Does SAS has procedure with automatic variable selection method for multinomial logit models?

[bncoxuk]

I plan to use PROC CATMOD to fit a multinomial logit models. But it seemed that this procedure does not have the automatic variable selection method (e.g. stepwise, forward). So it takes time to manually exclude a variable. the problem is that the data set has over 5 million cases. It takes over 3 hours to run the model, and then wait there So to finish the model, it probably needs to take a few days with waiting.

Please advise.

[lvm]

Instead of CATMOD, you could use PROC LOGISTIC for stepwise variable selection. I assume the selection= option on the model statement works for multinomial logit models in LOGISTIC. Note: automatic selection methods are controversial, and can be very misleading. At best, you should use the results as an exploratory guide -- never just accept the final (automatic) variables that are selected.

For some exploratory work, you could make faster progress by considering a random sample of your 5 million observations, and using the sample for some modeling (at least initially). I am guessing that fits will occur much faster with 1 million or 500,000 observations.

[bncoxuk]

Hi Ivm,

LOGISTIC cannot run Multinomial logit models. If it does, then the model means cumulative logit model which is different from multinomial one.

[trekvana]

bncoxuk-

proc logit can run multinomial models. use link=glogit in the model statement options

[lvm · Accepted solution]

The intro to the LOGISTIC procedure (User's Guide), and example 51.4 (v9.2) deal with generalized logits (not ordinal) and the multinomial problem. But I am guessing that you have a different modeling approach in mind. In that case, there is no variable selection method. But as I wrote above, I would be cautious of any automatic variable-selection method.