Using interaction terms in causal models

Question

I have the following model:

LOS (length of stay) = b0 + b1 Age + b2 Gender + b3 BMI where Age and Gender are control variables, and my main IV is BMI. And this model isn't made for predictive purpose; we want to know if high BMI causes more LOS.

Results showed insignificant effect of BMI (P > .05).

I then tried to create interaction effects (without any literature-based knowledge): Age x Gender; BMI x Gender; BMI x Age.

Adding these interaction terms results in a significant effect for BMI x Age.

My question is: is it ok to conclude that BMI has no significant effect on LOS, but it does a significant effect on LOS at certain age levels (interaction). I mean is it ok to dig for significant interactions once my effect initially fails to show significance, given no expert / literature knowledge exists about the existence of such interaction effects.

Answer 1

I mean is it ok to dig for significant interactions once my effect initially fails to show significance, given no expert / literature knowledge exists about the existence of such interaction effects.

In a word, "No".

Digging for statistical significance increases the probability you find at least one statistically significant result, even if there are no true effects to be found.

You should pre-specify all analyses and stick only to those analyses you intend to perform. If you do go ahead with this sort of "digging", I would encourage you to report those analyses exploratory, refrain from reporting p values, and be very modest in conclusions made therefrom.