For binary responses, you will still get a nonzero proportion due to interaction even when you do not fit the interaction.  This is because the model is non-linear; the effect attributable to interaction is on the additive difference scale, but the model is being fit on the log or logistic scale.  This is a known and derivable quality of Causal mediation analyses, but it is not clear how you would actually interpret when you don't fit the interaction explicitly.    ... View more