Covariate Selection in Propensity Scores Using Outcome Proxies.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Propensity scores (PSs) are crucial for reducing bias in observational studies.
• Covariate selection for PSs is challenging, especially without outcome data.
• Effective covariate selection impacts the validity of treatment effect estimation.

Purpose of the Study:

• To develop practical methods for covariate selection in propensity score analysis.
• To approximate covariate-outcome relationships without using observed outcomes.
• To enhance the scientific basis for causal inference through improved covariate selection.

Main Methods:

• Examined the use of proximal pretreatment outcome measures for covariate selection.
• Investigated the utility of cross-validation in informing covariate-outcome relationships.
• Focused on methods that retain propensity scores as a design tool.

Main Results:

• Outcome proxies and cross-validation can empirically inform covariate-outcome relationships.
• These methods aid in assessing covariates' bias reduction or amplification capacities.
• Substantive knowledge is augmented by empirical evidence for better covariate selection.

Conclusions:

• Propensity score covariate selection should prioritize bias reduction and a scientific basis for inference.
• Using outcome proxies or cross-validation improves covariate selection and estimation.
• These approaches provide an evidentiary basis for causal inference in observational research.