mapping disease risk in space and time, re-mapping

1
STA 517 – Introduction: Distribution and Inference
Topics
1. Introduction: Distributions and Inference for
Categorical Data
2. Describing Contingency Tables
3. Inference for Contingency Tables
4. Introduction to Generalized Linear Models
5. Logistic Regression
6. Building and Applying Logistic Regression Models
7. Logit Models for Multinomial Responses

2
Chapter 1 - Outline
1.1 Categorical Response Data
1.2 Distributions for Categorical Data
1.3 Statistical Inference for Categorical
Data
1.4 Statistical Inference for Binomial
Parameters
1.5 Statistical Inference for Multinomial
Parameters

3
1.1 CATEGORICAL RESPONSE DATA
 A categorical variable has a measurement scale
consisting of a set of categories.
 political philosophy: liberal, moderate, or
conservative.
 brands of a product: brand A, brand B, and brand C
 A categorical variable can be a response variable or
independent variable
 We consider primarily the CATEGORICAL RESPONSE
DATA in this course

4
1.1.1 Response–Explanatory Variable
Distinction
 Most statistical analyses distinguish between response
(or dependent) variables and explanatory (or
independent) variables.
 For instance, regression models:
selling price of a house = f(square footage, location)
 In this book we focus on methods for categorical
response variables.
 As in ordinary regression, explanatory variables can be
of any type.

5
1.1.2 Nominal–Ordinal Scale
Distinction
 Nominal: Variables having categories without a natural
ordering
 religious affiliation: Catholic, Protestant, Jewish,
Muslim, other.
 mode of transportation: automobile, bicycle, bus,
subway, walk
 favorite type of music: classical, country, folk, jazz,
rock
 choice of residence: apartment, condominium,
house, other.
 For nominal variables, the order of listing the
categories is irrelevant.
 The statistical analysis does not depend on that
ordering.

6
Nominal or Ordinal
 Ordinal: ordered categories
 automobile: subcompact, compact, midsize, large
 social class: upper, middle, lower
 political philosophy: liberal, moderate, conservative
 patient condition: good, fair, serious, critical.
 Ordinal variables have ordered categories, but
distances between categories are unknown.
 Although a person categorized as moderate is more
liberal than a person categorized as conservative, no
numerical value describes how much more liberal that
person is. Methods for ordinal variables utilize the
category ordering.

7
Interval variable
 An interval variable is one that does have numerical
distances between any two values.
 blood pressure level
 functional life length of television set
 length of prison term
 annual income
 An internal variable is sometimes called a ratio variable
if ratios of values are also valid. It has a clear
definition of 0:
 Height
 Weight
 enzyme activity

8
categories are not as clear cut as
they sound
 What kind of variable is color?
 In a psychological study of perception, different
colors would be regarded as nominal.
 In a physics study, color is quantified by
wavelength, so color would be considered a ratio
variable.
 What about counts?
 If your dependent variable is the number of cells in
a certain volume, what kind of variable is that. It
has all the properties of a ratio variable, except it
must be an integer.
 Is that a ratio variable or not? These questions just
point out that the classification scheme is appears
to be more comprehensive than it is

9
 A variable’s measurement scale determines which statistical
methods are appropriate.
 In the measurement hierarchy,
 interval variables are highest,
 ordinal variables are next,
 and nominal variables are lowest.
 Statistical methods for variables of one type can also be used
with variables at higher levels but not at lower levels.
 For instance, statistical methods for nominal variables can be
used with ordinal variables by ignoring the ordering of
categories.
 Methods for ordinal variables cannot, however, be used with
nominal variables, since their categories have no meaningful
ordering.
 It is usually best to apply methods appropriate for the actual
scale.

10
1.1.3 Continuous–Discrete Variable
Distinction
 according to the number of values they can take
 Actual measurement of all variables occurs in a discrete
manner, due to precision limitations in measuring
instruments.
 The continuous / discrete classification, in practice,
distinguishes between variables that take lots of values
and variables that take few values.
 Statisticians often treat discrete interval variables
having a large number of values, such as test scores,
as continuous

11
This class: Discretely measured
responses can be:
 Binary (two categories)
 nominal variables (unordered)
 ordinal variables (ordered)
 discrete interval variables having relatively few values,
and
 continuous variables grouped into a small number of
categories.

12
1.1.4 Quantitative–Qualitative
Variable Distinction
 Nominal variables are qualitative distinct categories
differ in quality, not in quantity.
 Interval variables are quantitative distinct levels have
differing amounts of the characteristic of interest.
 The position of ordinal variables in the quantitative or
qualitative classification is fuzzy.

13
 Analysts often utilize the quantitative nature of ordinal
variables by assigning numerical scores to categories or
assuming an underlying continuous distribution.
 This requires good judgment and guidance from
researchers who use the scale, but it provides benefits
in the variety of methods available for data analysis.

14
Summary
 Continuous variable
 Ratio
 Interval
 Discrete
 Categorical
 Binary
 Ordinal
 Nominal

15
Calculation:
OK to compute.... Nominal Ordinal Interval Ratio
frequency distribution Yes Yes Yes Yes
median and percentiles No Yes Yes Yes
add or subtract No No Yes Yes
mean, standard
deviation, standard
error of the mean
No No Yes Yes
ratio, or coefficient of
variation
No No No Yes

16
Example1: Grades measured
 pass/fail
 A,B,C,D,F
 3.2, 4.1, 5.0, 2.1, …
 86,71,99 … of 100

17
Example 2
o Did you get a flu? (Yes or No) – is a binary nominal
categorical variable
o What was the severity of your flu? (low, medium, or
high) – is an ordinal categorical variable
 Context is important. The context of the study and
corresponding questions are important in specifying
what kind of variable we will analyze.

18
1.2 DISTRIBUTIONS FOR
CATEGORICAL DATA
 Inferential data analyses require assumptions about the
random mechanism that generated the data.
 For continuous variable, Normal distribution
 For categorical variable
 Binomial
 hypergeometric distribution
 Multinomial
 Poisson

19
Overview of probability and
inference
 The basic problem we study in probability: Given a data
generating process, what are the properties of the
outcomes?
 The basic problem of statistical inference: Given the
outcomes (data), what we can say about the process
that generated the data?
Observed
data
Data
generating
process
probability
inference

20
Random variable
 A random variable is the outcome of an experiment
(i.e. a random process) expressed as a number.
 We use capital letters near the end of the alphabet (X,
Y , Z, etc.) to denote random variables.
 Just like variables, probability distributions can be
classified as discrete or continuous.

21
Continuous Probability Distributions
 If a random variable is a continuous variable, its
probability distribution is called a continuous probability
distribution.
 A continuous probability distribution differs from a
discrete probability distribution in several ways.
 The probability that a continuous random variable
will assume a particular value is zero.
 As a result, a continuous probability distribution
cannot be expressed in tabular form.
 Instead, an equation or formula is used to describe
a continuous probability distribution.

22
Normal
 Most often, the equation used to describe a continuous
probability distribution is called a probability density
function. Sometimes, it is referred to as a density
function, or a PDF.
 Normal N(µ, 2
) PDF
}
2
)
(
exp{
2
1
)
,
;
( 2
2
2
2








x
x
f

23
Chi-square distribution, PDF

24
Discrete random variables
 A discrete random variable is one which may take on only
a countable number of distinct values such as 0,1,2,3,4,........
 Discrete random variables are usually (but not necessarily)
counts. If a random variable can take only a finite number of
distinct values, then it must be discrete.
 Examples:
 the number of children in a family
 the Friday night attendance at a cinema
 the number of patients in a doctor's surgery
 the number of defective light bulbs in a box of ten.

25
discrete random variable
 The probability distribution of a discrete random variable
is a list of probabilities associated with each of its possible
values. It is also sometimes called the probability function or
the probability mass function.
 Suppose a random variable X may take k different values,
with the probability that X = xi defined to be P(X = xi) = pi.
The probabilities pi must satisfy the following:
 0 < pi < 1 for each i
 p1 + p2 + ... + pk = 1.

26
Example
 Suppose a variable X can take the values 1, 2, 3, or 4.
The probabilities associated with each outcome are
described by the following table:
Outcome 1 2 3 4
Probability 0.1 0.3 0.4 0.2
 The probability that X is equal to 2 or 3 is the sum of
the two probabilities: P(X = 2 or X = 3) = P(X = 2) +
P(X = 3) = 0.3 + 0.4 = 0.7.
 Similarly, the probability that X is
greater than 1 is equal to
1 - P(X = 1) = 1 - 0.1 = 0.9,
by the complement rule.
 This distribution may
also be described by
the probability histogram
shown to the right

27
Properties
 E(X)= x f(x)
 var(X)= (x-E(X))2
f(x)
 If the distribution depends on unknown parameters 
we write it as f(x; ) or f(x | )

28
1.2.0 Bernoulli Distribution
 the Bernoulli distribution is a discrete
probability distribution, which takes value 1 with
success probability  and value 0 with failure
probability 1 − . So if X is a random variable with this
distribution, we have:
or write it as
Then

29
1.2.1 Binomial Distribution
 Many applications refer to a fixed number n of binary
observations.
 Let y1 , y2 , . . . , yn denote responses for n
independent and identical trials (Bernoulli trials)
 Identical trials means that the probability of success 
is the same for each trial.
 Independent trials means that the Yi are independent
random variables.

30
 The total number of successes
 has the binomial distribution with index n and
parameter , denoted by bin(n,)
 The probability mass function
where

31
0 5 10 15 20 25
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
binomial pdf bin(25, )
=0.10
=0.25
=0.50

32
Moments
Because Yi=1 or 0, Yi=Yi
2
E(Yi)=E(Yi
2
)=1 x  + 0 x (1-)=
Skewness:

33
The distribution converges to
normality as n increases
0 5 10 15 20 25
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
Binomial(25, 0.25)
Normal

34
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Binomial(5, 0.25)
Normal(1.25,0.96825)

35
1.2.2 Multinomial Distribution
Multiple possible outcomes
 Suppose that each of n independent, identical trials can
have outcome in any of c categories.
if trial i has outcome in category j
= 0 otherwise

represents a multinomial trial, with
 Let denote the number of trials having
outcome in category j.
 The counts have the multinomial
distribution.

36
pdf:

37
1.2.3 Poisson Distribution
 count data do not result from a fixed number of trials.
 y=number of deaths due to automobile accidents on
motorways in Italy
 y>0
 Poisson probability mass function (Poisson 1837)
 It satisfies
 Skewness

38
Poisson
0 5 10 15 20 25 30 35
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
=5
=10
=15

39
Poisson Distribution
 used for counts of events that occur randomly over
time or space, when outcomes in disjoint periods or
regions are independent.
 an approximation for the binomial when n is large and
 is small with µ=n
 For example,
 n=50 million driving in Italy
 death rate/week =0.000002
 the number of deaths is bin(n, )
 Or approximately Poisson with µ=n=100

40
1.2.4 Overdispersion
 A key feature of the Poisson distribution is that its
variance equals its mean.
 Sample counts vary more when their mean is
higher.
 Overdispersion: Count observations often exhibit
variability exceeding that predicted by the binomial or
Poisson.

41
1.2.5 Connection between Poisson
and Multinomial Distributions
 Example, In Italy this next week, let
 y1=# of people who die in automobile accidents
 y2=number who die in airplane accidents
 y3=number who die in railway accidents
 (Y1, Y2, Y3) ~ independent Poisson ( )
 The total ~ Poisson ( )
 Here n is random variable rather than fixed
 If n is given, (Y1, Y2, Y3) is no longer independent and
Poisson, WHY

42
conditional distribution given that
let
~ multinomial distribution

43
multinomial distribution
vs. Poisson distribution
 Many categorical data analyses assume a multinomial
distribution.
 Such analyses usually have the same parameter
estimates as those of analyses assuming a Poisson
distribution, because of the similarity in the likelihood
functions.

44
1.3 STATISTICAL INFERENCE FOR
CATEGORICAL DATA(general)
 Once you choose the distribution of the categorical
variable, you need to estimate the parameters in the
distribution
 We first review general method
 Point estimate
 Confidence interval
 Section 1.4 MLE for binomial
 Section 1.5 MLE for multinominal

45
Likelihood
 Likelihood is a tool for summarizing the data's evidence
about unknown parameters. Let us denote the
unknown parameter(s) of a distribution generically by
.
 If we observe a random variable X = x from distribution
f (x|), then the likelihood associated with x, l(|x), is
simply the distribution f (x|) regarded as a function of
 with x fixed.
 For example, if we observe x from bin(n; ), the
likelihood function is
x
n
x
x
n
x
l 









 )
1
(
)
|
( 



46
Likelihood
 The formula for the likelihood looks similar algebraically
to the f (x|) but the distinction should be clear!
 The distribution function is defined over the support of
discrete variable x with  given, whereas the likelihood
is defined over the continuous parameter space for .
 Consequently, a graph of the likelihood usually looks
different from a graph of the probability distribution.
 In most cases, we work with loglikelihood
)
|
(
log
)
|
( x
l
x
L 
 
)
1
log(
)
(
log
)
1
log(
)
(
log
log
)
|
(
log
)
|
(
























x
n
x
x
n
x
x
n
x
l
x
L

47
Loglikelihood function bin(5,) and
we observe x=0, x=1, and x=2
0 0.5 1
-30
-20
-10
0

l
(

|
x
)
l (  | x )=x log +(n-x) log (1-)
0 0.5 1
-40
-30
-20
-10
0

l
(

|
x
)
l (  | x )=x log +(n-x) log (1-)
0 0.5 1
-20
-15
-10
-5
0

l
(

|
x
)
l (  | x )=x log +(n-x) log (1-)
0 0.2 0.4 0.6 0.8 1
-7000
-6000
-5000
-4000
-3000
-2000
-1000

l
(

|
x
)
l (  | x )=x log +(n-x) log (1-)
bin(842+982,)
x=842 (yes)

48
Likelihood
 In many problems of interest, we will derive our
loglikelihood from a sample rather than from a single
observation. If we observe an independent sample x1,
x2, …, xn from a distribution f (x|), then the overall
likelihood is the product of the individual likelihoods:
and loglikelihood is

 



n
i
i
n
i
i
n x
l
x
f
x
x
l
1
1
1 )
|
(
)
|
(
)
,
,
|
( 

 









n
i
i
n
i
i
n
i
i
n
x
L
x
f
x
f
x
x
L
1
1
1
1
)
|
(
)
|
(
log
)
|
(
log
)
,
,
|
(



 

49
Log likelihood
 In regular problems, as the total sample size n grows,
the loglikelihood function does two things:
(a) it becomes more sharply peaked around its
maximum,
and (b) its shape becomes nearly quadratic
 the loglikelihood for a normal-mean problem is exactly
quadratic.
 That is, if we observe y1, . . . , yn from a normal
population with known variance, the loglikelihood is
or in multi-dimension

50
MLE (maximum likelihood
estimation)
 ML estimate for θ is the maximizer of L(θ) or,
equivalently, the maximizer of l(θ). This is the
parameter value under which the data observed have
the highest probability of occurrence.
 In regular problems, the ML estimate can be found by
setting to zero the first derivative(s) of l(θ) with
respect to θ.

51
Transformations of parameters
 If l(θ) is a likelihood and φ = g(θ) is a one-to-one
function of the parameter with back-transformation θ =
g−1
(φ), then we can express the likelihood in terms of
φ as l( g−1
(φ) ).
 Transformations may help us to improve the shape of
the loglikelihood.
 If the parameter space for θ has boundaries, we may
want to choose a transformation to the entire real
space.
 For example, consider the binomial loglikelihood,
L

52
binomial loglikelihood
 If we apply the logit transformation
 whose back-transformation is
 the loglikelihood in terms of β is
L

53
 If we observe y = 1 from a binomial with n = 5, the
loglikelihood in terms of β looks like this.

54
 n=5; y=2;
 pi=0:0.01:1
 subplot(1,2,1)
 plot(pi, y*log(pi)+(n-y)*log(1-pi), '-')
 subplot(1,2,2)
 beta=-3:0.1:3
 plot(beta, y*beta+n*log(1./(1+exp(beta))), '-')
0 0.2 0.4 0.6 0.8 1
-14
-12
-10
-8
-6
-4
-2
-3 -2 -1 0 1 2 3
-10
-9
-8
-7
-6
-5
-4
-3

55
 Transformations do not affect the location of the
maximum-likelihood (ML) estimate.
 If l(θ) is maximized at ˆθ, then l(φ) is maximized at ˆφ
= g(ˆθ).

56
score function
 A first derivative of L(θ) with respect to θ is called a
score function or simply a score.
 In a one-parameter problem, the score function from
an independent sample y1, . . . , yn is
where
is the score contribution for yi.
 The ML estimate is usually the solution of the likelihood
equation L’(θ)=0.
L

57
Mean of the score function.
 A well known property of the score is that it has mean
zero.
 The score is an expression that involves both the
parameter θ and the data Y . Because it involves Y , we
can take its expectation with respect to the data
distribution f(y|θ). The expected score is no longer a
function of Y , but it’s still a function of θ. If we
evaluate this expected score at the “true value” of θ—
that is, at the same value of θ assumed when we took
the expectation—we get zero:
If certain differentiability conditions are met, the integral may be
rewritten as

58
 For example, in the case of the binomial proportion, we
have
which is zero because E(Y ) = n.
 If we apply a one-to-one transformation to the
parameter φ = g(θ), then the score function with
respect to the new parameter φ also has mean zero.

59
Estimating functions.
 This property of the score function—that it has an
expectation of zero when evaluated at the true
parameter θ—is a key to the modern theory of
statistical estimation.
 In the original theory of likelihood-based estimation, as
developed by R.A. Fisher and others, the ML estimate
ˆθ is viewed as the value of the parameter that, under
the parametric model, that makes the observed data
most likely.
 statisticians have begun to view ˆθ as the solution the
score equation(s). That is, we now often view an ML
estimate as the solution to
L’(θ)=0

60
estimating equations
 Any function of the data and the parameters having
mean zero at the true θ has this property as well.
Functions having the mean-zero property are called
estimating functions.
 Setting the estimating functions to zero is called
the estimating equations.
 In the case of the binomial proportion, for example,
Y − n
is a mean-zero estimating function, and so is
−1
[Y − n] .

61
Information and variance estimation.
 The variance of the score is known as the Fisher
information. In the case of a single parameter, the
Fisher information is
 If θ has k parameters, the Fisher information is the
k x k covariance matrix for scores

62
 Like the score function, the Fisher information is also a
function of θ. So we can evaluate it at any given value
of θ.
 Notice that i(θ) as we defined it is the square of a sum
which, in many problems, can be messy.
 To actually compute the Fisher information, we usually
make use of the well known identity

63
 In the multiparameter case, l(θ) is the k x k matrix of
second derivatives
whose (l,m)th element is

64
 why we care about the Fisher information?
 it provides us with a way (several ways, actually) of
assessing the uncertainty in the ML estimate.
 It is well known that, in regular problems, ˆθ is
approximately normally distributed about the true θ
with variance given by the reciprocal (or, in the
multiparameter case, the matrix inverse) of the Fisher
information.

65
two common ways to
approximate the variance of ˆθ.
 The first way is to plug ˆθ into i(θ) and invert,
this is commonly called the “expected information.”
 The second way is to invert (minus one times) the
actual second derivative of the loglikelihood at θ = ˆθ,
this is commonly called the “observed information.”

66
1.3.2 Likelihood Function and ML
Estimate for Binomial Parameter
 The binomial log likelihood is
 Differentiating with respect to  yields
 Equating this to 0 gives the likelihood equation, which
has solution
the sample proportion of successes for the n trials.

67
 Calculating , taking the expectation, and
we get
 Thus, the asymptotic variance of is
 Actually, from E(Y)=n and var(Y)=n (1- ), the
distribution if =Y/n has mean and standard error

68
Likelihood function and MLE
summary
 We use maximum likelihood estimate (MLE)
 asymptotically normal
 asymptotically consistent
 asymptotically efficient
 Likelihood function
 probability of those data, treated as a function of
the unknown parameter.
 maximum likelihood (ML) estimate
 parameter value that maximizes this function

69
MLE and its variance
 If y1; y2; … ; yn is a random sample from distribution
f(y|), then the score function is
 In regular problems, we can find the ML estimate by
setting the score function(s) to zero and solving for .
 The equations L’(θ)=0 are called the score equations.
More generally, they can be called estimating equations
because their solution is the estimate for θ.
 We defined the Fisher information as the variance of
the score function and

70
1.3.3 Wald–Likelihood Ratio–Score
Test Triad
 Three standard ways exist to use the likelihood function
to perform large-sample inference.
 Wald test
 Score test
 Likelihood ratio test
 We introduce these for a significance test of a null
hypothesis
H0:
and then discuss their relation to interval estimation.
 They all exploit the large-sample normality of ML
estimators.

71
Wald test
 With nonnull standard error SE of , the test statistic
has an approximate standard normal distribution when
 One- or two-sided P-value by z.
 Or z2
has a chi-squared null distribution with 1 df
The P-value is then the right-tailed chi-squared
probability above the observed value
This type of statistic, using the nonnull standard
error, is called a Wald statistic (Wald 1943).

72
Wald test
 For an .05-level two-side test, we reject H0 if
 Equivalently, if
where 3.84 is the 95th
percentile of 2
(1).
96
.
1
ˆ
0


SE


2
2
0
96
.
1
84
.
3
)
ˆ
var(
)
ˆ
(







73
Wald test
 The multivariate extension for the Wald test of
has test statistic
where is the inverse matrix of Information
matrix.
 W is an asymptotic chi-squared distribution with df =
rank of .
 Wald test is not invariant to transformations. That is, a
Wald test on a transformed parameter φ= g() may
yield a different p-value than a Wald test on the
original scale.

74
 uses the likelihood function through the ratio of two
maximizations:
(1). the maximum over the possible parameter values
under H0
(2). the maximum over the larger set of parameter values
permitting H0 or an alternative Ha to be true.
 The likelihood-ratio test statistic equals
where L0 and L1 denote the maximized log-likelihood
functions.
 is 2
distribution with df=dim(Ha U H0)-dim(H0)
 Reject H0 if
> 2
(=0.05)
Likelihood ratio test

75
 The score test is based on the slope and expected
curvature of the log-likelihood function L() at the null
value 0.
 Score function
The value tends to be larger in absolute value
when is farther from 0.
 Score statistic
has an approximate standard normal null distribution.
 The chi-squared form of the score statistic is
Score test
)
/
)
(
(
)
(
2
0
2
0





 L
E
u
)
/
)
(
(
)
(
2
0
2
0
2





 L
E
u

76
Why is score statistic reasonable?
 Recall that the mean of the score is zero and its
variance is equal to the Fisher information.
 In a large sample, the score will also be approximately
normally distributed because it's a sum of iid random
variables.
 Therefore, it will behave like a squared standard normal
[2
(1)] if H0 is true.

77
Wald–Likelihood Ratio–Score Test

78
 The three test statistics - Wald, LR and score are
asymptotically equivalent.
 The differences among them vanish in large samples if
the null hypothesis is true.
 If the null hypothesis is false, they may take very
different values. But in that case, all the test statistics
will be large, the p-values will be essentially zero, and
they will all lead us to reject H0.
 Score test does not require to calculate MLE.
 LR test is scale-invariant.
 LR statistic uses the most information of the three
types of test statistic and is the most versatile.

79
1.3.4 Constructing confidence
intervals
 In practice, it is more informative to construct
confidence intervals for parameters than to test
hypotheses about their values.
 For any of the three test methods, a confidence interval
results from inverting the test. For instance, a 95%
confidence interval for is the set of 0 for which the
test of H0: has a P-value exceeding 0.05.
 Let denote the z-score from the standard normal
distribution having right-tailed probability a; this is the
100(1-a) percentile of that distribution.
 Let denote the 100(1-a) percentile of the chi-
squared distribution with degrees of freedom df.

80
Tests and Confidence Intervals
At significant level ,
reject H0: , if
2
/
0
ˆ



z
SE


100(1-)%
confidence interval
2
/
0
ˆ



z
SE


 }
:
{ 0

 }
:
{ 0

0

 

81
Confidence Intervals
 The Wald confidence interval is most common in
practice because it is simple to construct using ML
estimates and standard errors reported by statistical
software.
 The likelihood-ratio-based interval is becoming more
widely available in software and is preferable for
categorical data with small to moderate n.
 For the best known statistical model, regression for a
normal response, the three types of inference
necessarily provide identical results.

82
1.4 STATISTICAL INFERENCE FOR
BINOMIAL PARAMETERS
 Recall log likelihood
 Score function
 MLE
SE=
)
1
log(
)
(
log
)
|
( 

 


 y
n
y
y
L
)
1
/(
)
(
/
)
( 

 


 y
n
y
u

83
1.4.1 Tests about a Binomial
Parameter
 Since H0 has a single parameter, we use the normal
rather than chi-squared forms of Wald and score test
statistics. They permit tests against one-sided as well
as two-sided alternatives.
 Wald statistic
 Evaluating the binomial score and information at 0
 The normal form of the score statistic simplifies to

84
 binomial log-likelihood under H0
 Under Ha
 The likelihood-ratio test statistic
or
has an asymptotic chi-squared distribution with df=1.
)
1
log(
)
(
log 0
0
0 
 


 y
n
y
L
)
ˆ
1
log(
)
(
ˆ
log
1 
 


 y
n
y
L

85
Test
 At significant level , two sided, reject H0, if
(Wald test)
(Score test)
(LR test)




86
1.4.2 Confidence Intervals for a
Binomial Parameter
 Inverting the Wald test,
 Unfortunately, it performs poorly unless n is very large
 The actual coverage probability usually falls below the
nominal confidence coefficient, much below when  is
near 0 or 1.
 An adjustment is needed. (Problem 1.24)

87
Simulation to calculate coverage prob.
%let n=1000; %let pi=0.5; %let simuN=10000;
data simu; drop i;
do i=1 to &simuN;
k=RAND('BINOMIAL',&pi,&n); output;
end;
run;
data res; set simu;
pihat=k/&n;
lci=pihat-1.96*sqrt(pihat*(1-pihat)/&n);
uci=pihat+1.96*sqrt(pihat*(1-pihat)/&n);
if lci>&pi or uci<&pi then cover=0; else cover=1;
proc sql;
select sum(cover)/&simuN as coverageprobabilty from res;

88
it performs poorly if 1) n is small; 2)
pi near 0 or 1.
 %let n=1000; %let pi=0.5; %let simuN=10000;

89
An adjustment is needed. (Problem
1.24)

90
%let n=20; %let pi=0.5; %let simuN=10000;
data simu; drop i;
do i=1 to &simuN;
k=RAND('BINOMIAL',&pi,&n); output;
end;
run;
data res; set simu;
pihat=(k+1.96**2/2)/(&n+1.96*1.96);
lci=pihat-1.96*sqrt(pihat*(1-pihat)/(&n+1.96*1.96));
uci=pihat+1.96*sqrt(pihat*(1-pihat)/(&n+1.96*1.96));
if lci>&pi or uci<&pi then cover=0; else cover=1;
proc sql;
select sum(cover)/&simuN as coverageprobabilty from res;

91
score confidence interval
 The score confidence interval contains 0 values for
which
 Its endpoints are the 0 solutions to the equations
 It is quadratic in 0. This interval is

92
LR-based confidence interval
 The likelihood-ratio-based confidence interval is more
complex computationally, but simple in principle.
 It is the set of 0 for which the likelihood ratio test has
a P-value exceeding  .
 Equivalently, it is the set of 0 for which double the log
likelihood drops by less than from its value at
the ML estimate.


93
1.4.3 Proportion of Vegetarians
Example

mapping disease risk in space and time, re-mapping

More Related Content

Similar to mapping disease risk in space and time, re-mapping (20)

More from yonas381043 (6)

Recently uploaded (20)

mapping disease risk in space and time, re-mapping