decentralise

Make (some) Stan programs non-centered automatically.

Inspired by reading this paper.

Installation

Requires python 3.6+. Install with pip:

pip install decentralise

Example

This package provides a command line tool: decentralise, which is used as follows:

decentralise <input-file> <output-destination>

It will take any normally distributed parameters that have variances that are themselves parameters, and reparameterise them to make them non-centered. For example, the canonical eight schools code is:

data {
  int<lower=0> J; // number of schools
  vector[J] y; // estimated treatment effects
  vector<lower=0>[J] sigma; // s.e. of effect estimates
}
parameters {
  real mu;
  real<lower=0> tau;
  vector[J] theta;
}
model {
  theta ~ normal(mu, tau);
  y ~ normal(theta, sigma);
}

Once decentralise is applied, this program becomes

data {
  int<lower=0> J; // number of schools
  vector[J] y; // estimated treatment effects
  vector<lower=0>[J] sigma; // s.e. of effect estimates
}
parameters {
  real mu;
  real<lower=0> tau;
  vector[J] theta_std;
}
transformed parameters {
  vector[J] theta = mu + theta_std * tau;
}
model {
  theta_std ~ normal(0, 1);
  y ~ normal(theta, sigma);
}

I wrote this tool for fun, and have not thoroughly tested it. For example, it will fail for any parameters whose sampling statements are not vectorised. e.g.:

x ~ normal(mu, tau);

will work, but

for i in (1:n) {
    x[i] ~ normal(mu, tau);
}

will not. There are probably a ton of other cases where it will break that I have not anticipated, as well!