Logistic Regression: A Self-learning Text, Third Edition (Statistics in the Health Sciences)

Parameter Estimates

Estimate Std. error Wald df Sig.

Threshold [GRADE¼.00] .511 .215 5.656 1 .017

[GRADE¼1.00] 1.274 .229 31.074 1 .000

Location RACE .427 .272 2.463 1 .117

ESTROGEN .776 .249 9.696 1 .002

Link function: Logit

95% Confidence interval

Lower bound Upper bound

.932 8.981E-02

.826 1.722

.106 .960

1.265 .288

A test of the parallel lines assumption is given in the table titled “Test of Parallel
Lines.” The null hypothesis is that the slope parameters are the same for the two
different outcome comparisons (i.e., the proportional odds assumption). The results
of the chi-square test statistic are not statistically significant (p¼0.638), suggesting
that the assumption is tenable.

The parameter estimates and resulting odds ratios are given in the next table. As
noted earlier, with the alternate formulation of the model, the parameter estimates
for RACE and ESTROGEN match those of the SAS output, but the signs of the
intercepts (labeled Threshold on the output) are reversed.

Modeling Correlated Dichotomous Data with GEE

The programming of a GEE model with the infant care dataset is demonstrated using
the GENLIN procedure. The model is stated as follows:

logit PðOUTCOME¼ 1 jXÞ¼b 0 þb 1 BIRTHWGTþb 2 GENDERþb 3 DIARRHEA

The dichotomous outcome variable (called OUTCOME) is derived from a weight-for-
height standardized score based on the weight-for-height distribution of a standard
population. The outcome is correlated since there are multiple measurements for
each infant. The code and output are shown for this model assuming an AR1 correla-
tion structure.

Open the datasetinfant.savin the Data Editor window. The corresponding com-
mand syntax is:

GET

FILE¼‘C:\infant.sav’.

To run GEE model using the GENLIN procedure, select Analyze!Generalized
Linear Models!Generalized Estimating Equations from the drop-down menus to

646 Appendix: Computer Programs for Logistic Regression