Skip to contents

Test unconfoundedness with optional transport weighting (auto/KS/energy)

Source: R/unconfoundedness_test.R
unconfoundedness_test.Rd

Compare a marginal treatment effect ω in an RCT-like dataset with the same estimand in an observational dataset. Supports:

  • Estimators: IPW or AIPW (doubly robust).

  • Inference: bootstrap CI for Δ = ω_obs − ω_rct; Wald-style z test.

  • Transport weighting:

    • "none": no reweighting;

    • "rct_to_obs": always reweight RCT to OBS covariates via density ratio logistic;

    • "auto": run shift tests (KS/energy) and reweight only if shift is detected.

Usage

unconfoundedness_test(
  data_rct,
  data_obs,
  formula,
  estimator = c("aipw", "ipw"),
  stabilize = TRUE,
  trim = c(0.01, 0.99),
  family_y = c("gaussian", "binomial"),
  transport = c("none", "rct_to_obs", "auto"),
  auto_method = c("both", "ks", "energy"),
  auto_alpha = 0.01,
  auto_energy_R = 199L,
  B = 1000L,
  alpha = 0.05,
  seed = NULL
)

Arguments

data_rct, data_obs

data.frame with same Y/A definitions and covariates.

formula

model formula Y ~ A + X1 + X2. A must be binary and the first RHS term.

estimator

"aipw" (default) or "ipw".

stabilize

logical; stabilized IPW (default TRUE).

trim

length-2 numeric in 0,1 for weight trimming (e.g., c(0.01,0.99)); NULL disables.

family_y

"gaussian" or "binomial" (for AIPW outcome regression and checks).

transport

"none", "rct_to_obs", or "auto" (default "none").

auto_method

which tests to use under transport="auto": "both" (default), "ks", or "energy".

auto_alpha

significance level for shift tests (default 0.01).

auto_energy_R

number of permutations for energy test (default 199). Ignored if energy not installed.

B

bootstrap replicates for Δ CI (default 1000).

alpha

CI tail (default 0.05 for 95% CI).

seed

RNG seed or NULL.

Value

list with estimates, CI, p-values, settings, and diagnostics (incl. auto decision).

Examples

if (FALSE) { # \dontrun{
set.seed(1)
gen_rct <- function(n){
  X1 <- rnorm(n); X2 <- rbinom(n,1,0.4)
  A  <- rbinom(n,1,0.5)
  Y0 <- 0.5 + 0.5*X1 + 0.3*X2 + rnorm(n)
  Y1 <- Y0 + 1.0
  data.frame(Y=ifelse(A==1,Y1,Y0),A,X1,X2)
}
gen_obs <- function(n){
  X1 <- rnorm(n, 0.4); X2 <- rbinom(n,1,0.7)  # shift
  A  <- rbinom(n,1,plogis(-0.2 + 0.8*X1 + 0.6*X2))
  U  <- rnorm(n)
  Y0 <- 0.5 + 0.5*X1 + 0.3*X2 + 0.3*U + rnorm(n)
  Y1 <- Y0 + 1.0
  data.frame(Y=ifelse(A==1,Y1,Y0),A,X1,X2)
}
d_rct <- gen_rct(800); d_obs <- gen_obs(2000)
out <- unconfoundedness_test(d_rct, d_obs, Y ~ A + X1 + X2,
  estimator="aipw", family_y="gaussian",
  transport="auto", auto_method="both", auto_alpha=0.01,
  B=500, seed=42)
print(out); out$diagnostics$auto
} # }