[R-sig-ME] question about lmer versus lme and evaluation of estimates of random parameters

[Paul Johnson]

Greetings!  I'm back for my once-a-semester adventure in mixed model
user support.

I'm asked to help a student with a problem involving surveys done in
about 50 countries.  The dependent variable is a 7 category ordinal
scale, and I've searched the archives on this (and other lists) and
see many of you recommend we try to use the linear/gaussian model
(with diagnostics after).  That's the only feasible option in R, as
far as I can tell. There are some commercial packages that claim to
support mixed model fitting for ordinal outputs (HLM), but I don't
have them and don't know if they are good.

We need to estimate a random intercept at the country level and the
researcher also supposes that there are variations across countries in
the slopes for 3 input variables.  I've estimated that with the newest
version of lmer in lme4.

However, I noticed when I tried to run the mcmcsamp routines to study
the distributions of estimates I noticed that the help page for
mcmsamp does not offer quite so much detail about how it is supposed
to be used.  Backtracking into this list again, I find that mcmcsamp
is "not yet ready" for non-gaussian models. Maybe it never will be, I
can't tell from the commentary.

But it is still supposed to be useful for linear models, right?   I
can still run mcmcsamp, but don't understand how to interact with the
result--the old stuff we were using last year with coda and the
HPDinterval function don't work any more.

Anyway, while searching in this list, I see some of you pointing
people in my situation to use routines from nlme, rather than lme4.

Is that right?  Should I push for use of lme (nlme) rather than lmer (lme4)?

I gather one advantage of using lmer is that it would allow correlated
random effects. However, if we just stay with a vanilla
specification--country level random intercepts and uncorrelated random
country level slopes, is there any advantage to lmer compared to lme?

Is lme expected to be faster because its random effects structure is simpler?

In either of these models, how do we gauge the accuracy of estimated
standard deviations for random parameters?

Can you tell me how to use mcmcsamp in current lme4

or

show me how people make that assessment in nlme ?



-- 
Paul E. Johnson
Professor, Political Science
1541 Lilac Lane, Room 504
University of Kansas