Doing a poisson regression like this: model<-glm(y~x*z,family=poisson)
with one predictor being a factor, I would use anova(model,test="Chisq")
to test the overall effect of the interaction rather than summary(model)
, which would give me two interaction terms (one for each of two levels in the factor variable that is not the intercept).
When I do this, I get an output with five columns:
Df, Deviance, Resid. Df, Resid.Dev, Pr(>Chi)
If I do the same with test="F"
instead I get an F value, so I assume there is a reason R doesn't report a chi-squared or LRT value.
Is this right? I would have thought that I should report a test statistic with my P value. Can I calculate it from the residual deviance and residual degrees of freedom? Should I report the residual deviance and df with my P value and leave it at that?
Best Answer
Ben Bolker answers this question in a response to another question here: Comparing nested GLMs via chi-squared and loglikelihood
Apparently, for the purposes of reporting anova(model, test="Chisq") results, the Deviance may be reported as the chi-squared value, though reporting it as Deviance would be more informative.
See also this GLM manual on stat.colombia.edu