I conducted a multilevel analysis using repeated measures across four time points. The model contains intercept, linear slope and quadratic slope. I am interested in examining the extent to which variable A interacts with other identified variables in my model, so I added the interaction terms to intercept, linear slope and quadratic slope to estimate the fixed effects of these interactions. I then eliminated the non-significant interactions and retan the model until the interaction terms in the model were all significant.
My questions are:
-
When delete non-significant interactions from the model, if the interaction was significant on linear slope but not quadratic slope, should I still keep the non-significant one on quadratic slope? Or should I only keep the significant one on linear slope?
-
How should I interpret an interaction on linear slope but not quadratic slope (or vice versa)?
Thank you for your response in advance! Any advice is appreciated.
Best Answer
My view is that you shouldn't delete variables because they are non-significant if they are of theoretical interest. Sometimes a small effect is more interesting than a large one (e.g. if previous studies have found a large effect). Also, if your variables are not measured with perfect reliability, the power for interactions will be less than for main effects.