[R-sig-ME] Fixed effects only model with lme4

[Emmanuel Charpentier]

Dear Pr Bates,

Le samedi 06 juin 2009 à 12:49 -0500, Douglas Bates a écrit :
> On Sat, Jun 6, 2009 at 12:03 PM, Jeroen Ooms<jeroenooms at gmail.com> wrote:
> > For my GUI, I would like the user to be able to compare a fixed effects
> > model with a random effects model. For example:
> > fm0 <- lmer(Reaction ~ Days, sleepstudy);
> > fm1 <- lmer(Reaction ~ Days + (1|Subject), sleepstudy);
> > anova(fm0,fm1);
> >
> > However, this returns the obvious "No random effects terms specified in
> > formula" error for the first model. I've also tried fitting the fixed
> > effects model with lm:
> >
> > fm0 <- lm(Reaction ~ Days, sleepstudy);
> > fm1 <- lmer(Reaction ~ Days + (1|Subject), sleepstudy);
> > anova(fm0,fm1);
> 
> Try listing them the other way around
> 
> anova(fm1, fm0)
> 
> If the first model in the call to anova is of class "lm" then the
> method for that class is the one chosen and that method doesn't know
> about models created by lmer.  You must list them so that the lmer
> model comes first.

You could do that ... with lme objects. As explained by the OP, this no
longer works with lmer objects.

One may always extract deviances and substract, and make rash hypotheses
in difference in numerator DFs, but, while this should work (= give
expected results) with lm and gaussian objects, nothing guarantees that
glm and lmer use the same parametrizations for non-gaussian models (the
GLM chapter in V&R4 states that deviances are computed up to an additive
constant. I tried to follow lmer code, but quickly got lost in C
code...).

May this join your (already well-grown) wishlist (along with
deviance/likelihood under arbitrary parameters, maybe :-) ?

Sincerely,

					Emmanuel Charpentier