Regression-Based Proximal Causal Inference.

Area of Science:

• Epidemiology
• Biostatistics
• Observational Studies

Background:

• Negative controls are crucial for assessing unmeasured confounding in observational research.
• Proximal causal inference (PCI) aims to reduce bias in causal effect estimates using control variables.
• Existing formal PCI methods involve complex, ill-posed integral equations, hindering practical application.

Purpose of the Study:

• To develop a simplified, regression-based proximal causal inference (PCI) method.
• To enable de-biasing of confounded causal effect estimates in observational studies.
• To provide an accessible implementation of PCI using generalized linear models (GLMs).

Main Methods:

• Developed a novel regression-based PCI approach utilizing two-stage generalized linear regression models (GLMs).
• This method bypasses the need for solving complex integral equations inherent in traditional PCI.
• The approach is designed for applicability to continuous, count, and binary outcome data.

Main Results:

• The regression-based PCI method is shown to be statistically sound.
• Demonstrated the approach's effectiveness through simulations and real-world empirical analyses.
• The method offers a practical alternative for de-biasing causal estimates.

Conclusions:

• Regression-based PCI provides a computationally tractable and broadly applicable method for causal inference.
• Its ease of implementation with standard GLM software facilitates wider adoption in observational research.
• This approach enhances the ability to obtain reliable causal effect estimates, even in the presence of confounding.