Inconsistency in `standata()` output between `resp_weights()` and `resp_trials()`

[fweber144]

I noticed an inconsistency in the standata() output between resp_weights() and resp_trials() if the data variable containing the weights is missing:

library(brms)

# Gaussian with resp_weights():
Puromycin$wobs <- rep_len(5:3, nrow(Puromycin))
bfit_gau <- brm(
  formula = rate | weights(wobs) ~ conc + state,
  data = Puromycin,
  chains = 1,
  iter = 500,
  seed = 2052109,
  refresh = 0
)
Puromycin_new <- head(Puromycin, 3)
Puromycin_new$wobs <- NULL
sdata_gau <- standata(bfit_gau, newdata = Puromycin_new)
## --> Runs through and gives weights which are constantly 1:
all(sdata_gau$weights == 1)
## --> TRUE

# Binomial with resp_trials():
N_bin <- 100
data_bin <- data.frame(wobs = rep_len(5:3, N_bin))
set.seed(57346)
data_bin$success <- rbinom(N_bin, size = data_bin$wobs, prob = 0.4)
bfit_bin <- brm(
  formula = success | trials(wobs) ~ 1,
  data = data_bin,
  family = binomial(),
  chains = 1,
  iter = 500,
  seed = 2052109,
  refresh = 0
)
data_bin_new <- head(data_bin, 3)
data_bin_new$wobs <- NULL
sdata_bin <- standata(bfit_bin, newdata = data_bin_new)
## --> Error: The following variables can neither be found in 'data' nor in 'data2':
##     'wobs'

Perhaps this is intentioned or not avoidable, but I thought it would be better to report. I guess the error thrown in the resp_trials() case is the safer solution compared to the silently used vector of 1s.

[paul-buerkner]

Weights are largely ignored in post-processing as they are relevant only during inference (and log_lik computations). Thus, valid_newdata currently does not care if weights are provided or not. The number of trials in binomial models always affects predictions and thus needs to be present. Does this answer your question?

[fweber144]

Ok, I was suspecting this to be rather a design issue, so I'm fine with leaving this as-is.

[paul-buerkner]

Okay. Thank you checking in about this!