I have a Poisson model with two categorical variables. I am wondering if the model can only include the interaction term without the main effect terms? And how can I interpret it? Thanks!


Mathematically this should not be a problem, but as you mention it is hard to interpret. Say you have two variables A and B with both two levels ($a_1$ and $a_2$, and $b_1$ and $b_2$). Define $X_a = I(A=a_1)$ and $X_b = I(B=b_1)$. Then your mean model would be:

$$\log(E(Y_i)) = \beta_0 + \beta_1X_aX_b$$

The coefficient $\beta_1$ would then represent the difference in the logged expectation of Y between an individual with $A=a_1$ and $B=b_1$ and the other individuals (all other terms kept equal, if any).

If you think this is a useful model you can go on with it, but note that you cannot disentangle the effects of variables A an B with this model.

  • $\begingroup$ Thanks for your reply! Just wondering if you know how to code this in R? $\endgroup$ – J. Zhu Oct 26 '17 at 13:41
  • $\begingroup$ See ?formula, something like glm(y~a:b, family="poisson") $\endgroup$ – Knarpie Oct 26 '17 at 13:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.