.extract_model_data(): Warn that arguments wrhs and orhs are ignored

[fweber144]

This PR adds warnings that arguments wrhs and orhs are currently ignored by .extract_model_data() (the workhorse function for extract_model_data() defined in get_refmodel.brmsfit()). This is relevant for projpred::proj_linpred(), projpred::proj_predict(), and projpred:::predict.refmodel(): When using an rstanarm reference model, it would be possible to specify weights and offsets in these 3 functions via arguments weightsnew and offsetnew, respectively. When using a brms reference model, these arguments are ignored (weights and offsets are fetched from newdata, as brms does at other places, too). So far, this has only been documented at ?get_refmodel.brmsfit. With this PR, warnings are thrown to increase awareness.

Illustration:

data("df_gaussian", package = "projpred")

df_gaussian_new <- df_gaussian[30:33, ]
dat_new <- data.frame(y = df_gaussian_new$y, df_gaussian_new$x)
dat_new$wcol <- rep_len(6:5, nrow(dat_new))
dat_new$ocol <- dat_new$X6

df_gaussian <- df_gaussian[1:29, ]
dat <- data.frame(y = df_gaussian$y, df_gaussian$x)
dat$wcol <- rep_len(4:1, nrow(dat))
dat$ocol <- dat$X6

devtools::load_all() # use `library(brms)` if PR version is installed
bfit <- brm(
  y | weights(wcol) ~ X1 + X2 + X3 + X4 + X5 + offset(ocol), data = dat,
  chains = 1, iter = 500, seed = 7403, refresh = 0
)

library(projpred)
prj <- project(bfit, predictor_terms = paste0("X", 1:3), ndraws = 25,
               verbose = FALSE, seed = 4678)

# Intended usage (no warnings):
pln <- proj_linpred(prj, newdata = dat_new)
# Unintended usage (for a brms reference model, not for an rstanarm reference
# model); throws warnings:
pln_vec <- proj_linpred(prj, newdata = dat_new,
                        weightsnew = dat_new$wcol,
                        offsetnew = dat_new$ocol)
pln_fml <- proj_linpred(prj, newdata = dat_new,
                        weightsnew = ~ wcol,
                        offsetnew = ~ ocol)

When running the tests with this PR, I received the following error, which doesn't seem to be related to projpred and this PR, though:

* checking tests ... [248s/248s] ERROR
  Running ‘testthat.R’ [248s/248s]
Running the tests in ‘tests/testthat.R’ failed.
Last 13 lines of output:
  
  ══ Failed tests ════════════════════════════════════════════════════════════════
  ── Error ('tests.make_stancode.R:2471:3'): threaded Stan code is correct ───────
  Error: Syntax error found! See the message above for more information.
  Backtrace:
      ▆
   1. └─brms::make_stancode(bform, dat, family = student(), threads = threads) at tests.make_stancode.R:2471:3
   2.   └─brms:::.make_stancode(...)
   3.     └─brms:::parse_model(scode, backend, silent = silent)
   4.       └─brms (local) .parse_model(model, ...)
   5.         └─out$check_syntax(quiet = TRUE)
  
  [ FAIL 1 | WARN 0 | SKIP 6 | PASS 3017 ]
  Error: Test failures
  Execution halted

[fweber144]

The conditioning on projpred version >= 2.8.0 in line https://github.com/fweber144/brms/blob/725590767c05711f4aa36412f348dc53eb53a3c8/R/projpred.R#L249 is done because in version 2.8.0 (more precisely, stan-dev/projpred#486), some adjustments were performed that allow brms to throw the warnings correctly.

[paul-buerkner]

Thanks!