Create priors on intercepts for cumulative brms ordinal regression model

cumulative-intercept-prior.R

cumulative_intercept_prior <- function(k, sd = 2, alpha = 1, beta = 1,
                                      shape = c("flat", "middle", "rightskewed", "leftskewed")) {
  ## Creates priors on intercepts for cumulative() family regression.
  ## Assumes that probability of response options follow cumulative beta 
  ## distribution specified by a and b or by "shape" argument.
  ##
  ## k = number of categories
  ## sd = std dev of normal over intercept
  ## a, b = alpha and beta specifying shape of distribution (defaults to uniform)
  ## shape = string specifying pre-defined distribution shape 
  shapes <- list("rightskewed" = c(2, 4), 
                 "leftskewed" = c(4, 2), 
                 "middle" = c(3, 3), 
                 "flat" = c(1, 1))

  if (length(shape) == 1) {
    alpha <- shapes[[shape]][1]
    beta <- shapes[[shape]][2]
  }

  intercepts <- seq(1, k - 1)
  prior_list <- lapply(intercepts, function(x) {
    center <- qlogis(pbeta(x / k, alpha, beta))
    p <- paste0("normal(", center, ",", sd, ")")

    return(set_prior(p, class = "Intercept", coef = as.character(x)))
  })

  return(Reduce(c,prior_list))
}

Generates priors (in the form of brmsprior object) on intercept terms for ordinal regressions using cumulative family and default logit link. For use with brms package.

To source in a script, run:

devtools::source_gist(id = "4880591f088623ee17d9637eed3aff1f", filename="cumulative-intercept-prior.R")