st: Significance of categorical variables in Logistic Regression

[Marcello Pagano <pagano@hsph.harvard.edu>]

For Michael:

I am running logistic regressions with a number of categorical iv's. I 
am building the model by starting out with variables that have a pval 
.15 when used in a single variable regression.

Then I put them all together, and weed out variables that are not 
significant in the multi-variable regression. When I get to a "core" 
model of significant predictors, I add back excluded variables one at a 
time to see if they are significant in the context of the smaller set of 
predictors.

If a categorical variable has at least one category that is significant, 
I keep the whole variable.

I have excluded a categorical variable, but noticed that if I base the 
variable on a different category than the default category, I suddenly 
see significant categories in the regression.

I.E.:

logistic yvar xvar1 xvar2 i.xvar3

results in every category of xvar having a high pval, but:

logistic yvar xvar1 xvar2 ib2.xvar3

results in several of xvar3's categories having a pval near 0.

From looking at marginplots I understand how this can happen, but I 
would like to know if there's a way of detecting this during the model 
building without looking at marginplots?

Many thanks,

Michael Cook

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/