On 11/29/06, Julien <martin.julien.2 at courrier.uqam.ca>
wrote:> Hi
> I know that p-values doesn't appear anymore in the summary of a linear
> mixed-model with lmer. However, if I do a mixed logistic regression with
> lmer using family=binomial, the summary includes a p-values for fixed
> effects.
> Is it normal, could I use those p-values to interpret the fixed effects
> or should I use mcmcsamp to obtain 95% confidence interval?
Those p-values are calculated from the standard normal quantile
function, which is likely to be a worse (more "anti-conservative")
though less controversial approximation that the p-values that were
reported for linear models.
I would recommend using a chain from mcmcsamp for more precise
evaluation of the variability of the parameter estimates except that
mcmcsamp for generalized linear mixed models isn't working very well
right now. The mcmcsamp function uses exact sampling in each stage of
the Gibbs sampling for linear mixed models. For generalized linear
mixed models there is one stage where I must use a Metropolis-Hastings
step and the surrogate distribution that I chose is less successful
than I had hoped it would be. The result is that fixed-effects
parameter values can get stuck at the same position for long periods
in the chain. I plan to fix this but there are several other tasks to
finish first.