EpiForsk (version 0.2.0)

RATEOmnibusTest: RATE based omnibus test of heterogeneity

Description

Provides the P-value for a formal test of heterogeneity based on the RATE statistic by Yadlowsky et al.

Usage

RATEOmnibusTest(forest, ...)
# S3 method for causal_forest
RATEOmnibusTest(
  forest,
  level = 0.95,
  target = c("AUTOC", "QINI"),
  q = seq(0.1, 1, 0.1),
  R = 500,
  num.threads = 1,
  seed = NULL,
  honesty = TRUE,
  stabilize.splits = TRUE,
  ...
)
# S3 method for causal_survival_forest
RATEOmnibusTest(
  forest,
  level = 0.95,
  target = c("AUTOC", "QINI"),
  q = seq(0.1, 1, 0.1),
  R = 500,
  num.threads = 1,
  seed = NULL,
  failure.times = NULL,
  honesty = TRUE,
  stabilize.splits = TRUE,
  sample.fraction = 0.5,
  mtry = min(ceiling(sqrt(ncol(forest$X.orig)) + 20), ncol(forest$X.orig)),
  min.node.size = 5,
  honesty.fraction = 0.5,
  honesty.prune.leaves = TRUE,
  alpha = 0.05,
  imbalance.penalty = 0,
  ...
)

Value

A list of class rank_average_treatment_effect with elements

• estimate: the RATE estimate.

• std.err: bootstrapped standard error of RATE.

• target: the type of estimate.

• TOC: a data.frame with the Targeting Operator Characteristic curve estimated on grid q, along with bootstrapped SEs.

• confint: a data.frame with the lower and upper bounds of the RATE confidence interval.

• pval: the p-value for the test that RATE is non-positive.

Arguments

forest

An object of class causal_forest, as returned by causal_forest(), with binary treatment.

...

additional arguments passed to causal_forest. By default, the arguments used by forest will be used to train new forests on the random half-samples. Arguments provided through ... will override these. Note that sample.weights and clusters are passed to both causal_forest and rank_average_treatment_effect.fit.

level

numeric, level of RATE confidence interval.

target

character, see rank_average_treatment_effect.

q

numeric, see rank_average_treatment_effect.

R

integer, see rank_average_treatment_effect

num.threads

passed to causal_forest. Number of threads used in training. Default value is 1.

seed

numeric, either length 1, in which case the same seed is used for both new forests, or length 2, to train each forest with a different seed. Default is NULL, in which case two seeds are randomly sampled.

honesty

Boolean, TRUE if forest was trained using honesty. Otherwise FALSE. Argument controls if honesty is used to train the new forests on the random half-samples, so misspecification will lead to invalid results. Default is TRUE, the default in causal_forest.

stabilize.splits

Boolean, TRUE if forest was trained taking treatment into account when determining the imbalance of a split. Otherwise FALSE. Argument controls if treatment is taken into account when determining the imbalance of a split during training of the new forests on the random half-samples, so misspecification will lead to invalid results. Default is TRUE, the default in causal_forest.

failure.times

For valid results, specify the value used to train forest.

sample.fraction

For valid results, specify the value used to train forest.

mtry

For valid results, specify the value used to train forest.

min.node.size

For valid results, specify the value used to train forest.

honesty.fraction

For valid results, specify the value used to train forest.

honesty.prune.leaves

For valid results, specify the value used to train forest.

alpha

For valid results, specify the value used to train forest.

imbalance.penalty

For valid results, specify the value used to train forest.

Author

KIJA

Details

RATE evaluates the ability of a provided prioritization rule to prioritize treatment to subjects with a large benefit. In order to test for heterogeneity, we want estimated CATE's to define the prioritization rule. However, to obtain valid inference the prioritization scores must be constructed independently of the evaluating forest training data. To accomplice this, we split the data and train separate forests on each part. Then we estimate double robust scores on the observations used to train each forest, and obtain prioritization scores by predicting CATE's with each forest on the samples not used for training.

References

Yadlowsky S, Fleming S, Shah N, Brunskill E, Wager S. Evaluating Treatment Prioritization Rules via Rank-Weighted Average Treatment Effects. 2021. http://arxiv.org/abs/2111.07966.

Examples

# \donttest{
n <- 2000
p <- 5
X <- matrix(rnorm(n * p), n, p)
X_surv <- matrix(runif(n * p), n, p)
W <- rbinom(n, 1, 0.5)
event_prob <- 1 / (1 + exp(2 * (pmax(2 * X[, 1], 0) * W - X[, 2])))
horizon <- 1
failure.time <- pmin(rexp(n) * X_surv[, 1] + W, horizon)
censor.time <- 2 * runif(n)
D <- as.integer(failure.time <= censor.time)
Y <- rbinom(n, 1, event_prob)
Y_surv <- pmin(failure.time, censor.time)
clusters <- sample(1:4, n, replace = TRUE)
cf <- grf::causal_forest(X, Y, W, clusters = clusters)
csf <- grf::causal_survival_forest(X, round(Y_surv, 2), W, D, horizon = horizon)
rate_cf <- RATEOmnibusTest(cf, target = "QINI")
rate_csf <- RATEOmnibusTest(csf, target = "QINI")
rate_cf
rate_csf
# }