[R-sig-ME] Partialing out the effect of a covariable in a Poisson GLMM

[Ben Bolker]

On 13-01-21 02:51 AM, v_coudrain at voila.fr wrote:
> Ben Bolker writes: 
> 
>> Don't know exactly what you want to do, but presumably fitting
>> a "null model" with the covariate and then building more complex models
>> on top of this -- adding the other covariates -- and doing sequential testing
>> (anova(full_model,model_with_only_first_covariate)). The linear models
>> idea of partialing out one variable and running a model on the residuals
>> doesn't work as well, unfortunately.
> 
> Oh yes, thank you I can do it this way. Sould I turn REML to FALSE for model comparison? 
> Best,
> Valérie

   REML is *ignored* when running glmer() [or lmer() with a 'family'
argument]; as discussed at http://glmm.wikidot.com/faq , REML is
somewhat poorly defined for GLMMs.  (The development version of lme4 at
least warns that the REML argument is ignored; the CRAN version just
silently ignores it.)

  Ben Bolker