[R-sig-ME] Binomial model variance and repeatability estimates with MCMCglmm

[Jarrod Hadfield]

Hi,

The simulation only had one trial so it was equivalent to categorical.  
If you up the number of trials then you can estimate the  
observation-level variance.

You might try parameter expanded priors to remedy the last problem.

Cheers,

Jarrod


Quoting Ned Dochtermann <ned.dochtermann at gmail.com> on Tue, 21 Oct  
2014 17:00:12 -0500:

> Thanks, I was aware of that for categorical and some of the other  
> families but thought I could get away with it here and I wasn't  
> quite sure otherwise how to calculate the relevant ratio (thanks for  
> providing that too).
> With smaller sample sizes repeatability still seems to get  
> misestimated and stuck close to zero even with long runs but running  
> multiple chains seem to resolve that.
>
> Thanks again,
> Ned
>
> On 10/21/2014 3:05 PM, Jarrod Hadfield wrote:
>> Hi,
>>
>> The residual variance of a binary response cannot be estimated, so use
>>
>> prior1 = list(R = list(V = 1, fix=1),
>>              G = list(G1 = list(V = 1, nu = 0.002)))
>>
>> In this example it is more efficient to aggregate success/failures  
>> of an individual into a multi-trial binomial response and use:
>>
>> prior2 = list(R = list(V = 1, nu=0.002))
>>
>> sim.mcmc2<-MCMCglmm(cbind(Fail,Success)~1,
>>                    family="multinomial2", prior=prior2,
>>                    nitt = 260000, thin = 200, burnin = 60000,
>>                    verbose=FALSE,data=ind.data)
>>
>> sim.mcmc2$VCV/(sim.mcmc2$VCV+pi^2/3)
>>
>> Cheers,
>>
>> Jarrod
>
>
>



-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.