In greta-dev/greta: Simple and Scalable Statistical Modelling in R

Regularized horseshoe prior

The regularized ('Finnish') horseshoe (doi.org/10.1214/17-EJS1337SI) remedies a problem of the original horseshoe: large, unregularized values for the coefficients. This is especially problematic in scenarios where the parameters are only weakly identified by the data, as in logistic regression with perfectly seperable data.

regularized_horseshoe <- function (tau = 1,  c = 1, dim = NULL) {
  stopifnot(c > 0)
  lambda <- cauchy(0, 1, truncation = c(0, Inf), dim = dim)
  lambda_tilde <- (c^2 * lambda^2) / (c^2 + tau^2 * lambda^2)
  sd <- tau ^ 2 * lambda_tilde ^ 2
  normal(0, sd, dim = dim)
}

# variables & priors
int <- variable()
sd <- inverse_gamma(1, 1)
coef <- regularized_horseshoe()

# linear predictor
mu <- int + coef * attitude$complaints

# observation model
distribution(attitude$rating) <- normal(mu, sd)