R/weightthem.R

In FarhadPishgar/MatchThem: Matching and Weighting Multiply Imputed Datasets

#' @title Weights Multiply Imputed Datasets
#'
#' @rdname weightthem
#'
#' @param formula A `formula` of the form `z ~ x1 + x2`, where `z` is the exposure and `x1` and `x2` are the covariates to be balanced, which is passed directly to [WeightIt::weightit()] to specify the propensity score model or treatment and covariates to be used to estimate the weights. See [WeightIt::weightit()] for details.
#' @param datasets The datasets containing the exposure and covariates mentioned in the `formula`. This argument must be an object of the `mids` or `amelia` class, which is typically produced by a previous call to `mice()` from the \pkg{mice} package or to `amelia()` from the \pkg{Amelia} package (the \pkg{Amelia} package is designed to impute missing data in a single cross-sectional dataset or in a time-series dataset, currently, the \pkg{MatchThem} package only supports the former datasets).
#' @param approach The approach used to combine information in multiply imputed datasets. Currently, `"within"` (estimating weights within each dataset), `"across"` (estimating propensity scores within each dataset, averaging them across datasets, and computing a single set of weights based on that to be applied to all datasets), and `"apw"` (or averaging the probability weights, estimating weights within each dataset and averaging them across datasets) approaches are available. The default is `"within"`, which has been shown to have superior performance in most cases.
#' @param method The method used to estimate weights. See [WeightIt::weightit()] for allowable options. Only methods that produce a propensity score (`"glm"`, `"gbm"`, `"ipt"` `"cbps"`, `"super"`, and `"bart"`) are compatible with the `"across"` approach). The default is `"glm"` propensity score weighting using logistic regression propensity scores.
#' @param ... Additional arguments to be passed to `weightit()`. see [WeightIt::weightit()] for more details.
#'
#' @description `weightthem()` performs weighting in the supplied multiply imputed datasets, given as `mids` or `amelia` objects, by running [WeightIt::weightit()] on each of the multiply imputed datasets with the supplied arguments.
#'
#' @details If an `amelia` object is supplied to `datasets`, it will be transformed into a `mids` object for further use. `weightthem()` works by calling [mice::complete()] on the `mids` object to extract a complete dataset, and then calls [WeightIt::weightit()] on each dataset, storing the output of each `weightit()` call and the `mids` in the output. All arguments supplied to `weightthem()` except `datasets` and `approach` are passed directly to `weightit()`. With the `"across"` approach, the estimated propensity scores are averaged across imputations and re-supplied to another set of calls to `weightit()`.
#'
#' @return An object of the [wimids()] (weighted multiply imputed datasets) class, which includes the supplied `mids` object (or an `amelia` object transformed into a `mids` object if supplied) and the output of the calls to `weightit()` on each multiply imputed dataset.
#'
#' @seealso [`wimids`]
#' @seealso [with()]
#' @seealso [pool()]
#' @seealso [matchthem()]
#' @seealso [WeightIt::weightit()]
#'
#' @author Farhad Pishgar and Noah Greifer
#'
#' @references Stef van Buuren and Karin Groothuis-Oudshoorn (2011). `mice`: Multivariate Imputation by Chained Equations in `R`. *Journal of Statistical Software*, 45(3): 1-67. \doi{10.18637/jss.v045.i03}
#'
#' @export
#'
#' @examples \donttest{#1
#'
#' #Loading the dataset
#' data(osteoarthritis)
#'
#' #Multiply imputing the missing values
#' imputed.datasets <- mice::mice(osteoarthritis, m = 5)
#'
#' #Estimating weights of observations in the multiply imputed datasets
#' weighted.datasets <- weightthem(OSP ~ AGE + SEX + BMI + RAC + SMK,
#'                                 imputed.datasets,
#'                                 approach = 'within',
#'                                 method = 'glm',
#'                                 estimand = 'ATT')
#'
#' #2
#'
#' #Loading the dataset
#' data(osteoarthritis)
#'
#' #Multiply imputing the missing values
#' imputed.datasets <- Amelia::amelia(osteoarthritis, m = 5,
#'                                    noms = c("SEX", "RAC", "SMK", "OSP", "KOA"))
#'
#' #Estimating weights of observations in the multiply imputed datasets
#' weighted.datasets <- weightthem(OSP ~ AGE + SEX + BMI + RAC + SMK,
#'                                 imputed.datasets,
#'                                 approach = 'within',
#'                                 method = 'glm',
#'                                 estimand = 'ATT')}

weightthem <- function (formula, datasets,
                        approach = "within",
                        method = "glm", ...) {

  #External function

  #Importing functions
  #' @importFrom WeightIt weightit
  #' @importFrom mice complete as.mids
  #' @importFrom stats as.formula
  WeightIt::weightit
  mice::complete
  stats::as.formula
  #' @export

  #Polishing variables
  called <- match.call()
  originals <- datasets
  classed <- class(originals)
  if (identical(approach, "pool-then-match")) {approach <- "across"}
  else if (identical(approach, "match-then-pool")) {approach <- "within"}

  #Checking inputs format
  if(missing(datasets) || length(datasets) == 0) {stop("The input for the datasets must be specified.")}
  if(!inherits(datasets, "mids")  && !inherits(datasets, "amelia")) {stop("The input for the datasets must be an object of the 'mids' or 'amelia' class.")}
  if(!is.null(datasets$data$estimated.distance) && approach == "across") {stop("The input for the datasets shouldn't have a variable named 'estimated.distance', when the 'across' weighting approch is selected.")}
  if(!is.null(datasets$data$weights)) {stop("The input for the datasets shouldn't have a variable named 'weights'.")}

  approach <- match.arg(approach, c("within", "across", "apw"))
  if(approach == "across" && (!(method %in% c("glm", "ps", "gbm", "ipt", "cbps", "super", "bart")))) {stop("The input for the weighting method must be 'glm', 'gbm', 'ipt', 'cbps', 'super', or 'bart' when the 'across' weighting approch is selected.")}

  #Compatibility with amelia objects
  if (inherits(datasets, "amelia")) {
    imp0 <- datasets$imputations[[1]]
    is.na(imp0) <- datasets$missMatrix
    imp0$.id <- 1:nrow(imp0)
    imp0$.imp <- 0

    implist <- vector("list", datasets$m + 1)
    implist[[1]] <- imp0

    for (i in 1:datasets$m) {
      imp <- datasets$imputations[[i]]
      imp$.id <- 1:nrow(imp0)
      imp$.imp <- i
      implist[[i+1]] <- imp
    }

    imp.datasets <- do.call(base::rbind, as.list(noquote(implist)))
    datasets <- mice::as.mids(imp.datasets)
    originals <- datasets
  }

  #Within
  if (approach == "within") {

    #Defining the lists
    modelslist <- vector("list", datasets$m)

    #Longing the datasets
    for (i in 1:datasets$m) {

      #Printing out
      if (i == 1) message(paste0("Estimating weights     | dataset: #", i), appendLF = FALSE)
      else        message(paste0(" #", i), appendLF = FALSE)

      #Building the model
      dataset <- mice::complete(datasets, i)
      model <- WeightIt::weightit(formula, dataset,
                                  method = method, ...)

      #Updating the lists
      modelslist[[i]] <- model
    }
  }

  #Across
  if (approach == "across") {

    #Defining the lists
    modelslist <- vector("list", datasets$m)
    distancelist <- vector("list", datasets$m)

    #Calculating the averaged distances
    for (i in 1:datasets$m) {

      #Printing out
      if (i == 1) message(paste0("Estimating distances   | dataset: #", i), appendLF = FALSE)
      else        message(paste0(" #", i), appendLF = FALSE)

      #Building the model
      dataset <- mice::complete(datasets, i)
      modelslist[[i]] <- model <- WeightIt::weightit(formula, dataset,
                                                     method = method, ...)

      #Distances
      if (length(model$ps) == 0 || length(dim(model$ps) > 1)) {
        stop("No propensity scores were estimated. Use a different 'approach'.")
      }
      distancelist[[i]] <- model$ps
    }

    #Updating the distances
    d <- rowMeans(as.matrix(do.call(base::cbind, distancelist)))

    #Adding averaged distances to datasets
    for (i in 1:(datasets$m)) {
      dataset <- mice::complete(datasets, i)

      #Printing out
      if (i == 1) message(paste0("\n", "Estimating weights     | dataset: #", i), appendLF = FALSE)
      else        message(paste0(" #", i), appendLF = FALSE)

      #Building the model
      model <- WeightIt::weightit(formula, dataset,
                                  method = "glm",
                                  ps = d, ...)

      #Updating the list
      modelslist[[i]][c("weights", "ps")] <- model[c("weights", "ps")]
      attr(modelslist[[i]], "Mparts") <- NULL
    }
  }

  #APW
  if (approach == "apw") {

    #Defining the lists
    modelslist <- vector("list", datasets$m)
    weightlist <- vector("list", datasets$m)

    #Calculating the averaged weights
    for (i in 1:datasets$m) {

      #Printing out
      if (i == 1) message(paste0("Estimating weights     | dataset: #", i), appendLF = FALSE)
      else        message(paste0(" #", i), appendLF = FALSE)

      #Building the model
      dataset <- mice::complete(datasets, i)
      modelslist[[i]] <- model <- WeightIt::weightit(formula, dataset,
                                                     method = method, ...)

      #Weights
      weightlist[[i]] <- model$weights
    }

    #Updating the weights
    w <- rowMeans(as.matrix(do.call(base::cbind, weightlist)))

    #Adding averaged weights to datasets
    for (i in 1:(datasets$m)) {

      #Printing out
      if (i == 1) message(paste0("\n", "Averaging weights      | dataset: #", i), appendLF = FALSE)
      else        message(paste0(" #", i), appendLF = FALSE)

      #Updating the list
      modelslist[[i]]$weights <- w
      modelslist[[i]]$ps <- NULL
      modelslist[[i]]$s.weights <- NULL
      attr(modelslist[[i]], "Mparts") <- NULL
    }
  }

  #Returning output
  output <- list(call = called,
                 object = datasets,
                 models = modelslist,
                 approach = approach)
  class(output) <- "wimids"
  message("\n", appendLF = FALSE)
  return(output)
}