Sensitivity analysis of unmeasured confounding in causal inference based on exponential tilting and super learner

Area of Science:

• Epidemiology
• Biostatistics
• Statistical Inference

Background:

• Causal inference heavily relies on the strongly ignorable treatment assumption, often violated in observational studies.
• Unmeasured confounding presents a significant challenge, potentially biasing causal effect estimates.
• Existing sensitivity analysis methods can be restrictive and difficult to implement due to complex parameterization.

Purpose of the Study:

• To develop a new sensitivity analysis method to assess the impact of unmeasured confounders in causal inference.
• To overcome limitations of existing methods by offering a more flexible and user-friendly approach.
• To improve the reliability of causal effect estimation in the presence of unmeasured confounding.

Main Methods:

• Leveraging doubly robust estimators, the exponential tilt method, and the super learner algorithm.
• The exponential tilting method avoids restrictive assumptions on the unmeasured confounder's structure.
• Incorporating the super learner algorithm for nonparametric estimation to mitigate modeling bias.

Main Results:

• The proposed method offers a sensitivity analysis without imposing structural constraints on unmeasured confounders.
• Utilizing super learner reduces parametric modeling bias inherent in traditional methods.
• The new approach features a univariate sensitivity parameter, simplifying interpretation and application.

Conclusions:

• The novel sensitivity analysis method provides a robust and flexible tool for evaluating unmeasured confounding.
• Its simplified parameterization enhances practical usability for researchers in observational studies.
• This method contributes to more reliable causal inference by systematically assessing potential biases.