A Sensitivity Analysis Framework Using the Proxy Pattern-Mixture Model for Generalization of Experimental Results.

Area of Science:

• Biostatistics
• Clinical Trials
• Epidemiology

Background:

• Generalizing findings from randomized controlled trials (RCTs) to broader populations is often hindered by unmeasured factors influencing participation and outcomes.
• Non-random selection mechanisms can introduce significant bias into treatment effect estimates from RCTs.

Purpose of the Study:

• To introduce a novel sensitivity analysis framework, the Proxy Pattern-Mixture Model in the context of RCTs (RCT-PPMM), for assessing the impact of unmeasured factors on treatment effect generalization.
• To quantify potential bias in treatment effect estimates arising from nonignorable selection mechanisms using proxy variables.

Main Methods:

• The RCT-PPMM framework utilizes proxy variables derived from baseline covariates to assess bias.
• It employs two bounded sensitivity parameters to quantify deviations from random sample selection.
• The method requires only summary-level baseline covariate data from the target population, enhancing applicability.

Main Results:

• Simulations demonstrate RCT-PPMM's ability to indicate the direction of bias and provide credible intervals that capture true treatment effects under nonignorable selection.
• The framework proved effective in various nonignorable selection scenarios.

Conclusions:

• RCT-PPMM offers a practical and interpretable tool for evaluating the generalizability of RCT findings.
• The method is particularly useful when individual-level data on non-participants are unavailable but summary-level covariate data are accessible.
• The study illustrates how conclusions from RCTs can be affected by plausible selection biases.