Date: Mon, 13 Dec 1999 17:24:46 -0500
Reply-To: Walter Davis <walter_davis@UNC.EDU>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Walter Davis <walter_davis@UNC.EDU>
Organization: The University of North Carolina at Chapel Hill
Subject: Re: Outliers in PROC LOGISTIC and IML
Content-Type: text/plain; charset=us-ascii
Eric VonDohlen wrote:
> Does anyone use the INFLUENCE option in PROC LOGISTIC to find outliers?
> I'm having problems
> for the following reasons:
> 1. My datasets are all at least 10,000 obs. with about 6-10 covariates, so
> looking point-by-point is
> tedious as hell
> 2. SAS won't let you output said diagnostics in OUTPUT OUT=
Ummm...sure it will, or at least p. 1084 of SAS/STAT v. 6 says it will.
You can use dfbetas, difchisq, difdev, reschi, resdev as keywords. If
you want "influence" aka "hat values", you can specify the model in proc
reg and get them that way. (Calculation of "hat values" only involves
the independent variables, though they're not really "hat values" in a
logistic model, just measures of distance from the mean).
Of course, with 10,000 cases, you're unlikely to have any outliers
unduly influencing your results. I guess I shouldn't be quite so bold,
maximum likelihood/asymptotic models may be a bit more sensitive than
regular regression. But I doubt you've got much to find here.
Institute for Research in Social Science
UNC -- Chapel Hill