I have a three level multilevel model (therapists, patients, repeated measures) in which I have included 4-way interactions, for instance:
therapist self-efficacy * treatment condition * patient on track (yes/no) * time
This tests the hypothesis that therapist who are in condition A and are higher no self-efficacy have a steeper slope with patients that are not on track
Most of the lower order interactions are non-significant and also not that relevant for interpretation. Should I still include all of them? I feel like I would put a lot of parameters in the model, whereas it would be difficult to interpret the lower order interactions that are not significant.
Most fora state that it would be adviceable to include lower order interaction, but there are also some places where it is stated that there is no statistical reason to include them. Any advice or references on this situation?
