Check negative-control validity across confounding scenarios

Focuses on whether the negative-control estimators (COCA, PGC) hold as the controls' coverage of the confounder subspace drops and as the number of latent confounders grows. Reports their bias/RMSE alongside IV2SLS, and flags scenarios where identification is not credible — in particular when the number of latent confounders exceeds the effective number of valid controls (the proximal-inference completeness condition).

Usage

nc_validity_check(
  trained_gan = NULL,
  coverage_grid = c(0.2, 0.5, 0.8, 1),
  k_grid = c(1, 2, 3),
  conf_strength = 0.8,
  n_valid_controls = 1,
  n_iter = 50,
  n_samples = 500,
  n_features = 20,
  nc_model = "proxy",
  base_seed = 800,
  n_cores = 1
)

Arguments

trained_gan

An iconic_gan (or NULL).

coverage_grid

Coverage values to sweep. Default c(0.2, 0.5, 0.8, 1).

k_grid

Numbers of confounders. Default c(1, 2, 3).

conf_strength

Fixed confounding strength. Default 0.8.

n_valid_controls

Number of distinct valid controls the design provides (for the identifiability flag). Default 1.

n_iter, n_samples, n_features, nc_model, base_seed, n_cores

As in gan_sensitivity().

Value

A list with summary (COCA/PGC/IV2SLS bias & RMSE per scenario, with an identified flag) and verdict (short per-scenario diagnosis).

Details

The matrix-bridge PGC is the estimator for which the completeness condition is binding: when k > n_valid_controls, PGC is under-identified while IV2SLS remains unbiased (it does not depend on NC completeness).

Examples

chk <- nc_validity_check(NULL, coverage_grid = 0.7, k_grid = 1,
  n_iter = 2, n_samples = 100, n_features = 5)
chk$verdict
#>       coverage k identified nc_abs_bias                 diagnosis
#> 0.7.1      0.7 1       TRUE       0.307 negative controls holding