binomial fixed-effect p-values by simulation
The other criterion is that the number of blocks has to be large. I have seen *no* rules of thumb for how large ... in the presentation that Doug Bates posted about recently ( http://www.stat.wisc.edu/~bates/PotsdamGLMM/GLMMD.pdf ) p. 32, 1934 observations, 60 groups, he uses LRT ... (you could try running your stuff on his example -- I think all the code etc is available from his web site) to see how well this works -- although here he gets p=0.796, so anti-conservative would just make that larger ... Ben
Daniel Ezra Johnson wrote:
We read, e.g. in Pinheiro and Bates, that one situation where fixed-effect LRTs are anti-conservative is when the number of fixed effect parameters being tested is large with respect to the number of groups. In the tests I'm doing, there's only a single binary fixed effect factor being tested - a between-subjects factor like gender, as noted earlier. I'm finding evidence for pretty serious anti-conservatism here too (e.g. LRT chi-square p=0.0001 vs. LRT simulation p=0.016), and I'm working on some reproducible code to demonstrate this.
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models