ray wrote: > Paige Miller wrote: > > No. In fact, there is no generally agreed upon method for identifying > > outliers, not is there any way to decide whether or not they should be > > dropped without using subject matter knowledge. > > Dear Paige: Thanks for the answer and I absolutely agree with it. In my > case, I am simply trying to replicate a previous result which dropped obs > for certain variable values that were greater than 3 stds from the mean.

[I re-arranged and trimmed so it was easier to read - blame me if anything is amiss.]

Ray, if all you want to do is check whether values are within 3 sd of the sample eman, you can do that with a PROC MEANS and a DATA step:

PROC MEANS DATA=yourdata NOPRINT; VAR yourvar; OUTPUT OUT=OUTVAR MEAN=SAMPMEAN STD=SAMPSTD; RUN;

DATA NEW; RETAIN SAMPMEAN SAMPSTD; IF _N_=1 THEN SET OUTVAR(KEEP = SAMPMEAN SAMPSTD); IF ABS(yourvar - SAMPMEAN) > 3*SAMPSTD THEN DELETE; RUN;

I think that's what you asked for. But that's fairly naive, and may have any number of drawbacks [as you agreed above]. If you decide to evaluate the performance of said previous result, you may want to look at some papers on outlier detection, like:

Rosner 1975 Technometrics #17 Tietjen & Moore 1972 Technometrics #14 Walsh 1950 Annals of Math. Stat. #21 Walsh 1958 Annals of the Inst. of Stat. Math #10

Those are the ones I found taking a fast look in my reference lists, but this is hardly comprehensive. The bottom line: if you have a mixture of distributions, nothing may do the job well.

David -- David Cassell, OAO cassell@mail.cor.epa.gov Senior Computing Specialist mathematical statistician