[avehtari]

The projpred PR stan-dev/projpred#591 adds support for monotonic terms mo().
This PR lets get_refmodel.brmsfit() pass mo() terms to projpred.

If the formula has mo() terms, get_refmodel.brmsfit() checks that projpred exports mo_effects(), which the projpred PR adds. The check does not use a version number, so development versions of projpred with this support also work.

This PR was assisted by Claude.

[paul-buerkner]

Thanks! I had one question regarding the added validation checks

[paul-buerkner]

You seem to only check for dpars$mu here. Are we, at this point of the code, sure that we only target mu with projpred?

[avehtari]

Well spotted. The checks looked only at mu, and some models have no mu. get_refmodel.brmsfit() stops earlier for distributional models, except for the families with conv_cats_dpars() (categorical, multinomial and the simplex families). In these models the dpars are mu2, mu3, ..., and there is no mu. With 3 or more categories, bterms$dpars$mu is an ambiguous partial match and gives NULL, so the term-type check and the sp-term check did nothing. With 2 categories, $mu partially matched mu2, and the checks gave the correct result. A categorical model with me() or gp() passed the checks and then failed with no applicable method for `@` applied to an object of class "NULL".

New commit uses get_effect(bterms, target = ...) in the term-type check and the sp-term check, so they look at the terms of all dpars. The commit adds categorical models with me() and gp() to the monotonic-terms test in tests/local/tests.models-5.R. It also removes the second require_package("projpred"), because get_refmodel.brmsfit() calls it at the start.

With the default latent = FALSE, projpred uses the augmented-data projection for categorical models, also with mo() terms. The monotonic projection in projpred (stan-dev/projpred#591, the default) does not support the augmented-data projection yet. It stops with a message that says to use options(projpred.mo_monotone = FALSE). With that option, project() and varsel() work for a categorical model with mo(x).

[paul-buerkner]

Looks good now. Thank you!