Causal Isotonic Regression.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Observational studies face challenges with confounders distorting exposure-outcome relationships.
• Classical methods for continuous exposures often rely on restrictive parametric assumptions, risking model misspecification.
• Nonparametric estimation of causal dose-response curves is difficult due to slow convergence rates and sensitivity to tuning parameters.

Purpose of the Study:

• To propose a novel nonparametric estimator for causal dose-response curves that are known to be monotone.
• To develop an estimation procedure that avoids tuning parameters and is invariant to transformations of the exposure variable.
• To provide a method for valid statistical inference in complex causal models.

Main Methods:

• Developed a nonparametric estimator for monotone causal dose-response curves.
• Demonstrated generalization of the classical least-squares isotonic regression estimator.
• Investigated theoretical properties, including irregular limit distributions and doubly-robust inference.

Main Results:

• The proposed estimator is free of tuning parameters and robust to monotone transformations of the exposure.
• Theoretical analysis revealed its irregular limit distribution and potential for doubly-robust inference.
• Numerical studies confirmed the estimator's performance.

Conclusions:

• The new nonparametric method offers a robust and parameter-free approach to estimating causal dose-response curves.
• This method enhances causal inference in observational studies, particularly when dealing with continuous exposures.
• Applied to HIV vaccine trials, it assessed the BMI-immune response relationship, demonstrating practical utility.