RdistNN: DNN Estimation for CATE and Nonparametric Regression

The initial public release focuses on conditional average treatment effects with cate_dnn() and nonparametric regression with np_dnn(), using distributional nearest-neighbor (DNN) weights and delete-1 jackknife standard errors. Nuisance predictions for CATE are supplied externally. The underlying toolkit combines:

• distance specifications (dnn_distance) for ranking observations around a target;

• rank-weight specifications (rank_weight) for converting ranks to weights;

• score objects (ustat_score) for flexible kernel-based moment evaluation.

Main User Workflows

• Compute DNN weights with weights_dnn(), weights_dnn_exact(), and weights_dnn_asymp().

• Resolve paper-default kernel orders with tuning_dnn_paper() and tuning_eval().

• Estimate local CATE targets with cate_dnn().

• Summarize/predict CATE fits with summary.cate_result(), confint.cate_result(), and predict.cate_result().

• Estimate local nonparametric means with np_dnn().

• Use two-scale signed-weight regression with np_tdnn(); see its help for order constraints and additional smoothness requirements.

• Inspect fitted local estimators with diagnostics().

• Plot CATE and regression estimates with plot_estimates() and inspect local weights with plot_weights() (requires ggplot2).

Experimental and Advanced Workflows

All PLQR functionality is experimental, including plqr_dnn(), its scores, demo helpers, methods, and plots. The current wrapper evaluates moments at a supplied coefficient; it does not solve for local coefficients. Local IV-GMM functionality is also experimental. These workflows are outside the CATE and regression focus of the initial public release.

• Build custom generalized U-statistics with new_score(), score_eval_batch(), and score_diff_path().

• Estimate uncertainty with jackknife_fast_rank() and jackknife_slow_rank().

Documentation Map

• Start with vignette("getting-started", package = "RdistNN").

• For step-by-step extension patterns for custom distances, rank weights, and scores, see vignette("extending", package = "RdistNN").

• For modular kernel-order tuning rules and registry usage, see vignette("kernel-order-tuning", package = "RdistNN").

• For fitted-object diagnostics and reporting, see vignette("diagnostics-reporting", package = "RdistNN").

• For distance/rank weighting details see vignette("dnn-weights", package = "RdistNN").

• For estimator workflows see vignette("conditional-z-estimation", package = "RdistNN"), including experimental PLQR details.

• For jackknife details see vignette("jackknife-gen-ustats", package = "RdistNN").

See also

weights_dnn(), cate_dnn(), np_dnn(), plqr_dnn(), diagnostics(), summary.cate_result(), predict.cate_result(), new_score(), distance_l2()

Author

Maintainer: Jakob Juergens info@jakobjuergens.com

Authors:

• Jakob Juergens info@jakobjuergens.com