Dear Brad,

You can use latent class analysis (LCA) for get more homogenous segments or classes.

For LCA in the structural portion of the model, the categorical latent variables of C represent mixture components that are related to covariates (X`s) through a multinomial logit regression model for an unordered polytomous response.

In multinomial logits, one of the segment (class) serves as the baseline or reference category and we calculate logits for the other categories relative to the reference category and then let the logits be a linear function of the predictors.

However, because latent class analysis uses a multinomial logit for determining differences between classes , this approach ignores interindividual differences that exist within classes. That is, response probabilities for all individuals of a given class are identical.

For SAS procedure see "A SAS Procedure Based on Mixture Models for Estimating Developmental Trajectories" Jones and at el.,

Hayrettin OKUT

>From: Brad Branford <b_branford@YAHOO.COM> >Reply-To: Brad Branford <b_branford@YAHOO.COM> >To: SAS-L@LISTSERV.UGA.EDU >Subject: Multinomial Logit ?? >Date: Fri, 1 Mar 2002 20:38:52 -0800 > >hi, > >I need some advice. > >I'm trying to build a predictive model for prospective account-holders >in my industry. Here's what I'm planning on doing: >1) Take a random sample of current account-holders and segment them (3 >or 4 groups) based on certain internal performance criteria. >2) Build a multinomial logit model with the segment being the >dependent variable and prospect external data (at the time of >acquisition) as explanatory variables. >3) Use the model to segment prospects at the time of acquisition. > >However, I had a few questions: >a) In a multinomial logit setting, how do I use the model predictions >to segment? >b) If I plan to use SAS, which procedure should I use? > >Thanks a lot in advance. > >brad

_________________________________________________________________ Chat with friends online, try MSN Messenger: http://messenger.msn.com