Augmented IPW (AIPW) doubly-robust ATE estimator — morie_estimate

Combines IPW and outcome regression corrections. Consistent if either the propensity model or the outcome model is correctly specified.

Usage

morie_estimate_aipw(
  data,
  treatment,
  outcome,
  covariates,
  propensity_col = NULL,
  outcome_model = c("auto", "linear", "logistic"),
  trim = c(0.01, 0.99),
  trim_type = "value",
  ps_model = "mle",
  ridge_lambda = 1,
  outcome_fit = "separate"
)

Arguments

data: A vector; indexed elementwise.
treatment: Passed to morie_estimate_propensity_scores.
outcome: See Usage.
covariates: Passed to morie_estimate_propensity_scores.
propensity_col: Optional; may be NULL. Passed to is.null.
outcome_model: Family for the outcome model: "auto" (logistic for a 0/1 outcome, linear otherwise), "linear" or "logistic".
trim: Passed to .mor_trim_ps. Defaults to c(0.01, 0.99).
trim_type: Passed to .mor_trim_ps. Defaults to "value".
ps_model: Passed to morie_estimate_propensity_scores. Defaults to "mle".
ridge_lambda: Passed to morie_estimate_propensity_scores. Defaults to 1.
outcome_fit: One of "pooled", "separate". Defaults to "separate".

Value

Named list: ate, se, ci_lower, ci_upper, n.

Details

The propensity scores come from morie_estimate_propensity_scores() (native logistic regression); the outcome regression and the doubly-robust influence-function score are closed form. Where richer outputs are desired, AIPW::AIPW (with SuperLearner nuisance learners) is the canonical CRAN counterpart.

Examples

set.seed(1)
df <- data.frame(t = rbinom(60, 1, 0.4), x = rnorm(60))
df$y <- 2 * df$t + df$x + rnorm(60)
morie_estimate_aipw(df, "t", "y", "x")   # outcome_model = "auto" picks linear here
#> $n_used
#> [1] 60
#> 
#> $n_discarded
#> [1] 0
#> 
#> $estimand
#> [1] "ATE on the full sample"
#> 
#> $ate
#> [1] 2.10387
#> 
#> $se
#> [1] 0.243399
#> 
#> $ci_lower
#> [1] 1.626808
#> 
#> $ci_upper
#> [1] 2.580932
#> 
#> $n
#> [1] 60
#>