Run one replicate of CATE estimators — sim_cate_one

Draws one dataset from the named DGP (via sim_cate_data()), then runs each requested method on the same dataset and returns per-replicate performance metrics aligned with the paper's evaluation criteria.

Usage

sim_cate_one_rep(
  n,
  kx,
  k,
  s,
  design,
  methods = "dnn_oracle",
  tuning = NULL,
  tuning_mode = c("default", "grid"),
  alpha = 0.05,
  seed = NULL,
  sigma = 0.5,
  clip = 0.001,
  n_folds = 5L
)

Arguments

n: Integer sample size.
kx: Integer covariate dimension ($\ge 2$).
k: Integer DNN neighbour count.
s: Integer kernel order (undersmoothing parameter). May be omitted when tuning is supplied.
design: Character DGP name; passed to sim_cate_data().
methods: Character vector of method names; see sim_cate_methods(). Default: "dnn_oracle".
tuning: Optional kernel-order tuning specification.
tuning_mode: One of "default" or "grid". "default" evaluates one scalar kernel order. "grid" requires tuning and emits one row per DNN candidate while non-DNN benchmarks remain one row per replicate.
alpha: Nominal coverage level (default 0.05).
seed: Optional integer seed for this replicate.
sigma: Noise SD forwarded to the DGP.
clip: Propensity clipping bound forwarded to the DGP and cross-fitting.
n_folds: Number of cross-fitting folds for the dnn_crossfit method.

Value

A data.frame with one or more rows per method and columns method, candidate_id, lambda, s, tau_hat, tau0, bias, sq_err, se, cover (0/1), ci_len, neff, and status. Methods based on GRF also record method_variant, forest_nuisance, and whether Y.hat and W.hat were supplied. The maintained causal_forest route uses externally supplied oracle nuisance inputs. with explicit localization weights also report the realized DNN-weighted truth, realized localization drift, estimator remainder, studentized error, physical 90-percent-mass radius, treatment-arm effective sample sizes, and the top-five score-contribution share. The localized_truth field is the finite-sample object $\sum_i W_{i,s}(x_0)\tau_0(X_i)$, not the deterministic population branch. With $x_0$ and $(n,s)$ fixed, expectation over the i.i.d. covariate sample and complete-DNN exchangeability give $$E(\mathrm{localized\_truth})=\alpha_s(x_0).$$ The historical branch_drift field name is retained for compatibility; it stores realized localization drift. The estimator remainder combines centered score variation with fitted-state effects. Failed methods have NA for all metric columns and status = "failed"; registered route metadata is retained when it is known from the selected method.

Examples

rep1 <- sim_cate_one_rep(
  n = 200, kx = 2, k = 1, s = 12,
  design = "dgp2_nonlinear",
  methods = c("dnn_fullsample", "ols_interact"),
  seed = 1
)
rep1
#>           method candidate_id lambda  s   tau_hat     tau0       bias
#> 1 dnn_fullsample         <NA>     NA 12 1.0398647 1.248121 -0.2082565
#> 2   ols_interact         <NA>     NA NA 0.7781338 1.248121 -0.4699874
#>       sq_err        se cover    ci_len     neff localized_truth branch_drift
#> 1 0.04337079 0.1810111     1 0.7095503 31.06479        1.115052   -0.1330695
#> 2 0.22088815 0.1239821     0 0.4860010       NA              NA           NA
#>   estimator_remainder studentized_error score_sd0 standardized_branch_drift
#> 1         -0.07518702         -1.150518  1.031413                -0.7190849
#> 2                  NA                NA        NA                        NA
#>   physical_radius_90 treated_ess control_ess top5_score_share method_variant
#> 1          0.7648358    20.04691    11.14627        0.3774768           <NA>
#> 2                 NA          NA          NA               NA           <NA>
#>   forest_nuisance y_hat_supplied w_hat_supplied status
#> 1            <NA>             NA             NA     ok
#> 2            <NA>             NA             NA     ok