Conditional cross-design synthesis estimators for generalizability in Medicaid.

Area of Science:

• Causal inference
• Health services research
• Biostatistics

Background:

• Internal and external validity are crucial for unbiased causal estimates in target populations.
• Generalizability methods are limited when target populations aren't well-represented by randomized studies alone.
• Combining randomized and observational data is needed for broader causal inference.

Purpose of the Study:

• To propose novel conditional cross-design synthesis estimators for generalizing causal quantities.
• To address biases from lack of overlap and unmeasured confounding when combining data sources.
• To estimate the causal effect of managed care plans on healthcare spending in New York City Medicaid beneficiaries.

Main Methods:

• Developed a novel class of conditional cross-design synthesis estimators.
• Incorporated outcome regression, propensity weighting, and double robust approaches.
• Utilized covariate overlap between randomized and observational data to mitigate unmeasured confounding.

Main Results:

• Found substantial heterogeneity in healthcare spending effects across different managed care plans.
• Demonstrated that unmeasured confounding is a greater concern than lack of overlap in this setting.
• Highlighted previously hidden spending variations within Medicaid managed care.

Conclusions:

• The proposed estimators effectively combine randomized and observational data for improved causal inference.
• Significant heterogeneity in spending effects has major implications for understanding Medicaid.
• Addressing unmeasured confounding is critical for accurate causal estimates in complex healthcare settings.