[R-sig-ME] Replicating type III anova tests for glmer/GLMM

[Brett Feltmate]

Hi Henrik,

I'm fairly new to mixed modelling, and I have a question regarding the output of afex::mixed() vs. glmer(), or more specifically the results of ANOVAs performed on them...

I'm modelling accuracy data in a within-Ss design with three factors, using afex I specify:

m1 <- afex::mixed( accuracy ~ x * y * z + ( 1 | subj ), method = 'LRT', family = binomial, data = dat)

and (eventually) using car:

m2 <- lme4::glmer( accuracy ~ x * y * z + ( 1 | subj ) , family = binomial, data = dat)

(I guess that bit was redundant...)

I get a discrepancy wherein if I run:

afex::nice(m1)

The effect of 'x' is reported as significant, but for:

afex::set_sum_contrasts()
car::Anova(m2)

The effect of 'x' doesn't even approached significance (according to the Chisq test)

Do the Chi-Square tests reported in both methods reflect the same thing? That is, the effect of 'x' on accuracy? Or are one of these methods testing something different?

Thank you,

Brett