Some interesting results: PROC GENMOD vs PROC LOGISTIC

[bncoxuk]

Hi everyone,

I ran the same logit model using two differnt procedures: PROC GENMOD and PROC LOGISTIC. The LOGISTIC procedure said:

NOTE: Convergence criterion (GCONV=1E-8) satisfied in Step 2.

WARNING: The information matrix is singular and thus the convergence is questionable

For the issue, the type 3 analysis results are not reasonable: some variables have 0 degrees of freedom.

In contrast, the PROC GENMOD does not have such issue. The type 3 results are normal without any unexpected values. Besides, the GENMOD is much quicker to produce the results which saved the running time by more than half.

Can anyone give some insight about the problem here? Does it mean that the PROC LOGISTIC should not be used if there is a warning messsage about the singular information matrix?

[Reeza]

Just because you don't get an error doesn't mean the results are statistically valid. If you're not meeting the assumptions of the model then your results will be flawed.

I'm assuming you're also using the link=logit in proc GENMOD?

Post the code used for each to make sure its a valid comparison.

[Rick_SAS · Accepted solution]

If you get a "convergence is questionable" warning, then I would not use that result. Although both GENMOD and LOGISTIC estimate parameters by maximizing a log-likelihood, they might not use the same optimization method or starting conditions, so one could converge and the other not. The LOGISTIC doc says "the maximum likelihood estimation is carried out with either the Fisher scoring algorithm or the Newton-Raphson algorithm" and the default technique is Fisher scoring, which is an iteratively reweighted least squares algorithm. For details, see http://support.sas.com/documentation/cdl/en/statug/63962/HTML/default/viewer.htm#statug_logistic_sec...