[R-sig-ME] Proportion response data in MCMCglmm (or INLA) phylogenetic mixed model

Pierre de Villemereuil p|erre@dev|||emereu|| @end|ng |rom ephe@p@|@eu
Fri Aug 20 13:57:05 CEST 2021


If you want to fit a phylogenetic mixed model with a beta distribution, you can use brms, which allows for both features:


Le mercredi 18 août 2021, 12:51:42 CEST r_1470 via R-sig-mixed-models a écrit :
> Hi all,
> I'm attempting to run a phylogenetic comparative analysis for which the response variable is a proportion, and I want to estimate the effect of one or more predictors while controlling for phylogenetic non-independence. I would like to use MCMCglmm, and the basic model would be (ignoring the 'family' argument for now):
> MCMCglmm(phenotype ~ x, random=~taxon + id + species, ginverse=list(taxon=Ainv),data=data, prior=prior)
> 'x' is an experimental treatment, but a more complex fixed-effects model with factors and continuous covariates is likely to be used. There's both a pre- and a post-treatment measure for most individuals ('id' refers to individual), and multiple individuals per species in many cases ('species' is intended to estimate non-phylogenetic among-species variation).
> However, MCMCglmm doesn't include beta family models. My question is: can I usefully run this analysis in MCMCglmm and interpret the coefficients, either by assuming gaussian family, or by logit-transforming the proportions, or by using another family? Most of the proportions are intermediate, which might be helpful, although there are a fair number between 0.1 - 0.01.
> I tried to replicate the model formula in other packages that allow beta-distributed variables (although I simulated data with gaussian errors for now, for direct comparison with MCMCglmm), but so far have been unable to either get reliable convergence (brms) or produce even qualitatively similar estimates on simulated datasets (INLA and phyr with bayes=T, although I only compared many datasets with INLA). For INLA, among other attempts I copied one of the model formulae used by animalINLA (the one using model="generic2"), but given the vastly different results from MCMCglmm for both the phylogenetic effect and the fixed effect, I'm left with the impression that INLA is doing something conceptually different to MCMCglmm.
> For INLA, I used the inverse matrix produced by MCMCglmm::inverseA, but only including tips and with rownames 1:N to match a numeric 'taxonN' variable in the data table.
> Back to my basic question: can I usefully analyse my proportion data with MCMCglmm?
> Best wishes,
> Richard.
> 	[[alternative HTML version deleted]]
> _______________________________________________
> R-sig-mixed-models using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models

More information about the R-sig-mixed-models mailing list