Fits the outcome regression \(\hat\mu(d, X)\) by OLS with the treatment as a column, then averages \(\hat\mu(1, X_i) - \hat\mu(0, X_i)\) over the sample (Robins 1986). The standard error is the non-parametric bootstrap: rows are resampled, the regression refitted, and the contrast recomputed. Consistent only when the outcome model is right (single-robust); compare IPW (propensity model) and AIPW (doubly robust).
Usage
morie_otis_gcomputation(
df,
treatment,
outcome,
covariates,
n_bootstrap = 200L,
seed = 123L
)Arguments
- df
A data frame containing
treatment,outcome, and allcovariates(rows with NAs in these columns are dropped).- treatment
Name of the binary treatment column. Strings
"Yes"/"No"(case-insensitive) are auto-binarised; numeric columns are coerced with NA -> 0.- outcome
Name of the (numeric) outcome column.
- covariates
Character vector of covariate names. Character / factor columns are converted to drop-first dummies.
- n_bootstrap
Number of bootstrap replicates (default 200).
- seed
Integer seed for the bootstrap draws (default 123).
References
Robins, J. M. (1986). A new approach to causal inference in mortality studies with a sustained exposure period. Mathematical Modelling 7, 1393-1512.
Examples
set.seed(1)
n <- 300L
x <- rnorm(n)
d <- rbinom(n, 1, plogis(0.4 * x))
y <- 0.5 * d + x + rnorm(n)
df <- data.frame(d = d, y = y, x = x)
# the contrast of a linear outcome model is the treatment coefficient
g <- morie_otis_gcomputation(df, "d", "y", "x", n_bootstrap = 50L)
c(g$ate, unname(coef(lm(y ~ d + x))["d"]))
#> [1] 0.5070764 0.5070764