[R-sig-ME] Binomial GLMM vs GLM question

[Ben Bolker]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Ken Beath wrote:
|> Dear Dr. Bates and other LMER experts,
|>     I am admittedly entry level in my R and mixed-model knowledge, but
|> I'm hoping that someone can help me and also forgive my lack of
|> insight.  Over 3 years, I monitored survival of 350 egg masses at two
|> ponds.  I thus have one continuous variable (rainfall) and two discrete
|> variables (year and pond).  My response variable, mortality, is coded as
|> a two column matrix featuring eggs survived and eggs dead. I'm primarily
|> interested in the effect of rain on survival, but also if rain has
|> different impacts at the different ponds and how much survival varied
|> over the three years.  Originally, I though I could tackle this with a
|> binomial GLM, but do I need a binomial GLMM instead, as rainfall and
|> year would be random and pond fixed?  The problem with this is trying to
|> make biological sense out of the results.  I've spent the last week
|> reading all the past posts about why p-values can't be calculated and
|> all that, which I'm fine with.  But what can I say about the effects of
|> rainfall or year on egg survival from the variance estimates?  Also,
|> doesn't LMER require that random factors be normally distributed,
|> because my rainfall measurements are far from it.  Is that a problem?
|> Thank you in advance for any advice you can give.
|> -Justin Touchon
|>
|
| I think your misunderstanding the idea of a random effect. This is
| something that is unobserved, causing correlation within a group. In your
| data this might be year or pond but definitely not rainfall which is
| simply a covariate. You have more than one measurement on a pond and more
| than one for each year, so it is likely that there will be correlation
| between them and one way of dealing with this is a random effect. The
| alternative is to use a fixed effects model. In your case, there are only
| 2 and 3 groups, so a fixed effects model is the best approach, so a GLM is
| appropriate. If there were say 20 ponds, a random effects model would be
| much more suitable.
|
| Ken

~  Agreed.
~  One other point to watch out for (not directly related to GLMMs)
is that you should check the scale parameter estimate: I don't know if
the model you have below is sensible or not, but if you do end up
with an estimated scale parameter of 4.6, you must take overdispersion
into account (e.g., with a quasibinomial or beta-binomial model).

~  Ben Bolker

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFILYmXc5UpGjwzenMRAimcAJ0ZhaBH8LEd4C44uk3n0Y7ul1HUtQCfazFY
sY6qlcU4psciD1VYXO9xeE4=
=S1x+
-----END PGP SIGNATURE-----