+ Introduction to Meta-analysis with Application in Clinical Trials (TBD, 2025)
+ Revolutionizing Drug Development with Bayesian Adaptive Designs: Simple, Efficient, Superior (TBD, 2025)
Zhaohui Liu, Bristol Myers Squibb (BMS).Dr Liu is a Director in Global Biometrics and Data Science (GBDS) of BMS. After graduating from University of Connecticut, where he witnessed the MCMC breakthrough from its cradle and was taught by a few Bayesian champions, Zhaohui has been working in the US pharmaceutical industry, first at Novartis, and then at BMS, for more than 20 years with experience in all phases of clinical trials.
How to integrate clinical evidence from various sources to make coherent and robust statistical inference so that sporadic data points can inform decision makers a unified story with statistical rigor? Meta-analysis arose with this backdrop. In this short course we will discuss its key concepts, methods, in both frequentist and Bayesian approaches, focusing mainly with examples in the context of clinical trials.
In Part 1 of the course, basic concepts, inference, examples from clinical trials setting, and programming implementation will be introduced. After brief layout of the scenario necessitating meta-analysis, we will introduce key concepts and inference methods such as continuous, binary, and time-to-event data, fixed-effect model, random effect model, measures of heterogeneity, and a few further topics, e.g., prediction, treatment difference, subgroup analysis, power consideration etc., if time permits.
In Part 2, we will shift gear into Bayesian paradigm with basic concepts, key methods, and worked examples. The advantage of Bayesian approach is its versatility in leveraging historical information by informed or non-informed priors to account for uncertainty. Meta-Analytic-Predictive (MAP) method for informative prior will also be discussed for the continuous, binary, and time-to-event data.
In the last part, we will delve into topics like model checking, diagnostics and influence analysis. Statistics is hard science, but statisticians performing data analysis need to work as a skillful physician treating patients. Mechanistical manual following can automatically produce results. However, are these conclusions valid or robust? Will the latest reported clinical trial readout derail the conclusion if it were included in the meta-analysis? Off-shelf ready-to-use statistical software packages and timeline pressure for significant p-value may entice people to jump into conclusions immediately without carefully checking the assumptions. This in turn may either lead to inflated results or to discard potentially meaningful findings. How should a statistician in the clinical trial setting facing these challenges proceed?topics like model checking, diagnostics and influence analysis.
J. Jack Lee, Ph.D. is Professor of Biostatistics and Kenedy Foundation Chair in Cancer Research. His areas of statistical research include design and analysis of clinical trials, Bayesian adaptive designs, statistical computation/graphics, drug combination studies, and biomarkers identification and validation. He is an elected Fellow of American Statistical Association, Society for Clinical Trials, and American Association for the Advancement of Science. He is a Statistical Editor of Cancer Prevention Research and serves on the Statistical Editorial Board of Journal of the National Cancer Institute. He has more than 500 publications in statistical and medical journals. He co-authored two books entitled: “Bayesian Adaptive Methods for Clinical Trials” and “Model-Assisted Bayesian Designs for Dose Finding and Optimization: Methods and Applications.”
Department of Biostatistics, University of Texas MD Anderson Cancer Center
It is well known that the current drug development process is lengthy, costly, and has a high failure rate. How can we improve this from a statistical point of view? Current medical research is dominated by the frequentist approach, which relies on p-values, confidence intervals, null hypothesis significance testing, etc. This approach is rigid, addresses key questions indirectly, and violates the likelihood principle.
On the other hand, the Bayesian approach directly models the parameter of interest by forming its prior belief and using observed data to update this belief by constructing the posterior distribution. It conforms to the likelihood principle and provides a better statistical inferential framework. The Bayesian approach is intuitive, innately adaptive, can incorporate all available information for efficient estimation, and is uniquely suitable for drug development.
In this short course, we will first compare and contrast the frequentist and Bayesian approaches. We will highlight the strengths of the Bayesian approach, such as the importance of incorporating the prior distribution and using posterior probability for statistical inference. We will introduce Bayesian adaptive designs, particularly model-assisted designs, which are a new class of designs developed to simplify the implementation of adaptive designs in practice. Model-assisted designs are derived from rigorous statistical theory, thus possessing superior operating characteristics and great flexibility, while being easy to implement with pre-calculated boundaries. These designs follow the new KISS principle: Keep It Simple and Smart, making them ideal for drug development.
The main application areas include adaptive dose-finding, adaptive toxicity and efficacy evaluation, dose optimization, posterior probability, and predictive probability for interim monitoring of study endpoints, outcome-adaptive randomization, hierarchical models, multi-arm, multi-stage designs, and platform designs, among others. Useful model-assisted designs for early-phase clinical trials to determine the maximum tolerated dose (MTD) and optimal biological dose (OBD) will be introduced. These include the Bayesian Optimal INterval (BOIN) design, Time-To-Event (TITE)-BOIN design, BOIN-COMB, U-BOIN design, BOIN12 design, and TITE-BOIN12 design. The Bayesian Optimal Phase 2 (BOP2) design and 2-Arm BOP2 design will be introduced to evaluate simple and complex efficacy and/or toxicity endpoints.
We will also discuss Bayesian adaptive platform designs with applications in master protocols, including umbrella trials and basket trials. Lessons learned from real trial examples and practical considerations for conducting adaptive designs will be shared. Additionally, easy-to-use Shiny applications on the web and downloadable standalone programs will be introduced to facilitate the study design and conduct of Bayesian adaptive designs. They are freely available at https://trialdesign.org and https://biostatistics.mdanderson.org/softwaredownload/.
References: