topic Re: Multinomial Logistic--no observations predicted in 3 of 6 dependent groups in Statistical Procedures

Multinomial Logistic--no observations predicted in 3 of 6 dependent groups

deleted_user — Thu, 06 May 2010 18:02:59 GMT

Hello. I am trying to build a multinomial logistic model using LINK=GLOGIT to predict customer membership in 6 groups. The customers are grouped into 6 groups based on survey responses to spend amount in the industry overall & with one particular vendor in the industry.
Behavioral data (retail purchase data) of these surveyed customers is what I'm using to try & develop a model to predict membership in one of the 6 groups. All solutions I've found result in nearly all modeled customers being placed into the largest segment (based on max score) & no customers being placed in some of the smaller segments. Basically, 93+% of those in each segment are scored to be placed into the largest segment, not the segment they actually fall in. The largest segment is 43% of all customers.

Any tips on how to alleviate this situation and reduce the very large rate of misclassifications? Different Proc options, data preparation & variable steps?

Thank you.

Message was edited by: c-cubed

Message was edited by: c-cubed

Re: Multinomial Logistic--no observations predicted in 3 of 6 dependent groups

plf515 — Fri, 07 May 2010 10:28:21 GMT

I don't have time for a full response right now, but look into the SCORE statement, in particular, look into PRIOR option within that.

HTH

Peter

Re: Multinomial Logistic--no observations predicted in 3 of 6 dependent groups

deleted_user — Fri, 07 May 2010 15:32:34 GMT

Thank you Peter, but I have tried that & it didn't change the results.

Re: Multinomial Logistic--no observations predicted in 3 of 6 dependent groups

Doc_Duke — Fri, 07 May 2010 19:46:16 GMT

There is always a chance that it is a non-predictive model. Have you looked at your model fit and discrimination statistics?