prior_summary() incorrect after updating a cumulative() fit using fewer response categories

[fweber144]

When update()-ing a cumulative() fit with a new dataset that contains fewer response categories than the original one, prior_summary() still displays all the thresholds from the initial model fit:

options(mc.cores = parallel::detectCores(logical = FALSE))

# Data --------------------------------------------------------------------

nobsv <- 150L
ncat <- 4L
yunq <- paste0("ycat", seq_len(ncat))
nthres <- ncat - 1L
thres <- seq(-1, 1, length.out = nthres)
npreds_cont <- 5L
coefs_cont <- seq(-4, 4, length.out = npreds_cont)
set.seed(3738)
dat_sim <- matrix(rnorm(nobsv * npreds_cont, sd = 0.5), nrow = nobsv,
                  dimnames = list(NULL, paste0("Xcont", seq_len(npreds_cont))))
eta <- drop(dat_sim %*% coefs_cont)
dat_sim <- as.data.frame(dat_sim)
thres_eta <- sapply(thres, function(thres_k) {
  thres_k - eta
})
yprobs <- brms:::inv_link_cumulative(thres_eta, link = "logit")
dat_sim$Y <- factor(sapply(seq_len(nobsv), function(i) {
  sample(yunq, size = 1, replace = TRUE, prob = yprobs[i, ])
}), ordered = TRUE)

# Model fit ---------------------------------------------------------------

stopifnot(identical(levels(dat_sim$Y), yunq))
bfit <- brms::brm(
  formula = Y ~ Xcont1 + Xcont2 + Xcont3 + Xcont4 + Xcont5,
  data = dat_sim,
  family = brms::cumulative(),
  prior = brms::prior(normal(0, 1), class = "Intercept"),
  seed = 878232
)
brms::prior_summary(bfit)

## Update model fit -------------------------------------------------------

# Update using a dataset which lacks a response category:
dat_upd <- dat_sim[dat_sim$Y != "ycat3", , drop = FALSE]
dat_upd$Y <- droplevels(dat_upd$Y)
stopifnot(identical(levels(dat_upd$Y), setdiff(yunq, "ycat3")))
bupd <- update(
  bfit,
  formula. = Y ~ Xcont1 + Xcont2 + Xcont3 + Xcont4 + Xcont5,
  newdata = dat_upd,
  ### Actually not needed (only added to check if this makes a difference, but
  ### it doesn't):
  prior = brms::prior(normal(0, 1), class = "Intercept"),
  ###
  seed = 54543
)
brms::prior_summary(bupd)

where the last line gives

       prior     class   coef group resp dpar nlpar lb ub       source
      (flat)         b                                         default
      (flat)         b Xcont1                             (vectorized)
      (flat)         b Xcont2                             (vectorized)
      (flat)         b Xcont3                             (vectorized)
      (flat)         b Xcont4                             (vectorized)
      (flat)         b Xcont5                             (vectorized)
normal(0, 1) Intercept                                            user
normal(0, 1) Intercept      1                             (vectorized)
normal(0, 1) Intercept      2                             (vectorized)
normal(0, 1) Intercept      3                             (vectorized)

just like for the initial model fit.

I tested this with brms v2.17.0. As shown in the update() call above, I tried to specify argument prior explicitly a second time, but that didn't help.

Apart from this prior_summary() thing, all other post-processing functions seem to work correctly, at least as far as I have checked them.

[paul-buerkner]

This is because priors are not updated automatically by update in the sense that they are not recomputed from newdata. This of course only concerns the priors that are data-dependent. I will see what options we have in the case you presented.

[paul-buerkner]

Thank you for opening this issue. Should now be fixed.

[fweber144]

Yes, thank you. The example from above works now as expected.