Effects of missing observations on

International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.6, December 2014
DOI:10.5121/ijcsa.2014.4601 1
EFFECTS OF MISSING OBSERVATIONS ON
PREDICTIVE CAPABILITY OF CENTRAL
COMPOSITE DESIGNS
Yisa Yakubu1*
, Angela Unna Chukwu2
, Bamiduro Timothy Adebayo2
, Amahia
Godwin Nwanzo2
1
Department of Mathematics and Statistics, Federal University of Technology, Minna,
Niger State, Nigeria
2
Department of Statistics, University of Ibadan, Ibadan, Oyo state, Nigeria
Abstract
Quite often in experimental work, many situations arise where some observations are lost or become
unavailable due to some accidents or cost constraints. When there are missing observations, some
desirable design properties like orthogonality,rotatability and optimality can be adversely affected. Some
attention has been given, in literature, to investigating the prediction capability of response surface
designs; however, little or no effort has been devoted to investigating same for such designs when some
observations are missing. This work therefore investigates the impact of a single missing observation of the
various design points: factorial, axial and center points, on the estimation and predictive capability of
Central Composite Designs (CCDs). It was observed that for each of the designs considered, precision of
model parameter estimates and the design prediction properties were adversely affected by the missing
observations and that the largest loss in precision of parameters corresponds to a missing factorial point.
Keywords
Central Composite Designs, Missing Observations, Predictive Capability, Precision of parameter
estimates
1.Introduction
Central composite design (CCD) is the most popular class of second-order response surface
designs, which was introduced by Box and Wilson (1951). This design consists of factors with
five levels that involve three categories:
i. complete (or a fraction of) 2k
factorial design with factor levels coded as -1,
1(called the factorial portion),
ii. an axial portion consisting of 2k points arranged so that two points are chosen on
the coordinate axis of each control variable at a distance of α from the design
center,
iii. n0 center points.
iv. Thus the total number of points in a CCD is 022 nkn k
++= . The second-order
response surface model for these designs is

2
‫ݕ‬ = ߚ଴ + ෍ ߚ௜‫ݔ‬௜
௞
௜ିଵ
+ ෍ ߚ௜௜‫ݔ‬௜
ଶ
௞
௜ିଵ
+ ෍ ෍ ߚ௜௝‫ݔ‬௜‫ݔ‬௝
௞
௝ୀ௜ାଵ
௞ିଵ
௜ୀଵ
+ ߝ (1.1)
Where y is the response variable, x is the input variable, ߚ is a model coefficient, and ࢋ is a
random error component.
One of the functions of the fitted response surface model, and of course, the primary goal of
many designed experiments, is to allow for good prediction of response values at various points
of interest throughout the experimental region. In response surface methodology, interest is more
on prediction than parameter estimation since the points on the fitted surface are predicted
responses. Thus the prediction variance criteria are essential tools for selecting response surface
designs as the researcher is provided with variance information regarding the worst prediction
scenario. Box and Hunter (1957) noted that consideration of only the variances of the individual
model coefficient estimates does not, for the case of second or higher-order models, lead to any
unique class of “best” designs. Their argument is that the precision of the coefficient estimates
should be studied simultaneously. Several measures of prediction performance exist for
comparing designs of which the most commonly considered is the scaled prediction variance
(SPV).
For the second-order response surface model in (Equation 1.1), the scaled prediction variance of
the expected response is given by
‫)ݔ(ݒ‬ =
ே୚ୟ୰ൣ௬ො(ೣ)൧
ఙమ = ݂ܰ′(‫ܺ()ݔ‬ᇱ
ܺ)ିଵ
݂(‫)ݔ‬ (1.2)
where N is the design size, ߪଶ
is the observation error, and ݂(‫)ݔ‬ is the general form of the 1‫݌ݔ‬
model vector.
Desirable designs are those with the smallest maximum SPV, and with reasonably stable SPV
(i.e. smallest range) in the design region. The SPV allows the practitioner to measure the
precision of the predicted response on a per observation basis and it penalizes larger designs over
small designs.
Directly associated with the prediction variance ‫)ݔ(ݒ‬ are the maximum (G) and the integrated (V)
prediction variance optimality criteria. The G-optimality and the corresponding G-efficiency
emphasize the use of designs for which the maximum scaled prediction variance ‫)ܠ(ݒ‬ in the
design region is not too large. Thus, a G-optimal design )(ξ is one that
Min఍ൣmax఍ ‫(ݒ‬x)൧
( 1.3)
The G-efficiency is determined as
‫ܩ‬௘௙௙ =
௣
୫ୟ୶ೣചೃ ௩(୶)
(1.4)

3
The V-criterion attempts to generate a single measure of prediction performance through an
average process, that is, ࢜(‫)ܠ‬ is averaged over some region of interest R (Myers et al., 2009 ). For
a design )(ξ , the V-optimality criterion can be expressed as






=
→
∫
∫
−
R
R
xfXXxfMin
vMinV
)dx()'()'(
K
1
(x)dx
K
1
1
ξ
ξ
(1.5)
Since ݂(‫ݔ‬)′(ܺᇱ
ܺ)ିଵ
݂(x) is a scalar, (1.5) can be written as
[ ]
[ ]
[ ]{ }{ }dxxfxfKXXtrMin
dxXXxfxftr
K
Min
dxxfXXxftr
K
MinV
R
R
R
∫
∫
∫
−
−
−
=
=
=
'1'
1''
1''
)()()/1()(
)()()(
1
)()()(
1
ξ
ξ
ξ
(1.6)
Where ∫=
R
dxK is the volume of the region R, and the matrix dx)x(f)x(f)K/1( R
'
∫ is a
‫ ݌‬x ‫݌‬ matrix of region moments.
The A-criterion is also often used to quantify the precision of parameter estimates, which is a way
of assessing the design’s capability to estimate the underlying model. The A-criterion minimizes
the trace of the covariance matrix, that is, an A-optimal design is one in which we have
min఍ ‫݁ܿܽݎݐ‬(X′X))ିଵ
(1.7)
There has been much discussion concerning the distance, α, of the axial points from the design
center, in response surface designs. The choice of the axial distance (α) and the number of center
runs (nc) have substantial influence on the performance of CCDs. When the scaled prediction
variance (SPV) is constant for any location of a fixed distance from the design center, the design
is rotatable (Box and Hunter, 1957).
Missing observations adversely affect the orthogonality, rotatability, and optimality properties of
a design. These properties are design performance criteria used for choosing “good” response
surface designs. Extensive studies have been carried out in literature on predictive capability of
full central composite designs both for spherical and cuboidal regions. For instance, Box and
Draper (1959, 1963) proposed designs that are robust to model misspecification in terms of
integrated average prediction variance (V).
Ahmad and Gilmour (2010) study the robustness of subset response surface designs to a missing
value in terms of prediction variance by computing the ratio of prediction variances for the design
with a missing observation to the prediction variance for the full design. For each of the designs
considered, the authors plotted the maximum and minimum ratios of variances against 1 <
‫ݏݑ݅݀ܽݎ‬ < 1.8. They observed that the minimum ratio of prediction variances were quite robust

4
to missing design points for almost all designs and for all types of missing design points except
few. The authors also noted that the maximum ratio decreases gradually with the increase of
radius for some designs with some type of missing points.
However, not much work has been done in literature on predictive capability of completely
randomized central composite designs with some observations missing. This study therefore
examines the predictive capability of standard central composite designs in the presence of
missing observations.
For each of the CCDs we considered in this work, the effect of missing a single observation of the
different design points on the precision of estimates of the model parameters is first examined,
then we investigate the same effect on the maximum and average prediction variance criteria
under different values of the axial point distance (α) from the design center.
2.Materials and Methods
Consider a second order response surface model in k variables and n design points as given in
equation (1.1) above. Over the n design points, this model can be expressed in matrix form as
ܻ = ܺߚ + ߝ
(2.1)
Where ܻ is the n x 1 response vector, ܺ is the n x p model matrix, ߚ is the p x 1 vector of
partameters, and ߝ is the n x 1 vector of random error terms.
When all the observations are available, the standard least squares estimate of the parameter ߚ is
given as
ߚመ = (ܺᇱ
ܺ)ିଵ
ܺᇱ
‫ݕ‬
The fitted model of the response variable ܻ
ܻ෠ = ܺߚመ (2.2)
which is used to estimate the response variable at any point in the design region.
Now, the variance-covariance matrix of the estimator ߚመ is
Var ቀߚመቁ = (ܺ′
ܺ)ିଵ
ܺ′
Var ቀ‫ݕ‬ቁ ܺ(ܺ′
ܺ)ିଵ
= (ܺ′
ܺ)ିଵ
ܺ′
Var(ܺߚ + ࢿ)ܺ(ܺ′
ܺ)ିଵ
= ߪఌ
ଶ(ܺ′
ܺ)ିଵ
(2.3)

5
The A-criterion aims to improve the precision of estimates of the model parameters by
minimizing the trace of this variance-covariance matrix.
The variance of the predicted response is
ܸ൫ܻ෠൯ = ܸܺ ቀߚመቁ ܺ′
= ߪఌ
ଶ
ܺ(ܺ′
ܺ)ିଵ
ܺ′
= ߪଶ
‫ܪ‬ (2.4)
Where ‫ܪ‬ = ܺ(ܺ′
ܺ)ିଵ
ܺ′
is the ‘hat’ matrix of the design with the property that trace(‫)ܪ‬ = ‫,݌‬ p
being the number of model parameters.
If any m observations are missing, we partition the response vector ܻ and the model matrix X as
൥
ܻ௠
⋯
ܻ௥
൩ and ൥
ܺ௠
⋯
ܺ௥
൩
respectively, where ܻ௠ consists of m missing observations and ܺ௠ consists of m corresponding
rows. Thus the information matrix can be expressed as
ܺᇱ
ܺ = ܺ௠
ᇱ
ܺ௠ + ܺ௥
ᇱ
ܺ௥
The ordinary least squares estimator of the parameters of the residual design is
ߚመ∗
= (ܺ௥
′
ܺ௥)ିଵ
ܺ௥
′
‫ݕ‬௥
And the variance-covariance matrix of this estimator is
Var ቀߚመ∗
ቁ = (ܺ௥
′
ܺ௥)ିଵ
The desire is always to minimize the variance of ߚመ (‫ܣ‬ − ܿ‫)݊݋݅ݎ݁ݐ݅ݎ‬ so as to increase the
precision of the parameter estimates, but whenever some observations are missing, the variance is
increased (i.e., the estimates become less precise) as a result and such increase in the variance is
given by
݅. ‫.ݒ‬ = trace(ܺ௥
′
ܺ௥)ିଵ
− trace(ܺ′
ܺ)ିଵ
(2.5)
2.1.Loss function in terms of the Trace Criterion
Expression (2.5) measures the amount of reduction in the precision of the estimates when an
observation is missing. Therefore, we define the loss function in terms of the trace(‫)ܣ‬ criterion
when some observations are missing as the relative reduction in the precision of the parameter
estimates. This is given by
‫.ݎ‬ ݅. ‫.ݒ‬ =
୲୰ୟୡୣ൫௑ೝ
′ ௏ೝ
షభ௑ೝ൯
షభ
୲୰ୟୡୣ(௑′௏షభ௑)షభ
− 1 (2.6)

6
This expression is used in this work to estimate the loss in precision of estimates of the model
parameters when some observations are missing.
2.2Efficiency in terms of Maximum and Average Prediction Variances Criteria
In order to examine the effects of missing observations on the predictive capability of response
surface designs, efficiency of each of the resulting residual design was examined relative to the
full design in terms of these criteria and thus we compute the relative efficiencies
ܴ‫ܧ‬ீ =
ெ஺௑೉ചೃ௩(௫)
ெ஺௑೉ചೃ௩(௫)ೝ೐೏ೠ೎೐೏
(2.7)
and
( )
( )
reduced
R
R
V
dxxfxf
K
XXtrace
dxxfxf
K
XXtrace
RE






















=
∫
∫
−
−
)()(
1
)()(
1
'1'
'1'
(2.8)
Where ࢌ(࢞), ‫,)ݔ(ݒ‬ and ∫R
'
dx)x(f)x(f
K
1
are as defined above.
And we note that
(i) Relative efficiency greater than 1 indicates that the missing observation has little or no adverse
effect on the design in terms of this criterion, which means that the criterion is quite robust to the
missing point.
(ii) Relative efficiency smaller than 1 indicates that the missing point has large adverse effect on
the design in terms of the criterion.
Four different central composite designs were considered in this work. These designs are given in
the table below, where k is the number of design variables, ݊௙ is the number of factorial points,
݊ఈ is the number of axial points, ݊଴ is the number of center points, and ݊ is the total number of
design points.
Table 2.1: Candidate CCDs considered.
k ݊௙ ݊ఈ ݊଴ ݊
2 4 4 4 12
3 8 6 4 18
4 16 8 4 28
5 32 10 4 46

7
3.Results and Discussions
In this section, the losses in precision of parameter estimates, the maximum and average
prediction variances computed under various α values for full and residual designs were
presented. The plots of the losses and that of the relative efficiencies were also presented.
TABLES 1a, 2a, 3a, and 4a below consist of the losses in precision of the model parameter
estimates for the respective CCDs with one observation missing. The corresponding loss curves
were given by FIGURES 1a, 2a, 3a, and 4a. We can see from these tables that the trace-criterion
values for the full designs decreases as the distance of the axial points from the design center
increases. We observed that missing observations of each of the design point categories (factorial,
axial, and center) has adverse effect on the parameters.
We can observe from TABLES 1a and 2a and also from the corresponding figures (1a and 2a)
that the loss due to missing a factorial point was the highest at ߙ = 1.00. In FIGURE 1a, the
precision continues to improve sharply as ߙ increases and becomes stable at ߙ > 1.41, while in
2a, it decreases as ߙ increases up to 1.21 after which it improves and becomes slightly stable at
ߙ > 1.68. As we can see from FIGURES 3a and 4a, this criterion becomes slightly robust to a
missing factorial point as the number of design variables increases.
Also from Tables and Figures 1a and 2a, we observed that the loss in precision due to a missing
axial point was the second largest. From Figure 1a, we observed a slight improvement in
precision as ߙ exceeds 1.00 and the precision becomes stable for 1.21 < ߙ < 1.50 after which it
decreases sharply for the rest ߙ values. From Figure 2a, the precision improves sharply as ߙ
increases and a sharp loss was again observed as ߙ goes beyond 1.68. From Figures 3a and 4a,
missing axial point causes the highest loss in precision of parameter estimates at ߙ = 1.00 as we
can directly observe. However, the loss in precision due to this missing (axial) point continues to
decrease sharply as ߙ increases to 2.00 in Figure 3a and to 2.24 in Figure 4a after which the loss
increases for the rest ߙ values in both cases.
We observed from Figure 1a that the loss due to the missing center point is the lowest and makes
a bell-shaped curve with the ߙ values while in Figures 2a, 3a and 4a, the bell shape increases as
the number of design variables increases. In Figure 2a, the loss curve due to this missing (center)
point intersects that due to the missing axial point at two distance points and has its peak at
ߙ ≅ 1.71. In Figures 3a and 4a, this loss curve intersects the other two curves and has its peak at
ߙ ≅ 2.11 and ߙ ≅ 2.31 respectively.
In general, the loss in precision of parameter estimates due to a missing factorial point continues
to decrease as the number of design factors increases.
TABLES 1b, 2b, 3b and 4b give the values of the maximum and average prediction variances for
the complete designs and for designs with a single observation missing. These tables show that
the location of the maximum prediction variance is restricted to the factorial portion of the
designs at low values of α and to the axial portion as α increases.
Figures 1b, 2b, 3b, and 4b give the relative G-efficiency plots respectively for the designs
considered here. We can observe from each of these figures that the G-criterion is quite robust to
the missing center point. In figure 1b, efficiency loss was only due to the missing axial point at
low values of α after which it continues to decrease sharply. In figure 2b, the largest loss in
efficiency corresponds to the missing factorial point and this loss remains slightly stable for
1.00 ≤ ߙ ≤ 1.732, after which it continues to decrease as α increases. In this figure also, loss in

8
efficiency due to the missing axial point was observed as α approaches 1.732, which continues to
increase and then decreases slowly as α exceeds 2.00. In figure 3b, we observed a loss in
efficiency corresponding to the missing axial point and also to the missing factorial point. The
loss due to the missing factorial point increases gradually at low values of α and then suddenly
decreases as α increases beyond 2.00. The loss corresponding to the missing axial point is slightly
stable as α goes beyond 2.00.In figure 4b, we observe the largest loss in efficiency corresponding
to the missing axial point at α = 1.00, which reduces gradually as α increases, and as α exceeds
2.236, the loss becomes slightly robust to changes in α values. We also observed the smallest loss
in efficiency from this figure, which corresponds to the missing factorial point at α = 1.00, which
suddenly decreases as α increases.
Figures 1c, 2c, 3c, and 4c give the respective V-efficiency plots for the designs under discussion.
We observed the largest efficiency loss corresponding to missing a factorial point in figures 1c
and 2c, which continues to decrease as α increases. In figure 1c, we observed a slightly stable
efficiency loss corresponding to the missing axial point for 1.00 ≤ ߙ ≤ 1.50, after which it starts
to increase, while that of figure 2c makes a bell-shaped curve with α values. The loss due to
missing axial point in figure 3c and 4c also decreases up to a certain value of α and then starts to
increase again. The loss in efficiency due to missing center point is only observed in figure 1c as
α exceeds 1.10 and increases gradually as α increases up to 1.414 after which it becomes stable as
α continues to increase. The smallest efficiency loss is recorded for the missing center point at α =
1.210 in figure 2c, which gradually increases as α increases and then starts to decreases again. In
figure 3c, we observed the smallest efficiency loss due to the missing center point as α exceeds
1.210; the same loss was observed in figure 4c as α exceeds 1.00, which continues to increase as
α increases up to the point ߙ ≈ 2.11 in figure 3c and ߙ ≈ 2.25 in figure 4c, and then decreases
again.
3.1 Two-factor CCD
This design consists of nf = 4 factorial points, nα = 4 axial points, n0 = 4 center points, and thus N
= 12 total design points.
Table 1a:Loss in precision of parameter estimates due to single missing observations in 2-variable CCD
Axial
point
distance
A-criterion
value for full
design
Loss in precision (lp) due to
missing
factorial(f) axial(ߙ) center(c)
1.000 1.5416 0.4702906 0.2072522 0.06117
1.210 1.2440 0.3397106 0.1685691 0.098553
1.414 1.0626 0.2550348 0.176454 0.117636
1.500 0.9967 0.2362797 0.1902278 0.118491
2.000 0.7187 0.2319466 0.3015166 0.090024
Table 1b: Scaled prediction Variance and average prediction variance for complete design and for designs
with one observation missing
No. of variables k = 2. Total design points n = 12
No. of parameters p = 6. No. of centre points = 4
ߙ Missing point ‫)ݔ(ݒ‬ at V
Factorial Axial Center

9
point point point
1.000 None 9.500 6.000 2.500 3.633
Factorial 9.533 5.866 2.383 4.913
Axial 8.861 6.111 2.444 3.931
Center 8.732 5.596 2.894 3.586
1.21 None 8.464 6.669 2.866 3.318
Factorial 8.370 6.252 2.661 3.930
Axial 9.819 6.712 2.670 3.535
Center 7.772 6.139 3.451 3.410
1.414 None 7.500 7.499 2.999 3.166
Factorial 7.334 6.952 2.749 3.483
Axial 6.954 7.332 2.749 3.361
Center 6.875 6.8741 3.666 3.351
1.5 None 7.159 7.861 2.979 3.109
Factorial 6.995 7.283 2.737 3.364
Axial 6.652 7.710 2.737 3.311
Center 6.566 7.209 3.633 3.312
2.00 None 6.000 9.500 2.500 2.766
Factorial 6.111 8.861 2.444 2.943
Axial 5.866 9.533 2.383 3.098
Center 5.596 8.732 2.894 2.931
3.2.Three-factor CCD
This design consists of nf = 8 factorial points, nα = 6 axial points, and n0 = 4 center points, with
ܰ = 18 total design points.
Table 2a: Loss in precision of parameter estimates due to single missing observations in 3-
variable CCD
Axial
point
distance
A-criterion
value for full
design
missing
1.000 1.9369 0.2111622 0.1953637 0.021116
1.210 1.4227 0.233078 0.1447951 0.044071
1.681 1.0814 0.1827261 0.088034 0.103292
1.732 1.0575 0.1780615 0.088227 0.105059
2.000 0.9333 0.1692918 0.1031823 0.094718
2.250 0.8322 0.1760394 0.1237683 0.07366
2.500 0.7549 0.1855875 0.1413432 0.056431
3.000 0.6572 0.1982654 0.1653987 0.037279
Table 2b: Scaled prediction Variance and average prediction variance for complete design and
for designs with a single observation missing

10
No. of variables k = 3 Total design points n = 18
No. of parameters p = 10 No. of center points = 4
Axial point
distance (ߙ)
Missing point ‫)ݔ(ݒ‬ at IV
Factorial
point
Axial
point
Center
point
1.000 None 14.292 9.085 2.785 5.607
Factorial 16.606 9.060 2.677 6.647
Axial 13.698 11.769 2.942 6.098
Center 13.510 8.763 3.112 5.497
1.21 None 13.685 9.502 3.374 4.817
Factorial 16.091 9.341 3.249 5.575
Axial 13.111 11.409 3.425 5.110
Center 12.943 9.113 3.922 4.855
rotatable (1.681) None 12.059 10.929 4.486 4.543
Factorial 14.124 10.465 4.239 4.868
Axial 11.509 11.938 4.240 4.632
Center 11.390 10.324 5.643 4.944
1.732 None 11.893 11.142 4.499 4.527
Factorial 13.932 10.660 4.249 4.828
Axial 11.354 12.142 4.249 4.609
Center 11.285 10.562 5.663 4.948
2.00 None 11.175 12.300 4.200 4.344
Factorial 13.263 11.769 4.016 4.590
Axial 10.736 13.421 4.026 4.448
Center 10.578 11.641 5.173 4.745
2.25 None 10.729 13.247 3.670 4.078
Factorial 13.089 12.713 3.601 4.332
Axial 10.442 14.536 3.594 4.247
Center 10.201 12.554 4.354 4.377
2.50 None 10.416 13.989 3.183 3.811
Factorial 13.138 13.461 3.205 4.081
Axial 10.309 15.350 3.160 4.042
Center 9.940 13.253 3.652 4.018
3.00 None 9.972 15.012 2.536 3.392
Factorial 13.401 14.474 2.650 3.678
Axial 10.213 16.235 2.530 3.713
Center 9.550 14.204 2.788 3.497
3.3.Four-factor CCD
This design consists of nf = 16 factorial points, nα = 8 axial points, and n0 = 4 center
points, with ݊ = 28 total design points.
Axial
point
distance
A-criterion
value for full
design
missing
1.000 2.2675 0.0480706 0.1753032 0.009261
1.210 1.4871 0.0689933 0.1374487 0.020846

11
2.000 0.9583 0.0771157 0.047793 0.108734
2.250 0.8747 0.0784269 0.0533897 0.09649
2.500 0.7866 0.0845411 0.0652174 0.069286
3.000 0.6602 0.0960315 0.0802787 0.033475
with a single observation missing
(ߙ) Missing point ‫)ݔ(ݒ‬ at IV
Factorial
point
Axial
point
Center
point
1.000 None 18.446 13.932 3.347 7.477
factorial 21.424 13.619 3.237 7.559
Axial 17.913 21.461 3.634 8.365
Center 17.791 13.666 3.666 7.400
1.21 None 18.104 14.294 3.992 5.628
factorial 21.289 13.988 3.868 5.728
Axial 17.605 20.769 4.245 6.140
Center 17.465 14.010 4.489 5.699
rotatable (2.00) None 16.333 16.333 7.000 5.211
factorial 19.800 15.862 6.750 5.190
Axial 15.862 19.800 6.750 5.257
Center 15.750 15.750 9.000 6.075
2.25 None 15.815 17.623 6.491 4.983
factorial 19.400 17.122 6.288 4.971
Axial 15.400 21.255 6.344 5.048
Center 15.266 17.035 8.148 5.759
2.50 None 15.431 18.913 5.447 4.512
factorial 19.278 18.410 5.342 4.535
Axial 15.118 22.907 5.455 4.651
Center 14.929 18.319 6.521 5.050
3.00 None 14.878 20.885 3.713 3.613
factorial 19.398 20.399 3.741 3.684
Axial 14.825 25.121 3.815 3.855
Center 14.434 20.208 4.128 3.838
3.4.Five - factor CCD
A full replicate of this design consists of nf = 32 factorial points, nα = 10 axial points, and
n0 = 4 center points, i.e., there are a total of ݊ = 46 design points.
Axial
point
distance
A-criterion
value for full
design
missing
1.000 2.5839 0.0109137 0.1580557 0.004915

12
1.500 1.0301 0.0260169 0.0940685 0.029803
2.236 0.8163 0.0285434 0.0292785 0.122504
2.378 0.7828 0.029254 0.0297649 0.117782
2.500 0.7460 0.0302949 0.033378 0.104155
2.750 0.6646 0.0341559 0.0430334 0.068914
3.000 0.5963 0.0379004 0.0489686 0.042428
with a single observation missing
(ߙ) Missing point ‫)ݔ(ݒ‬ at IV
Factorial
point
Axial
point
Center
point
1.000 None 22.570 22.592 4.456 12.587
factorial 25.492 22.165 4.362 12.447
Axial 22.144 38.628 4.878 14.157
Center 22.080 22.393 4.827 12.520
1.500 None 22.063 23.307 6.720 6.801
factorial 25.242 22.897 6.583 6.764
Axial 21.685 36.209 7.095 7.288
Center 21.590 23.101 7.698 7.140
2.236 None 20.946 24.971 11.499 7.609
factorial 24.391 24.499 11.249 7.527
Axial 20.576 33.571 11.249 7.614
Center 20.491 24.428 14.999 9.177
rotatable (2.378) None 20.724 25.817 11.158 7.468
factorial 24.225 25.331 10.9208 7.392
Axial 20.369 34.493 10.968 7.475
Center 20.278 25.286 14.411 8.969
2.500 None 20.554 26.663 10.407 7.152
factorial 24.120 26.169 10.199 7.089
Axial 20.219 35.637 10.335 7.197
Center 20.120 26.167 13.158 8.454
2.75 None 20.259 28.442 8.315 6.223
factorial 24.008 27.942 8.187 6.195
Axial 19.993 38.248 8.479 6.371
Center 19.855 27.984 9.930 7.026
3.00 None 20.009 30.006 6.406 5.314
factorial 23.972 29.511 6.344 5.315
Axial 19.830 40.414 6.664 5.524
Center 19.625 29.514 7.281 5.767

(b)
Fig.1: Loss curves due to single missing observation (a), and relative
factor CCD under various values of
(a)
(c)
Loss curves due to single missing observation (a), and relative G, and V-efficiency plots for a 2
factor CCD under various values of α.
13
efficiency plots for a 2-

(b)
for a 3-factor CCD under various values of
(a)
(c)
Loss curves due to single missing observation (a), and relative G, and V-efficiency plots ((b) and (c))
14
efficiency plots ((b) and (c))

(b)
for a 4-factor CCD under various values of
(a)
(c)
15

(b)
Fig.4: Loss curves due to single missing observation (a),
for a 5
(a)
(c)
for a 5-factor CCD under various values of α.
16

17
4.Conclusions
This study has shown that missing a single observation of any of the three different categories of
points in a central composite design (factorial, axial, and center points) adversely affects the
designs predictive and estimation capabilities. The trace criterion is a decreasing function of α as
can be seen directly from the presented tables; thus we can obtain a good design in terms of this
criterion by taking large values for α during experimentation. The tables of variances of
parameter estimates corresponding to the missing observations show adverse effects on the
precision of the parameters (intercept, linear, quadratic, and interaction) due to these missing
points. It was observed that the extent of the effect depends on the type of the missing point and
also on the distance (ߙ) of the axial points from the design center. It was also shown in the tables
of the maximum and the average prediction variances that the location of maximum lies only in
the factorial portion of the designs for low values of α and then in the axial portion as α increases.
The average prediction variances vary with the type of missing point and also with the value of α.
It was observed that for each of the four designs under discussion, the G-efficiency is robust to a
missing center point but is highly affected by a missing axial point as the number of factors of
the design increases. We also observed that the effect of missing a single factorial point on the
design’s predictive capability decreases as the number of factors of the design increases. For
each of the designs, it was observed that the smallest loss in A-efficiency corresponds to a
missing center point and at α = 1.00. Therefore, we recommend here that the practitioner should
always endeavor to investigate the effect of missing observations on CCDs in terms of more than
just one optimality criterion when faced with the challenges of the design choice. Knowledge of
all these will alert the practitioner prior to data collection as to where in the design region he/she
should expect the worst prediction and where to collect more data if necessary.
References
[1]Akhtar, M. and Prescott, P. (1986): Response Surface Designs Robust to Missing Observations,
Communications in Statistics, Simulations 15(2), pp 345-363.
[2] Ahmad,T. and Gilmour, S. G. (2010): Robustness of Subset Response Surface Designs to missing
Observations, Journal of Statistical Planning and Inference, 140, 92-103.
[3] Box, G. E. P. and Wilson, K.B. 1951, On the Experimental Attainment of Optimum Conditions (with
discussions). Journalof the Royal Statistical Society Series B 13(1):1-45.
[4] Box, G. E. P. and Hunter, J. S. 1957, Multi-Factor Experimental Designs for Exploring Response
Surfaces. The Annals of Mathematical Statistics, 28, pp. 195-241.
[5] Box,G.E.P. and Draper, N.R. 1959. A basis for the selection of a response surface designs. Journal of
American Statistical Association, 54: 622-654.
[6] Box,G.E.P. and Draper, N.R. 1963. The choice of a second order rotatable design. Biometrika 50:
335-352.
[7] Myers, R. H., Montgomery, D. C. and Christine, M. A. (2009): Response Surface Methodology:
Process and Product Optimization Using Designed Experiments, 3rd
Ed., John Wiley and Sons inc.
[8] Ahmad, T., Akhtar, M., and Gilmour, S. G. (2012): Multilevel Augmented Pairs Second-order
Response Surface Designs and their Robustness to Missing Data. Communications in Statistics-Theory
and Methods, 41(3): 437- 452.

18
Authors
Mr Yisa Yakubu holds an M.Sc. degree in statistics, from the University of Ibadan, Nigeria and is
currently a Ph.D. student in the Department of Statistics, University of badan, Oyo State, Nigeria. His
research interests are experimental design and response surface methodology.
Dr. Angela U. Chukwu, is a Senior Lecturer in the Department of Statistics,University of
Ibadan, Oyo state, Nigeria. She holds a Ph.D. degree in statistics from the University of
Ibadan. Her research interests are biometrics and environmental statistics.

Effects of missing observations on

More Related Content

What's hot (19)

Viewers also liked (16)

Similar to Effects of missing observations on (20)

Recently uploaded (20)

Effects of missing observations on