[R-sig-ME] sanity-checking plans for glmer

[Ben Bolker]

Mitchell Maltenfort wrote:
> Having briefly fallen for the notion that the negative.binomial family
> in MASS could be used in glmer, I want to use these lists for a sanity
> check on my final (?) plans.
> 
> I want to use glmer for logistic regression and for poisson regression
> on a data set of 10,000 items.  There will be two crossed random
> effects.
> 
> For the logistic regression, I want odds ratios with confidence
> intervals.For the poisson regression, I'd want multiplying factors,
> again with confidence intervals. Either way, my co-authors will want
> p-values.  I'd also like to examine whether the two random effects
> have comparable variation, and whether any of them are provide odds
> ratios or multipliers that are are statistically distinct from 1.0 .
> 
> Can I use the s.e's and p-values (from z-statistics) provided by
> glmer?  Or should I use the pvals.fnc function from the language R
> package?
> 
> Are there any holes in my plans I'm not seeing?
> 

  [cross-posting to multiple R lists is discouraged.]

  How difficult this is, practically, depends on how many random-effects
levels you have (in the random effect with the minimum number of
random-effect levels).  If this number is large (>50?  Angrist and
Pischke give "42" as the rule of thumb in their Douglas-Adams-themed
econometrics book ...), then you can get away with Wald-Z or likelihood
ratio inferences without worrying about adjusting for finite-sample
effects (i.e., use anova() or the s.e.'s and p-values provided by
glmer).  For the random effects, use anova() [or possibly the RLRsim
package, although I don't know if it will work in this case].
Confidence intervals on the random effects are harder; see the "standard
 errors of variance estimates" section in <http://glmm.wikidot.com/faq>.
 (It says there that you can use lme4a, the bleeding-edge R-forge
version of lme4, for profiles, but as of a couple of days ago profiles
weren't available for GLMMs ...)

   pvals.fnc from the languageR function hasn't worked for a while, as
far as I know, because it uses the mcmcsamp() function which has been
disabled for some months (due to difficulty getting the sampling
algorithm to mix reliably).

  MCMCglmm would be a reasonable alternative -- I don't know how slow it
would be with 10,000 items, I have used it successfully for a problem
with 1600, so it should definitely work given that you're willing to be
patient.

  good luck,
    Ben Bolker