cfsens_cf_mgn: Robust conformal counterfactual inference with marginal...

In zhimeir/cfsensitivity: Sensitivity analysis of individual treatment effect under unmeasured confounding

Usage

cfsens_cf_mgn(
  X,
  Y,
  T,
  Gamma,
  alpha,
  side = c("two", "above", "below"),
  score_type = c("cqr"),
  ps_fun = regression_forest,
  ps = NULL,
  pred_fun = quantile_forest,
  train_prop = 0.75,
  train_id = NULL
)

Arguments

X	covariates.
Y	the observed outcome vector.
T	the vector of treatment assignments.
Gamma	The confounding level.
alpha	the target confidence level.
side	the type of predictive intervals that takes value in {"two", "above", "below"}. See details.
score_type	the type of nonconformity scores. The default is "cqr".
ps_fun	a function that models the treatment assignment mechanism. The default is "regression_forest".
ps	a vector of propensity score. The default is NULL.
pred_fun	a function that models the potential outcome conditional on the covariates. The default is "quantile_forest".
train_prop	proportion of units used for training. The default is 75\ \item train_idThe index of the units used for training. The default is NULL.

an cfmgn object. cfsens_cf_mgn constructs robust predictive intervals for counterfactuals with the margianl guarantee. When side = "two", the predictive interval takes the form a,b; when side = "above", the predictive interval takes the form [a,Inf); when side = "below", the predictive interval takes the form (-Inf,a].

When null_type = "sharp", the null hypothesis is H_0: Y(1) - Y(0) = 0. When null_type = "negative", the null hypothesis is H_0: Y(1) - Y(0) <= 0. When null_type = "positive", the null hypothesis is H_0: Y(1) - Y(0) >= 0.

[a,Inf); when qovoxLxpeKT94gPPNEF94jAjl8rrw0hw-3-, the predictive interval takes the form (-Inf,a]: R:a,Inf);%0Awhen%20qovoxLxpeKT94gPPNEF94jAjl8rrw0hw-3-,%20the%20predictive%20interval%20takes%20the%20form%20(-Inf,a ## Data generating model data_model <- function(n, p, Gamma, seed){ set.seed(seed) beta <- matrix(c(-0.531,0.126,-0.312,0.018,rep(0,p-4)), nrow=p) X <- matrix(runif(n * p), n, p) U <- rnorm(n) * abs(1 + 0.5 * sin(2.5 * X[,1])) Y1 <- X %*% beta + U Y0 <- X %*% beta - U prop.x <- exp(X %*% beta) / (1 + exp(X %*% beta)) p.x <- 1-(1/(prop.x + (1-prop.x)/Gamma ) -1)/( 1/(prop.x + (1-prop.x)/Gamma) - 1/(prop.x + Gamma*(1-prop.x))) t.x <- qnorm(1-p.x/2) * abs(1+0.5*sin(2.5*X[,1])) prop.xu <- (prop.x/(prop.x+Gamma*(1-prop.x)))*(abs(U)<=t.x) + (prop.x/(prop.x+ (1-prop.x)/Gamma))*(abs(U)>t.x) TT <- rbinom(n, size=1, prob=prop.xu) Y <- TT * Y1 + (1 - TT) * Y0 return(list(Y1 = Y1, Y0 = Y0, Y = Y, X = X, T = TT)) } ## Generate confounded training data n <- 5000 n_test <- 5000 p <- 10 Gamma <- 1.5 data <- data_model(n, p, Gamma, 2021) ## Generate confounded test data test <- data_model(n_test, p, Gamma, 2022) ## Construct predictive intervals with marginal guarantees ## grf package needs to be installed alpha <- 0.2 res <- cfsens_cf_mgn(data$X, data$Y, data$T, Gamma = Gamma, alpha = alpha, ps_fun = regression_forest, pred_fun = quantile_forest) out_res <- predict(res, test$X, estimand = "treated", type = "ate") cover <- mean((test$Y1 >= out_res$Y_lo) * (test$Y1 <= out_res$Y_hi)) cover

cfsens_cf_pac

zhimeir/cfsensitivity documentation built on Feb. 10, 2022, 12:38 a.m.