testing significance of fixed factors in a mixed model
Can you show us some of your output? In general, an F-test will be more accurate than an LRT in cases where you can figure out how to do it, because an LRT doesn't make any finite-size corrections (what df values are you getting from KR/Satterthwaite)? The thing you will have to think hardest about is which variables are conditioned on when testing, especially when you have interactions in your model and are considering testing main effects. On Thu, Dec 14, 2017 at 2:25 PM, Julia Chacon Labella
<juliachacon at gmail.com> wrote:
I am sorry if I am asking a na?ve question or if that has been already asked many times. I am not an expert in mixed models at all, but want to understand and be confident about what I am doing. Actually, I am analysing a data set of a experiment. I have several treatments, a balance design, randomized by blocks, etc., and generally following the recommendations of the FAQs by Bolker, what are really usefull (Thanks!). http://bbolker.github.io/mixedmodels-misc/glmmFAQ.html But, I am pretty confused about how to test the significance of fixed factors in a mixed model. I find huge differences in the significance output when computing 1) Anova from the car package or 2) anova from lmerTest with KR or satterthwaite approaches (both, KR or Satterthwaite have similar results), or 3) a LRT via anova. The biggest differences are found for car::Anova! * First, I am not sure if I should employ a LRT or a conditional F-test (lmerTest::anova, KR or satterthwaite approaches). Both results are pretty similar in my case, but conditional F-test seem to slightly be more conservative in my case. Can the F-test be somehow problematic in terms of statistical power? * Second, in case of using lmerTest::anova, I am not completely sure if I should used a type I or type III anova. There is any reason for using type I anova in balanced designs? Thanks in advance, Julia [[alternative HTML version deleted]]
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models