Search code examples
rdata-structuresstanrstan

How to have a variable in the data block of Stan be an array of length J >= 1?

I'm using the following very simple Stan model saved as model.stan,

data {
  int<lower=1> J;
  real x[J];
}

parameters {
  real mu[J];
  real<lower=0> sigma[J];
}

model {
  sigma ~ inv_gamma(1, 1);
  mu ~ normal(0, 10);
  x ~ normal(mu, sigma);
}

In this model I have a single data point x[j] that I model as coming from J different Normal distributions.

When J > 1, the following R code works perfectly:

library(rstan)
model <- stan_model('~/model.stan')
data <- list(J = J, x = runif(J))
stan.fit <- rstan::sampling(model, data=data)

However, when J = 1 I get the following error:

failed to create the sampler; sampling not done

How can I write this Stan model so it works for all J >= 1?

Solution

  • This is a common problem (and one that is compounded by a bug where the description of the problem is suppressed in rstan 2.18.1). If you declare a real array in the data block of a Stan program, the corresponding R object must have a dimension attribute. Thus,

    stan.fit <- sampling(model, data = list(J = J, x = as.array(runif(J))))

    does run (although there are many divergent transitions) because x has a dim attribute that is J.