Consider implementing covariance regressions

[shannonajw]

Distributional models are already well-implemented in brms (models allowing prediction of mean and variance parameters). Multivariate models are also well-implemented. It would be great if brms could also make it easy to implement covariance regressions, such as in this paper: https://doi.org/10.1177/0049124119882449. Example Stan code (for two outcomes) is provided by the authors in the Supplemental Material section (https://journals.sagepub.com/doi/suppl/10.1177/0049124119882449).

[paul-buerkner]

So this effectively means predicting (residual) correlations?

[shannonajw]

Yes, that's right.

[shannonajw]

Not sure how feasible this is (or if it is already implemented), but on top of modeling the residuals, it may also be worth considering extending this to mixed models - where the variance of the random effects (and possibly even the covariances) can be modeled as a function of covariates. See https://www.jstatsoft.org/article/view/v052i12 for an implementation in R.

Thanks again for considering these features!

[tfjaeger]

Paul, have you by chance started on this? I'm also working on a research project that would reduce dozens of analyses into one if that feature was added (assessing differences in talker-level statistics of phonetic categories, which draws brms's multivariate and distributional model abilities). I'm wondering whether I should look into the articles linked here or whether this will be a feature 'soon' to be released in brms.

(How you manage to do all of these things is beyond me! The conceptual clarity of brms is truly amazing.)

[paul-buerkner]

I have not and it will take a few months until I can start working on this unfortunately.

[tfjaeger]

Thanks for the quick response!

[paul-buerkner]

This seems to be effectively a duplicate of issue #254. Closing this issue here.