REML vs ML in lmerTest

 If the anova with lmerTest works like the method for merMod objects, it
will refit (but has the option refit=TRUE). See



http://127.0.0.1:25203/library/lme4/html/merMod-class.html



I don't use the lmerTest package, so hopefully another reader with
knowledge of it can respond if it works differently.



*From:* Bradley Carlson [mailto:carbrae at gmail.com]
*Sent:* Tuesday, May 05, 2015 1:34 PM
*To:* Daniel Wright
*Cc:* r-sig-mixed-models at r-project.org
*Subject:* Re: [R-sig-ME] REML vs ML in lmerTest



The package is lmerTest, the function is anova within that package. I
can't find anyway to determine whether it is re-estimating the model (it
doesn't report anything about REML or ML). If it was re-estimating, though,
then the sums of squares shouldn't be different between the two - but they
are.



On Tue, May 5, 2015 at 2:19 PM, Daniel Wright <Daniel.Wright at act.org>
wrote:

Check if the lmerTest function re-estimates the model with ML.  Not sure
what lmerTest function you are using (which package?).


-----Original Message-----
From: R-sig-mixed-models [mailto:r-sig-mixed-models-bounces at r-project.org]
On Behalf Of Bradley Carlson
Sent: Tuesday, May 05, 2015 1:15 PM
To: r-sig-mixed-models at r-project.org
Subject: [R-sig-ME] REML vs ML in lmerTest

Hi everyone,

I'm a little confused about the use REML and ML. I fit models in lmer,
which were pretty straightforward (a few continuous and a few nominal
predictors, plus random intercepts for clusters of data). I tested the
fixed effects using the lmerTest anova function with Kenward-Roger df (I
have no interest in testing random effect significance). I get the same F
values, df, and p values regardless of whether the models were fit with
REML or ML, but the actual sums of squares in the anova output differ
modestly. Given that it didn't matter at all for the results, it doesn't
seem I should particularly care whether I use REML and ML in the lmer. But,
I want to report which I used.

So my questions:

-Why do I get the same statistical values except for SS with REML and ML?
-Which would be more appropriate - REML or ML? I'm thinking REML because I
have an unbalanced sample sizes for each level of the random effect (based
on Bolker et al. 2008), but I wanted to double check that this makes sense.

Thank you!
Brad

--

Bradley Evan Carlson
Assistant Professor of Biology
Wabash College, Crawfordsville IN

Email: *carlsonb at wabash.edu* <+carlsonb at wabash.edu>
Website: https://sites.google.com/site/bradleyecarlson/home

        [[alternative HTML version deleted]]