[R-sig-ME] covariances between non normal traits
Emmanuel Charpentier
emm.charpentier at free.fr
Sat Mar 24 10:39:32 CET 2012
Dear Celine, dear list,
That kind of multivariable modeling is (relatively) easy to do in
(Win)BUGS/JAGS, by programming a logistic model for each trait, whose
linear predictor could be modeled as (say) a multivariate normal. This
could provide some kind of comparison standard to which you could compare
other results obtained by (say) maximum likelihood (see for example the
dclone package for a possible approach).
I'd recommend you take a look at Gelman's & Hill's (2007) book on
multilevel regression, whose part 2 discusses that kind of Bayesian
modeling. Using "weakly informative" priors (e. g. inverse-Wishart priors
for variances) should give you estimations (point estimates and
credibility intervals) close to those of a frequentist analysis (but no p-
values : for that, you'll have to do that yourself (not easily : ask
Douglas Bates...) or resort to Bayes factors).
HTH,
Emmanuel Charpentier
On Fri, 23 Mar 2012 12:22:57 +0100, Celine Teplitsky wrote :
> Dear all,
>
> I have seen this post by Arthur Gilmour in an ASReml related forum:
>> GLMM models have only been developed for the case of a single GLMM
>> trait. It is difficult to conceive what is the appropriate error
>> structure for bivariate GLMM traits when the variances are defined as
>> functions of the mean.
>
> I would like to know what people on this list think about these issues
> e.g. if we want to estimate correlations between 2 traits with a poisson
> distribution, are there some special issues to take into account?
>
> Thanks a lot in advance for your help,
>
> All the best
>
> Celine
More information about the R-sig-mixed-models
mailing list