Skip to content
Prev 15016 / 20628 Next

MCMCglmm phylogenetically controlled categorical R structure and priors help

Hi Ben,

If each species can only belong to one category (?) the choice of 
R-structure is arbitrary since observation-level variation is not 
identifiable in the likelihood.

You probably want to add trait as a main effect, and interact it with 
your other predictors too. Currently the model assumes that the 
probability of  being Bicoloured/Mottled/Plain compared to the base-line 
category (not sure what that is) is the same and that they are all 
affected by diet in the same way. This may make sense depending on what 
the base-line category is, but I doubt it.

The phylogenetic variance looks large, but it is hard to say because you 
have specified scale=FALSE. If it really is large it is possible that 
you will run into numerical errors. If you have pl=TRUE in your call to 
MCMCglmm, and then do a histogram of my_model$Liab, make sure the 
distribution is contained within +/-25. If this is not so then you will 
have to use software that assumes the phylogenetic heritability is 1. I 
have developed algorithms under this scenario but only for ordered 
categorical traits.

I would just use the ginverse argument (the pedigree argument is 
redundant) and use the default scale=TRUE and  nodes="ALL" not 
nodes="TIPS". The model is the same, but the former allows sparse linear 
solvers to exploit the special sparse structure of phylogenies. It will 
probably be orders of magnitude faster for large phylogenies.

Note that this type of analysis is VERY data demanding and you will need 
many hundreds of species to get precise estimates, particularly if some 
wing pattern types are rare. If the data set is modest in size expect 
the results to be sensitive to your choice of prior for the phylogenetic 
species effects.

Cheers,

Jarrod
On 14/11/2016 23:48, ben hogan wrote: