Predictive Probabilities Made Simple: A Fast and Accurate Method for Clinical Trial Decision Making

TL;DR

This paper introduces a fast approximation method for Bayesian predictive probabilities in clinical trials, significantly reducing computational effort while maintaining accuracy across various endpoint types and analysis strategies.

Contribution

The authors propose a novel approximation technique using p-values or posterior probabilities that simplifies predictive probability calculations in clinical trial monitoring.

Findings

High concordance with Monte Carlo methods

Effective across multiple endpoint types

Reduces computational burden substantially

Abstract

Bayesian predictive probabilities are commonly used for interim monitoring of clinical trials through efficacy and futility stopping rules. Despite their usefulness, calculation of predictive probabilities, particularly in pre-experiment trial simulation, can be a significant challenge. We introduce an approximation for computing predictive probabilities using either a p-value or a posterior probability that significantly reduces this burden. We show the approximation has a high degree of concordance with standard Monte Carlo imputation methods for computing predictive probabilities, and present five simulation studies comparing the approximation to the full predictive probability for a range of primary analysis strategies: dichotomous, time-to-event, and ordinal endpoints, as well as historical borrowing and longitudinal modeling. We find that this faster method of predictive probability approximation works well in all five applications, thus significantly reducing the computational burden of trial simulation, allowing more virtual trials to be simulated to achieve greater precision in estimating trial operating characteristics.