PROC GAM, PROC ADAPTIVEREG, MULTINOMIAL DEPENDENT VARIABLE

[davis]

Can either Proc GAM, Proc Adaptivereg, or both handle a multinomial dependent variable (e.g., likely status of a credit card account next month as current, delinquent, or paid off)? 

[Rick_SAS]

For categorical responses, PROC ADAPTIVEREG supports binomial, negative binomial, and Poisson responses.

PROC GAMPL supports the same distribution and a binary distribution.

It sounds like you want a nonparametric (or semi-parametric) fit. One possible alternative is to use the EFFECT statement in GLMSELECT to generate spline effects for the explanatory variables and output the design matrix.

You can then use PROC FMM or HPGENSELECT to model the response in terms of the columns of the design matrix.

[davis]

Thanks for your reply.

You are correct.

I am searching for either a nonparametric or semi-parametric fit, where the dependent variable can be specified as a multinomial.

I will try your suggestion.

[StatDave]

It might be easier to just use PROC LOGISTIC which can fit a multinomial response model and also supports the EFFECT statement. For example, these statements fit a model to a nominal multinomial response and use splines to represent the predictor.

proc logistic;
effect edspline=spline(ed / naturalcubic basis=tpf(noint));
model occ(ref="1")=edspline / link=glogit;
run;

[davis]

Thanks.

I understand PROC LOGISTIC is an alternate path, at least theoretically.

However, I am trying to assess the feasibility of whether either PROC GAM, PROC ADAPTIVEREG or both can handle a multinomial nominal dependent variable.

[Rick_SAS]

I already answered your question. Neither procedure supports a multinomial response.

I think StatDave_sas's suggestion is the best you can do with one call to a SAS/STAT procedure.

By the way, don't use PROC GAM any more. Switch to GAMPL, which is faster and scales to larger data sets.

[StatDave]

The main procedures that do support the multinomial distribution are the LOGISTIC, HPLOGISTIC, GLIMMIX, HPGENSELECT, and GEE (for repeated measures data only) procedures, as well as the older and seldom used PROC CATMOD . For ordinal multinomial responses only, PROC GENMOD is also available. The highly versatile but more difficult to use PROC NLMIXED could also be used. It is summarized in this note. Of these procedures, flexible models similar to what GAM and the newer GAMPL do can be fit in LOGISTIC and GLIMMIX since these procedures support the EFFECT statement which allow you to use splines similar to GAM and GAMPL. So, as Rick says, they are your best bet for what you seem to want to do.

[davis]

Thank you.