Leveraging historical data into oncology development programs: Two case studies of phase 2 Bayesian augmented control trial designs. (5th January 2020)

• Record Type: Journal Article
• Authors: Smith, Claire L.; Thomas, Zachary; Enas, Nathan; Thorn, Katharine; Lahn, Michael; Benhadji, Karim; Cleverly, Ann
• Is Part Of: Pharmaceutical statistics. Volume 19:Number 3(2020)
• Page Start: 276
• Page End: 290
• Publication Date: 2020-01-05
• Subjects: Bayesian -- borrowing -- covariates -- historical data; Pharmacy -- Statistical methods -- Periodicals; Pharmacy -- Statistics -- Periodicals; 615.10727
• DOI: 10.1002/pst.1990
• Languages: English
• ISSNs: 1539-1604

Abstract

SUMMARY: Leveraging historical data into the design and analysis of phase 2 randomized controlled trials can improve efficiency of drug development programs. Such approaches can reduce sample size without loss of power. Potential issues arise when the current control arm is inconsistent with historical data, which may lead to biased estimates of treatment efficacy, loss of power, or inflated type 1 error. Consideration as to how to borrow historical information is important, and in particular, adjustment for prognostic factors should be considered. This paper will illustrate two motivating case studies of oncology Bayesian augmented control (BAC) trials. In the first example, a glioblastoma study, an informative prior was used for the control arm hazard rate. Sample size savings were 15% to 20% by using a BAC design. In the second example, a pancreatic cancer study, a hierarchical model borrowing method was used, which enabled the extent of borrowing to be determined by consistency of observed study data with historical studies. Supporting Bayesian analyses also adjusted for prognostic factors. Incorporating historical data via Bayesian trial design can provide sample size savings, reduce study duration, and enable a more scientific approach to development of novel therapies by avoiding excess recruitment to a control arm. Various sensitivity analyses are necessary to interpret results. Current industry efforts for data transparency have meaningful implications for access to patient‐level historical data, which, while not critical, is helpful to adjust for potential imbalances in prognostic factors.