Variable selection and estimation in causal inference using Bayesian spike and slab priors.

Area of Science:

• Biostatistics
• Causal Inference
• Observational Data Analysis

Background:

• Accurate causal effect estimation from observational data requires adjusting for confounding variables.
• Standard methods may fail to identify all relevant confounders, impacting unbiased estimation.
• Covariate selection often prioritizes outcome prediction over treatment association.

Purpose of the Study:

• To introduce a novel bilevel spike and slab causal estimator (BSSCE) for unbiased causal effect estimation.
• To develop a Bayesian approach that simultaneously models outcome and treatment assignment.
• To improve upon existing methods by minimizing the mean squared error of treatment effect estimation.

Main Methods:

• Utilizing a Bayesian framework with spike and slab priors on covariate coefficients.
• Simultaneously fitting models for both the outcome and treatment assignment.
• Deriving theoretical properties of the treatment effect estimator to justify the chosen priors.

Main Results:

• BSSCE demonstrated substantial reduction in mean squared error compared to existing methods.
• The method performs robustly, particularly with a high number of covariates, even exceeding sample size.
• Causal effect of vasoactive therapy versus fluid resuscitation on hypotensive episode length was estimated.

Conclusions:

• BSSCE offers a reliable and straightforward method for causal inference from observational data.
• The bilevel approach effectively addresses confounding by considering both outcome and treatment models.
• This method shows significant promise for complex datasets, including critical care data.