anova.merMod recalculates models using REML = FALSE, where to find the reference?

Question

I have been using the anova.merMod function from lme4 package to obtain p-values for fixed effects through likelihood ratio tests for a scientific publication (most reviewers still demand p-values in my field). I noticed that the anova.merMod function recalculates the lmer functions using REML = FALSE (see the example below), which is an incredibly nice feature forcing less acquainted users to do the test right. However, I have been trying to read most of the documentation for lme4 package and cannot find that a notation of this feature (for instance, see ?anova.merMod, which directs the user to ?vcov.merMod). This makes me confused.

Question: Why is this feature not clearly mentioned in the documentation? Have I understood it wrong, perhaps?

Ps. there seems to be a question about this on the R-mailing lists, but the answers make me even more confused.

library(lme4)
data(sleepstudy)

reml <- lmer(Reaction ~ Days + (Days | Subject), sleepstudy)
noreml <- lmer(Reaction ~ Days + (Days | Subject), sleepstudy, REML = F)

reml0 <- lmer(Reaction ~ (Days | Subject), sleepstudy)
noreml0 <- lmer(Reaction ~ (Days | Subject), sleepstudy, REML = F)

## Returns similar likelihood ratio test statistics:
(a <- anova(reml, reml0))
(b <- anova(noreml, noreml0))

## Not identical though
identical(a, b)
[1] FALSE

EDIT: sessionInfo: R version 3.0.2 (2013-09-25), lme4_1.0-5

Answer 1

At least as of recent versions of lme4, this is documented (albeit way down in the details) in ?anova.merMod (emphasis added):

‘anova’: returns the sequential decomposition of the contributions of fixed-effects terms or, for multiple arguments, model comparison statistics. For objects of class ‘lmerMod’ the default behavior is to refit the models with ML if fitted with ‘REML = TRUE’, this can be controlled via the ‘refit’ argument. See also ‘anova’.

Answer 2

With thanks to Roland for doing the legwork, I'm posting my comment as the answer.

I'm not convinced the answers aren't the same: identical will return FALSE if any floating-point numbers aren't exactly the same or if any name of any variable differs. Can you take a look at the actual values of the returned elements of interest and see if they differ by more than machine precision? –

Roland ran the test and found that the only difference is the name attributes.