Revision 1b89ec86b6a914c4d45bee998d08eed1ef23f57f authored by Dominique Makowski on 20 July 2020, 08:30:03 UTC, committed by cran-robot on 20 July 2020, 08:30:03 UTC
1 parent d2eac42
Raw File
weighted_posteriors.R
#' Generate posterior distributions weighted across models
#'
#' Extract posterior samples of parameters, weighted across models.
#' Weighting is done by comparing posterior model probabilities, via \code{\link{bayesfactor_models}}.
#'
#' @param missing An optional numeric value to use if a model does not contain a parameter that appears in other models. Defaults to 0.
#' @param prior_odds Optional vector of prior odds for the models compared to the first model (or the denominator, for \code{BFBayesFactor} objects). For \code{data.frame}s, this will be used as the basis of weighting.
#' @param iterations For \code{BayesFactor} models, how many posterior samples to draw.
#' @inheritParams bayesfactor_models
#' @inheritParams bayesfactor_parameters
#'
#' @details
#' Note that across models some parameters might play different roles. For example,
#' the parameter \code{A} plays a different role in the model \code{Y ~ A + B} (where it is a main effect)
#' than it does in the model \code{Y ~ A + B + A:B} (where it is a simple effect). In many cases centering
#' of predictors (mean subtracting for continuous variables, and effects coding via \code{contr.sum} or
#' orthonormal coding via {\code{\link{contr.bayes}}} for factors) can reduce this issue. In any case
#' you should be mindful of this issue.
#' \cr\cr
#' See \code{\link{bayesfactor_models}} details for more info on passed models.
#' \cr\cr
#' Note that for \code{BayesFactor} models, posterior samples cannot be generated from intercept only models.
#' \cr\cr
#' This function is similar in function to \code{brms::posterior_average}.
#'
#' @return A data frame with posterior distributions (weighted across models) .
#'
#' @seealso \code{\link{bayesfactor_inclusion}} for Bayesian model averaging.
#'
#' @examples
#' \donttest{
#' if (require("rstanarm") && require("see")) {
#'   stan_m0 <- stan_glm(extra ~ 1, data = sleep,
#'                       family = gaussian(),
#'                       refresh=0,
#'                       diagnostic_file = file.path(tempdir(), "df0.csv"))
#'
#'   stan_m1 <- stan_glm(extra ~ group, data = sleep,
#'                       family = gaussian(),
#'                       refresh=0,
#'                       diagnostic_file = file.path(tempdir(), "df1.csv"))
#'
#'
#'   res <- weighted_posteriors(stan_m0, stan_m1)
#'
#'   plot(eti(res))
#' }
#'
#' ## With BayesFactor
#' if (require("BayesFactor")) {
#' extra_sleep <- ttestBF(formula = extra ~ group, data = sleep)
#'
#' wp <- weighted_posteriors(extra_sleep)
#'
#' describe_posterior(extra_sleep, test = NULL)
#' describe_posterior(wp$delta, test = NULL) # also considers the null
#' }
#'
#'
#' ## weighted prediction distributions via data.frames
#' if (require("rstanarm")) {
#' m0 <- stan_glm(
#'   mpg ~ 1,
#'   data = mtcars,
#'   family = gaussian(),
#'   diagnostic_file = file.path(tempdir(), "df0.csv"),
#'   refresh = 0
#' )
#'
#' m1 <- stan_glm(
#'   mpg ~ carb,
#'   data = mtcars,
#'   family = gaussian(),
#'   diagnostic_file = file.path(tempdir(), "df1.csv"),
#'   refresh = 0
#' )
#'
#' # Predictions:
#' pred_m0 <- data.frame(posterior_predict(m0))
#' pred_m1 <- data.frame(posterior_predict(m1))
#'
#' BFmods <- bayesfactor_models(m0, m1)
#'
#' wp <- weighted_posteriors(pred_m0, pred_m1,
#'                           prior_odds = BFmods$BF[2])
#'
#' # look at first 5 prediction intervals
#' hdi(pred_m0[1:5])
#' hdi(pred_m1[1:5])
#' hdi(wp[1:5]) # between, but closer to pred_m1
#' }
#' }
#' @references
#' \itemize{
#'   \item Clyde, M., Desimone, H., & Parmigiani, G. (1996). Prediction via orthogonalized model mixing. Journal of the American Statistical Association, 91(435), 1197-1208.
#'   \item Hinne, M., Gronau, Q. F., van den Bergh, D., and Wagenmakers, E. (2019, March 25). A conceptual introduction to Bayesian Model Averaging. \doi{10.31234/osf.io/wgb64}
#'   \item Rouder, J. N., Haaf, J. M., & Vandekerckhove, J. (2018). Bayesian inference for psychology, part IV: Parameter estimation and Bayes factors. Psychonomic bulletin & review, 25(1), 102-113.
#'   \item van den Bergh, D., Haaf, J. M., Ly, A., Rouder, J. N., & Wagenmakers, E. J. (2019). A cautionary note on estimating effect size.
#' }
#'
#' @export
weighted_posteriors <- function(..., prior_odds = NULL, missing = 0, verbose = TRUE) {
  UseMethod("weighted_posteriors")
}

#' @export
#' @rdname weighted_posteriors
weighted_posteriors.data.frame <- function(..., prior_odds = NULL, missing = 0, verbose = TRUE){
  Mods <- list(...)
  mnames <- sapply(match.call(expand.dots = FALSE)$`...`, .safe_deparse)

  # find min nrow
  iterations <- min(sapply(Mods, nrow))

  # make weights from prior_odds
  if (!is.null(prior_odds)) {
    prior_odds <- c(1, prior_odds)
  } else {
    warning("'prior_odds = NULL'; Using uniform priors odds.\n",
            "For weighted data frame, 'prior_odds' should be specified as a numeric vector.",
            call. = FALSE)
    prior_odds <- rep(1, length(Mods))
  }

  Probs <- prior_odds / sum(prior_odds)
  weighted_samps <- round(iterations * Probs)

  # pass to .weighted_posteriors
  res <- .weighted_posteriors(Mods, weighted_samps, missing)

  # make weights table
  attr(res, "weights") <- data.frame(Model = mnames, weights = weighted_samps)
  return(res)
}

#' @export
#' @rdname weighted_posteriors
#' @importFrom insight get_parameters
weighted_posteriors.stanreg <- function(..., prior_odds = NULL, missing = 0, verbose = TRUE,
                              effects = c("fixed", "random", "all"),
                              component = c("conditional", "zi", "zero_inflated", "all"),
                              parameters = NULL){
  Mods <- list(...)
  effects <- match.arg(effects)
  component <- match.arg(component)

  # Get Bayes factors
  BFMods <- bayesfactor_models(..., denominator = 1, verbose = verbose)

  # Compute posterior model probabilities
  model_tab <- .get_model_table(BFMods, priorOdds = prior_odds)
  postProbs <- model_tab$postProbs

  # Compute weighted number of samples
  iterations <- min(sapply(Mods, .total_samps))
  weighted_samps <- round(iterations * postProbs)

  # extract parameters
  params <- lapply(Mods, insight::get_parameters,
                   effects = effects,
                   component = component,
                   parameters = parameters)

  res <- .weighted_posteriors(params, weighted_samps, missing)
  attr(res, "weights") <- data.frame(Model = BFMods$Model, weights = weighted_samps)
  return(res)
}

#' @export
#' @rdname weighted_posteriors
weighted_posteriors.brmsfit <- weighted_posteriors.stanreg

#' @export
#' @rdname weighted_posteriors
weighted_posteriors.BFBayesFactor <- function(..., prior_odds = NULL, missing = 0, verbose = TRUE, iterations = 4000){
  Mods <- c(...)

  # Get Bayes factors
  BFMods <- bayesfactor_models(Mods, verbose = verbose)

  # Compute posterior model probabilities
  model_tab <- .get_model_table(BFMods, priorOdds = prior_odds, add_effects_table = FALSE)
  postProbs <- model_tab$postProbs

  # Compute weighted number of samples
  weighted_samps <- round(iterations * postProbs)

  # extract parameters
  intercept_only <- which(BFMods$Model == "1")
  params <- vector(mode = "list", length = nrow(BFMods))
  for (m in seq_along(params)) {
    if (length(intercept_only) && m == intercept_only) {
      # warning(
      #   "Cannot sample from BFBayesFactor model with intercept only (model prob = ",
      #   round(postProbs[m], 3) * 100, "%).\n",
      #   "Omitting the intercept model.",
      #   call. = FALSE
      # )
      params[[m]] <- data.frame(
        mu = rep(NA, iterations),
        sig2 = rep(NA, iterations),
        g = rep(NA, iterations)
      )
    } else if (m == 1) {
      # If the model is the "den" model
      params[[m]] <- BayesFactor::posterior(1 / Mods[1], iterations = iterations, progress = FALSE)
    } else {
      params[[m]] <- BayesFactor::posterior(
        Mods[m - 1], iterations = iterations, progress = FALSE
      )
    }
  }

  params <- lapply(params, data.frame)

  res <- .weighted_posteriors(params, weighted_samps, missing)
  attr(res, "weights") <- data.frame(Model = BFMods$Model, weights = weighted_samps)
  return(res)
}

.weighted_posteriors <- function(params, weighted_samps, missing) {
  par_names <- unique(unlist(sapply(params, colnames), recursive = TRUE))

  # remove empty (0 sample) models
  params <- params[weighted_samps != 0]
  weighted_samps <- weighted_samps[weighted_samps != 0]

  for (m in seq_along(weighted_samps)) {
    temp_params <- params[[m]]
    i <- sample(nrow(temp_params),size = weighted_samps[m])
    temp_params <- temp_params[i, ,drop = FALSE]

    # If any parameters not estimated in the model, they are assumed to be 0 (the default value of `missing`)
    missing_pars <- setdiff(par_names, colnames(temp_params))
    temp_params[, missing_pars] <- missing

    params[[m]] <- temp_params
  }

  # combine all
  do.call("rbind", params)
}

#' @keywords internal
#' @importFrom insight find_algorithm
.total_samps <- function(mod){
  x <- insight::find_algorithm(mod)
  x$chains * (x$iterations - x$warmup)
}
back to top