topic Re: Mixed logit model with a weight statement in Statistical Procedures

Mixed logit model with a weight statement

Demographer — Mon, 03 Aug 2015 19:47:23 GMT

Hello,

Is is possible to perform a mixed logit with a weight statement? The procedure I know to perform mixed logit models is proc MDC with the option type=mixedlogit, which doesn't support the weight statement. Otherwise, I think the estimated parameters will be biaised, because the weight of the individual in my database varies from 0.8 to 10.3.

Re: Mixed logit model with a weight statement

SteveDenham — Mon, 03 Aug 2015 20:03:06 GMT

This is a shot in the dark. Try looking at PROC LOGISTIC or PROC SURVEYLOGISTIC. Both support a WEIGHT statement, but it needs to be considered carefully. PROC GLIMMIX also supports a WEIGHT statement, and a weight= option in the RANDOM statement.

Steve Denham

Re: Mixed logit model with a weight statement

Demographer — Mon, 03 Aug 2015 20:23:55 GMT

Unfortunately, for a mixed logit model, we have to use the STRATA statement. Thus, the weight statement is not supported neither with PROC LOGISTIC.

GLIMMIX and SURVEYLOGISTIC don't work on my SAS version. I'll try them later.

Re: Mixed logit model with a weight statement

SteveDenham — Tue, 04 Aug 2015 12:46:09 GMT

In particular, you would probably need access to SAS/STAT 13.2 or higher for what I was looking at for GLIMMIX. In the Details part of the documentation, there is a section titled 'Pseudo-likelihood Estimation for Weighted Multilevel Models' where a three level model, with different weights at each level is fit. This may very well parallel what you are doing.

One thing I did notice in PROC MDC is that the residual error is fit as having a Gumbel distribution for TYPE=MIXEDLOGIT. I don't believe that is available in the procs I recommended, although an alternative link function could be written for this distribution in GLIMMIX.

Steve Denham