The problem
These results clearly differ. The B:C
interaction is no longer significant and the P-value for the A:B
interaction is quite a bit higher. Both models should be computing the P-values in similar ways and so it's hard to imagine them being so different.
Why are they different?
###This was a part of the original question, but it can be misleading, see the answer below.
In fact, it seems that the anova(model, type = 3)
function is actually using type 2 SS, which we can verify by running anova(model, type = 2)
.