Skip to content
Back to formatted view

Raw Message

Message-ID: <1288331857.4800.130.camel@PortableToshiba>
Date: 2010-10-29T05:57:38Z
From: Emmanuel Charpentier
Subject: Numerical integration for cross-classified random effects in lme4
In-Reply-To: <4CC87FF7.7090106@u.washington.edu>

Le mercredi 27 octobre 2010 ? 12:39 -0700, David Atkins a ?crit :
> Jeremy--
> 
> Just to chime in here, I think the devil is almost always in the details 
> of the specific data and model.  Thus, "Stata can 'do' AGQ for multiple 
> random-effects" really does not mean too much outside of a specific set 
> of data and model.
> 
> I recently assisted with an Epidemiology colleague who was trying to run 
> a cross-classified GLMM with logit outcome and ~45K participants.  After 
> 24 hours Stata had made it through 4-5 iterations.  lmer() fit the model 
> in 5 minutes.
> 
> At the same time, that example does not mean lmer() is universally 
> superior for all models and datasets.  However, if you're talking 
> cross-classified data... it probably is. ;)
> 
> As for MCMC, I would strongly recommend taking a look at MCMCglmm, which 
> is a fully Bayesian package for generalized linear mixed models and very 
> good.

Or biting the whole bullet (cannonball ?), building a BUGS model, lauch
it in JAGS and planning  long nice weekend trip... after having
validated the model on a "reasonable" random subset of the data against
a "first aproximtion" given by lme4, Stata or whatever is your fad of
the week.

					Emmanuel Charpentier