[R-sig-ME] zero-inflated models in MCMCglmm
Jarrod Hadfield
j.hadfield at ed.ac.uk
Thu Dec 1 07:48:56 CET 2016
Hi Gustaf,
I don't have dat, so I can't run the final bit of code.
However, these are my thoughts.
1/ In the zap model you are allowing X1 to X3 to effect the level of
zero alteration, whereas in the zip model you are just fitting a
constant level of zero inflation across X1 to X3. In this sense the zap
model will provide a superior fit because it has more parameters. The
warning message about dropping terms is fine, although the default
contrast for the zip model is a bit annoying: I would have preferred the
main effect for X1 to be yes rather than no, but I guess its no big deal.
2/ You have fitted a single nested_plot effect in the random effects.
This is a little odd, except in the case where the data are not
zero-inflated. In this case having a single nested_plot term in the zap
model is equivalent to fitting a nested_plot term in the standard
Poisson. If the data are zero-inflated, and/or the model is not a zap
model, then the case for having a single term is not well justified. I
would use us or idh and in the latter case consider fixing the second
variance (associated with zero-inflation/alteration) close to zero.
3/ The marginal predictions do not take into account the covariances
between traits. This is generally OK, but its not ideal when the traits
refer to the multiple parameters of a single distribution as with
za/zi/hu models. I should put a warning in. You can also use the
simulate function on the model and then average over the draws to get
the predictions, and this will take into account any covariances. Note
that if you use idh(trait):nested_plot there are no covariances so
predict should be fine.
Cheers,
Jarrod
On 30/11/2016 23:30, Gustaf Granath wrote:
> cbind(aggregate(y ~ X1*X2*X3_nest, zero.dat, mean),
> zip = aggregate(p.zip ~ fire*retention*micro_hab.two, dat,
> mean)$V1,
> zap = aggregate(p.zap ~ fire*retention*micro_hab.two, dat,
> mean)$V1,
> pois = aggregate(p.pois ~ fire*retention*micro_hab.two, dat,
> mean)$V1)
--
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
More information about the R-sig-mixed-models
mailing list