Bayesian Mixed Models - Prior Specification and Bayes Factors

Whirly123

I have been running Bayesian Logistic Mixed Models for model comparison in R for a current project and was very excited to see these are now implemented in JASP.

1. A couple of questions: What are the default priors used here? Are they adjustable in any way (or will they be in due time)
2. Since I am wanting to do model comparison I will need to calculate Bayes Factors but that doesn't seem to be an option yet in the current implementation. Will this be possible in the future or …
3. I think I read that this is doable with the Bayesian Anova implementation (hypothetically, if this was linear rather than generalised), but there doesn't seem to be a way to specify things like random slopes for specific factors.

Thanks, JASP's development is looking very cool!

EJ

Hi Whirly123,

1. I'll ask our expert.
2. BF are our goal for the future, but it requires additional work so we wanted to provide the estimation framework now.
3. Yes, the ANOVA is linear, not generalized. It was implemented as a linear mixed model (in the R package BayesFactor), but I'll ask whether this is possible in JASP.

Cheers,

E.J.

JohnnyB

Hi Whirly 123,

As for point 3, this is not yet possible to do within JASP given its current interface. In the BayesFactor package, it is possible, ideally when there are multiple observations per person per condition (so multiple items or trials, for instance). The JASP RM ANOVA for now only includes random intercepts (this is default behaviour for Bayesian RM ANOVA) and no random slopes, but we are working on a project that will hopefully lead to this feature being implemented in JASP some time soon.

Cheers

Johnny

Whirly123

Understood! And great to hear :) Cheers!

František

Hi Whirly123,

I will try to answer the 1st question. As of now, we are using the default rstanarm priors that are autoscaled according to the predictors and response (normal distributions) and that are designed to be weakly informative - in order to provide moderate regularization and help stabilize computation. We are planning to allow users to specify priors, but that is still under development.

Cheers,

Frantisek