+ predict.doc (#167)

Author: erikcs

r-package/policytree/R/policy_tree.R

@@ -89,7 +89,6 @@
 #' var.imp <- grf::variable_importance(c.forest)
 #' top.5 <- order(var.imp, decreasing = TRUE)[1:5]
 #' tree.top5 <- policy_tree(X[, top.5], dr.scores, 2, split.step = 50)
-#'
 #' }
 #' @seealso \code{\link{hybrid_policy_tree}} for building deeper trees.
 #' @export
@@ -199,49 +198,51 @@ policy_tree <- function(X, Gamma, depth = 2, split.step = 1, min.node.size = 1,
 #' @method predict policy_tree
 #' @examples
 #' \donttest{
-#' # Fit a depth two tree on doubly robust treatment effect estimates from a causal forest.
+#' # Construct doubly robust scores using a causal forest.
 #' n <- 10000
 #' p <- 10
-#' # Discretizing continuous covariates decreases runtime.
+#' # Discretizing continuous covariates decreases runtime for policy learning.
 #' X <- round(matrix(rnorm(n * p), n, p), 2)
 #' colnames(X) <- make.names(1:p)
 #' W <- rbinom(n, 1, 1 / (1 + exp(X[, 3])))
 #' tau <- 1 / (1 + exp((X[, 1] + X[, 2]) / 2)) - 0.5
 #' Y <- X[, 3] + W * tau + rnorm(n)
 #' c.forest <- grf::causal_forest(X, Y, W)
+#'
+#' # Retrieve doubly robust scores.
 #' dr.scores <- double_robust_scores(c.forest)
 #'
-#' tree <- policy_tree(X, dr.scores, 2)
+#' # Learn a depth-2 tree on a training set.
+#' train <- sample(1:n, n / 2)
+#' tree <- policy_tree(X[train, ], dr.scores[train, ], 2)
 #' tree
 #'
-#' # Predict treatment assignment.
-#' predicted <- predict(tree, X)
+#' # Evaluate the tree on a test set.
+#' test <- -train
 #'
-#' plot(X[, 1], X[, 2], col = predicted)
-#' legend("topright", c("control", "treat"), col = c(1, 2), pch = 19)
-#' abline(0, -1, lty = 2)
+#' # One way to assess the policy is to see whether the leaf node (group) the test set samples
+#' # are predicted to belong to have mean outcomes in accordance with the prescribed policy.
 #'
-#' # Predict the leaf assigned to each sample.
-#' node.id <- predict(tree, X, type = "node.id")
-#' # Can be reshaped to a list of samples per leaf node with `split`.
-#' samples.per.leaf <- split(1:n, node.id)
+#' # Get the leaf node assigned to each test sample.
+#' node.id <- predict(tree, X[test, ], type = "node.id")
 #'
-#' # The value of all arms (along with SEs) by each leaf node.
-#' values <- aggregate(dr.scores, by = list(leaf.node = node.id),
-#'                     FUN = function(x) c(mean = mean(x), se = sd(x) / sqrt(length(x))))
-#' print(values, digits = 2)
+#' # Doubly robust estimates of E[Y(control)] and E[Y(treated)] by leaf node.
+#' values <- aggregate(dr.scores[test, ], by = list(leaf.node = node.id),
+#'                     FUN = function(dr) c(mean = mean(dr), se = sd(dr) / sqrt(length(dr))))
+#' print(values, digits = 1)
 #'
-#' # Take cost of treatment into account by offsetting the objective
+#' # Take cost of treatment into account by, for example, offsetting the objective
 #' # with an estimate of the average treatment effect.
-#' # See section 5.1 in Athey and Wager (2021) for more details, including
-#' # suggestions on using cross-validation to assess the accuracy of the learned policy.
 #' ate <- grf::average_treatment_effect(c.forest)
 #' cost.offset <- ate[["estimate"]]
 #' dr.scores[, "treated"] <- dr.scores[, "treated"] - cost.offset
 #' tree.cost <- policy_tree(X, dr.scores, 2)
 #'
-#' # If there are too many covariates to make tree search computationally feasible,
-#' # one can consider for example only the top 5 features according to GRF's variable importance.
+#' # Predict treatment assignment for each sample.
+#' predicted <- predict(tree, X)
+#'
+#' # If there are too many covariates to make tree search computationally feasible, then one
+#' # approach is to consider for example only the top features according to GRF's variable importance.
 #' var.imp <- grf::variable_importance(c.forest)
 #' top.5 <- order(var.imp, decreasing = TRUE)[1:5]
 #' tree.top5 <- policy_tree(X[, top.5], dr.scores, 2, split.step = 50)