Variable selection in Brms

Continuing the discussion from Projective variable selection for an Ordered Logistic model, I would like to know if there is the opportunity to perform it in brms whit a model like this:

fit3 <- brm(SCORE ~ ., data = df, family = cumulative())

SCORE is an ordered factor with 4 levels, nrow(df) ≈ 100, ncol(df) ≈ 30 (and represents the total number of variables from which I would like to extract the most influential ones.

brms version: 2.9.3

Not yet, but we are actively working on extending projpred in various directions and hopefully we will eventually have a solution for ordinal models as well.

1 Like

Thanks @paul.buerkner, I’m happy to hear that

@paul.buerkner do you have anything in the pipeline regarding this feature for projpred?

Not yet. This is because the cumulative distribution is not from the exponential family which makes things much harder.

That said, improved projpred for multilevel models (and certain families) will soon be entering beta developmenet and we will ask users for feedback.


I tag @jpiironen because I am still very interested in the evolution of this feature. I see in github that there is buzz around supporting HM. There are plan to support cumulative likelyhood and Robust Regression with student-t likelihood?

@jpiironen has moved to a company and doesn’t have much time for developing projpred and most of the code development is now made by @AlejandroCatalina and @paul.buerkner, and right now your best hope for these specific features is to ask @paul.buerkner and me :)

That’s right. I still follow what’s going on but don’t have much time to participate in the development unfortunately.


thanks @avehtari and @jpiironen for clarifying! I hadn’t noticed the “handover” in the development. What I can say is that Stan, LOOCV, and projection predictive feature selection (projpred) made the difference (for the better) in my work.
So, @jpiironen, @avehtari and @paul.buerkner, thank you all for all the work done, and please keep on rocking!

1 Like

I would be happy to learn more if you have something you can share!

I will be very happy to share my findings, as soon as I will have “edible” results.
At the moment I am involved in two investigations: one is related to variable selection in observational studies, the other is long term (and I still have problems in understanding how to formalize it) is GP with non-gaussian outcomes. As usual, when I will have a solution - I will post my findings like in this case

1 Like