10.8. GENMOD versus CATMOD

I’ve used PROC GENMOD exclusively in this chapter, but you can also estimate loglinear models with PROC CATMOD. I prefer GENMOD for three reasons:

• PROC GENMOD uses a dummy variable parameterization for CLASS variables while PROC CATMOD uses “effect coding.” As I explained in Section 5.8, I find PROC CATMOD’s parameterization difficult to interpret.

• PROC GENMOD can fit a wider range of loglinear models than PROC CATMOD. Most importantly, PROC CATMOD does not allow the device of treating a variable as both quantitative and qualitative in the same model.

• PROC GENMOD can correct for overdispersion when appropriate.

• PROC GENMOD can optionally produce likelihood-ratio hypothesis tests.

  PROC CATMOD only produces Wald tests.

There are two other noteworthy differences:

• PROC CATMOD can operate on individual-level data by constructing the minimal contingency table required for each model it estimates. PROC GENMOD requires the contingency table as input, although this is easily produced with PROC FREQ by using the OUT= option in the TABLES statement.

• PROC CATMOD assumes that any zeros it encounters in the contingency table are structural zeros and, therefore, deletes them before fitting the model. If you want to treat them as random zeros, you must replace the zeros with a very small number (see the SAS/STAT User’s Guide for details).