[R-sig-ME] MCMCglmm variance estimates Poisson distribution

[Hodsoll, John]

Dear Jarrod 

Thanks for your reply. I was using 4, but all give a similar answer

1. 4.796014

2. 4.792395

3. 4.798754

4  4.790677

As I said I'm not sure I'm missing something obvious?

Cheers
John


-----Original Message-----
From: Jarrod Hadfield [mailto:j.hadfield at ed.ac.uk] 
Sent: 04 March 2014 12:07
To: Hodsoll, John
Cc: 'r-sig-mixed-models at r-project.org'
Subject: Re: [R-sig-ME] MCMCglmm variance estimates Poisson distribution

Dear John,

How are you calculating the posterior expectation:

1/
posterior.mode(exp(mcmc.c11.cf2$Sol+mcmc.c11.cf2$VCV/2))
2/
mean(exp(mcmc.c11.cf2$Sol+mcmc.c11.cf2$VCV/2))
3/
exp(posterior.mode(mcmc.c11.cf2$Sol)+posterior.mode(mcmc.c11.cf2$VCV/2))
4/
exp(mean(mcmc.c11.cf2$Sol)+mean(mcmc.c11.cf2$VCV/2))

If it is not by method 1/ try that and see if there is less of a discrepancy.

Cheers,

Jarrod

Quoting "Hodsoll, John" <john.hodsoll at kcl.ac.uk> on Tue, 4 Mar 2014
11:25:17 +0000:

> Dear list
>
> I'm trying to get some unadjusted estimates and 95% CI for a set of 
> correlated count data (due to repeated measures on the same cluster) . 
> To do this I was trying to run an over-dispersed poisson model using a 
> glmer and MCMCglmm.
>
> I want to use MCMCglmm as that's the package I wish to use for my main 
> analysis. However, it seems to over-estimate the variance meaning that 
> the mean value I get from the intercept only model y = XB + Var/2 (ch2 
> jarrod hadfield's course notes) is slightly greater than the actual 
> mean. For example, if I fit the model
>
> priortr <- list(R=list(V=1, nu=0.001))
>
> mcmc.c11.cf2 <- MCMCglmm(totflct ~ 1, family="poisson", prior = 
> priortr, data=uc11,
>                          nitt = 100000, burnin = 10000, thin = 90)
> summary(mcmc.c11.cf2)
>
> I'm ignoring the random effect and assuming the additive 
> over-dispersion term will capture all the extra variance. For a count 
> rate of 4.69 in the data I get 4.79 and for a count of 5.2 I get 5.52. 
> On the other hand, if I use glmer including a per observation random 
> effect I get the correct means
>
> re.uc12.cf <- glmer(totflct ~ (1|obs), family=poisson, data=uc12)
> summary(re.uc12.cf)
>
> Is there something I missing here?
>
> Regards
> John Hodsoll
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list 
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>
>



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