Interaction analysis - Problematic Wald test results

[Isabel Mansfield]

-

[Leonardo Guizzetti]

I'm not sure why you are cagey with the presentation of the full regression results, but nevertheless, enough is presented to answer your question. Please not that in future, it is best to present data, code and/or results using the CODE tags as you are asked to do when you read the FAQ.

The p-value for the interaction term belongs to just that -- the interaction term alone. The use of multiple -test- statements (or -testparm-) shows you doing a simultaneous hypothesis test of 3 conditions: that discrimination, sex and their interaction are all equal to zero. This is not the same as your intended question. This answers the question, is it plausible that the interaction and both main effects are all exactly equal to zero? To look at just the interaction effect the p-value can be read directly from the regression table, or alternatively just use the interaction term in testpram. Note that I have used just one hash mark (#), which instructs Stata to just include the the interaction term and not the main effects.

testparm i.w1discr#i.sex_recoded   //   <-- what you want
testparm i.w1discr##i.sex_recoded   //   <-- what you did ...
testparm i.w1discr i.sex_recoded i.w1discr#i.sex_recoded   //   <-- ...which is equivalent to

See -help fvvarlist- for more information on factor notation.