topic Re: Outlier detection in SAS Procedures

Outlier detection

deleted_user — Fri, 19 Feb 2010 02:59:55 GMT

Hi ,

could you suggest soem useful commands and their interpretations if I want to find if there are any influential observations / outliers in my data.

Kind Regards ,
markc

Re: Outlier detection

Cynthia_sas — Fri, 19 Feb 2010 03:17:20 GMT

Hi:
To investigate and explore your data, PROC UNIVARIATE is a good overall procedure to start with:
http://support.sas.com/documentation/cdl/en/procstat/63032/HTML/default/procstat_univariate_sect008.htm
http://support.sas.com/documentation/cdl/en/procstat/63032/HTML/default/procstat_univariate_sect003.htm

cynthia

Re: Outlier detection

Doc_Duke — Fri, 19 Feb 2010 03:28:19 GMT

Whether or not an observation is an outlier or influential observation (they are often thought of differently) is a function of your putative model. The UNIVARIATE procedure that Cynthia mentioned will look at several different single variable distributions.

All of the regression type statistical models in SAS 9.2 have very good ODS graphics to assist in outlier detection. There is a section with each procedure describing the diagnostic models available in the ODS Graphics.

Re: Outlier detection

Parker — Fri, 19 Feb 2010 13:07:10 GMT

For the "interpretations" part of your question, I use the mean and standard deviation to identify thresholds for univariate outliers.

For samples < 500, outliers are values that exceed [mean + or - 1.96 standard deviations]. This is equivalent to a 95% confidence interval.

If 500 < n < 5,000, [mean + or - 2.576 standard deviations]... 99% CI.

For n > 5,000, [mean + or - 3.291 standard deviations]... 99.9% CI.

For multivariate outliers, I look at Cook's D & DFFits most often.

Good luck,

Parker

Re: Outlier detection

SAPPER — Fri, 19 Feb 2010 17:24:40 GMT

just as an addition to above answers.....To find outliers with character data you can use PROC FREQ to find all the values of a particular character variable.
To find outliers in Numeric data we can use PROC Means.