Testing multiple interaction variables in logistic regression

Hi everyone,

For a study that I'm currently conducting I want to know the effect of some of my variables on other variables, which I think should be done by creating interaction variables.

However, I'm not sure how I should test for an effect of variable A on a number of other variables (let's say B up to G). The textbook literature I read explains how to test an interaction variable, but in their example they only test 1 in the model..

So my question is, should I test the impact of variable A on variable B-G by creating a new model every time (one with A*B, then one with A*C and so forth) or can I create one model including all the interaction terms A*B, A*C up to A*G in one go?

I ask because the results differ in both approaches. This seems weird to me because if I was only interested in the interaction effect of, say, A*B, I could conclude that there is an impact, where in the full model there actually doesn't seem to be an impact (or the other way around).

From my previous thread I've learned I should normally include all the variables in one model :p but because it's about interaction here I'm not so sure how to go about it.

Thanks for any and all help!! :)


Less is more. Stay pure. Stay poor.
I would use both approaches. Probably not uncommon for someone to through them all in, However I would imagine that approach may mask some information. I also cannot recalling seeing any etiquette on this topic.
I am currently trying both approaches but I'm getting lost at which effects I should consider. In the one by one models, some interactions are significant that aren't significant in the 'all-in-one' model; and the same happens the other way around (significant interactions in the 'all-in-one' model but not in the one by one model). Then there are some interactions that are significant in both approaches, and some that are significant in one approach and significant at the 10% level in the other.

Should I conclude that an interaction is significant if it appears in one of both approaches? It feels weird to test it in two different ways with different results and throwing the results on a single pile..


Less is more. Stay pure. Stay poor.
Its like when building a model, some variables are significant in bivariate analyses but do you keep them in the final multiple regression model if they are insignificant - probably not but those early analyses helped you build and find candidate terms.

You also incorporate what you know about the content.

Mean Joe

TS Contributor
Variables A-G are your explanatory variables, for some other outcome variable?
You think that A could be affecting B, and B could be affecting the outcome?
@hlsmith, ok thanks now I know what you're getting at. The problem is that some interaction variables that aren't significant in a model where they are the only interaction variable become significant in a model with other interaction variables, so I'm not sure how to justify their selection into the final model. As I said, some interactions are significant in a model with all normal variables and one interaction, others in a model with all normal variables and all interactions of one variable with the rest, and some in both. I just don't know what to make of it..

@Mean Joe, let's say A-G are explanatory variables for the (binominal) outcome variable. I've collected additional data, let's say variables H-J. I want to test if scoring high on variable H will affect the way variable A affects the outcome, if scoring high on variable H will affect the way variable B affects the outcome, and so forth. Then the same for variables I and J.

Following this explanation, I don't understand that if variable H affects variable C in a model where I also include H*A, H*B etc; that variable H doesn't affect variable C in a model where I only include variable A-J and the interaction H*C. And because of these discrepancies I'm not sure whether to use models with one interaction variable at a time or a complete model, and if the latter, how exactly to build it if not to include all interactions for one of the variables H, I or J...


Less is more. Stay pure. Stay poor.
I am not saying you are establishing causality, but perhaps drawing a path/causality pathway may help you put the scope and context in order. Perhaps Mean Joe is getting at mediating effects. Perhaps you should better describe your scenario, so we can understand fully the question you are posing.