Propensity score methods for merging observational and experimental datasets.

Area of Science:

• Biostatistics
• Epidemiology
• Health Services Research

Background:

• Estimating causal treatment effects often requires large datasets.
• Randomized controlled trials (RCTs) provide high-quality data but are often limited in size.
• Observational data bases (ODBs) are large but prone to confounding.

Purpose of the Study:

• To develop and evaluate methods for merging limited RCT data into large ODBs.
• To improve the estimation of average causal treatment effects by leveraging both data sources.
• To address bias issues in combining data from different study designs.

Main Methods:

• Stratification based on effect moderators and propensity scores from the ODB.
• A "spiked-in" method inserting RCT data into ODB strata.
• A convex combination estimator blending ODB and RCT treatment effect estimates within strata.
• Bias assessment using the delta method and simulation studies.

Main Results:

• The "spiked-in" method can introduce bias.
• The convex combination estimator effectively ameliorates this bias.
• Application to hormone therapy (HT) data from the Women's Health Initiative.
• The combined approach yielded lower mean squared error (MSE) for estimating HT's causal impact on coronary heart disease compared to using either data source alone.

Conclusions:

• Merging RCT and ODB data can enhance causal effect estimation.
• The convex combination method offers a less biased approach than simple data insertion.
• This integrated strategy provides more accurate estimates of treatment effects, as demonstrated in the context of hormone therapy and cardiovascular health.