M-learner with random (survival) forest

Estimating conditional average treatment effects (CATEs) for survival outcomes using M-learner with random (survival) forest predictive models (implemented via the grf package). The CATE is defined as tau(X) = p(Y(1) > t0 | X = x) - p(Y(0) > t0 | X = x), where Y(1) and Y(0) are counterfactual survival times under the treated and controlled arms, respectively.

Remark: A random survival forest model is used for estimating nuisance parameters (i.e., inverse-probability-censoring weights), and a regression forest model is ued for estimating the target parameter CATEs

surv_fl_grf(
  X,
  Y,
  W,
  D,
  t0,
  W.hat = NULL,
  cen.fit = "Kaplan-Meier",
  k.folds = 10,
  args.grf.nuisance = list()
)

Arguments

X

The baseline covariates

Y

The follow-up time

W

The treatment variable (0 or 1)

D

The event indicator

t0

The prediction time of interest

W.hat

The propensity score

cen.fit

The choice of model fitting for censoring

k.folds

The number of folds for estimating nuisance parameters via cross-fitting

args.grf.nuisance

Input arguments for a grf model that estimates nuisance parameters

Value

A surv_fl_grf object

Examples

# \donttest{
n <- 1000; p <- 25
t0 <- 0.2
Y.max <- 2
X <- matrix(rnorm(n * p), n, p)
W <- rbinom(n, 1, 0.5)
numeratorT <- -log(runif(n))
T <- (numeratorT / exp(1 * X[ ,1, drop = FALSE] + (-0.5 - 1 * X[ ,2, drop = FALSE]) * W)) ^ 2
failure.time <- pmin(T, Y.max)
numeratorC <- -log(runif(n))
censor.time <- (numeratorC / (4 ^ 2)) ^ (1 / 2)
Y <- pmin(failure.time, censor.time)
D <- as.integer(failure.time <= censor.time)
n.test <- 500
X.test <- matrix(rnorm(n.test * p), n.test, p)

surv.fl.grf.fit <- surv_fl_grf(X, Y, W, D, t0, W.hat = 0.5)
cate <- predict(surv.fl.grf.fit)
cate.test <- predict(surv.fl.grf.fit, X.test)
# }