www.cmm.bristol.ac.uk/forum

Hi cmm
I have a 2 level discrete response model. 3 of the explanatory variables seem to have a strange relationship and i just want to check if its ok.
Variable A is a level 1 polynomial variable (A1 + A2), it also has a random slope on A1.
Variables B and C are level 2 variables. They interact with each other.
B also interacts with A1 (but not A2), while C interacts with A2 but not A1.
Everything runs ok & all SEs are fine
Is this ok or should i force the the non-significant B*A2 and C*A1 interactions in?

Thanks

I passed on your question to Prof. Jones and his response was as follows:

My understanding is that it is good practice to keep the lower order terms in the model; so that main effects A and B are included when the interaction A*B is included. The same applies to cross level interactions. There is a long FAQ at

http://www.ats.ucla.edu/stat/stata/faq/ ... effect.htm

If the interactions are only significant when the main effects are not in the model, it may be that the main effects are significant and the interactions not.