rvar conversion drops dimensions

[wds15]

The rvar creation utilities do drop dimensions from the input variables whenever just 1 entry lives in the last index as it seems. Here is a minimal example:

> library(posterior)
+ mu <- rdo(matrix(rnorm(1, mean = 1, sd = 1), 1, 1), ndraws=10)
+ mu
+ mur1 <- as_draws_rvars(as_draws_array(mu))
rvar<10>[1,1] mean ± sd:
     [,1]     
[1,] 0.53 ± 1 
> > as_draws_rvars(as_draws_array(mur1))
# A draws_rvars: 10 iterations, 1 chains, and 1 variables
$x: rvar<10>[1] mean ± sd:
[1] 0.53 ± 1 

You see that mu is a matrix of dimension 1x1, which I can create with rdo. Converting it to an array and then back to an rvar gets rid of the last dimension. This is with posterior 1.3.0.

The offending code in posterior seems to be here:
https://github.com/stan-dev/posterior/blob/master/R/as_draws_rvars.R#L78-L82

This misbehaviour triggers a problem in OncoBayes2 in a special case:

library(OncoBayes2)

groups <- "one_group"
combo_data <- tibble(
  group_id = factor(groups, groups),
  drug_A = 1,
  drug_B = 1,
  num_patients = 3,
  num_toxicities = 1
)

num_groups <- 1

fit <- blrm_exnex(
  cbind(num_toxicities, num_patients - num_toxicities) ~
    1 + I(log(drug_A)) |
    1 + I(log(drug_B)) |
    0 + I(2 * drug_A * drug_B / (1 + drug_A * drug_B)) |
    group_id,
  data = combo_data,
  prior_EX_mu_mean_comp = matrix(
    c(log(1/4), 0,
      log(1/4), 0),
    nrow = 2,
    ncol = 2,
    byrow = TRUE
  ),
  prior_EX_mu_sd_comp = matrix(
    c(1, 0.7,
      1, 0.7),
    nrow = 2,
    ncol = 2,
    byrow = TRUE
  ),
  prior_EX_mu_mean_inter = 0,
  prior_EX_mu_sd_inter = 1,
  prior_EX_tau_mean_comp = matrix(
    c(0, 0),
    nrow = 2,
    ncol = 2
  ),
  prior_EX_tau_sd_comp = matrix(
    c(1, 1),
    nrow = 2,
    ncol = 2
  ),
  prior_EX_tau_mean_inter = matrix(0),
  prior_EX_tau_sd_inter = matrix(1),
  prior_is_EXNEX_comp = c(FALSE, FALSE),
  prior_is_EXNEX_inter = FALSE,
  prior_EX_prob_comp = matrix(1, nrow = num_groups, ncol = 2),
  prior_EX_prob_inter = matrix(1, nrow = num_groups, ncol = 1),
  prior_tau_dist = 0,
  prior_PD = FALSE
)

dd <- posterior_predict(fit)

## Error in aeta[, g, , drop = FALSE] : incorrect number of dimensions

Tagging @mjskay here... seems to be another corner case to catch?

[wds15]

@mjskay would be great if you could have a look at this.... or may I also attempt to fix this issue myself and provide a PR?

[mjskay]

Happy to look, though the absolute earliest would likely be next week or the week after. Alternatively if you wanted to work on a PR that sounds good to me! 😊