Bayes Factors for Linear Mixed Models in JASP
Dear all,
Tl;dr: Is it possible to get (or calculate/derive) Bayes Factors for linear mixed models, for planned contrasts or at all?
The output seems to provide 95% credible intervals for the highest posterior density, but no BF.
A little more detail:
I’m running a mixed model with Condition (2 levels) and Group (3 levels) as fixed factors, and Subject as a random factor.
I would like to unpack the Group*Condition interaction to compare the Condition effect between groups. Essentially, the difference between the conditions for Group A is different from that for Group B and from the condition difference for Group C, while the latter two do not significantly differ from one another. (says the frequentist analysis)
I was hoping to use the Bayesian mixed models to obtain something akin to a Bayes factor for these three contrasts, that could indicate the strength of evidence for these three comparisons. Especially for the non-difference between groups B and C it would be nice to have an indication of how robust this non-difference is. (e.g. can I state that they are the same, or should I say that the study was underpowered to detect the difference).
I have specified the three contrasts and for these I get 95% credible intervals. Can I somehow use the credible intervals to arrive a conclusion about the evidence for H0? Alternatively, there are some fit indices (WAIS, LOO) for the entire model; could I somehow run different models and compare based on those?
Many thanks for your help,
Elise
Comments
Hi Elise,
This is presently not possible within the new mixed model functionality. However, the Bayesian ANOVA in JASP is based on the BayesFactor package in R, and this implements a mixed model. So it should be possible, but honestly I haven't checked this. I will ask around.
Cheers,
E.J.
I would be very much interested in this too. Currently I am not sure how to interpret the results of the bayesian linear mixed models analysis.
Thanks for your work!
Hi all, any latest updates on BF for mixed(Multilevel model, MLM) model?
For Randomised Clinical Trial with longitudinal design including baseline and followups (continuous outcomes), the European Medicine agency recommend quantifying the treatment effects using ANCOVA models. due to the missingness, Multilevel modelling (MLM, mixed model called here) was regarded as a better approach to analyze data as repeated ANOVA will remove the case with missingness. IT will be great to have BF output for each margin and contrast so the Bayesian parameter and hypothesis test could from one model which accounts for all data information simultaneously. thanks for your attention. Boliang