Background: The higher cost of confirmatory trials and ethical considerations for patients with the existing treatment options led to the investigation of new trial designs. There have been multiple efforts to develop statistical methodologies to evaluate the impact of borrowing external and historical randomized clinical trials (RCT) data sources in a hybrid control design setting. However, few studies used real-world data and historical RCT while comparing Bayesian and frequentist approaches under similar conditions. We evaluated Bayesian (commensurate prior model) and frequentist (Frailty Cox model) hierarchical methods in hybrid control simulation. In addition, dynamic borrowing between the control arms of two historical RCTs in previously treated non-small cell lung cancer (NSCLC) was tested.
Methods: Bayesian and frequentist models were carried out and evaluated through computational simulations to mimic 300 repeated clinical trial experiments. Estimation of bias, power, and type I error was conducted. The critical value for testing of 5% (two-sided type I error) was used. Half-Cauchy hyperprior was chosen for the Bayesian commensurate approach. Normal prior for heterogeneity between RCT and the external data was used for Bayesian commensurate prior and Frailty Cox model.
Results: Model performance, in terms of estimator bias, power, and Type I error, was similar between Bayesian and frequentist approaches. Estimator bias of the two models was within 0.07 (Log(HR), |β^-β|). Type I errors estimated by the two methods ranged from 3% and 10%. Slight type I error inflation was observed when there were differences between data sources. Power to detect treatment effect using various scenarios ranged from 60% to 85%. Power gain was demonstrated with sufficient commensurability between RCT and external control (81% to 82% compared to 75% without borrowing) and potential power loss observed with more differences between data sources (58%-66% compared to 75% without borrowing). When applying this methodology with historical RCTs, results were consistent with simulation scenarios.
Conclusions: We demonstrated similar model performance for Bayesian and Frailty Cox models in estimating treatment effects in hybrid control simulation and existing RCT. While some differences in models’ setup, availability of standard software, and flexibility exists, a successful construct of hybrid control can lead to accelerated internal decision making and streamlined clinical trial design. However, many factors can affect treatment effect estimates and careful consideration of the differences between data sources, distribution of estimates, and direction of the bias has to be taken before implementing this approach in a clinical trial setting.
Citation Format: Fu-Chi Hsieh, Natalia Sadetsky, Xiao Li, Ruilin Li, Jiawen Zhu. Real-world data in hybrid clinical trial design: Considerations and impact on treatment effect [abstract]. In: Proceedings of the AACR Special Conference on Advancing Precision Medicine Drug Development: Incorporation of Real-World Data and Other Novel Strategies; Jan 9-12, 2020; San Diego, CA. Philadelphia (PA): AACR; Clin Cancer Res 2020;26(12_Suppl_1):Abstract nr 05.