I have a made a linear mixed model using lmer().
I began with a maximal model containing 6 terms and all two-way interactions (15 total), I also have a random effect of subject ID. Of these terms, 11 interactions were significant and so were kept in the final model. I found this by dropping interactions and comparing to previous model using anova(). As all 6 fixed terms are part of a significant interaction they remained in the model. My problem is that when I tried to test the fixed terms to find a chi-squared value for them by removing the term and comparing to the full model, the anova output said chisq = 0, df=0 but that p< 2.2e-16. I don't understand how p can be so small if chisq = 0. I also don't understand why df=0 which would suggest I am comparing the same model when actually I have removed a term. What does this mean? Have I fit too many interactions to my data (I only have 160 data points). Is there any point in finding the chi-squared value for these fixed effects?
Best Answer
Probably because you write the wrong code. I got df=0 too, and found out why.
For example: I used the formula: y ~ x1 + x2 +x1*x2 + (1|x3) in lmer, and then: y ~ x2 +x1*x2 + (1|x3).
Then I used anova to get the p-value for x1 using the above two lmer outputs. It gives result df=0, chisq=0, and very small p.
I found out that I misunderstood *; x1*x2 means x1 + x2 + x1:x2, whereas x1:x2 is the interaction of x1 and x2.