# A Sensitivity Analysis Framework Using the Proxy Pattern–Mixture Model for Generalization of Experimental Results

## TL;DR

This paper introduces a new method to assess how unmeasured factors might bias the results of clinical trials when generalizing to a broader population.

## Contribution

The Proxy Pattern-Mixture Model (RCT-PPMM) provides a novel sensitivity analysis framework for evaluating generalizability in RCTs.

## Key findings

- RCT-PPMM uses proxy variables to quantify bias from nonignorable selection mechanisms.
- The method requires only summary-level baseline covariate data for the target population.
- Simulations show RCT-PPMM can reveal the direction of bias and credible intervals under various selection scenarios.

## Abstract

Generalizing findings from randomized controlled trials (RCTs) to a target population is challenging when unmeasured factors influence both trial participation and outcomes. We propose a novel sensitivity analysis framework to assess the impact of such unmeasured factors on treatment effect estimates called the Proxy Pattern‐Mixture Model in the context of RCTs (RCT‐PPMM). By leveraging proxy variables derived from baseline covariates, our framework quantifies the potential bias in treatment effect estimates due to nonignorable selection mechanisms. The RCT‐PPMM relies on two bounded sensitivity parameters that capture the deviation from sample selection at random and that can be varied systematically to determine how robust trial results are to a departure from ignorable sample selection. The approach only requires summary‐level baseline covariate data for the target population (not individual‐level data), thus increasing its applicability. Through simulations, we demonstrate that RCT‐PPMM can provide information about the potential direction of bias and provide credible intervals that capture the true treatment effect under various nonignorable selection scenarios. We illustrate the use of the method using a yoga intervention RCT for breast cancer survivors, illustrating how conclusions may shift under plausible selection biases. Our approach offers a practical and interpretable tool for evaluating generalizability, particularly when individual‐level data on nonparticipants are unavailable, but summary‐level covariate data are accessible.