Causal isotonic calibration for heterogeneous treatment effects

Area of Science:

• Causal inference
• Machine learning
• Statistical modeling

Background:

• Accurate estimation of heterogeneous treatment effects is crucial for personalized decision-making.
• Existing calibration methods often require large datasets or separate validation sets, limiting their practical application.
• Nonparametric approaches are needed to handle complex, unknown functional forms in treatment effect prediction.

Purpose of the Study:

• To propose a novel nonparametric method, causal isotonic calibration, for calibrating predictors of heterogeneous treatment effects.
• To introduce a data-efficient variant, cross-calibration, that eliminates the need for hold-out calibration sets.
• To establish theoretical guarantees for the proposed calibration methods.

Main Methods:

• Causal isotonic calibration: A nonparametric method for calibrating treatment effect predictors.
• Cross-calibration: A data-efficient variant using cross-fitting to avoid hold-out sets.
• Doubly-robust analysis: Theoretical framework to establish calibration rates under weak assumptions.

Main Results:

• Causal isotonic calibration and cross-calibration achieve fast, doubly-robust calibration rates.
• Performance relies on the accurate estimation of either the propensity score or outcome regression.
• The calibrator can be integrated with any black-box learning algorithm.

Conclusions:

• The proposed causal isotonic calibrator offers strong, distribution-free calibration guarantees.
• It preserves predictive performance while enhancing the reliability of heterogeneous treatment effect estimates.
• Cross-calibration provides a data-efficient alternative for practical implementation.