Unnecessary variables required in newdata for predictions in projpred

[fweber144]

When performing predictions with new data in projpred, brms reference models require newdata to contain the response (and in fact all variables from the original dataset), which is not necessary after projecting onto a smaller model:

library(brms)
options(mc.cores = parallel::detectCores())
ref_fit <- brm(
  formula = rate ~ conc + state,
  data = Puromycin,
  backend = "cmdstanr",
  seed = 2052109,
  refresh = 0
)

library(projpred)
soltrms <- "conc"
prj <- project(ref_fit, solution_terms = soltrms, ndraws = 40)
Puromycin_new <- as.data.frame(sapply(Puromycin[soltrms], quantile))
prj_linpred <- proj_linpred(prj, newdata = Puromycin_new)

with the last line throwing

Error: Response variables must be specified in 'newdata'.
Missing variables: 'rate'

(Tested with projpred 2.4.0 and brms's GitHub version, commit b31fcee.)

This also came up in The Stan Forums and seems to be a long-standing bug, probably related to the extract_model_data() function defined in brms and passed to projpred::init_refmodel() (the extract_model_data argument was introduced in projpred 2.0.0, so quite some time ago). I will try to fix this and then create a PR.

[paul-buerkner]

thank you for checking this out! Frank Weber ***@***.***> schrieb am Di., 14. Feb. 2023, 13:29:

…

When performing predictions with new data in projpred, brms reference models require newdata to contain the response (and in fact *all* variables from the original dataset), which is not necessary after projecting onto a smaller model: library(brms) options(mc.cores = parallel::detectCores())ref_fit <- brm( formula = rate ~ conc + state, data = Puromycin, backend = "cmdstanr", seed = 2052109, refresh = 0 ) library(projpred)soltrms <- "conc"prj <- project(ref_fit, solution_terms = soltrms, ndraws = 40)Puromycin_new <- as.data.frame(sapply(Puromycin[soltrms], quantile))prj_linpred <- proj_linpred(prj, newdata = Puromycin_new) with the last line throwing Error: Response variables must be specified in 'newdata'. Missing variables: 'rate' (Tested with projpred 2.4.0 and brms's GitHub version, commit b31fcee <b31fcee> .) This also came up in The Stan Forums <https://discourse.mc-stan.org/t/projpred-model-with-horseshoe-prior-does-not-converge/29935> and seems to be a long-standing bug, probably related to the extract_model_data() function defined in brms and passed to projpred::init_refmodel() (the extract_model_data argument was introduced in projpred 2.0.0, so quite some time ago). I will try to fix this and then create a PR. — Reply to this email directly, view it on GitHub <#1457>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADCW2ADPQJ6BHI5UGYZTHPTWXN3BPANCNFSM6AAAAAAU3P5R7U> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>