Interpreting main effects in a Poisson model with a binary covariate, a continuous covariate and their interaction

[Michael Monuteaux]

Hello,

I need help interpreting the main effects in a Poisson model with a binary covariate, a continuous covariate and their interaction. I'm using STATA 13 for Windows.

My understanding is that, in general, the estimate corresponding to a main effect in such a model only applies to the baseline values of the other main effect. For instance, if there is a binary covariate for sex and a binary covariate for employment, then in a model with sex, employment and the sex-by-employment interaction, the main effect of sex only applies to those subjects who have a value of zero for the employment variable. Similarly, the main effect of employment only applies to those subjects who have a value of zero for the sex variable. When interpreting such a model, I can verify that this is true by running other models within strata of the covariates and compare the stratum-specific estimates to the main effects of the interaction model.

When one of the covariates is continuous, the main effect of the continuous covariate is still the same as what is produced by a model restricted to the subjects with the baseline value of the binary covariate. However, I'm not sure how to interpret the main effect of the binary covariate. Any guidance would be most appreciated.

Thanks

[Carlo Lazzaro]

Michael:
welcome to the list.
The easiest way to get helpful replies is posting what you typed and what Stata gave you back (as per FAQ). Thanks.

[Rich Goldstein]

following your own description, it is the effect of being non-zero when the continuous covariate is equal to zero - this may or may not make sense depending on how you have coded your continuous covariate (e.g., is it "centered"? is zero a possible value? does it actually exist if it is a possible value?)

[Phil Bromiley]

With non-linear models, it is often advisable to use predictive margins to look at the effects and margins to do the tests for specific values of the interacting variables. The simple interpretation available in linear regression is not necessarily correct with non-linear models. This will also give you an idea of the substantive importance of the effect.