Compute diagnostics for fitted localized estimators

Returns a compact diagnostics table for supported fit objects. Per-target DNN diagnostics report fixed top-1, top-5, and top-10 weight concentration shares. Distance diagnostics such as radius_rank_k and radius_rank_s are ordinary nearest-rank scale summaries; radius_rank_s is not the DNN support radius. For exact DNN fits the finite-population support edge is rank n - s + k.

Usage

# S3 method for class 'cate_result'
diagnostics(object, candidates = "default", ...)

diagnostics(object, candidates = "default", ...)

# Default S3 method
diagnostics(object, candidates = "default", ...)

# S3 method for class 'dnn_diagnostics'
print(x, ...)

# S3 method for class 'np_result'
diagnostics(object, candidates = "default", ...)

# S3 method for class 'plqr_result'
diagnostics(object, candidates = "default", ...)

Arguments

object: A fitted estimator object, currently an object returned by np_dnn(), cate_dnn(), or plqr_dnn().
candidates: Candidate rows to report for grid fits. The default "default" preserves the per-target diagnostics table for the selected fit. Use "all" or a character vector of candidate ids to report candidate-level grid diagnostics from object$tuning_grid. Candidate grid diagnostics include top5_weight_share; the cutoff is recorded as top_n = 5 and is independent of the estimator's k.
...: Not used.
x: A dnn_diagnostics object.

Value

An object of class dnn_diagnostics.

x, invisibly.

Details

PLQR fits additionally report residualized-design Gram conditioning (gram_min_eig, gram_condition_number) and, when direct-sandwich coefficient inference is available, target-level Jacobian conditioning: jacobian_available, jacobian_min_singular_value, jacobian_condition_number, jacobian_rank, density summaries over the weight support (density_min_support, density_max_support, plus the weighted density_mean_weighted and density_median_weighted), and theta_vcov_available. A target can have an acceptable residualized Gram but an ill-conditioned density-weighted Jacobian, so coefficient-inference reliability should be judged from the Jacobian block.

Experimental methods

Methods for plqr_result objects are experimental, as is all PLQR functionality in this release; see plqr_dnn().

Examples

set.seed(1)
x_mat <- matrix(rnorm(40 * 2), 40, 2)
y_vec <- x_mat[, 1] + rnorm(40, sd = 0.2)
fit_np <- np_dnn(X = x_mat, x0 = c(0, 0), Y = y_vec, se = FALSE)
diagnostics(fit_np)
#> <dnn_diagnostics>
#>   targets: 1
#>   n_eff  : min=11.4527, median=11.4527, max=11.4527
#>   weights: max(top5 share)=0.5771, max(weight)=0.15
#>   dist   : available for 0/1 target(s)
#>   outcome: available for 1/1 target(s)

demo_cate <- cate_demo_data(n = 50, p = 2, seed = 2)
fit_cate <- cate_dnn(
  data = demo_cate$data,
  x_cols = c("X1", "X2"),
  y_col = "Y",
  t_col = "D",
  m_col = "m_oracle",
  e_col = "e_oracle",
  x0 = demo_cate$x0,
  se = FALSE
)
diagnostics(fit_cate)
#> <dnn_diagnostics>
#>   targets: 1
#>   n_eff  : min=12.4648, median=12.4648, max=12.4648
#>   weights: max(top5 share)=0.5457, max(weight)=0.14
#>   dist   : available for 0/1 target(s)
#>   overlap: available for 1/1 target(s)

demo_plqr <- plqr_demo_data(n = 60, p = 2, p_d = 1, seed = 3)
fit_plqr <- plqr_dnn(
  data = demo_plqr$data,
  x_cols = c("X1", "X2"),
  D_cols = "D1",
  g_col = "g_oracle",
  mD_cols = "mD1",
  y_col = "Y",
  theta_eval = demo_plqr$truth$theta_true,
  x0 = demo_plqr$x0,
  se = FALSE
)
diagnostics(fit_plqr)
#> <dnn_diagnostics>
#>   targets: 1
#>   n_eff  : min=13.2383, median=13.2383, max=13.2383
#>   weights: max(top5 share)=0.5241, max(weight)=0.1333
#>   dist   : available for 0/1 target(s)
#>   gram   : available for 1/1 target(s)