topic Re: dummy variables in model selection in Statistical Procedures

dummy variables in model selection

deleted_user — Tue, 04 Aug 2009 18:05:50 GMT

hi,

i've been stuck on this forever. I've created dummy variables and need to run a stepwise regression using proc reg. do i keep my referent variable in the model? or out of the model?
for example if age is outcome and race is what i created dummies for and 'other' is referent do i use brackets? and leave out the referent? or include all + the referent?

using referent:
proc reg data = XYZ
model age = { asian american dutch } edu SES / selection = stepwise ........groupnames='race'; run;

including referent:

proc reg data = XYZ;
model age = asian american dutch other edu / selection = stepwise.....;run;?????

if it's the first and edu is included and SES is not? does anyone know how to interpret that interms of the referent group???? should i be using proc logistic instead when using dummy variables?

Re: dummy variables in model selection

Doc_Duke — Tue, 04 Aug 2009 21:16:04 GMT

If you have SAS 9.2, use the GLMSELECT procedure.

If you are using earlier versions, and the model is that straight-forward, I would use GLM and manually implement the stepwise algorithm by looking at the criteria.

Your referent approach is the only one that really works in REG (otherwise you have a singular design matrix, AKA over-specified model), but you don't have the option of grouping the other levels and that can become a problem.

Doc Muhlbaier
Duke

Re: dummy variables in model selection

Doc_Duke — Tue, 04 Aug 2009 21:17:51 GMT

See this thread for a more complete discussion

http://support.sas.com/forums/thread.jspa?messageID=23573尕