Robust estimation of heterogeneous treatment effects using electronic health record data

Area of Science:

• Causal Inference
• Precision Medicine
• Biostatistics

Background:

• Estimating heterogeneous treatment effects is crucial for precision medicine.
• Existing causal inference methods often struggle with data irregularity and high dimensionality common in electronic health record (EHR) data.
• The performance of traditional learners under square error loss in complex EHR data is understudied.

Purpose of the Study:

• To develop a unified framework for estimating heterogeneous treatment effects that accommodates data challenges in EHR analysis.
• To incorporate robust regression techniques and dimension reduction into causal inference learners.
• To propose novel estimators suitable for high-dimensional and irregular EHR data.

Main Methods:

• A general formulation unifying existing learners via a common score function.
• Incorporation of least absolute deviation (LAD) regression and dimension reduction techniques.
• Development of two specific estimators using weighted LAD with penalties for sparsity and smoothness.

Main Results:

• The proposed unified formulation and estimators demonstrate asymptotic normality under mild regularity conditions.
• Simulation studies confirm the robustness of the new methods to outliers compared to existing approaches.
• The methods were successfully applied to assess antihypertensive therapies' blood pressure-lowering effects.

Conclusions:

• The novel framework and proposed estimators offer a robust and flexible approach for heterogeneous treatment effect estimation in EHR data.
• These methods enhance the reliability of causal inference in complex, real-world healthcare datasets.
• The findings have implications for advancing precision medicine through improved analysis of clinical data.