I am running a multinomial logistic regression and have my final model, but now want to check for interactions between my two exposure variables and my independent variables. When I run this, one of my exposure variables is no longer significant, but some of the interactions between the levels of that variable are significant. I was wondering what the rule was for keeping or removing interaction terms in my model. Do I remove the insignificant variables and consequently any interactions involving that variable? All of my variables are categorical and relevant so there is not a question about whether or not insignificant interactions should be included. THANKS!