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

Horseshoe prior

A simple, one-variable Bayesian linear regression model using a horseshoe prior. The horseshoe, just as the LASSO, can be used when the slopes are assumed to be sparse. According to the original publication:

its flat, Cauchy-like tails allow strong signals to remain large [...] a posteriori. Yet its infinitely tall spike at the origin provides severe shrinkage for the zero elements

horseshoe <- function (tau = 1, dim = NULL) {
  lambda <- cauchy(0, 1, truncation = c(0, Inf), dim = dim)
  sd <- tau ^ 2 * lambda ^ 2
  normal(0, sd, dim = dim)
}

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

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

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