"No data for the selected variables." with optional variables

[mhollanders]

Hi,

I've got a model with lots of optional variables but SBC throws an error when I try to plot their ranks. Here's a reproducible example with the following Stan program called test.stan.

data {
  int N, indicator;
  array[N] real y;
}

parameters {
  vector[indicator ? 1 : 0] mu1;
  vector[indicator ? 0 : 1] mu2;
  real<lower=0> sigma;
}

model {
  if (indicator) {
    target += normal_lpdf(y | mu1[1], sigma);
  } else {
    target += normal_lpdf(y | mu2[1], sigma);
  }
}

And the following R script reproduces the error:

# packages
library(cmdstanr)
library(SBC)

# compile model
mod <- cmdstan_model("test.stan")

# generator function and data
generator <- function(N, indicator) {
  mu <- rnorm(2)
  sigma <- rexp(1)
  if (indicator) {
    y <- rnorm(N, mu[1], sigma)
  } else {
    y <- rnorm(N, mu[2], sigma)
  }
  list(variables = list(mu1 = mu[1],
                        mu2 = mu[2],
                        sigma = sigma),
       generated = list(N = N,
                        indicator = indicator,
                        y = y))
}
gen <- SBC_generator_function(generator, N = 100, indicator = 1)
n_sim <- 10
data <- generate_datasets(gen, n_sim)

# backend and SBC
backend <- SBC_backend_cmdstan_sample(mod, chains = 3, 
                                         iter_warmup = 100, iter_sampling = 100)
sbc <- compute_SBC(data, backend)

# rank errors
plot_rank_hist(sbc)

Thanks for taking a look!

[mhollanders]

Update: this is due to some indexing stuff that occurs with vectors created in R and Stan. If I explicitly use the right indexing there's no problem. My apologies.

[martinmodrak]

Hi,
no problem - this is in fact related to issue stan-dev/posterior#187 - R does not distinguish between scalars and vectors of length 1, while Stan does.

I've recently made some improvements (still experimental) to the package, that once merged should allow for better handling of those situations (e.g. let you explicitly say that a variable is array, so it won't be treated as scalar, even if it is length 1 and also let you say that a variable is optional and should not be expected to be present in all fits).

[mhollanders]

Good stuff Martin. Also congrats to you and everyone else on making such a useful package. It's really a joy to use.