Default weakly informative priors for BFs in generalised linear mixed model (log link)

Mattscott92

Hi there,

I'm currently trying to compute Bayes factors for a generalised linear mixed model with logistic link function. I plan to compare my final model to reduced models to obtain BFs for my main fixed and random effects. I'd like to be as 'objective' as possible and have used the default priors in Rouder et al. (2012) for linear models to good effect. I'm struggling to find a similarly 'objective' approach to specifying my priors for this type of model. 

I've thought of adopting the default 'weakly informative' priors within the rstanarm package, and I know this is the approach also used in JASP for Bayesian GLMMs. However, as these are autoscaled and optimised for analysis of posterior distributions I'm unsure if this would be appropriate for computing BFs. To be honest I'm struggling to get my head around how this autoscaling function works!

 My question is, is there any consensus on default priors for these models?

Thanks a lot, and all the best,

Matt

EJ

Hi Matt,

You are correct that these priors are not appropriate for BFs. We have not developed this. I recall a paper by Overstall & Forster on Bayes factors for GLMMs, where they employ a unit-information prior. That is an appealing idea.

Cheers,

E.J.