Partial Ordering Bayesian Logistic Regression Model for Phase I Combination Trials and Computationally Efficient Approach to Operational Prior Specification

TL;DR

This paper introduces a flexible Bayesian logistic model for Phase I combination trials, improving performance with randomization and proposing a computationally efficient calibration method that reduces simulation costs significantly.

Contribution

It develops the POBLRM, a more adaptable model for combination trials, and introduces cyclic calibration to drastically cut computational efforts.

Findings

POBLRM performs comparably to POCRM in non-randomized settings.

Randomization between experimental and control groups improves performance.

Cyclic calibration reduces computational cost by over 500 times.

Abstract

Recent years have seen increased interest in combining drug agents and/or schedules. Several methods for Phase I combination-escalation trials are proposed, among which, the partial ordering continual reassessment method (POCRM) gained great attention for its simplicity and good operational characteristics. However, the one-parameter nature of the POCRM makes it restrictive in more complicated settings such as the inclusion of a control group. This paper proposes a Bayesian partial ordering logistic model (POBLRM), which combines partial ordering and the more flexible (than CRM) two-parameter logistic model. Simulation studies show that the POBLRM performs similarly as the POCRM in non-randomised settings. When patients are randomised between the experimental dose-combinations and a control, performance is drastically improved. Most designs require specifying hyper-parameters, often chosen from statistical considerations (operational prior). The conventional "grid search'' calibration approach requires large simulations, which are computationally costly. A novel "cyclic calibration" has been proposed to reduce the computation from multiplicative to additive. Furthermore, calibration processes should consider wide ranges of scenarios of true toxicity probabilities to avoid bias. A method to reduce scenarios based on scenario-complexities is suggested. This can reduce the computation by more than 500 folds while remaining operational characteristics similar to the grid search.