Feature request: non-binary categorical variables

[jferrerobiol]

Hi @Ax3man!
Thanks for a really useful package! I was wondering if you are planning to provide support for the inclusion of non-binary categorical variables.
Thank you!
Joan

[Ax3man]

Hi Joan!

Unfortunately, I think this is unlikely to happen. Here's why. To perform PPA we need to able to model all variables as both explanatory variables, and as response variable. While we can use the phylolm package to include non-binary categorical variables as explanatory variables, there is no frequentist method (to my knowledge) that allows us to perform phylogenetic regression with more than 2 categories as the response. (Then there's probably also some other issues, like getting single p-values for categorical variables.)

I'm pretty sure one could fit such a model (a model with non-binary categorical outcome) using Bayesian methods though, e.g. using brms/stan or MCMCglmm, but there is (currently) no Bayesian framework for PPA (although such a thing has been rumored for a while now). So maybe one day, but not any time soon.

Cheers,
Wouter

[jferrerobiol]

Thanks for your rapid and detailed answer Wouter!
For the time being I will limit analyses to binary and continuous variables then.
Cheers,
Joan

[MarcRieraDominguez]

Hi! I was having a similar question. I wonder whether we could model categorical variables in PPA in a sort of "dummy-variable" approach. Suppose I want to model the effect of "plant life form" (annual herbaceous, perennial herbaceous, shrub, tree) on "abundance" (continuous). What about splitting this categorical variable into 4 yes/no binary variables. It's not equivalent to a multi-state categorical variable, but maybe it could be a work-around?
Best,
Marc

[Ax3man]

Well, you could fit such a model. But I'm not sure how to interpret it. E.g. how PPA is currently defined, you would "test" the conditional independence between the different dummy variables, but by definition they are not independent.

[MarcRieraDominguez]

I see, it is best then to limit oneself to continuous/binary variables. Thanks! :)