[R-sig-ME] Likelihood Ratio tests and fixed effects with LMER

[David Duffy]

On Tue, 16 Dec 2008, Rafael Maia wrote:

> Hello,
>
> I am quite new to all this approach, but since it's somewhat different from 
> what I've seen in classes and I have to rely on what I have been learning by 
> myself and on lists such as this one, it's easy to get confused. Anyway, I 
> have seen in several textbooks which take this anova table / LRT approach the 
> opposite "direction" of effect testing: starting with the full model, 
> removing terms (interactions first, then according to effect size, for 
> example) and comparing to the previous one.
>
> This may even be kind of "off topic" for this list, but since considerable 
> discussion has been going on here about hypothesis testing and LRT 
> previously, and the question was originally asked here, I thought it wouldn't 
> hurt to continue the discussion...

It's a generic model building question.  If you have large numbers of 
variables, fitting the all K-way interactions model as a base model can be 
expensive, so people try to build upwards from simpler models.  I quite 
like the idea of Bayesian Model Averaging...

>
> There was some discussion here previously about how it wouldn't be good to 
> compare likelihoods of GLMM and GLM, because they are generated differently 
> or something like that...

That was just whether particular constants are included or excluded in the 
likelihood expressions used in the different codes.  My comment was 
to the effect that the parameter estimates and LRTs from simple fixed 
effects models are a reality check for the results coming from the GLMMs, 
especially if you are worried that something isn't working properly.

David Duffy.
-- 
| David Duffy (MBBS PhD)                                         ,-_|\
| email: davidD at qimr.edu.au  ph: INT+61+7+3362-0217 fax: -0101  /     *
| Epidemiology Unit, Queensland Institute of Medical Research   \_,-._/
| 300 Herston Rd, Brisbane, Queensland 4029, Australia  GPG 4D0B994A v