[R-sig-ME] Multiple random and cross-classified factors specification and caveats in a generalized linear model (with MCMCglmm)

[Ned Dochtermann]

Dear list,
Since I tend to be verbose, the question eventually being asked below is 
whether the model specification seems appropriate and, being unfamiliar 
to this sort of categorical cross-classified model, whether there are 
glaring problems I'm missing.

I'm trying to analyze a dataset of paired contest outputs with winners 
and losers. I'm wanting to use a mixed model as opposed to an 
alternative approach because I'm particularly interested in whether 
there are different among-individual variances for a particular grouping 
variable (handedness). Individuals show up as both focal individual and 
opponent across multiple contests so there are a few ways in which 
pseudoreplication enters the dataset.

The data basically looks like:
Focal.ID        Opp.ID                  Contest.ID Handedness        Won
Bob                Jack                      1    L                    
         W
Jack                Bob                      1   R                    
         L
Bob                Sam                      2 L                            W
Sam               Bob                       2 L                            L
Matt              Sam                      3 R                           W
Sam               Matt                     3 L                            L
...
John              Steve                    880                   R    
                        W
Steve             John                     880 L                        
     L


We have 880 contests with 588 unique focal id's, with 58% of individuals 
competing in at least one contest. Unfortunately some individuals had to 
be excluded so both individuals from a contest aren't always included. 
The fixed portion of the analysis includes only an intercept (based on a 
published analysis someone else did and because of the specific 
question). The random part of the model is then:

random=~idh(Handedness):Focal.ID+Opp.ID+Contest.ID

I'm then looking at the posterior distribution difference in latent 
scale variances for lefties and righties. I don't actually care about 
the variances of the other effects, they're just included because they 
seemingly should be. The thought was that this sort of model structure 
would at least partially encompass the data structure.
Does this structure seem reasonable?
I know detecting variance differences between groups is going to be 
tough, but this is the data we could obtain.

Thanks,
Ned

-- 
Ned A. Dochtermann
Assistant Professor / Department of Biological Sciences
*NORTH DAKOTA **STATE UNIVERSITY*
p: 701.231.7353 / f: 701.231.7149 / www.ndsu.edu

https://sites.google.com/site/neddochtermann/
ned.dochtermann at ndsu.edu