I have a statistical problem in connection to interaction effects and main effects in the social sciences.

I am doing a regression where I have two interaction terms with one common variable. The interaction terms are used in order to prove or falsify two hypotheses. If I include all main effects and interaction effects, all effects become insignificant.

My approach is to leave the main effects out. This will leave with only significant interaction effects. Is this a valid approach?

  • 1
    $\begingroup$ As Peter Flom suggests, I think you should read what's been written on this elsewhere. But in brief, I don't think this is a valid approach to hypothesis testing. $\endgroup$ Nov 18, 2012 at 19:14
  • $\begingroup$ Thank you for the answers. I have read all the other threads with connection to this topic. I hope someone can give me an explanation why it is necessary to have insignificant terms in the regression. I still haven't found an answer for this. $\endgroup$
    – Jake
    Nov 18, 2012 at 19:57
  • 1
    $\begingroup$ I think this question is addressed in some detail at the post Peter linked to. There are a number of issues. One fundamental one is that if you are choosing how to specify a model based on what shows up as significant, then your p-values no longer represent fair tests of the null hypothesis (if they ever did before, which I couldn't judge without far more information." There are many assumptions involved in regression and hypothesis testing. Learn them well. $\endgroup$ Nov 18, 2012 at 20:32