Using projpred with monotonic variables

I am part of a research team studying the amount of fish caught in UK seawaters. I have some ordered variables in my model, fitted with brms support for monotonic effects. I have found projpred useful for other projects and would like to use it here too.

I see there are ~year-old discussions scattered through the forum where @avehtari has suggested support for monotonic effects in projpred may be on the horizon. I would love to know: has there been any development with projpred supporting monotonic variables?

Refactored projpred is hopefully soon ready, which may make it easier to add what you are asking. Pinging @AlejandroCatalina and you can discuss more how easy or difficult it could be to add.


Indeed, we are now finishing our work in the refactoring of projpred and should be ready soon. This refactoring would make it reasonably easy to add support for new optimization methods, such as required for GAMs and ultimately for monotonic effects.

I myself haven’t worked with monotonic effects before but I believe the only missing component right now is a maximum likelihood estimation method that supports monotonic effects estimation. In the worst case we could find maximum a posteriori estimates with brms, although that would incur in some extra cost for compiling the model and such. We haven’t worked in this yet but it is not too far in our horizon as we will add support for GAMs first.

Hope this helps.


Great! It sounds like I will have to leave projpred for this current project but hopefully when I’m at this point next time, the refactored projpred will be released and I can work out a way to incorporate group and monotonic effects. Looking forward to it :)

Thanks for your work!