[R-sig-ME] fitting models with poisson distributed data

[Page E. Van Meter]

Thanks, Ken. I am coming to a similar conclusion. My data is very zero 
inflated and I have considered using negative binomial, which also does 
not seem to work with lmer.
I will try working on both quasipoisson and a binomial version of my 
data. Thanks,
-Page

Ken Beath wrote:
> On 25/10/2008, at 9:37 AM, Page E. Van Meter wrote:
>
>> Hi,
>> Now that I have the code figured out, I hoping for some help on 
>> defining my model. I might be guilty of trying to fit an overly 
>> complex model to my data, although my model seems very simple in 
>> comparison to what has been discussed here. I'm hoping for feedback 
>> on my model design. Thanks in advance!
>>
>> I have some pretty ugly longitudinal data measuring hormones and 
>> behaviors from individual hyenas over many years (355 samples from 39 
>> individuals). We collect hormone samples based on opportunity and 
>> have several samples from each individual (3-9 samples per hyena). My 
>> ultimate goal is to see if my hormone data explains any of the 
>> variation we see in the behavior data (aggression, I'll call it 
>> aggs). My dependent measurement is a behavior rate, count of aggs 
>> over time just prior to hormone sample collection. It is very zero 
>> heavy (lots of individuals did not aggress prior to hormone sample 
>> donation) and resistant to transformation to normality, but seems to 
>> be a pretty poisson distribution. My predictors are hormones and 
>> reproductive state (pregnant or lactating, which effect both 
>> aggression and hormones).
>>
>
> From the output the estimated scale (I don't see this in the version 
> of lmer I'm using?)  is 7.7 so data is definitely not Poisson. 
> Assuming Poisson will give incorrect p values.
>
> Seeing the quasi Poisson doesn't seem to work properly I'm not certain 
> what is a good choice. I haven't tried it but maybe quasi Poisson 
> works in one of the GEE packages.
>
> It may be Ok to limit the analysis to no aggression/aggression 
> allowing fitting as binomial data.
>
> Ken
>
>
>> m2<-lmer(aggs~reprostate+hrm1+hrm2+(1|id), family=poisson, aggs)
>>
>> Generalized linear mixed model fit using Laplace
>> Formula: aggs ~ reprostate + hrm1 + hrm2 + (1 | id)
>>  Data: aggs
>> Family: poisson(log link)
>>  AIC   BIC logLik deviance
>> 12369 12387  -6179    12359
>> Random effects:
>> Groups Name        Variance Std.Dev.
>> id     (Intercept) 4.9353   2.2216 number of obs: 307, groups: id, 39
>>
>> Estimated scale (compare to  1 )  7.682887
>>
>> Fixed effects:
>>           Estimate Std. Error z value Pr(>|z|)   (Intercept) 
>> -3.07625    0.39575  -7.773 7.65e-15 ***
>> reprostate 2.32056    0.07724  30.044  < 2e-16 ***
>> hrm1        0.12575    0.04020   3.128  0.00176 **
>> hrm2       -0.80434    0.04770 -16.862  < 2e-16 ***
>> ---
>>
>> Correlation of Fixed Effects:
>>        (Intr) statct ecent
>> statecat -0.381             ecent    -0.085  0.271      acent     
>> 0.136 -0.339 -0.079
>>
>> -- 
>> ************************************
>> Page E. Van Meter
>> Michigan State University
>> Department of Zoology
>> vanmete7 at msu.edu
>> **http://msu.edu/~vanmete7/*
>>
>> _______________________________________________
>> R-sig-mixed-models at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>>
>
>

-- 
************************************
Page E. Van Meter
Michigan State University
Department of Zoology
vanmete7 at msu.edu
**http://msu.edu/~vanmete7/*