Causal inference with cross-temporal design.

Area of Science:

• Causal inference
• Health services research
• Biostatistics

Background:

• Randomized trials face challenges with participant non-compliance.
• Randomized encouragement designs offer a solution but are not always feasible for policy interventions.
• Observational data often requires alternative causal inference methods.

Purpose of the Study:

• To propose a cross-temporal design that mimics randomized encouragement experiments using observational data.
• To develop and evaluate Bayesian procedures for estimating causal effects under this design.
• To assess the impact of Medicare Advantage enrollment on skilled nursing facility re-hospitalization.

Main Methods:

• Developed a cross-temporal design using time to simulate randomized encouragement.
• Replaced exclusion restrictions with temporal assumptions to address confounding trends.
• Implemented Bayesian procedures for causal effect estimation and compared with instrumental variables and matching methods.
• Applied the method to analyze Medicare Advantage program expansion (2011-2017).

Main Results:

• The proposed Bayesian approach demonstrated superior estimation accuracy compared to instrumental variables and matching methods in simulations.
• The Bayesian method showed robustness to violations of common trends assumptions.
• Analysis of Medicare Advantage expansion indicated its effect on 30-day re-hospitalization risk for skilled nursing facility residents.

Conclusions:

• The cross-temporal design provides a viable alternative for estimating causal effects from observational data when randomized encouragement is not possible.
• Bayesian procedures offer an accurate and robust method for causal inference in such settings.
• The study provides valuable insights into the effects of Medicare Advantage on healthcare utilization patterns.