Bayesian Nonparametric Model for Heterogeneous Treatment Effects With Zero-Inflated Data.

Area of Science:

• Biostatistics
• Precision Medicine
• Causal Inference

Background:

• Precision medicine aims to personalize treatments using individual patient data.
• Existing statistical models for treatment effect heterogeneity are sensitive to model specification and covariate selection.
• Zero-inflated outcome data are common in health studies, posing challenges for causal effect estimation.

Purpose of the Study:

• To propose a new Bayesian nonparametric (BNP) method for estimating heterogeneous causal effects in studies with zero-inflated outcome data.
• To address limitations of existing parametric and other BNP methods in handling covariate-dependent treatment effects.
• To evaluate the performance of the proposed method against existing approaches using simulation studies.

Main Methods:

• Developed a novel BNP method utilizing an enriched Dirichlet process (EDP) mixture.
• Linked outcome and covariate Dirichlet process mixtures for concurrent posterior distribution estimation.
• Applied the method to analyze the relationship between heart radiation dose and cardiac troponin T levels.

Main Results:

• The proposed BNP method demonstrated superior performance over two other BNP methods in simulations, reducing bias and mean squared error (MSE) for conditional average treatment effect estimates.
• The model effectively reflects uncertainty in regions with violated overlap conditions.
• The application to heart radiation dose data showed the method's utility in real-world health research.

Conclusions:

• The novel BNP method offers a robust approach for estimating heterogeneous causal effects, particularly in health studies with zero-inflated data.
• This method provides more reliable inference for individual causal effects and uncertainty quantification.
• The proposed approach advances precision medicine by enabling more accurate treatment effect evaluation across diverse patient subgroups.