An embedded and charset-unspecified text was scrubbed... Name: not available URL: <https://stat.ethz.ch/pipermail/r-sig-mixed-models/attachments/20120223/0e7a918a/attachment-0001.pl>
fitting a distribution to zero-inflated catch per unit effort mixed model
2 messages · Karla Letto, Ben Bolker
Karla Letto <karla.letto at ...> writes:
I am having trouble fitting a distribution to my mixed model for meadow vole catch per unit effort (CPUE) data. I have tried several families and cannot find one that does not violate both the homogeneity and normality assumptions. NOTE: The data set is zero-inflated (no captures). Here is my study design: I am trying to determine if the CPUE of meadow voles differ among lines (line = near,mid, or far) at increasing distances from a linear feature (type = road, trail or powerline corridor) in two different habitat types (habitat=forest or barren). Response variable: catch per unit effort (non-integer values) Fixed explanatory variables: line (3 categories), habitat (2 categories), type (3 categories) Random explanatory variable: site (8 categories), cycle (2 categories)
The random variable site is for the 8 different sites I sampled in (4 barren and 4 forest) and the cycle is there because I visited each site twice.
Practically speaking you probably can't use cycle as a random effect; you can include cycle as a fixed effect (specifying the difference between first & second visits), and possibly nesting it within site as a random effect (if you have more than one observation per site/sample combination). What is the total size (number of observations) in your data set?
Here is an excerpt of my data set: CE2 catch effort site line cycle habitat type 0.000000 0 57.5 A near first forest trail 3.278689 2 61.0 A mid first forest trail 0.000000 0 60.5 A far first forest trail 0.000000 0 66.5 G near first barren road 0.000000 0 74.5 G mid first barren road 0.000000 0 74.0 G far first barren road 1.449275 1 69.0 E near second barren powerline 0.000000 0 73.0 E mid second barren powerline 0.000000 0 71.5 E far second barren powerline I tried the lme4 package using the following syntax:
[snip]
I then tried using a poisson error structure using catch (the actual number of animals) as the response and incorporated effort as an offset. Effort as an offset is commonly used for analysis of CPUE data. Model2<- glmer(catch~line+habitat+type+(1|site)+(1|cycle)+ offset(effort),family=poisson)
This is a good way to do it, but you need to incorporate the LOG of effort. You may also need to account for overdispersion and/or zero-inflation; the former via incorporating an observation-level random effect (in glmer, glmmadmb, or MCMCglmm) or negative binomial distribution (in glmmadmb), the latter (if necessary) via zero-inflation or hurdle models (in glmmadmb or MCMCglmm). [snip snip snip]
Does anyone have any suggestions on how I can analyze zero-inflated CPUE data? I have been trying to figure out how to do a Monte Carlo permutation test for a mixed model but I am having trouble figuring out the syntax. Any help would be greatly appreciated.
What are you using to assess homogeneity of variance and normality? Normality of residuals can only be expected approximately (and in the case of large mean counts) in this case. I would start off this way: mydata$obs <- factor(seq(nrow(mydata))) glmer(catch~line+habitat+type+cycle+(1|site/cycle)+(1|obs)+ offset(log(effort)),family=poisson, data=mydata) You can use the simulate() method to simulate data sets, count the proportion of zeros expected, and see if your observed proportion of zeros is off ...