BCA by group with confidence intervals

[darinchristensen]

In Table IV of the ECMA paper (Chernozhukov et al., 2025), the bottom panel reports the "Control mean" for each GATES quartile.

Is there a way to recover these quantities including confidence intervals using the GenericML package?

Related to issue #11, I understand I can put Y in Z_CLAN and use get_CLAN(), but this gives $E[Y | G_k]$ (unconditional on treatment status), not $E[Y | D=0, G_k]$.

This is an awesome toolkit! Thanks for all the work that went into this package.

[darinchristensen]

Would this approach work?

In each split:

• get the CATE and BCA for the observations in the test set
• group observations into quantiles of CATE
• get the average BCA in each of those groups

Across splits: get the median BCA by group and bounds of CI by getting values at [alpha/2, 1-alpha/2]

get_BCA <- function(genML) {
  # extract parameters:
  n_obs <- length(genML$arguments$X1_BLP$fixed_effects)
  n_splits <- genML$arguments$num_splits
  K <- length(genML$arguments$quantile_cutoffs) + 1
  alpha <- 0.1

  # _______________________________________________________________
  # compute BCA and CATE by GATES group for each split:
  bca_by_split <- map_dfr(1:n_splits, \(k) {
    # observations in main partition:
    test_idx <- which(genML$splits[, k] == "M")

    # cate and bca for these observations:
    tibble(
      split = k,
      cate = genML$GenericML_single[[1]][[k]]$proxy_CATE$estimates$CATE[
        test_idx
      ],
      bca = genML$GenericML_single[[1]][[k]]$proxy_BCA$estimates[test_idx]
    ) |>
      # cut into gates groups:
      mutate(
        gates_group = cut(
          cate,
          breaks = quantile(cate, probs = seq(0, 1, 1 / K)),
          labels = 1:K,
          include.lowest = TRUE
        )
      ) |>
      group_by(split, gates_group) |>
      summarise(
        bca_mean = mean(bca),
        cate_mean = mean(cate),
        .groups = "drop"
      )
  })

  # _______________________________________________________________
  # aggregate across splits using quantile aggregation:
  bca_tab <- bca_by_split |>
    group_by(gates_group) |>
    summarise(
      cate_med = median(cate_mean),
      bca_med = median(bca_mean),
      bca_ub = quantile(bca_mean, probs = 1 - alpha / 2),
      bca_lb = quantile(bca_mean, probs = alpha / 2),
      .groups = "drop"
    ) |>
    rename(group = gates_group)
  return(bca_tab)
}

get_BCA(genML)