Dont use programs to build models unless you have thousands of covariates, no hypotheses and theory.
Just build a model yourself and see what turns out.
Hello!
Can anyone explain why my Breslow-Day test for interaction was significant (p=0.0005) but my forward selection logistic regression test for the same interaction was not significant (p=0.10). Should I increase my p threshold for the logistic regression test? If yes, should I not be concerned that my interaction may be a false positive?
Thanks!
Confused
Dont use programs to build models unless you have thousands of covariates, no hypotheses and theory.
Just build a model yourself and see what turns out.
Stop cowardice, ban guns!
Thanks for the reply hismith! Can you elaborate?
The data I am using is from a large national survey. I used SAS to do the above tests.
Naxaris
First describe by what you mean by forward selection?
Stop cowardice, ban guns!
What I mean is that I compare a reduced model (main effects) with a model that contains the interaction I am interested in. If the Likelihood Ratio Test has a p value of less than 0.05 then I include that interaction and use the new model as my reduced model when testing the next interaction.
In my situation I first did a BD test (which was yes to interaction) and then I did a logistic regression test for interaction and got a p value of 0.10. I don't understand the opposing results. What should I conclude?
I am doing a graduate level course in Epidemiology - which is why I am using multiple tests to test for interaction.
What variables are in your logistic model besides x1, x2, x3 (the interaction of two binary variables) ?
I thought your first post was saying you were doing stepwise or something like that.
What university are you at?
Stop cowardice, ban guns!
Its a uni in Canada which you may not know since it looks like your in the US.
My covariates are not all binary, for example the interaction I am interested in would be between a binary exposure and a categorical predictor (income).
Would you know why the two tests do not align?
Figured it out!
Tweet |