User manual | Recent Advances in Categorical Data Analysis Maura E. Stokes SAS Institute Inc.

Statistics, Data Analysis, and Modeling
Paper 268
Recent Advances in Categorical Data Analysis
Maura E. Stokes
SAS Institute Inc.
Cary, North Carolina, USA
Abstract
The last fifteen years have brought many changes to
the practice of categorical data analysis. This paper reviews some of the major changes and shifts
of emphasis and discusses several examples using
SAS software procedures. Topics include the use
of exact methods, Generalized Estimating Equations,
conditional logistic regression, and current uses of
weighted least squares modeling. Applications provide illustrations for many topics. This paper describes software currently available in the SAS System and indicates the areas where users can expect to find strategies implemented in the next few
releases. The references include some pertinent
methodology and review papers.
Introduction
Fifteen years ago, the SUGI paper describing the categorical data analysis enhancements in the upcoming Version 5 release described the new CATMOD
procedure that would replace PROC FUNCAT and
also described how the FREQ procedure would now
include the Mantel-Haenszel nonparametric statistics
then present in PROC TFREQ. The paper then continued with a discussion directed at weighted least
squares methods for statistical modeling and MantelHaenszel methods for testing association in a contingency table or sets of contingency tables.
modeling repeated measurements data that can include continuous covariates, missing data, and timedependent covariates. Highly stratified data can be
handled with conditional logistic methods, which also
provide a way of analyzing crossover data. When
asymptotic assumptions are not appropriate due to
sparse data or small sample sizes, exact methods
can provide a way to produce appropriate p-values
for tests, valid confidence limits for odds ratios, and
parameter estimates and standard errors in logistic
regression. Weighted least squares, a very important
strategy in the 70s and 80s, still provides a useful way
of modeling functions of interest such as rank measures of association and incidence densities.
What follows are descriptions of these newer strategies and illustrations of their application.
Exact p-Values
Since that time, new developments in categorical data
analysis and new computing strategies have changed
the typical strategies employed by the data analyst
facing categorical data. Once safely armed with
her CATMOD documentation and PROC FREQ, then
patiently learning the difference between the supplemental procedure PROC LOGIST contributed by
Frank Harrell and PROC LOGISTIC, the analyst now
contends with a number of different procedures such
as LOGISTIC, GENMOD, and PHREG, and keeps a
scorecard of the growing number of exact tests available in the FREQ and NPAR1WAY procedures.
Exact p-values provide an alternative strategy when
data are sparse, skewed, or unbalanced so that the
assumptions required for standard asymptotic tests
are violated. Advances in computer performance
and developments in network algorithms over the last
decade have made exact p-values accessible for a
number of statistical tests. In Release 6.11, exact p-values were added for the simple linear rank
statistics produced by the NPAR1WAY procedure.
In Release 6.12, exact p-values are produced for
many of the statistics computed by the FREQ procedure. You are now able to request exact p-values
for the following chi-square statistics: Pearson’s chisquare, likelihood-ratio chi-square, Mantel-Haenszel
chi-square, Fisher’s exact test and r by c exact test,
Jonckheere-Terpstra test, and McNemar’s test. In
addition, you can also obtain exact p-values for hypothesis tests that the following statistics are equal to
0: Pearson correlation coefficient, Spearman correlation coefficient, simple kappa statistic, and weighted
kappa statistic. Exact confidence bounds are also
available for the odds ratios produced for 2 by 2 tables.
Mantel-Haenszel strategies are now employed in the
analysis of repeated measurements and crossover
studies. GEE methods provide a convenient way of
In Version 7, a test for the binomial proportion is
available along with an exact p-value, and exact pvalues are available for the Pearson chi-square statis-
Statistics, Data Analysis, and Modeling
tic for two-way tables and the goodness-of-fit statistic for one-way tables. In addition, the Monte Carlo
method of computing exact p-values is included in
the NPAR1WAY procedure and will be included in the
FREQ procedure in Version 8.
The following example illustrates the use of the new
EXACT statement to produce an exact p-value for the
simple kappa statistic. Researchers studied two scoring systems for evaluating fitness in fifth grade students. Forty-three students were classified into one
of four fitness categories. Interest lies in determining
whether there is agreement between the two scoring
systems, which can be assessed by testing whether
the kappa coefficient is equal to 0.
data fitness;
input score1 $ score2 $ count;
datalines;
poor poor 5
average average 4
good good 4
superior superior 3
poor average 3
average poor 1
average good 6
good average 5
good superior 1
superior average 10
superior good 1
;
TABLE OF SCORE_1 BY SCORE_2
SCORE_1
SCORE_2
Frequency|
Percent |
Row Pct |
Col Pct |average |good
|poor
|superior| Total
---------+--------+--------+--------+--------+
average |
4 |
6 |
1 |
0 |
11
|
9.30 | 13.95 |
2.33 |
0.00 | 25.58
| 36.36 | 54.55 |
9.09 |
0.00 |
| 18.18 | 54.55 | 16.67 |
0.00 |
---------+--------+--------+--------+--------+
good
|
5 |
4 |
0 |
1 |
10
| 11.63 |
9.30 |
0.00 |
2.33 | 23.26
| 50.00 | 40.00 |
0.00 | 10.00 |
| 22.73 | 36.36 |
0.00 | 25.00 |
---------+--------+--------+--------+--------+
poor
|
3 |
0 |
5 |
0 |
8
|
6.98 |
0.00 | 11.63 |
0.00 | 18.60
| 37.50 |
0.00 | 62.50 |
0.00 |
| 13.64 |
0.00 | 83.33 |
0.00 |
---------+--------+--------+--------+--------+
superior |
10 |
1 |
0 |
3 |
14
| 23.26 |
2.33 |
0.00 |
6.98 | 32.56
| 71.43 |
7.14 |
0.00 | 21.43 |
| 45.45 |
9.09 |
0.00 | 75.00 |
---------+--------+--------+--------+--------+
Total
22
11
6
4
43
51.16
25.58
13.95
9.30
100.00
Figure 1.
Exact Test for Simple Kappa
The resulting exact p-value for the hypothesis that the
simple kappa statistic is equal to 0 is p=0.055, which
may be considered to have marginal significance at
best. Note the value p=0.038 for the asymptotic test.
Using exact p-values for this analysis leads to a very
different conclusion than using the asymptotic test.
STATISTICS FOR TABLE OF SCORE_1 BY SCORE_2
Statistic = 11.091
To request the exact p-value for the kappa statistic,
you specify the keyword KAPPA in the EXACT statement. The AGREE option in the MODEL statement
requests the measures of agreement.
proc freq;
weight count;
tables score1 * score2 / agree;
exact kappa;
run;
Kappa = 0.167
Prob = 0.086
Simple Kappa Coefficient
-----------------------95% Confidence Bounds
ASE = 0.102
-0.032
0.366
Asymptotic P-Values
(Right-sided) = 0.019
(Two-sided)
= 0.038
Exact P-Values
(Right-sided) = 0.034
(Two-sided)
= 0.055
Weighted Kappa Coefficient
-------------------------95% Confidence Bounds
Kappa = 0.100
ASE = 0.100
-0.096
0.297
Sample Size = 43
Figure 2.
The following figure displays the contingency table
form of the data. Note the number of zero cells, which
makes the use of the asymptotic test questionable.
Test of Symmetry
---------------DF = 6
Exact Test for Simple Kappa
Generalized Estimating Equations
Weighted least squares (WLS) modeling of repeated
categorical data was described by Landis et al. (1977)
and provides a large sample asymptotic method that
works nicely for data that have adequate sample size,
a small number of response points, a small number
of categorical explanatory variables measured at the
subject level, and no missing data. The CATMOD procedure introduced the REPEATED statement to provide this analysis.
While this method is still useful for data that meet
these conditions, most data that are collected with
clustered or repeated responses do not. Longitudinal data are usually plagued with missing responses,
Statistics, Data Analysis, and Modeling
explanatory variables include continuous variables as
well as categorical, and there is often interest in timedependent covariates such as blood pressure in a
clinical trial. In addition, as you start to increase the
number of explanatory variables, you often don’t meet
the asymptotic requirements for the WLS approach.
Generalized Estimating Equations (GEE) provides a
nonlikelihood based approach to modeling repeated
or clustered data that applies to a broader set of
data situations that are frequently encountered. It
handles missing data, continous explantory variables,
and time-dependent explanatory variables. While responses can be either continuous or categorical, it
is especially useful for data that are binary or discrete counts. An extension of the generalized linear model (GLM) first suggested by Nelder and Wedderburn (1972), the GEE approach was outlined in
work by Zeger and Liang (1986) and Liang and Zeger
(1986) that describe a quasi-likelihood approach for
modeling correlated responses. Besides using the
linear predictor set-up of the GLM, you model the covariance matrix of the responses. The GEE approach
produces population-averaged estimates. With quasilikelihood, you can pursue statistical models by making assumptions about the link function and the relationship between the first two moments, but without
specifying the complete distribution of the response.
Say that Yij (j = 1; : : : ; ni ; i = 1; : : : ; K ) represent the
jth measurement on the ith subject. There are ni meaPK
surements on subject i and i=1 ni total measurements.
The generalized estimating equation for estimating is an extension of the GLM estimating equation:
K
X
@ 0
=1
i
,1
@ V (Y , ())= 0
i
i
is the corresponding vector of means =
1 ; : : : ; ]0 and V is an estimate of the covariance
i
Y.
i
ini
matrix of
Y
The covariance matrix of
i
V = A R ()A
i
i
1
2
i
i
is modeled as
1
2
A
where i is an ni ni diagonal matrix with v (ij ) as
the jth diagonal element.
The working correlation matrix
r
ij
=
yp , v( )
ij
ij
ij
There are many choices for the working correlation matrix. The independent working correlation
matrix includes 1s on the diagonals and 0s on
the off-diagonals. Other choices are the unstructured, exchangeable (compound symmetry), autoregressive(1), and m-dependent.
Finding the GEE solution requires these steps:
Relate the marginal response ij =E(Yij ) to
with a link function. For example, the logit.
ij
Specify the variance function.
Choose a working correlation matrix
R ().
i
Compute an initial estimate of , for example
with an ordinary generalized linear model assuming independence.
Compute the working correlation matrix
R.
i
Compute an estimate of the covariance matrix:
V = A R^ ()A
i
x0 i
1
2
i
i
1
2
Update :
+1 = ,
r
r
@ 0 V,1 @ ,1 X @ 0 V,1 (Y , )
@
=1 @ =1 @ X
K
K
i
i
i
i
i
i
i
Compute residuals and update
V.
i
Iterate until convergence.
i
where
[
using the current value of the parameter vector to
compute appropriate functions of the Pearson residual.
R () is estimated as
i
The GEE parameter estimates have many important
properties. They are the generalized linear model estimating equations when you have one measurement
per cluster. They are the maximum likelihood score
equations for multivariate Gaussian data. And, most
importantly, the GEE parameter estimates are consistent as the number of clusters becomes large, even if
you have misspecified the working correlation matrix,
as long as the mean model is correct.
^ ) is given by
The model-based estimator of Cov(
^) =
CovM (
I0,1
where
I0 =
K
X
@ 0 ,1 @ =1 @ i
V @
i
Statistics, Data Analysis, and Modeling
This is a consistent estimator if the model and working
correlation matrix are correctly specified.
^ ) is given
The empirical, or robust, estimator of Cov (
by
M = I0,1I1 I,0 1
where
I1 =
K
X
@ 0
V
=1 @ i
,1 Cov(Y)V,1 @ @
^ ) as the numThis is a consistent estimator of Cov(
ber of clusters become large, even if the working correlation matrix is not specified correctly.
The GEE approach produces a marginal model. It
models a known function of the marginal expectation of the dependent variable as a linear function of
explanatory variables. The resulting parameter estimates are population-averaged. GEE relies on independence across subjects to consistently estimate
the variance. Compare this to a mixed model where
you are estimating subject-specific parameter estimates, but you are heavily leveraging the correlation
assumption.
Exercise Study
One of the applications of GEE methods is to
crossover studies. In a crossover study, subjects have
their response measured at different periods under
different conditions. In a classic two period, two treatment crossover study, some subjects get a Treatment
A during Period 1 and Treatment B during Period 2,
while other subjects get the sequence Treatment B
during Period 1 and Treatment A during Period 2.
Usually, there is some sort of washout period so that
whatever effects of the treatment during the first period have washed out before the second period begins. Also, the nature of the response is such that
the subject is able to have the indicated response a
few times within a relatively short period of time. In a
crossover study, the subject acts as his own control.
More complicated designs, including sequences that
can draw from more than two possible treatments
(for example, A, B, and Placebo) and additional periods, can provide a better framework for estimating
the treatment effect. In this example, GEE methods
are used to analyze a three-period crossover study in
which patients with a chronic respiratory condition are
exposed to different levels of air pollution while exercising and measured for respiratory distress on a four
point ordinal scale, ranging from 0 for none to 3 for
severe. A dichotomous baseline distress measurement was taken at the beginning of the study. Six sequences were studied: HML, HLM, MHL, MLH, LHM,
and LMH, where ’H’ means High, ’M’ means medium,
and ’L’ means low.
In this analysis, the subject is the cluster and there
may be a maximum of three response corresponding
to the three periods. Missing responses occurred at
each of the three periods. Interest lies in determining
whether there was a pollution effect, baseline effect,
period, and carryover effects.
The following DATA step inputs the exercise data.
There is one observation per subject per period.
The variable Sequence contains the sequence information, for example, observations with the value
’HML’ received the sequence High in the first period,
Medium in the second period, and Low in the third period. The indicator variables High and Medium take
the value ‘1’ is the exposure is High or Medium, respectively, for that period. ID is the subject ID within
sequence group, Period1 and Period2 are indicator
variables for whether the observation is from Period 1
or Period 2, and CarryHigh and CarryMedium are indicator variables for whether the previous period was
High exposure or Medium exposure. The variable
Baseline takes the value ‘1’ for respiratory distress at
the beginning of the study.
data Exercise;
input Sequence $ ID $ Period1 Period2 High Medium Baseline
Response CarryHigh CarryMedium @@;
strata=sequence||id;
DichotResponse= (Response >0);
datalines;
HML 1 1 0 1 0 0 3 0 0 HML 1 0 1 0 1 0 1 1 0
HML 1 0 0 0 0 0 0 0 1
HML 2 1 0 1 0 0 3 0 0 HML 2 0 1 0 1 0 2 1 0
HML 2 0 0 0 0 0 0 0 1
HML 3 1 0 1 0 1 3 0 0 HML 3 0 1 0 1 0 2 1 0
HML 3 0 0 0 0 0 . 0 1
HML 4 1 0 1 0 0 2 0 0 HML 4 0 1 0 1 0 0 1 0
HML 4 0 0 0 0 0 2 0 1
...
The following statements produce a listing of the number of subjects in each of the sequences.
proc freq;
tables Sequence Response;
run;
The FREQ Procedure
Cumulative
Cumulative
Sequence
Frequency
Percent
Frequency
Percent
------------------------------------------------------HLM
72
16.00
72
16.00
HML
78
17.33
150
33.33
LHM
72
16.00
222
49.33
LMH
72
16.00
294
65.33
MHL
60
13.33
354
78.67
MLH
96
21.33
450
100.00
Figure 3.
Frequencies of Exercise Sequences
Statistics, Data Analysis, and Modeling
The GEE analysis is performed with the GEE facility in the GENMOD procedure. This has been made
much more comprehensive in Version 7 with the inclusion of Type III tests, the CONTRAST, ESTIMATE,
and LSMEANS statement, and the capability of handling the ordinal response with the proportional odds
model. PROC GENMOD also now provides the alternating logistic regression method for binary data.
The following statements request the analysis. The
crossclassification of the variables Sequence and Id
uniquely identify each cluster (subject), so that effect
is specified with the SUBJECT= option in the REPEATED statement. The model consisting of all the
main effects is specified, and the proportional odds
model is requested with the DIST=MULTINOMIAL
and LINK=CLOGIT specifications.
proc genmod;
class Id Sequence;
model Response = Period1 Period2 High Medium
Baseline CarryHigh CarryMedium
/ dist=multinomial
link=clogit;
repeated subject= Sequence*Id /type=ind corrw;
contrast ’Carryover Effect’ CarryHigh 1,
CarryMedium 1;
contrast ’Period Effect’ period1 1,
period2 1 ;
run;
In order to assess joint effects for Period and Carryover, two sets of two-row contrasts are specified.
Figure 4 tells you that the link function and distribution
have been specified correctly and that there are 406
total period measurements. Missing values for the response occurs 44 times.
Model Information
Data Set
Distribution
Link Function
Dependent Variable
Observations Used
Missing Values
WORK.EXERCISE
Multinomial
Cumulative Logit
Response
406
44
Figure 6 displays the information for the GEE analysis. The data includes 150 clusters, for which 37 have
missing values. Cluster size ranges from 1 (only one
period measured) to 3 (all periods represented).
GEE Model Information
Correlation Structure
Subject Effect
Number of Clusters
Clusters With Missing Values
Correlation Matrix Dimension
Maximum Cluster Size
Minimum Cluster Size
Figure 6.
Analysis Of GEE Parameter Estimates
Empirical Standard Error Estimates
Parameter
Estimate
Standard
Error
Intercept1
Intercept2
Intercept3
Period1
Period2
High
Medium
Baseline
CarryHigh
CarryMedium
-0.8959
0.9478
3.2798
0.2609
-0.0287
-3.1225
-0.4693
0.4932
0.3721
0.4265
0.3074
0.3290
0.3524
0.2973
0.2380
0.3032
0.2641
0.3708
0.3041
0.2968
Figure 7.
95% Confidence
Limits
Lower
Upper
-1.4983
0.3030
2.5891
-0.3219
-0.4953
-3.7167
-0.9869
-0.2335
-0.2240
-0.1551
Z
Pr > |Z|
-0.2934 -2.91
1.5926
2.88
3.9705
9.31
0.8436
0.88
0.4378 -0.12
-2.5283 -10.30
0.0484 -1.78
1.2199
1.33
0.9682
1.22
1.0081
1.44
0.0036
0.0040
<.0001
0.3803
0.9040
<.0001
0.0756
0.1835
0.2211
0.1507
Parameter Estimates
The contrast results provide the 2 degree of freedom
tests for both the Carryover and Period effects.
CONTRAST Statement Results for GEE Analysis
Contrast
Carryover Effect
Period Effect
DF
ChiSquare
Pr > ChiSq
2
2
2.55
1.03
0.2799
0.5976
Type
Score
Score
Results of Contrasts
GLM Information
Figure 5 displays the internal ordering of responses
values, which is from 0 to 3, for no distress to severe
distress.
Response Profile
Ordered
Level
1
2
3
4
Figure 5.
GEE Information
Figure 7 contains the parameter estimates. Neither
the Carryover nor Period effects appear to be influential. The Medium exposure appears to be marginally
influential with a parameter estimate of ,0:4693 and
a p-value of 0.0756; the High exposure appears to be
very significant with a parameter estimate of ,3:1225
and a p-value of less than 0.0001.
Figure 8.
Figure 4.
Independent
ID*Sequence (150 levels)
150
37
3
3
1
Ordered
Value
0
1
2
3
With both p-values greater than 0.25, these joint tests
are non-significant. There appears to be neither Carryover nor Period effects for these data. Note that the
default test for the CONTRAST statement used for the
GEE analysis is a score test; you can also request a
Wald statistic. The score statistics are generally more
suitable for smaller sample sizes.
Count
87
130
127
62
Ordered Values
The reduced model was fit with the following MODEL
statement. Baseline was retained as a covariate.
model response = high medium baseline
/ dist=multinomial
link=clogit;
Statistics, Data Analysis, and Modeling
Figure 9 contains the parameter estimates for the final
model.
Analysis Of GEE Parameter Estimates
Empirical Standard Error Estimates
Parameter
Intercept1
Intercept2
Intercept3
High
Medium
Baseline
Figure 9.
Estimate
-0.5404
1.2933
3.6171
-3.2523
-0.6204
0.5006
Standard
Error
0.1740
0.1949
0.2586
0.3057
0.2293
0.3381
95% Confidence
Limits
Lower
Upper
-0.8814
0.9114
3.1102
-3.8513
-1.0698
-0.1620
-0.1994
1.6752
4.1239
-2.6532
-0.1711
1.1631
Z Pr > |Z|
-3.11
6.64
13.99
-10.64
-2.71
1.48
0.0019
<.0001
<.0001
<.0001
0.0068
0.1387
Reduced Model Results
Note that goodness-of-fit statistics are still being researched for GEE analyses. See work by Barnhart
and Williamson (1998) and Preisser and Quaqish
(1996) for some recent discussions of goodness of
fit and deletion diagnostics. Given adequate sample
size, it may be advantageous to assess model adequacy by including additional terms in the model such
as pairwise and possibly higher interactions and then
performing a joint test on those effects. If the test is
non-significant, it leads credence to the corresponding reduced model under consideration.
Conditional Logistic Regression
Conditional logistic regression has long been used in
epidemiology where a retrospective study matched
subjects, or cases, with an event of interest with a
similar subject, or control, who didn’t have the event.
You determine whether the case and control had the
risk factors being investigated, and, by using a conditional likelihood, you can predict the event given the
explanatory variables. You set up the probabilities for
having the exposure given the event and then apply
Bayes’ theorem to determine a relevant conditional
probability.
More recently, conditional logistic regression has also
been applied in the situation of highly stratified data
and crossover studies. When you have highly stratified data, you may have a small number of subjects
pers stratum, and thus you have a small number of
subjects relative to the number of parameters you are
estimating because you will need to estimate stratification effects. Sample size requirements for the usual
maximum likelihood approach to unconditional logistic regression may not be met.
You have a similiar situation with crossover studies, in
which subjects are acting as their own controls.
Highly Stratified Data
Stokes et al. (1995) include an example of a clinical trial in which researchers studied the effects of a
new treatment for a skin condition. A pair of patients
participated from each of 79 clinics. One person received the treatment and another person received the
placebo. Age, sex, and an initial score for the skin
condition (ranging from 1 to 4 for mild to severe) were
recorded. The response was whether the skin condition improved. Note that because there are only
two observations per clinic, it would not be possible
to estimate properly a clinic effect. Generally speaking, you would want to have at least five observations
per clinic in order for that type of estimation.
However, by conditioning away the clinic effects as
nuisance parameters, you can perform a logistic regression that results in far fewer parameters. In
Stokes et al., this analysis is performed by recognizing that in the case of pairs within strata, you can create a response that is a within-stratum difference and
analyze those differences with the LOGISTIC procedure. However, it’s more straightforward to use the
PHREG procedure to perform this analysis. While
designed for proportional hazards regression analysis, through computational equivalences the procedure can also be used for conditional logistic regression.
The data have the following form, where each line
consists of two observations. Indicator variables are
created for various interactions and to make treatment
into a numerical variable. (A CLASS statement is on
the list for future PROC PHREG work).
data trial;
input center treat $ sex $ age improve initial @@;
/* compute model terms for each observation */
trt=treat=(’t’);
i_sex=(sex=’m’);
i_trt=(treat=’t’);
trtsex=i_sex*i_trt;
trtinit=i_trt*initial;
trtage=i_trt*age;
isexage=i_sex*age;
isexinit=i_sex*initial;iageinit=age*initial;
cards;
1 t f 27 0 1 1 p f 32 0 2
2 t f 41 1 3 2 p f 47 0 1
3 t m 19 1 4 3 p m 31 0 4
4 t m 55 1 1 4 p m 24 1 3
. . .
The following statements request the conditional logistic regression. The variable center is the stratification variable. The TIES=DISCRETE option is required. The first four variables in the MODEL statement are automatically included in the model, and
then the procedure produces a score statistic for the
joint inclusion of the remaining variables. This serves
as a goodness-of-fit check.
Statistics, Data Analysis, and Modeling
proc phreg data=trial;
strata center;
model improve = trt initial age i_sex
isexage isexinit iageinit
trtsex trtinit trtage / ties=discrete
selection=forward include=4 details;
run;
Figure 10 displays the parameter estimates for this
analysis.
Analysis of Maximum Likelihood Estimates
Variable DF
Parameter Standard
Wald
Hazard
Estimate
Error Chi-Square Pr > ChiSq Ratio
trt
initial
age
i_sex
-0.70244
-1.09148
-0.02483
-0.53115
1
1
1
1
0.36009
0.33508
0.02243
0.55451
3.8052
10.6104
1.2252
0.9175
0.0511
0.0011
0.2683
0.3381
0.495
0.336
0.975
0.588
Figure 10. Parameter Estimates for Clinical Trial
Data
Figure 11 and Figure 12 contain information about entering variables.
Analysis of Variables Not in the Model
Variable
isexage
isexinit
iageinit
trtsex
trtinit
trtage
Score
Chi-Square Pr > ChiSq
0.6593
0.1775
2.9194
0.2681
0.0121
0.4336
0.4168
0.6736
0.0875
0.6046
0.9125
0.5102
Figure 11. Analysis of Entering Variables
Residual Chi-Square Test
Chi-Square
4.7211
DF
6
Pr > ChiSq
0.5800
Figure 12. Score Test for Other Variables
First, consider the statistics displayed in Figure 11
and Figure 12. The residual score statistic value of
4.7211 with 6 df (p-value= 0.5800) indicates that the
additional terms are of little consequence. Age and
sex appear not to be influential, and you may choose
to keep them in the model as covariates or further
perform model reduction by fitting the model without
them.
Note that since the alphanumeric ordering of the response values is (0,1), the model is based on predicting the probability of no improvement. For dichotomous logistic regression, you can simply switch the
signs of the parameter estimates to obtain the estimates for the model based on the probability of improvement. Thus, those with treatment have an odds
e:70244
times higher of improving than those
2 =:019
patients receiving the placebo. This is true even after initial grade is adjusted for in the model. And, the
odds of improvement are e1:0918 = 2:980 times higher
per unit increase in initial grade. Note that the conditional logistic analysis has also taken into account any
effect of clinic.
Crossover Data
Conditional logistic regression also provides a useful analysis strategy for the crossover design. When
you apply conditional logistic regression in this setting, you are creating a strata for each subject. You
are conditioning out subject to subject variability and
focusing on intrasubject information. Thus, you can
often perform analyses that would not be possible
with population-averaging methods due to small sample size. Note that the odds ratios resulting from the
conditioning approach apply to subjects individually
instead of to subjects on average. This may be an
important consideration depending on your analysis
objectives. For example, in a study whose objective is
to produce a model that can be used for patient protocol prediction, the conditional logistic model may be
more appropriate.
The exercise data above are reanalyzed with the conditional logistic model, using the PHREG procedure.
For this analysis, the response is dichotomized into
1, for severe response, and 0, for other responses.
In addition, a new variable Strata has been defined,
which is a unique identifier for each subject based on
a combination of Sequence and Id.
The STRATA statement defines the strata; note that
the specification TIES=DISCRETE is required in order to produce the correct estimates. You use the
TEST statement to specify tests concerning the parameter estimates: here, joint tests for both the Carryover and Period effects are requested.
proc phreg data=Exercise;
strata Strata;
model dichotresponse = period1 period2 high
medium baseline CarryHigh CarryMedium
/ ties=discrete;
Carryover: test CarryHigh=CarryMedium=0;
Period:
test Period1=period2=0;
run;
Results of the parameter estimation are displayed in
Figure 13 and are similar to those obtained in the GEE
analysis. High and Medium exposures are influential,
and it doesn’t appear that there are Carryover or Period effects.
Statistics, Data Analysis, and Modeling
interest; another is incidence densities.
Analysis of Maximum Likelihood Estimates
Variable
Period1
Period2
High
Medium
Baseline
CarryHigh
CarryMedum
DF
1
1
1
1
1
1
1
Parameter
Estimate
-0.15440
-0.26202
-1.70458
-0.68624
0.65094
0.40777
0.28252
Standard
Wald
Hazard
Error
Chi-Square Pr > ChiSq Ratio
0.44683
0.31225
0.35000
0.35899
0.52766
0.44493
0.53024
0.1194
0.7041
23.7198
3.6542
1.5219
0.8400
0.2839
0.7297
0.4014
<.0001
0.0559
0.2173
0.3594
0.5942
0.857
0.769
0.182
0.503
1.917
1.503
1.326
Figure 13. Parameter Estimates for Exercise Data
The results for the joint tests for Carryover and Period are displayed in Figure 14. The Wald statistics
indicate that neither effect is important.
A reduced model seemed to fit the data adequately.
Linear Hypotheses Testing Results
Label
CARRYOVER
PERIOD
Wald
Chi-Square DF
Pr > ChiSq
0.8723
0.7083
0.6465
0.7018
2
2
Figure 14. Test Results for Carryover and Period
Exact Logistic Regression
Sometimes, sample sizes are simply not appropriate
for the usual logistic regression or conditional logistic regression strategies to be appropriate. In those
cases, exact logistic regression provides a means of
producing regression estimates and standard error
that are statistically valid. You use conditioning principles similar to those used in conditioning on observed
margins of contingency tables to obtain exact tests.
You eliminate nuisance parameters by conditioning on
the observed values of their sufficient statistics and
then use the exact permutational distribution of the
sufficient statistics for the parameters of interest. You
can apply conditional inference to both the unstratified
and stratified logistc regression settings. See Mehta
and Patel (1995). LogXact software, Cytel Software
Corporation (1993), provides this capability currently.
Weighted Least Squares Modeling of Categorical Response Functions
While weighted least squares may not longer be the
workhorse of categorical data modeling, it still plays
an important role in providing a strategy for modeling the variation among various functions that can be
computed for categorical data. Essentially you are
applying noniterative generalized least squares to response functions that are of interest and using an
observed covariance matrix as the weights. If you
have adequate sample sizes, the response functions
have a approximate multivariate normal distribution
and you can carry out hypothesis tests concerning linear combinations of them. Rank meaures of association are one type of response function that may be of
Modeling Rank Measures of Association Statistics
Many studies include outcomes that are ordinal in nature. When the treatment is also ordinal, rank measures of correlation can be modeled using WLS methods to investigate various treatment effects and interactions; such an analysis can often complement statistical models such as the proportional odds model.
Refer to Carr et al (1989) for an example of such an
analysis applied to Goodman-Kruskal rank correlation
coefficients, also known as gamma coefficients.
The Mann-Whitney rank measure of association
statistics are also useful statistics for assessing the
association between an explanatory variable and an
ordinal outcome. Consider the data from a randomized clinical trial of chronic pain (Stokes et al,
1995). Investigators compared a new treatment with
a placebo and assessed the response for a particular
condition. Patients were obtained from two investigators whose design included stratification relative to
four diagnostic classes.
Table 1 displays these data.
Table 1. Chronic Pain Data
Diagnostic
Class
I
I
I
I
II
II
II
II
III
III
III
III
IV
IV
IV
IV
Researcher
A
A
B
B
A
A
B
B
A
A
B
B
A
A
B
B
Treatment
Active
Placebo
Active
Placebo
Active
Placebo
Active
Placebo
Active
Placebo
Active
Placebo
Active
Placebo
Active
Placebo
P
3
7
1
5
1
1
0
3
2
5
2
2
8
5
1
3
Patient Status
F
M
G
2
2
1
0
1
1
6
1
5
4
2
3
0
1
2
1
0
1
1
1
1
1
1
5
0
3
3
0
0
8
4
1
10
5
1
4
1
3
4
0
3
3
5
2
3
4
3
4
E
0
1
3
3
2
1
6
0
2
1
3
2
0
0
1
2
You may be interested in computing the MannWhitney rank measure of association as a way of assessing the extent to which patients with active treatments are more likely to have better response status than those with placebo. You may then be interested in seeing whether diagnostic status and investigator influence this association through model-fitting.
You can perform such modeling by first computing the
Mann-Whitney statistics and their standard errors and
then using these estimates as input to the CATMOD
procedure to perform modeling.
You can compute the Mann-Whitney measures as
functions of the Somer’s D measures, which are produced by the FREQ procedure.
U
i
=
fSomer’s D C|R + 1g and S
2
i
=
SE
2
Statistics, Data Analysis, and Modeling
Si is the standard error of
statistic.
U,
i
the Mann-Whitney
The following statements produce measures of association for the eight 2 4 tables formed for the combination of investigator and treatment.
data cpain;
input dstatus $ invest $ treat $
status $ count @@;
datalines;
I A active poor
3 I A active fair 2
I A active moderate 2 I A active good 1
I A active excel
0 I A placebo poor 7
I A placebo fair
0 I A placebo moderate 1
I A placebo good
1 I A placebo excel 1
I B active poor
1 I B active fair 6
I B active moderate 1 I B active good 5
I B active excel
3 I B placebo poor 5
I B placebo fair
4 I B placebo moderate 2
I B placebo good
3 I B placebo excel 3
II A active poor
1 II A active fair 0
II A active moderate 1 II A active good 2
II A active excel 2
...
proc freq;
weight count;
tables dstatus*invest*treat*status/ measures;
run;
Figure 15 displays the table for Diagnostic Status I
and Investigator A. Figure 16 displays the measures
of association for that table.
The FREQ Procedure
Table 1 of treat by status
Controlling for dstatus=I invest=A
treat
status
Frequency|
Row Pct |poor
|fair
|moderate|good
| excel | Total
---------|--------|--------|--------|--------|--------|-------active
|
3 |
2 |
2 |
1 |
0 |
8
---------|--------|--------|--------|--------|--------|-------placebo |
7 |
0 |
1 |
1 |
1 |
10
| 70.00 |
0.00 | 10.00 | 10.00 | 10.00 |
---------|--------|--------|--------|--------|--------|-------Total
|
10|
2|
3|
2 |
1|
18
Figure 15. Frequency Table
Statistics for Table 1 of treat by status
Controlling for dstatus=I invest=A
Statistic
Value
ASE
-----------------------------------------------------Gamma
-0.2857
0.3515
Kendall’s Tau-b
-0.1763
0.2253
Stuart’s Tau-c
-0.1975
0.2485
Somers’ D C|R
Somers’ D R|C
-0.2000
-0.1553
0.2514
0.2026
Pearson Correlation
Spearman Correlation
-0.0866
-0.1900
0.2331
0.2416
Figure 16. Measures of Association
Table 2 displays the calculated values.
Table 2. Mann Whitney Statistics
Diagnostic
Class
I
I
II
II
III
III
IV
IV
Researcher
A
B
A
B
A
B
A
B
Somer’s
.2000
.2002
.2083
.6778
.0260
.1893
.0000
,:0156
ASE
.3515
.1915
.3622
.1834
.2271
.1923
.2007
.2116
Ui
.6000
.6001
.6042
.8389
.5130
.5947
.5000
.4922
Si
.1758
.0958
.1811
.0917
.1136
.0962
.1004
.1058
You compute the covariances and then create a data
set that contains the estimates and the covariance
matrix. The following DATA step creates the MannWhitney data set.
data MannWhitney;
input b1-b8 _type_ $
datalines;
.6000 .6011 .6042
.03091 .0000 .0000
.0000 .00918 .0000
.0000 .0000 .3280
.0000 .0000 .0000
.0000 .0000 .0000
.0000 .0000 .0000
.0000 .0000 .0000
.0000 .0000 .0000
;
_name_ $8.;
.8389
.0000
.0000
.0000
.0084
.0000
.0000
.0000
.0000
.5130
.0000
.0000
.0000
.0000
.0129
.0000
.0000
.0000
.5947
.0000
.0000
.0000
.0000
.0000
.0093
.0000
.0000
.5000
.0000
.0000
.0000
.0000
.0000
.0000
.0101
.0000
.4922
.0000
.0000
.0000
.0000
.0000
.0000
.0000
.0112
parms
cov b1
cov b2
cov b3
cov b4
cov b5
cov b6
cov b7
cov b8
This data set is then input into the CATMOD procedure. Thus, instead of generating functions from
an underlying contingency table, the CATMOD procedure does modeling directly on the input functions
using the input covariance matrix as the weights. You
define the profiles for each function with the PROFILE
option in the FACTORS statement. You also define
your factors, along with the number of levels for each,
and describe the effects you want to include in your
model with the – RESPONSE– option.
proc catmod data=MannWhitney;
response read b1-b8;
factors diagnos $ 4 , invest $ 2 /
_response_ = diagno invest
profile = (I
A,
I
B,
II A,
II B,
III A,
III B,
IV A,
IV B);
model _f_ = _response_ / cov;
run;
The ANOVA table results follow. The residual Wald
test is a test of the diagnostic class and investigator
interaction, which is non-significant with a p-value of
0.78. Neither diagnostic class nor investigator appear
to explain significant variation, with diagnostic class
appearing to be modestly influential with a p-value of
0.093.
Statistics, Data Analysis, and Modeling
Analysis of Variance
Source
DF
Chi-Square
Pr > ChiSq
-------------------------------------------Intercept
1
213.14
<.0001
diagnos
3
6.42
0.0930
invest
1
0.58
0.4469
Residual
3
1.06
0.7862
Figure 17. Main Effects Output
By submitting another MODEL statement like the following
model _f_ =
/ cov; run;
You can obtain a test of the hypothesis that the measures have the same value for each diagnostic class
and investigator combination. This is the seven degree test that is labeled ‘residual.’
Analysis of Variance
Source
DF
Chi-Square
Pr > ChiSq
-------------------------------------------Model|Mean
7
9.41
0.2247
Residual
0
.
.
Figure 18. Intercept-Only Model
You can’t reject this hypothesis with these data.
Modeling Incidence Densities
Incidence densities are defined as the ratio of the
number of episodes of a disease or illness to the total
person-time at risk. These measures are often of interest in epidemiological work. Lavange et al (1994)
studied the effect of passive smoking exposure on the
incidence of lower respiratory illness in young children. Measurements were made over time. The investigators used WLS to model the incidence densities; because the study was a complex survey design,
the covariance structure was determined with sample survey methods. Interest was in comparing the
marginal rates of lower respiratory disease between
exposed and unexposed groups. The ratio estimates
and their covariances were obtained with SUDAAN
software, and then the response functions and covariances wered used as input to the CATMOD procedure
for WLS modeling. The exposed group had a significantly higher rate of illness.
Summary
This paper describes recent enhancements in the
area of categorical data analysis and discusses several applications of the recent methodology. Exact
methods and quasi-likelihood methods provide ways
to analyze data that previously had many data analysis limitations. The SAS System includes software
for many of these newer methodologies and should
contain additional features that implement the recent
methodological advances in the next several years.
References
Barnhart, H. and Williamson, J (1998). Goodness-ofFit Tests for GEE Modeling with Binary Responses,
Biometrics, 54, 720–729.
Cytel Software Corporation (1993), LogXact: Software for Exact Logistic Regression, Cytel Software
Corporation, Cambridge, MA.
Carr, G. J., Hafner, K. B., and Koch, G. G., (1989),
“Analysis of rank measures of association for ordinal
data from longitudinal studies”, Journal of the American Statistical Association, 84, 797–804.
Diggle, P.J., Liang, K.-Y. and Zeger, S.L. (1994). Analysis of Longitudinal Data, Oxford: Oxford Science
Liang, K.-Y. and Zeger, S.L. (1986). Longitudinal Data Analysis Using Generalized Linear Models,
Biometrika, 13–22
LaVange, L. M., Keyes, L. L., Koch, G. G., and Margolis, P. A. (1994). Application of sample survey methods for modelling ratios to incidence densities, Statistics in Medicine, 13, 343–355.
Koch, G. G., Landis, J. R., Freeman, J. L., Freeman, D. H., and Lehnen, R. G. (1977). A general
methodology for the analysis of experiments with repeated measurement of categorical data, Biometrics,
33, 133–158.
Mehta, C. R. and Patel, N. R. (1995), “Exact logistic regression: theory and examples”, Statistics in
Medicine, 14, 2143–1260
Nelder, J.A., and Wedderburn, R.W.M. (1972), Generalized Linear Models, Journal of the Royal Statistical
Society A, 135, 370–384.
Preisser, J. S., and Quaqish, B. F. (1996), Deletion diagnostics for generalised estimating equations,
Biometrika, 83, 3, 551–562
Preisser, J. S., and Koch, G. G., (1997), “Categorical Data Analysis in Public Health”, Annual Review of
Public Health, 18, 51–82
Stokes, M. E., Davis, C. S., and Koch, G. G (1995).
Categorical Data Analysis Using the SAS System,
Cary: SAS Institute, Inc.
Zeger, S.L. and Liang, K.-Y. (1986). Longitudinal Data
Analysis for Discrete and Continuous Outcomes, Biometrics, 42 121–130
Statistics, Data Analysis, and Modeling
Author
Maura E. Stokes, SAS Institute Inc., SAS Campus
Drive, Cary, NC 27513. FAX (919) 677-4444 Email
[email protected]
SAS is a registered trademarks of SAS Institute Inc.
in the USA and other countries.  indicates USA
registration.
Other brand and product names are registered trademarks or trademarks of their respective companies.
Version 1.0

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project