[R-sig-ME] Switch to ML during model simplification

[Douglas Bates]

On Tue, Mar 15, 2011 at 3:29 PM, Iker Vaquero Alba <karraspito at yahoo.es> wrote:
>
>    Hello, list members:
>
>    I am fitting a mixed effects model with a split-plot analysis. I have read that in order to be able to use anova to compare models with a different fixed-effects structure, it's not correct to use the default REML method, so it must be changed to ML.
>
>    - When using "lme", that can be done just writing: method="ML", whereas when using "lmer" you achieve that by writing "REML=false".

Actually, that's REML=FALSE (case is important).

>    - My questions are:
>       1. Is it necessary to specify the ML method when fitting the maximum model, or is it better when simplifying the first term (with the "update" function), so that the maximum model is analized by REML and the subsequent simplified models by ML?

Not quite sure what you mean there but whenever you compare two model
fits they should both be fit by ML.

>       2. Once you change the method to "ML" either in the maximum model or in the first simplified model, is it necessary to specify "method="ML"" or or "REML=false" (depending on the function) on each of the subsequent model simplifications or is it enough with writing it once?

If you use the update() function it will preserve the setting of
method="ML" or REML=FALSE when generating the new call from the
previous model fit.