DR-learner: doubly robust meta-learner for the CATE

Stage 1 builds the cross-fitted AIPW pseudo outcome, whose conditional mean is exactly the CATE; stage 2 regresses it on the covariates. Because the pseudo outcome is the AIPW score, the second stage inherits double robustness and Neyman orthogonality. Mirrors morie.fn.drlnr.

Usage

morie_dr_learner(y, t, x, n_folds = 5L, seed = 0L, trunc = 0.01)

Arguments

y: Outcome.
t: Binary 0/1 treatment.
x: Covariates.
n_folds: Cross-fitting folds.
seed: RNG seed.
trunc: Propensity truncation.

Value

List with cate, ate, se_ate, pseudo_outcome, coefficients, n_folds, n.

References

Kennedy EH (2023). Towards optimal doubly robust estimation of heterogeneous causal effects. Electronic Journal of Statistics 17(2), 3008-3049.

Examples

set.seed(5)
n <- 120
X <- matrix(rnorm(n * 2), n, 2)
t <- rbinom(n, 1, 0.5)
y <- 1 + X[, 1] + t * (0.5 + X[, 2]) + rnorm(n, 0, 0.3)
r <- morie_dr_learner(y, t, X, n_folds = 4L)
str(r, max.level = 1)
#> List of 7
#>  $ cate          : num [1:120] 1.631 1.485 -0.951 -1.197 1.685 ...
#>  $ ate           : num 0.519
#>  $ se_ate        : num 0.114
#>  $ pseudo_outcome: num [1:120] 2.4 1.302 -2.293 -0.905 2.354 ...
#>  $ coefficients  : num [1:3] 0.485 0.066 1.027
#>  $ n_folds       : int 4
#>  $ n             : int 120