Bounds and E-values for Marginal Causal Effects.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Unmeasured confounding poses a significant challenge in estimating causal effects from observational data.
• Existing methods, like Ding and VanderWeele's E-value, are primarily for conditional effects and can be impractical for marginal effects in high-dimensional settings.
• There is a need for more accessible and robust methods to quantify unmeasured confounding for marginal causal effects.

Purpose of the Study:

• To propose novel bounds for marginal causal effects that are more practical and less conservative than previous methods.
• To develop an E-value analog for marginal causation that is easier to implement.
• To demonstrate the estimation and application of these new bounds using standard statistical techniques.

Main Methods:

• Developed new bounds for marginal causal effects utilizing sensitivity parameters from Ding and VanderWeele.
• Reduced dimensionality by requiring only maximal sensitivity parameter values across confounder levels.
• Proposed a natural E-value analog for marginal causation.
• Demonstrated estimation using standard regression techniques.

Main Results:

• The proposed bounds are often narrower than existing bounds.
• The method effectively reduces dimensionality by simplifying sensitivity parameter specification.
• The bounds translate naturally into an E-value for marginal causation.
• The approach is applicable to high-dimensional data and can be estimated using standard regression.

Conclusions:

• The novel bounds offer a more practical and less conservative approach to assessing unmeasured confounding in marginal causal effect estimation.
• This method provides a valuable tool for researchers using observational data, enhancing the reliability of causal inference.
• The developed E-value for marginal causation simplifies interpretation and application in epidemiological studies.