Summarise raw per-replicate CATE simulation results

Aggregates a data.frame of raw replicate-level results (as returned by sim_cate_run()) to one summary row per (method, design, n, kx, candidate_id, lambda, s) cell and, when present, per GRF nuisance configuration. Reports bias, RMSE, mean SE, empirical coverage, mean CI length, empirical sampling dispersion of tau_hat, the ratio of mean SE to empirical dispersion, and (optionally) mean effective sample size. When the raw rows carry localization diagnostics, the summary also reports mean realized localization drift, the mean estimator remainder, physical neighborhood radius, arm-specific effective sample sizes, and shape diagnostics for the studentized statistic. Failed or skipped rows are excluded from all averages and dispersion calculations; their count is reported in n_failed. Here localized_truth is the realized DNN-weighted truth, not the deterministic population branch. Under complete-DNN exchangeability its expectation equals \(\alpha_s(x_0)\), so mean realized localization drift estimates the population branch displacement. The estimator remainder contains centered score variation and fitted-state effects.

Usage

summarise_sim_results(raw)

Arguments

raw

A data.frame with at least the columns method, design, n, kx, candidate_id, lambda, s, bias, sq_err, se, cover, ci_len, and status. Optional columns neff, tau_hat, and tau0 (returned by sim_cate_run()) are handled gracefully: neff is averaged into the summary; tau_hat is summarized through its empirical standard deviation; tau0 remains a per-rep quantity and is not aggregated. When present, method_variant, forest_nuisance, y_hat_supplied, and w_hat_supplied extend the grouping key. Older raw objects without these fields receive typed missing metadata; no nuisance route is inferred.

Value

A data.frame with one row per unique group and columns method, design, n, kx, candidate_id, lambda, s, n_reps, n_failed, bias, rmse, se_mean, cover, ci_len, emp_sd, se_to_sd, bias_to_se, localization, support, and studentized-shape diagnostics (and neff when present in raw). The output always includes the four GRF route fields, with typed missing values when they were absent from raw or do not apply.

See also

sim_cate_run()

Examples

if (FALSE) { # \dontrun{
raw <- sim_cate_run(designs = "dgp1_linear", n_vals = 500L, n_reps = 10L,
                    methods = c("dnn_oracle", "ols_interact"))
summarise_sim_results(raw)
} # }