Novel Bayesian approaches for the planning and monitoring phases of clinical studies

Randomized clinical trials are of paramount importance in medical research and are particularly valuable in understanding causal relationships, as confounding is removed through the randomization process. With new designs being developed and the use of multicenter studies increasingly common, trials are growing in complexity, and their associated cost has been increasing year on year for the past two decades. Unfortunately, many trials fail, perhaps through poor planning (underestimating sample size needs) or the inability to meet accrual targets. This thesis considers aspects of these challenges in planning and monitoring, developing new Bayesian approaches to sample size calculations for a multi-stage randomization design and accrual monitoring for multicenter studies.A novel trial design that has earned the spotlight in the growing field of precision medicine is the sequential multiple assignment randomized trial (SMART). Within this design, patients are randomized at two or more key treatment stages accounting for a small set of characteristics or responses to previous interventions. This structure allows for the development and comparison of adaptive treatment strategies. Most of the primary analyses performed on SMARTs are based on the comparison of two means or strategies, and while the frequentist sample size formulae are similar to traditional randomized controlled trials (RCTs), their estimation relies on additional assumptions. In the first manuscript, I developed a more robust sample size methodology in the Bayesian framework by adapting the `two priors' approach to the SMART design while incorporating estimates of the variance components and their uncertainty from pilot studies, resulting in a methodology that relies on fewer assumptions, is more robust to model misspecification, and allows for the incorporation of pre-trial knowledge. The performance of this approach is compared to the frequentist formulae in a simulation study. Its properties are further displayed in a case study where I used data from a SMART pilot to estimate the sample size of its full-scale version.In the second manuscript, I turn my attention from the planning to the monitoring phase, developing a novel approach to forecasting enrollments in multicenter studies applicable to both cohort and trial designs. The forecasting of recruitments is a topic of rapidly increasing interest, however, most models used in practice are either deterministic or rely on often unrealistic assumptions, such as the constant recruitment intensity over time. The most popular methodology that belongs to the second group is the Poisson-Gamma (PG) model. I extended this methodology by allowing the enrollment rates to vary with time after the opening of recruitment centers up to a stabilization point. I illustrate the accuracy of this methodology compared to the standard PG model in a simulation study and by forecasting the enrollments in the Canadian Co-infection Cohort study.The third manuscript aims to further validate the proposed recruitment model on data from randomized trials and offer a practical guide for its use. One of the main hurdles to the adoption of statistical models to forecast enrollments in practice lies in the difficulty of their implementation. In this manuscript, I outline how to implement the time-dependent PG model to predict the recruitment process via the newly developed tPG R package. The model is further validated on the recruitment data from two HIV trials

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.