High-Dimensional Confounding Adjustment Using Continuous Spike and Slab Priors.

Area of Science:

• Biostatistics
• Epidemiology
• Causal Inference

Background:

• Estimating causal effects in observational studies requires adjusting for confounding variables.
• High-dimensional confounders (p > n) pose challenges for traditional statistical methods.
• Existing dimension reduction techniques often prioritize outcome prediction over causal effect estimation.

Purpose of the Study:

• To develop a novel statistical approach for robust causal effect estimation in the presence of high-dimensional confounding.
• To address limitations of standard penalization methods that can introduce bias when confounders have differential associations with treatment and outcome.
• To improve the accuracy of causal inference in complex observational datasets.

Main Methods:

• Proposed a Bayesian approach using continuous spike and slab priors on regression coefficients for potential confounders.
• Developed a prior distribution designed to avoid excessive shrinkage of coefficients for variables strongly associated with treatment but weakly with outcome.
• Compared the proposed method against state-of-the-art techniques using simulations and real-world data.

Main Results:

• The proposed method effectively reduces confounding bias in high-dimensional settings.
• Demonstrated the ability to appropriately shrink coefficients of instrumental variables towards zero.
• Achieved good coverage rates for causal effect estimates, even with small sample sizes.

Conclusions:

• The novel spike and slab prior approach offers a powerful tool for causal inference in high-dimensional observational studies.
• This method provides a more reliable estimation of causal effects compared to existing penalization techniques.
• Application to NHANES data highlights its utility in assessing environmental exposures, like pesticide exposure, on health outcomes such as triglyceride levels.