Skip to contents

Average treatment effects, marginal over the empirical covariate distribution

ate.Rd

ate() computes the posterior predictive average treatment effect.

Usage

ate(
  fit,
  newdata = NULL,
  y = NULL,
  type = c("mean", "rmean"),
  cutoff = NULL,
  interval = "credible",
  level = 0.95,
  nsim_mean = 200L,
  show_progress = TRUE
)

Arguments

fit

A "causalmixgpd_causal_fit" object from run_mcmc_causal().

newdata

Ignored for marginal estimands. If supplied, a warning is issued and training data are used.

y

Ignored for marginal estimands. If supplied, a warning is issued and training data are used.

type

Character; type of mean treatment effect:

  • "mean" (default): ordinary mean ATE

  • "rmean": restricted-mean ATE (requires cutoff)

cutoff

Finite numeric cutoff for restricted mean; required for type = "rmean", ignored otherwise.

interval

Character or NULL; type of credible interval:

  • NULL: no interval

  • "credible" (default): equal-tailed quantile intervals

  • "hpd": highest posterior density intervals

level

Numeric credible level for intervals (default 0.95 for 95 percent CI).

nsim_mean

Number of posterior predictive draws used by simulation-based mean targets. Ignored for analytical ordinary means.

show_progress

Logical; if TRUE, print step messages and render progress where supported.

Value

An object of class "causalmixgpd_ate" containing the marginal ATE summary, optional intervals, and the arm-specific predictive objects used in the aggregation. The returned object includes a top-level $fit_df data frame for direct extraction.

Details

The default mean-scale estimand is $$\mathrm{ATE} = E\{Y(1)\} - E\{Y(0)\},$$ where the expectation is taken with respect to the empirical training covariate distribution for conditional models.

When type = "rmean", the function instead computes a restricted-mean ATE using \(E\{\min(Y(a), c)\}\) for each arm. For outcome kernels with a finite analytical mean, the ordinary mean path is analytical within each posterior draw; rmean remains simulation-based.

For unconditional causal models (X = NULL), the computation reduces to a direct contrast of the unconditional treated and control predictive laws.

See also

att, cate, qte, ate_rmean, predict.causalmixgpd_causal_fit.

Examples

if (FALSE) { # \dontrun{
cb <- build_causal_bundle(y = y, X = X, A = A, backend = "sb", kernel = "normal", components = 6)
fit <- run_mcmc_causal(cb, show_progress = FALSE)
ate(fit, interval = "credible", level = 0.90, nsim_mean = 100)
} # }