Model assisted sensitivity analyses for hidden bias with binary outcomes.

Area of Science:

• Medical and Health Sciences
• Epidemiology
• Biostatistics

Background:

• Observational studies are crucial for inferring causal relationships in health sciences but are susceptible to bias from unmeasured confounders.
• Sensitivity analysis is essential for evaluating the potential impact of unmeasured covariates on estimated causal effects.
• Existing frameworks for sensitivity analysis in matching designs can be complex to implement and interpret.

Purpose of the Study:

• To propose a novel model-assisted sensitivity analysis for binary outcomes in 1:k matching designs.
• To simplify the implementation and interpretation of sensitivity analyses for causal inference.
• To provide a closed-form representation for sensitivity parameters that ensure non-significant maximum p-values.

Main Methods:

• Developed a model-assisted sensitivity analysis incorporating a conditional logistic outcome model for binary data.
• Applied the method to a 1:k matching design, demonstrating equivalence to nonparametric approaches in large samples.
• Introduced sensitivity parameters to quantify the association between unmeasured covariates and exposure/outcome.

Main Results:

• The proposed method simplifies sensitivity analysis implementation and interpretation compared to conventional approaches.
• A closed-form solution was derived for the set of sensitivity parameters yielding non-significant results.
• The methodology demonstrated robustness and applicability in a real-world U.S. trauma care database analysis.

Conclusions:

• The model-assisted sensitivity analysis offers a more accessible and interpretable tool for assessing unmeasured confounding in matching designs.
• This approach enhances the reliability of causal effect estimates derived from observational health data.
• The method is adaptable for extensions to multilevel treatments and various matching designs.