Skip to content
Prev 5369 / 20628 Next

Interpretation of lmer output in R

[I had started to answer this when Doug's answer came through, so I
will interleave my answers where they expand or differ ...]
On 11-02-19 10:04 AM, Douglas Bates wrote:
If you have a reasonably thorough understanding of both linear mixed
models (LMMs) and generalized linear models (GLMs) you can usually
triangulate to get a good idea of what GLMMs are doing.  I would
strongly recommend Zuur et al's book on mixed models: I don't agree with
absolutely everything says, but it's the most accessible treatment for
ecologists that I have seen.  Faraway 2006 is pretty good too (although
Zuur covers many more of the complexities that ecologists end up dealing
with).
Yes, although even the likelihood ratio test (although better than
the default Wald test printed by summary()) is an approximation based on
large sample sizes; if you're interested in seeing the effect of this
approximation, you can try the methods shown in the examples of the
"simulate-mer" help page on the latest release of lme4 (should be on
CRAN now)
I'm afraid I have to disagree with Doug here. This kind of model
reduction, while seemingly sensible (and not as abusive as large-scale,
automated stepwise regression), is a mild form of data snooping.  Don't
do it ...
Hmmm. My interpretation is a little bit different here too.  I
wouldn't try to remove it: I would say that there is indeed a fairly
large remaining variability in the probability that different birds of
the same sex will stay at the site even given the same breeding history.
You can compare the magnitude of the standard deviation to the
magnitudes of the fixed effects -- they're on the same scale -- so the
between-bird variability is fairly large relative to the difference
between sexes, but small relative to the effect of breeding history.