--- On Tue, 1/6/09, AgEconomist <matttbogard@GMAIL.COM> wrote:

> From: AgEconomist <matttbogard@GMAIL.COM> > Subject: PROC LOGISTIC tests > To: SAS-L@LISTSERV.UGA.EDU > Date: Tuesday, January 6, 2009, 9:15 AM > Are there any commands in SAS that would test a logit model in PROC > LOGISTIC for multicollinearity, heteroskedasticity, or serial > correlation ? PROC REG has the VIF, DW options in the model statement > but not in PROC LOGISTIC. I could probably write a routine, but > frankly, I’m not even sure about how to get the ‘residuals’ necessary > for some of these tests. I know that RESDEV and RESCHI in the model > statement gives residuals, but can I treat these the way I would treat > residuals from OLS?

Matt,

First of all, collinearity does not depend on the left-hand side variable. Residuals have no bearing on the issue of collinearity. Thus, you can use the collinearity diagnostics available in PROC REG to assess the magnitude of any collinearity issues. By the way, I would suggest use of the COLLIN option available on the MODEL statement in PROC REG rather than VIF for assessing problems due to collinearity. To be explicit, you can assess collinearity issues by fitting the "wrong" model through code like the following:

proc reg data=mydata; model <binary response> = <predictors> / collin; run;

Of course, if any of your predictor variables are categorical with K levels, then you will need to construct a set of K-1 dummy variables beforehand which are named in the <predictors> list.

Heteroskedasticity and serial correlation are issues related to the response. But please observe that for a binary response, heteroskedasticity is not an issue. Overdispersion can be an issue. An overdispersed binomial occurs when there is variation in the probability of a success across units (subjects). There are a number of ways of dealing with this overdispersion. Basically, an overdispersed binomial is similar to a binomial with serial correlation in that there is some form of nonindependence among the observations.

This brings us to the topic of serial correlation. Serial correlation could be an issue. But if you have correlated responses, you should not be using the LOGISTIC procedure to begin with. You should be using one of the many procedures available in SAS for dealing with correlated binary response values. These include:

1) the GENMOD procedure with a REPEATED statement 2) the GLIMMIX procedure 3) the NLMIXED procedure

This is an incomplete list. No doubt there are other procedures for fitting a logistic regression model where there are correlated responses.

So, you might want to post again to SAS-L with a description of the data you have at hand and the problems that you believe are present in those data. A more helpful response may be offered if you give a more careful description of the problems that are presented to you in the data you must analyze.

Dale

--------------------------------------- Dale McLerran Fred Hutchinson Cancer Research Center mailto: dmclerra@NO_SPAMfhcrc.org Ph: (206) 667-2926 Fax: (206) 667-5977 ---------------------------------------