Addressing Non-Exchangeability in Hybrid Control Studies: A Variable Selection Approach.

Area of Science:

• Biostatistics
• Clinical Trials
• Epidemiology

Background:

• Hybrid control designs merge randomized controlled trials (RCTs) with external data for treatment evaluation.
• While efficient, these designs risk bias due to potential non-exchangeability between internal and external control groups.
• Adjusting for baseline covariates can mitigate bias, but exchangeability assumptions must be carefully handled.

Purpose of the Study:

• To propose a variable selection approach for addressing non-exchangeability in hybrid control studies.
• To identify and adjust for covariate interactions that violate the exchangeability assumption.
• To improve the efficiency of hybrid control designs by appropriately incorporating external data.

Main Methods:

• Utilized an outcome regression model to represent non-exchangeability as covariate-external control indicator interactions.
• Employed the adaptive lasso for variable selection to identify significant interactions requiring adjustment.
• Applied g-computation with the fitted model to estimate treatment effects.

Main Results:

• Simulation results demonstrated the approach's ability to improve efficiency under specific conditions.
• The method successfully incorporated external control data even when full exchangeability was absent.
• Variable selection effectively distinguished between null and non-null interactions, guiding model adjustment.

Conclusions:

• The proposed variable selection method effectively addresses non-exchangeability in hybrid control studies.
• This approach allows for the efficient use of external control data while mitigating potential bias.
• The adaptive lasso and g-computation provide a robust framework for hybrid trial analysis.