Dear Thierry,
Thank you for your advice. So, using (1|site) should suffice, even
though it was always the same sites sampled? I really worry about not
specifying the random part correctly.
Finally, do you know of a good tutorial on how to speciffy contrast
coefficients to do the comparisons I am interested in?
Thanks again.
Ellen
2015-10-02 3:03 GMT-05:00 Thierry Onkelinx <thierry.onkelinx at inbo.be>:
Dear Ellen,
You're using the Poisson distribution. There is no error (noise) term in
glmm with Poisson distribution.
1) The random part seems to be quite complicated given the sample size.
(1|site) is probably sufficient. Note that your design is not nested but
crossed.
2) Overdispersion is likely in bird abundance. You could use a negative
binomial distribution instead of a Poisson distribution. Then the
overdispersion is modeled. Use the glmer.nb() function.
3) Have a look at the glht() function in the multcomp package. That
you to test specific contrasts of your model parameters.
Note that the r-sig-mixedmodels list is more appropriate for follow-up
questions.
Best regards,
ir. Thierry Onkelinx
Instituut voor natuur- en bosonderzoek / Research Institute for Nature
Forest
team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance
Kliniekstraat 25
1070 Anderlecht
Belgium
To call in the statistician after the experiment is done may be no more
asking him to perform a post-mortem examination: he may be able to say
the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of
~ John Tukey
2015-10-01 16:59 GMT+02:00 Ellen Andresen <eandresens at gmail.com>:
Hello,
I studied the effect of a hurricane in Cozumel on understory birds. I
have bird abundances (i.e. counts) registered always on the SAME six
sites (i.e. blocks). I have data for: before the hurricane, first year
after the hurricane, second year after the hurricane. I each of these
time periods, I also have data for summer season and for winter
season. I do not have a balanced design, in one of the time periods I
only have data for 5 of the six sites, and for another period I only
have data for 3 of the six sites.
I am defining Poisson error distrubution for the response variable.
I am using 'glmer' with two fixed factors, and I am interested in
their interaction:
- factor hurricane (three levels: before, after 1 y, after 2 y)
- factor season (two levels: summer, winter)
I am also specifying a random factor (sites), and I am specifying the
nested structure of the design. However, I don't know if I am
specifying the random part of the model in the correct way; this is
what I am doing:
abundance ~ hurricane*season + (1|site/hurricane/season)
I have three questions:
1. Is the random part specified correctly?
2. How do I check for overdispersion, and how can I correct for it?
(for each site I only have one observation; sites are my replicates)
3. How do I make the following comparisons: I am interested in testing
for each season separately, after 1 y vs. before the hurricane, and
after 2 years vs. before the hurricane.
Thank you so much!
Ellen Andresen
UNAM-Mexico