[R-sig-ME] lmer vs glmmPQL
bolker at ufl.edu
Wed Jun 24 18:48:01 CEST 2009
Daniel Ezra Johnson wrote:
> Regarding this thread, what about the method of fitting nested models
> and using anova() to estimate a p-value.
> For example:
> mod2 = lmer(y ~ genotype + (1|block), family = binomial, data)
> mod0 = lmer(y ~ (1|block), family = binomial, data)
> How does that p-value (which is one value for the whole term
> "genotype") relate to the individual coefficient p-values derived from
> the Z-scores inside summary(mod2)?
The Z-scores are Wald tests. The anova results are likelihood
ratio tests. The latter are in general more reliable but are known
to be *un*reliable for small-sample (i.e.
small-number-of-random-effects-levels) LMMs (Pinheiro and Bates).
Wald tests are not *known* to be unreliable in the small-sample
case, but I believe that is a statement of ignorance rather than
a statement that they're OK ...
Associate professor, Biology Dep't, Univ. of Florida
bolker at ufl.edu / www.zoology.ufl.edu/bolker
GPG key: www.zoology.ufl.edu/bolker/benbolker-publickey.asc
More information about the R-sig-mixed-models