R/bootstrap.R

In incidence: Compute, Handle, Plot and Model Incidence of Dated Events

#' Bootstrap incidence time series
#'
#' This function can be used to bootstrap `incidence` objects. Bootstrapping is
#' done by sampling with replacement the original input dates. See `details` for
#' more information on how this is implemented.
#'
#' @author Thibaut Jombart \email{thibautjombart@@gmail.com}
#'
#' @md
#'
#' @export
#'
#' @details As original data are not stored in `incidence` objects, the
#'   bootstrapping is achieved by multinomial sampling of date bins weighted by
#'   their relative incidence.
#'
#' @param x An `incidence` object.
#'
#' @param randomise_groups A `logical` indicating whether groups should be
#'   randomised as well in the resampling procedure; respective group sizes will
#'   be preserved, but this can be used to remove any group-specific temporal
#'   dynamics. If `FALSE` (default), data are resampled within groups.
#'
#' @return An `incidence` object.
#'
#' @seealso [incidence::find_peak] to use estimate peak date using bootstrap
#'
#' @examples
#'
#' if (require(outbreaks) && require(ggplot2)) { withAutoprint({
#'   i <- incidence(fluH7N9_china_2013$date_of_onset)
#'   i
#'   plot(i)
#'
#'   ## one simple bootstrap
#'   x <- bootstrap(i)
#'   x
#'   plot(x)
#'
#' })}
#'

bootstrap <- function(x, randomise_groups = FALSE) {
  if (!inherits(x, "incidence")) {
    stop("x is not an incidence object")
  }

  ## `counts` is a vector of event counts, meant to be a column of x$counts
  boot_one_group <- function(counts) {
    sample_(x$dates, size = sum(counts), replace = TRUE, prob = counts)
  }

  new_dates <- do.call(c,
                       lapply(seq.int(ncol(x$counts)),
                              function(i) boot_one_group(x$counts[, i])))
  group_sizes <- colSums(x$counts)
  new_groups <- rep(colnames(x$counts), group_sizes)

  if (randomise_groups) {
    new_groups <- sample_(new_groups)
  }

  incidence(new_dates, interval = x$interval, groups = new_groups)

}