Semiparametric sensitivity analysis: unmeasured confounding in observational studies.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Observational studies often face challenges in establishing cause-effect relationships due to potential unmeasured confounding.
• Assessing the robustness of conclusions from non-experimental studies to unmeasured confounders is critical for reliable scientific evidence.

Purpose of the Study:

• To generalize existing sensitivity analysis methods for estimating the average causal effect (ACE).
• To develop a robust statistical framework for causal inference from observational data, addressing unmeasured confounding.

Main Methods:

• Utilized semiparametric theory to derive the non-parametric efficient influence function for the ACE.
• Developed a one-step, split-sample, truncated estimator based on the derived influence function.
• The proposed estimator accommodates semiparametric models without restricting sensitivity parameters and ensures $\sqrt{n}$ asymptotics under sufficient conditions.

Main Results:

• The methodology provides a generalized sensitivity analysis framework for the average causal effect (ACE).
• A novel one-step estimator was developed and shown to have $\sqrt{n}$ asymptotic properties.
• The approach was applied to investigate the causal effect of smoking during pregnancy on birth weight, with performance evaluated via simulation.

Conclusions:

• The developed methodology offers a robust approach to causal inference in the presence of unmeasured confounding.
• The new estimator enhances the reliability of conclusions drawn from observational studies.
• The application to smoking and birth weight demonstrates the practical utility of the proposed causal inference technique.