R Anova() for backwards model selection: should test = F or chisq?

Question

When performing an anova comparing two models (glm or lm, sometimes (g)lmer) should I use test="Chisq" or test="F"? Or what criteria should I use to chose the test?

Answer 1

Don't do stepwise regression at all.

---

(See here for a general motivation for an answer that essentially says that "the question is wrong".)

(See here for a motivation for short answers. Longer answers are always welcome.)