[R-sig-ME] feedback: Anova (type III-tests) table based on LRT for glmmTMB models (drop1, anova, mixed)

[Guillaume Adeux]

Sorry, my intent was definetely not to shortcut anyone or anything.

I thought I would give you all a little feedback on your propositions.

The problem with drop1 (when an interaction is in the model) is when you
want it to behave like a type III anova. This results in one of the
variables having zero degree of freedom, hence LRT of 0 and no p-value.

Here is the example on an output:



Model: mod=glmmTMB(ratio ~ block + trt * time
+(1|plot)+(1|year)+(1|year:plot),family=list(family="beta",link="logit"),data=biomass)

drop1(mod,.~block+trt*time,test=”Chisq”)

Single term deletions

                               Df                           AIC
LRT                      Pr(>Chi)

<none>                                                -5092.2

block                    1                             -5093.5
0.7345                  0.391415

trt                         4                             -5082.7
17.5018                0.001544

time                      0                             -5092.2
0

trt:time                  4                             -5100.0
0.2169                  0.994527


If I understand correctly, drop1 is incapable of comparing a model A+A:B
with A+B+A:B.

Anova.III.glmmTMB indeed works but only yields Wald tests which I thought
were not ideal for glmms.

I contacted the developper of the {afex} package to see if the mixed
function could be adapted to run on glmmTMB objects but no luck as of now.
Considering the framework of glmmTMB is similar to glmer, maybe this can be
easily adapted.

Thanks for your interest. Don't hesitate if you have any input.

Guillaume ADEUX

	[[alternative HTML version deleted]]