Interactions in logistic regression (moderation analysis)

[rebekah shao]

Dear all,

Am I right in thinking that you would only test for interaction effects in the presence of main effects? So even if you decided in advance that you wanted to test for an interaction (between two IVs), if you found that there wasn't two main independent effects when they were both included in the model (they had significant effects on the outcome univariately but only one remained a significant effect when together), then you wouldn't go on to test for an interaction? I am doing a logistic regression by the way on the effect of life stress on depression, and I am considering including anxiety as a potential moderator of that relationship. But I have found that when life events and anxiety are predictors together, only the main effect of anxiety remains, hence my question above.

Many thanks in advance for all your help!!

[Erick Turner]

I have a similar question, though with P value changes in the opposite direction from NS to significant.

Term X₁, as the sole term in a logistic regression, shows a significant effect on Y.
Term X₂ as the sole term is NS (though it approaches .05).
When used together in a bivariate logistic regression model, both X₁ and X₂ are associated with P values that are highly significant.

Using Fisher Exact to test the above two univariate relationships, one gets the very same P values.
And in the 2nd (NS) case, when one parses according to level of X₁, one 2x2 --> P=1.000) while the other 2x2 is highly significant.

Is there a way to determine from this information alone whether an interaction term is justified?

Alternatively, if anyone can suggest a primer on interaction terms, when to use them and how to interpret the associated results, that would be helpful.

Thanks.