: Mou- ton, pp. 345–356. Cited on p. 164. Osborne, J. W. and Costello, A. B. (2004), “Sample size and subject to item ratio in principal components analysis,” Practical Assessment, Research &
- 283. References 267 Evaluation, 9, URL http://www.pareonline.net/getvn.asp?v=9&n=11. Cited on p. 93. Paradis, E., Bolker, B., Claude, J., Cuong, H. S., Desper, R., Durand, B., Dutheil, J., Gascuel, O., Heibl, C., Lawson, D., Lefort, V., Legendre, P., Lemon, J., Noel, Y., Nylander, J., Opgen-Rhein, R., Schliep, K., Strimmer, K., and de Vienne, D. (2010), ape: Analyses of Phylogenetics and Evolution, URL http://CRAN.R-project.org/package=ape, R package version 2.6-2. Cited on p. 121. Pearson, K. (1901),“LIII. on lines and planes of closest fit to systems of points in space,” Philosophical Magazine Series 6, 2, 559–572. Cited on p. 101. Pett, M., Lackey, N., and Sullivan, J. (2003), Making Sense of Factor Analysis: The Use of Factor Analysis for Instrument Development in Health Care Research, Thousand Oaks, CA: Sage Publications. Cited on p. 146. Pinheiro, J., Bates, D., DebRoy, S., Sarkar, D., and R Development Core Team (2010), nlme: Linear and Nonlinear Mixed Effects Models, URL http: //CRAN.R-project.org/package=nlme, R package version 3.1-97. Cited on p. 235. Pinker, S. (1997), How the Mind Works, London, UK: The Penguin Press. Cited on p. 163. Pledger, S. and Phillpot, P. (2008), “Using mixtures to model heterogeneity in ecological capture-recapture studies,” Biometrical Journal, 50, 1022–1034. Cited on p. 186. Prim, R. C. (1957), “Shortest connection networks and some generalizations,” Bell System Technical Journal, 36, 1389–1401. Cited on p. 121. Proudfoot, J., Goldberg, D., Mann, A., Everitt, B. S., Marks, I., and Gray, J. A. (2003), “Computerized, interactive, multimedia cognitive-behavioural program for anxiety and depression in general practice,” Psychological Medicine, 33, 217–227. Cited on p. 248. Rencher, A. (2002), Methods of Multivariate Analysis, New York: John Wiley & Sons. Cited on p. 63, 71. Rencher, A. C. (1995), Methods of Multivariate Analysis, New York: John Wiley & Sons. Cited on p. 148. Rocke, D. and Woodruff, D. (1996), “Identification of outliers in multivariate data,”Journal of the American Statistical Association, 91, 1047–1061. Cited on p. 23. Romesburg, H. C. (1984), Cluster Analysis for Researchers, Belmont, CA: Lifetime Learning Publications. Cited on p. 123. Rubin, D. (1976), “Inference and missing data,” Biometrika, 63, 581–592. Cited on p. 253. Rubin, D. (1987), Multiple Imputation for Survey Nonresponse, New York: John Wiley & Sons. Cited on p. 7. Rubin, D. B. and Thayer, D. T. (1982), “EM algorithms for ML factor anal- ysis,” Psychometrika, 47, 69–76. Cited on p. 142. Sarkar, D. (2008), Lattice: Multivariate Data Visualization with R, New York: Springer-Verlag. Cited on p. 51, 52, 57, 180.
- 284. 268 References Sarkar, D. (2010), lattice: Lattice Graphics, URL http://CRAN.R-project. org/package=lattice, R package version 0.19-13. Cited on p. 51, 180. Schafer, J. (1999), “Multiple imputation: A primer,” Statistical Methods in Medical Research, 8, 3–15. Cited on p. 7. Scheipl, F. (2010), RLRsim: Exact (Restricted) Likelihood Ratio Tests for Mixed and Additive models., URL http://CRAN.R-project.org/ package=RLRsim, R package version 2.0-5. Cited on p. 236. Scheipl, F., Greven, S., and K¨uchenhoff, H. (2008),“Size and power of tests for a zero random effect variance or polynomial regression in additive and linear mixed models,” Computational Statistics & Data Analysis, 52, 3283–3299. Cited on p. 236. Schmid, C. F. (1954), Handbook of Graphic Presentation, New York: Ronald Press. Cited on p. 25. Schumaker, R. E. and Lomax, R. G. (1996), A Beginner’s Guide to Structural Equation Modeling, Mahwah, NJ: Lawrence Erlbaum Associates. Cited on p. 202. Shepard, R. N. (1962a),“The analysis of proximities: Multidimensional scaling with unknown distance function, Part I,”Psychometrika, 27, 125–140. Cited on p. 121. Shepard, R. N. (1962b), “The analysis of proximities: Multidimensional scal- ing with unknown distance function, Part II,” Psychometrika, 27, 219–246. Cited on p. 121. Sibson, R. (1979), “Studies in the robustness of multidimensional scaling. Per- turbational analysis of classical scaling,” Journal of the Royal Statistical Society, Series B, 41, 217–229. Cited on p. 109. Silverman, B. (1986), Density Estimation, London, UK: Chapman & Hall/CRC. Cited on p. 42, 43, 46. Skrondal, A. and Rabe-Hesketh, S. (2004), Generalized Latent Variable Model- ing: Multilevel, Longitudinal and Structural Equation Models, Boca Raton, FL: Chapman & Hall/CRC. Cited on p. 186, 232, 235, 257. Sokal, R. R. and Rohlf, F. (1981), Biometry, San Francisco, CA: W. H. Free- man, 2nd edition. Cited on p. . Spearman, C. (1904), “General intelligence objectively determined and mea- sured,” The American Journal of Psychology, 15, 201–293. Cited on p. 136. Stanley, W. and Miller, M. (1979), “Measuring technological change in jet fighter aircraft,” Technical Report R-2249-AF, Rand Corporation, Santa Monica, CA. Cited on p. 171. Steinley, D. (2008), “Stability analysis in k-means clustering,” British Journal of Mathematical and Statistical Psychology, 61, 255–273. Cited on p. 176. Thomsen, O. O., Wulff, H. R., Martin, A., and Singer, P. A. (1993), “What do gastronenterologists in Europe tell cancer patients?” The Lancet, 341, 473–476. Cited on p. 189. Thorndike, R. M. (2005), “History of factor analysis: A psychological perspec- tive,” in Encyclopedia of Statistics in Behavioral Science, eds. B. S. Everitt and D. Howell, Chichester, UK: John Wiley & Sons. Cited on p. 136.
- 285. References 269 Thurstone, L. L. (1931), “Multiple factor analysis,” Psychology Review, 39, 406–427. Cited on p. 145. Titterington, D., Smith, A., and Makov, U. (1985), Statistical Analysis of Finite Mixture Distributions, John Wiley & Sons. Cited on p. 186. Tubb, A., Parker, N. J., and Nickless, G. (1980), “The analysis of Romano- British pottery by atomic absorption spectrophotometry,” Archaeometry, 22, 153–171. Cited on p. . Tufte, E. R. (1983), The Visual Display of Quantitative Information, Cheshire, CT: Graphics Press. Cited on p. 25, 26. van Hattum, P. and Hoijtink, H. (2009), “Market segmentation using brand strategy research: Bayesian inference with respect to mixtures of log-linear models,” Journal of Classification, 26, 297–328. Cited on p. 186. Vanisma, F. and De Greve, J. P. (1972), “Close binary systems before and after mass transfer,” Astrophysics and Space Science, 87, 377–401. Cited on p. 46. Velleman, P. and Wilkinson, L. (1993), “Nominal, ordinal, interval, and ratio typologies are misleading,” The American Statistician, 47, 65–72. Cited on p. 5. Venables, W. and Ripley, B. D. (2010), MASS: Support Functions and Datasets for Venables and Ripley’s MASS, URL http://CRAN.R-project. org/package=MASS, R package version 7.3-9. Cited on p. 123. Venables, W. N. and Ripley, B. D. (2002), Modern Applied Statistics with S, New York: Springer-Verlag, 4th edition. Cited on p. 46, 123. Verbyla, A., Cullis, B., Kenward, M., and Welham, S. (1999), “The analysis of designed experiments and longitudinal data by using smoothing splines,” Journal of the Royal Statistical Society, Series C, 48, 269–311. Cited on p. 50. Wand, M. and Ripley, B. D. (2010), KernSmooth: Functions for Kernel Smoothing, URL http://CRAN.R-project.org/package=KernSmooth, R package version 2.23-4. Cited on p. 46. Wand, M. P. and Jones, M. C. (1995), Kernel Smoothing, London, UK: Chap- man & Hall/CRC. Cited on p. 42. Wheaton, B., Muthen, B., Alwin, D., and Summers, G. (1977), “Assessing reliability and stability in panel models,” Sociological Methodology, 7, 84– 136. Cited on p. 216. Wolfe, J. (1970), “Pattern clustering by multivariate mixture analysis,” Mul- tivariate Behavioral Research, 5, 329–350. Cited on p. 187. Wright, S. (1934), “The method of path coefficients,” The Annals of Mathe- matical Statistics, 5, 161–215. Cited on p. 202. Young, G. and Householder, A. S. (1938), “Discussion of a set of points in terms of their mutual distances,” Psychometrika, 3, 19–22. Cited on p. 107. Zerbe, G. O. (1979), “Randomization analysis of the completely randomized design extended to growth and response curves,” Journal of the American Statistical Association, 74, 215–221. Cited on p. 226.
- 287. Index adjusted goodness of fit index, 205 airdist data, 112 ape, 121 apply(), 19 available-case analysis, 6 bandwidth, 43, 46 Best linear unbiased predictions, 247 bimodal, 47 biplot, 92 bivariate boxplot, 28 bivariate data, 27 bivariate density estimate, 42 bivariate density estimation, 45 bivariate Epanechnikov kernel, 46 bkde2D(), 46 boxplot(), 27 boxplots(), 27 BtheB data, 248, 251 BtheB_long data, 251 bubble plot, 34 canonical correlation analysis, 95 canonicaml correlation analysis, 96 characteristic root, 64 characteristic vector, 64 chi-plot, 34 chi-squared distance, 129 chi-squared distribution, 18 chiplot(), 34 city block distance, 24 classical multidimensional scaling, 106 cluster analysis, 7, 14 cmdscale(), 110, 130 common factors, 135 communality, 138 complete-case analysis, 5 compound symmetry, 234 confirmatory factor analysis, 135 confirmatory factor analysis models, 201 convex hull trimming, 32 cor(), 14 correlation coefficient, 14 correlation matrix, 6 correlations, 138 correspondence analysis, 127 covariance, 12 covariance matrix, 6, 13 covariance structure modelling, 201 crime data, 176 curse of dimensionality, 61 CYGOB1 data, 46 data mining, 3 depression data, 103 disattenuated correlation, 209 discrepancy function, 202 disparities, 122 dissimilarity, 105 dist(), 15, 110 distance, 14 dropout at random, 254 dropout completely at random, 253 eigenvalue, 64 eigenvector, 64 empirical Bayes estimates, 247
- 288. 272 Index Euclidean distance, 14 exactLRT(), 236 exam data, 9 exploratory factor analysis, 9, 63, 135 factanal(), 149 factor analysis, 6, 135 factor loadings, 70, 135, 136 factor pattern coefficients, 146 factor rotation, 144 factor structure coefficients, 146 fastICA, 102 fixed effect, 234 flexclust, 194 gardenflowers data, 134 Gaussian kernel, 43 generalised distance, 18 glyph plots, 37 goodness-of-fit index, 205 hclust(), 169 headsize data, 74, 96, 103 heptathlon data, 78 Heywood case, 142 hist(), 27 House of Representatives voting, 123 HSAUR2, viii hybrid models, 106 hypo data, 4 image plot, 180 imputation, 6 intelligence, 135 intra-class correlation, 234 jet data, 171 kernel density estimators, 42–46 kernel function, 43 KernSmooth, 46 LAdepr data, 99 Lagrange multipliers, 64 latent, 255 latent variables, 135 lattice, 51, 153, 180, 240 layout(), 27, 171 levelplot(), 153 life data, 148, 149 linear combinations, 16, 61 linear mixed-effects models, 232 lm(), 39, 233, 236, 243 lme(), 235, 236, 251 local independence, 232 log-eigenvalue diagram, 72 longitudinal data, 5, 12 magnitude criterion, 110 mahalanobis(), 115 Mahalanobis distance, 115 Manhattan distances, 112 manifest variables, 135 MASS, 123 match(), 32 maximum likelihood factor analysis, 141 Mclust(), 188 mclust, 188, 189, 191 measure data, 7 minimum spanning tree, 120 missing completely at random, 6 missing values, 5 monotonic transformations, 121 multidimensional scaling, 14, 105 multiple conditioning, 50 multiple imputation, 7 multivariate normal, 2 multivariate normal density function, 16 multivariate normal distribution, 15 MVA, viii, 191 nlme, 235 non-ignorable, 254 non-metric multidimensional scaling, 106 non-parametric estimation, 42 normal probability plot, 17 normed fit index, 205 normed residuals, 211 nuisance parameters, 232 oblimin, 146 oblique rotation, 145 orthogonal rotation, 145 outliers, 27 pairs(), 39 parallel(), 240 path diagram, 209
- 289. Index 273 path.diagram(), 209 plasma data, 228, 240 plot(), 27 posterior probability, 247 pottery data, 8 principal components, 61 principal components analysis, 7, 61 principal coordinates, 109 principal factor analysis, 141 probability plot, 16 probability-probability plot, 16 promax, 146 proximity, 105 qchisq(), 19 quantile-quantile plot, 16 quartimax, 146 random effects, 232 random intercept, 232 random intercept and slope, 232 random intercept model, 233 random slope and intercept model, 234 random variables, 2 random.effects(), 246 rectangular kernel function, 43 reduced covariance matrix, 141 repeated-measures data, 12 reshape(), 226 resid(), 246 restricted maximum likelihood, 235 RLRsim, 236 robust estimation, 32 root-mean-square residual, 205 rug(), 27 sapply(), 20 scale(), 15 scatterplot, 26 scatterplot matrix, 39 scatterplot3d, 47 sem, 207, 208 Shepard diagram, 124 similarity matrices, 105 single-factor model, 136 skulls data, 114 smoothing parameter, 43 social class, 135 spatial models, 106 specific variates, 137 spectral decomposition, 70 structural equation modelling, 6, 201 structural relationships, 65 subject-specific random component, 233 subset(), 13 teensex data, 130 text(), 22 timber data, 227 time-varying covariates, 232 trace criterion, 109 tree models, 106 Trellis graphics, 50 triangular kernel function, 43 Tucker-Lewis index, 205 unidentifiable, 203 unobserved, 255 USairpollution data, 10 var(), 13 variance component model, 234 variance-covariance matrix, 13 varimax, 146 voting data, 124 watervoles data, 118 Windows, in kernel density estimation, 43 with(), 27 WWIIleaders data, 124