I have used lme4 for mixed effects models of reaction times and accuracy rates. I could not use lmerTest because the type of model I was using are not yet implemented there (problem with predictors that are factors). I was able to get p-values for the models ran on accuracy rates (based on Wald z-values) but not for the models built on reaction times. In order to get p-values for all models, I used Anova in the car package which gives me Wald chisquare values and probability of significance based on those chisquares. My concern is that sometimes for the accuracy rates, the effects indicated as significant in the analysis of deviance table (with the Wald chisquare values) are not significant in the mixed effect models. That is even for main effects of a factor with only 2-levels. Does anyone know why this could be the case?
Asked
Active
Viewed 632 times
1
-
can we see a reproducible example please? Consider also the `mixed()` function in the `afex` package. – Ben Bolker Jul 22 '14 at 02:13
-
Sorry for not answering this earlier. I got an answer elsewhere indicating how this can happen. Basically the tests reported by default by Anova(), are formulated assuming that terms to which a particular term is marginal are absent (i.e., have coefficients that are 0) which can lead to discrepancies with the estimated significance given by summary(). – stephanie Aug 20 '14 at 18:27