Demo dataset
Observational
Synthetic credit-risk study with modeled confounding and known treatment-effect heterogeneity.
Data and study
Run settings
42, 43, 44
2 folds × 3 runs
Fixed seeds provide reproducible repeated estimates and confidence intervals.
Study configuration
Quick evaluates one configuration per model. Larger grids provide broader performance or interpretability searches.
Causal graph and assumptions
Baseline covariates may affect treatment assignment and outcome; treatment may affect outcome.
T-HUG methodology
Control and treatment outcome models share adaptive bins and feature semantics. Utilities, patterns, coefficients, and the downstream LR/RPTE branch remain arm-specific.
Repeated out-of-fold predictions support stability intervals. A cross-fitted propensity model provides the DR/AIPW estimate and trimming sensitivity.
Analysis complete · 3,200 rows · 3 models
Credit intervention · Grid: quick · Selection: ROC AUC
T-HUG ATE
−1.39%
Primary analysis
Mean control risk
26.66%
Mean treatment risk
25.27%
Estimated to benefit
50.9%
Treatment-effect distribution
Repeated cross-fitted estimates
Repeated OOF ATE
−1.47%
95% CI (−2.12%, −0.82%)
DR/AIPW ATE
−1.53%
95% CI (−2.75%, −0.31%)
Every row receives out-of-fold predictions in each seeded run.
Rows
3,200
Control
1,771
Treatment
1,429
Overlap .05–.95
96.8%
Treatment assignment and overlap
Overlap sensitivity
| Range | Retained | DR ATE (95% CI) |
|---|---|---|
| Untrimmed | 100.0% | −1.53% (−2.75%, −0.31%) |
| .01–.99 | 99.4% | −1.49% (−2.66%, −0.32%) |
| .05–.95 | 96.8% | −1.42% (−2.51%, −0.33%) |
| .10–.90 | 88.2% | −1.38% (−2.42%, −0.34%) |
T-HUG structural regions
Shared thresholds keep control and treatment patterns directly comparable.
| Region / HUG pattern | Population | Mean P0 | Mean P1 | Mean CATE |
|---|---|---|---|---|
| credit_util=[0.72,0.91) & credit_score=[540,620) | 318 | 35.8% | 27.4% | −8.4% |
| prior_delinquency=1 & dti=[0.41,0.58) | 226 | 42.1% | 35.9% | −6.2% |
| annual_income=[42000,61000) & age=[28,42) | 401 | 24.7% | 23.3% | −1.4% |
Baseline comparison
All methods use the same covariates, split policy, scoring rule, and T-learner estimand.
| Model | Held-out AUC | Held-out Brier | Fit seconds | Oracle CATE RMSE | Oracle CATE corr |
|---|---|---|---|---|---|
| T-HUG | 0.688 | 0.173 | 3.35 | 0.076 | 0.104 |
| T-LR | 0.653 | 0.231 | 1.84 | 0.089 | 0.221 |
| T-XGB | 0.696 | 0.172 | 18.89 | 0.087 | −0.070 |
Best values are shown in bold; second-best values are italicized. Ranking is applied only where higher or lower has an objective interpretation.
Repeated-split stability and doubly robust estimates
| Model | Repeated OOF ATE (95% CI) | ATE repeat SD | DR/AIPW ATE (95% CI) | DR repeat SD |
|---|---|---|---|---|
| T-HUG | −1.47% (−2.12%, −0.82%) | 0.0026 | −1.53% (−2.75%, −0.31%) | 0.0049 |
| T-LR | −9.95% (−27.76%, 7.86%) | 0.0717 | 0.81% (0.07%, 1.55%) | 0.0030 |
| T-XGB | −0.83% (−3.05%, 1.38%) | 0.0089 | −2.01% (−3.54%, −0.48%) | 0.0061 |