[Feature request] Multivariate Laplace as random effect distribution?

[trinhdhk]

Albeit unofficial, brms supports Multivariate Student T as a random effect distributional family. Please, is it possible to add Multivariate Laplace to a future version, as the pdf is already well-defined in R [1]? My guess is that it might be a bit slow using this implementation as it involves the calculation of $\Sigma$ , but maybe there is a different formulation elsewhere.

Laplace is nice to have, especially when one wants to check if some groups diverge from the normative curve.

[1] https://search.r-project.org/CRAN/refmans/LaplacesDemon/html/dist.Multivariate.Laplace.html

Edit: A second look, it may not involve $\Sigma$ after all, since every $\Sigma$ is encapsulated inside a square root, so theoretically we can get a Cholesky decomposition, but I don't have the math capability to factorise it out.

[paul-buerkner]

It should be possible to add it. What are the advantages of using the MV laplace distribution (vs. normal or student-t) for random effects?

[trinhdhk]

It should be possible to add it. What are the advantages of using the MV laplace distribution (vs. normal or student-t) for random effects?

While Student's T captures over-dispersion, Laplace can do stronger shrinkage. In the same manner as variable selection, Laplace can be more powerful than Normal in detecting which group-specific random effect diverge from the average. It might not be relevant when looking at marginal effects but is extremely useful when we want to detect which particular, say, region of the brain, the disease is targeting, in one model.

[paul-buerkner]

got it thanks! Trinh Dong ***@***.***> schrieb am Mo., 13. Jan. 2025, 22:36:

…

It should be possible to add it. What are the advantages of using the MV laplace distribution (vs. normal or student-t) for random effects? While Student's T captures over-dispersion, Laplace can do stronger shrinkage. In the same manner as variable selection, Laplace can be more powerful than Normal in detecting which group-specific random effect diverge from the average. It might not be relevant when looking at marginal effects but is extremely useful when we want to detect which particular, say, region of the brain, the disease is targeting, in one model. — Reply to this email directly, view it on GitHub <#1723 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADCW2AAUBXYQUCRMRYI643T2KQWVFAVCNFSM6AAAAABVCZQ5W2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKOBYGI3DSMJRGU> . You are receiving this because you commented.Message ID: ***@***.***>