R/generate_data.R

In inbo/inlatools: Diagnostic Tools for INLA Models

#' Generate dummy data with several distributions
#'
#' All distributions share the same latent variable \eqn{\eta_{ij} = a + b_i}
#' with \eqn{b_i = N(0, \sigma_r)}
#'
#' - The Poisson distribution uses \eqn{\lambda = e^{\eta_{ij}}}
#' - The negation binomial distribution uses \eqn{\mu = e^{\eta_{ij}}}
#' - The binomial distribution uses \eqn{\pi_{ij} =
#' e^{\eta_{ij}}/(e^{\eta_{ij}}+ 1)}
#' @param a the intercept of the latent variable
#' @param sigma_random The standard error for the random effect \eqn{\sigma_r}
#' @param n_random the number of random effect levels (groups)
#' @param n_replicate the number of observation per random effect level
#' @param nb_size the size parameter of the negative binomial distribution.
#' Passed to the `size` parameter of \code{\link[stats]{NegBinomial}}
#' @param b_size the size parameter of the binomial distribution. Passed to the
#' `size` parameter of `\link[stats]{Binomial}`
#' @param zero_inflation the probability the the observed value stems for the a
#' point mass in zero
#' @return A `data.frame`
#'
#' - `ìd` the id of the random effect
#' - `eta` the latent variable
#' - `zero_inflation` use the point mass in zero
#' - `poisson` the Poisson distributed variable
#' - `zipoisson` the zero-inflated Poisson distributed variable
#' - `negbin` the negative binomial distributed variable
#' - `zinegbin` the zero-inflated negative binomial distributed variable
#' - `binom` the binomial distributed variable
#' @importFrom assertthat assert_that is.number is.count
#' @importFrom dplyr %>% mutate n
#' @importFrom rlang .data
#' @importFrom stats rnorm rpois rnbinom rbinom plogis
#' @export
#' @family utils
#' @examples
#' set.seed(20181202)
#' head(generate_data())
generate_data <- function(
  a = 0,
  sigma_random = 0.5,
  n_random = 20,
  n_replicate = 10,
  nb_size = 1,
  b_size = 5,
  zero_inflation = 0.5
) {
  assert_that(is.number(a))
  assert_that(
    is.number(sigma_random),
    sigma_random > 0
  )
  assert_that(is.count(n_random))
  assert_that(is.count(n_replicate))
  assert_that(
    is.number(nb_size),
    nb_size > 0
  )
  assert_that(is.count(b_size))

  re_random <- rnorm(n_random, mean = 0, sd = sigma_random)
  data.frame(
    group_id = rep(seq_len(n_random), n_replicate)
  ) %>%
    mutate(
      observation_id = seq_along(.data$group_id),
      eta = a + re_random[.data$group_id],
      zero_inflation = rbinom(n(), size = 1, prob = zero_inflation) %>%
        as.logical(),
      poisson = rpois(n(), lambda = exp(.data$eta)),
      zipoisson = ifelse(.data$zero_inflation, 0, .data$poisson),
      negbin = rnbinom(n(), size = nb_size, mu = exp(.data$eta)),
      zinegbin = ifelse(.data$zero_inflation, 0, .data$negbin),
      binom = rbinom(n(), size = b_size, prob = plogis(.data$eta))
    )
}