[R-sig-ME] using Poisson glmer for non-integer data

[Ben Bolker]

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


 [cc'ing to r-sig-mixed-models]

  I don't think it's crazy to use a Poisson distribution with
non-integer response values in some cases, but you're correct that
non-integer response values don't work in lme4 at present.  In
principle we could dig down and fix the problem (feel free to post an
issue at https://github.com/lme4/lme4/issues, with a simple
reproducible example ...) but I have to say it's not very high on our
list, because these are (usually? often?) cases where the original
model is somewhat suspect anyway.  I can suggest the following
workarounds:

  * try another package such as glmmML or glmmPQL
  * use an offset that characterizes the total lifespan; if you use
offset(log(lifespan)) that will effectively model success per unit
lifespan, and if you include log(lifespan) as a predictor (with the
standard log link) that will effectively model success as proportional
to (lifespan)^b, where b is a parameter to be estimated.

  Other discussion of Poisson with non-integer values

http://www.r-bloggers.com/poisson-regression-on-non-integers/
http://stats.stackexchange.com/questions/38530/how-does-a-poisson-distribution-work-when-modeling-continuous-data-and-does-it-r/38588#38588
http://stats.stackexchange.com/questions/70054/how-is-it-possible-that-poisson-glm-accepts-non-integer-numbers

  sincerely
    Ben Bolker

On 14-08-10 09:08 PM, Christina Painting wrote:
> Dear Prof Bolker,
> 
> I'm a behavioural ecologist at the University of Auckland in New 
> Zealand, and I currently have a masters student who is tackling
> some lifetime mating success data for the NZ giraffe weevil.
> 
> We were hoping you might be able to offer us some advice on an
> issue we are having using the /lme4/ R package. Our response
> variable is the average mating success of a giraffe weevil for its
> lifetime (total success/lifespan) and we are looking at this in
> relation to body size and time of year. Using your 2008 TREE paper
> on using GLMMs we figured out that the best method to use was a
> model with Poisson distribution with Laplace approximation because
> av. mating success is non-normally distrubited, can't be fixed with
> standard transformations and has a mean <5. However, because the
> data are not integers we have run into problems, with the models
> returning warnings about the data being non-integer, and then we
> can't get log-lik and AIC values.
> 
> Reading online on various forums that you have been part of
> suggests we aren't the only ones having this problem, and I
> wondered if you had any solutions to this problem, or could suggest
> another method to use that would be robust to our average measure
> of mating success?
> 
> We would greatly appreciate any advice you can offer, and thank you
> in advance
> 
> Kind regards, Chrissie Painting
> 
> *Dr Chrissie Painting* Post Doctoral Researcher School of
> Biological Sciences University of Auckland 
> cpai015 at aucklanduni.ac.nz <mailto:cpai015 at aucklanduni.ac.nz> 
> https://sites.google.com/site/paintingchristina/ Mobile: +64 27 306
> 1610

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)

iQEcBAEBAgAGBQJT6Dm0AAoJEOCV5YRblxUH7WQH/RYNGkliTnUkVyzCrnA0MNl9
sPTR6Vkz7Hdxm8YdHzuSHa1pzyyVmkuVMQP4IW9nD2II872nJxjTJgmC13Ji2bCj
dgV+HPTg2yQeYZqh6NQXWGvTWjrn8rQwzKF/2sxn5l3YWk5T13egy6lw2DI5FQVi
+xnOFMBkcnTQ3eysuO+8zaJFN6550IKQ6UT79D8hQ2TiHEXGrq2DeHVmp8Uqr2UY
f4Mx4G+NBRjmdtYoeQpgK3baRw1uoOI40GcQRPUHzUVaVJVXxk1QfrDiReK8yErs
Mg+lxV7ylMgv0J0+prqBsS7wibeSN8k31ghXWQ9MrnxKusdlsbcicKQQ5dbmgTU=
=5+8D
-----END PGP SIGNATURE-----