Difference in proportions from logistic contrast estimates

[lcmichael_unc]

I want to compute differences in proportions for a two-level outcome using the logistic procedure with contrasts, according to:

proc logistic data=&library..SVR_GP_mITT;
class ACTARMCD HCVGTCAT;
model SVR12= TREATMENT HCVGTCAT LOGHCVBASE;

contrast 'a - b' HCVGTCAT. 1 / estimate=parm;
contrast 'Treatment A vs B' TREATMENT 1;

run;

TREATMENT is a two-level predictor (A/B); HCVGTCAT is also two-level (a/b); LOGHCVBASE is continuous.

Unless I’m missing something, the contrast estimates do not, themselves, provide the differences in proportions between treatment arms directly. Is it legitimate to use the contrast (treatment arm, HCV sub-genotype) and parameter estimates (LogHCVBASE) obtained from the logistic model to compute predicted SVR12 proportions for each subject, controlling for all predictors in the model, with the difference in mean proportions (and confidence intervals) subsequently computed by arm group (A vs. B )?

Many thanks for all insights.

[StatDave]

That is correct. The CONTRAST and ESTIMATE statements provide tests and estimates of linear combinations of model parameters. In the case of a logistic model, such linear combinations are estimates of log odds. If you add the ILINK option, the ESTIMATE statement provides estimates of probabilities. But this does not extend to differences or other functions of probabilities which generally are nonlinear combinations of model parameters. To estimate or test such nonlinear combinations and marginal effects, you can use the Margins, NLEST, or NLMeans macros. Further discussion and an example of using these can be found in this note. The NLEST and NLMeans macros are immediately available for calling in the latest release of SAS (SAS 9.4 TS1M6). Otherwise, they (and the Margins macro) can be downloaded using the links provided in the note.