- Thread starter ks1983
- Start date

is it ok to interact that terms when one of them is not statistically significant on its own?

(Example: imagine an interaction plot where the line cross each other like an X.)

But if you include an interaction term the rule says that you should also include each of the two variables as linear terms.

General rule, don't add an interaction term to model unless you have a theoretical justification.

It is natural that people try different models, and try to find something that fits to the data.

(it is also very good to plot the the data and have a look at it.)

What if I tamed down the language from "theoretical" - to something that makes intuitive/justifiable sense. Otherwise what would stop me from looking at every combination of variables as interactions. Then also given this, what would stops me from looking at all higher order interactions. Eventually, I could find something that isn't generalizable and but spurious. Especially if I am not penalizing the process or evaluating the effect in a holdout set.

I get that each field is different, but I would question exploratory interaction findings. I have only discerned a couple interaction across my career and both were related to explainable behavioral phenomenon.