[R-sig-ME] zipoisson in MCMCglmm
j.hadfield at ed.ac.uk
Tue May 12 00:14:31 CEST 2009
As with the random effect specification you probably want to form
interactions with the reserved variable "trait". These interactions
allow you to model fixed effects for both processes. Something like:
~ trait+trait:fixed1 + trait:fixed2-1
I usually use -1 so that the contrasts are within traits, rather than
trait2-trait1. For zipossion models I would also save the posterior
distribution of latent variables (pl=TRUE) to make sure its mixing
properly by plotting the mcmc traces.
Most importantly, you should probably fix the residual variance of the
zero-inflation process to something (I use 1) because it cannot be
identified from the data:
priors <- list(R=list(V=diag(2),n=2, fix=2), G=list(G1=list(V=diag(2),
n=2), G2=list(V=diag(2), n=2)))
fix=2 fixes the bottom right diagonal matrix starting at [2,2] (in
this case the zero-inflation variance).
Quoting Josh Van Buskirk <jvb at zool.uzh.ch>:
> Dear all,
> Does anyone have experience working with Jarrod Hadfield's MCMCglmm
> package with a zero-inflated Poisson distribution? After fitting the
> model, I'm having trouble obtaining the fixed effect coefficients from
> the logistic (inflated) and Poisson parts of the model. I'm interested
> in estimating how the fixed effects influence both processes.
> In this example, many random genotypes are sampled within many random
> populations. There are two fixed effects. The response variable is
> highly zero-inflated.
> priors <- list(R=list(V=diag(2),n=2), G=list(G1=list(V=diag(2), n=2),
> G2=list(V=diag(2), n=2)))
> model <- MCMCglmm(
> response ~ fixed1 + fixed2 ,
> random = ~idh(trait):Population + idh(trait):Genotype,
> family = "zipoisson",
> prior = priors,
> rcov = ~idh(trait):units,
> data = mydata )
> After fitting the model, the object called model$VCV contains 8
> variance components, which makes a little bit of sense: zero-inflated
> and Poisson parts of two random effects (Population and Genotype), plus
> the same for the residual.
> However, the object model$Sol contains estimates for three fixed
> effects (intercept, fixed1, fixed2). I expected there to be twice as
> many, because fixed effects can influence both the logistic and Poisson
> parts of the model. In fact, I'm not sure which process these estimates
> refer to (Poisson or logistic).
> Any insight here?
> Many thanks,
> Josh Van Buskirk
> University of Zurich
> jvb at zool.uzh.ch
> R-sig-mixed-models at r-project.org mailing list
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
More information about the R-sig-mixed-models