MCMC Issue relating to the State of lme4

Hi folks,
Here are code and files for the analysis.
-------------- next part --------------

Hi folks,
Repost/summary of previous correspondence.  Thanks to Ben Bolker,
Spencer Graves and Ken Beath for questions and comments.

I have doubts about the MCMC results with the most recent version of
lme4 (versions ....-20 on CRAN, and -21 on R-Forge).

I recently tried analyzing a moderately sized (n=~500) unbalanced data
set where lme and lmer gave similar F ratios.  lmer simulations gave
the same F-tests (via simulation) as lme, but the MCMC output (using
Helmert and sum contrasts) depended wildly on terms that were retained
in the model. Unfortunately, it seems to depend on the data, and I
cannot make the data (widely) publicly available, although I could
send it to individuals.

My original question and compilation of other emails:

Are there general situations in which we might expect very different
answers from F tests vs. mcmcpvalue with orthogonal contrasts (Helmert)?

I helping someone with a normal linear model with a moderate sized,
noisy data set, and I am getting very different probabilities between
F-tests and mcmcpvalue for some interactions.

(Output appended.)

I get similar F-test results whether I use lm (and ignore the random
effect of subject), lme, and lmer with an DDF approximation and via
simulation (lmer::simulate,  of the null hypothesis indicate that F-
stats as large (or larger) than my observed F-stats are VERY unlikely,
under the null hypothesis).

When I use mcmcpvalue, I get huge changes in P-value of a main effect
(0.6 to 0.01) when I remove its interactions. In contrast, the F-test
(using trace of the hat matrix DF's) are much more consistent when I
change the fixed effect structure.

Is mcmcpvalue much more sensitive to overfitting the model? In
some cases, removing the interactions results in a lower AIC (with ML
fits). The MCMC sample traces look (in my limited experience)
withoutpeculiarities, and the densityplots are all quite symmetrical
and normal-ish.

In the full model, we have 28 fixed coefs (22 continuous variables or
slope interactions) and about 500 obs of about 200 subjects (2-3
observations per subject). The data are VERY unbalanced. The QQ plots
look normal, but highlight the lack of balance (from one to dozens of
reps per treatment combo).

OUTPUT:OUTPUT:OUTPUT:OUTPUT:OUTPUT:OUTPUT:OUTPUT:OUTPUT:OUTPUT:OUTPUT:

The following tables are output of factors/covariates from a function
I wrote to use with lmer. "P.sum" is an F-test using one hack for
denominator degrees of freedom (these agree qualitatively with lme).
"P(H|D)" is the empirical P-value generated from Bates' function
mcmcpvalue, with my hack for the new structure of the latest mcmcsamp
output (NOTE that I checked that these P-values also agree with direct
HPDinterval and quantile assessments of individual coefficients).

FULL MODEL