Slowly but surely, Bayesian ideas revolutionize medical research
Donald A. Berry (University of Texas M.D. Anderson Cancer Center, USA)
Bayesian theory is elegant and intuitive. But elegance may have little value in practical settings. The “Bayesian Revolution” of the last half of the 20th century was irrelevant for biostatisticians. They were busy changing the world in another way, and they neither needed nor wanted more methodology than they already had. The randomized controlled trial (RCT) came into existence in the 1940s and it changed medical research from an art into a science, with biostatisticians guiding the process. To make matters worse for the reputation of Bayesians, we seemed to be anti-randomization, and medical researchers feared we wanted to return them to the dark ages.
The standard approach to clinical experimentation is frequentist, which has advantages and disadvantages. One disadvantage is that unit of statistical inference is the entire experiment. As a consequence, the RCT has remained largely unchanged. It is still the gold standard of medical research, but it can make research ponderously slow. And it is not ideally suited for the “personalized medicine” approach of today, identifying which types of patients benefit from which therapies.
In this presentation, I’ll chronicle the increased use of the Bayesian perspective in medical research over this period. An important niche regards adaptive design. I’ll describe a variety of approaches, most of which employ randomization, and all employ Bayesian updating. Accumulating trial results are analyzed frequently with the possibility of modifying the trial’s future course based on the overall theme of the trial. It is possible to have many treatment arms. Including combination therapies enables learning how treatments interact with each other as well as the way they interact with biomarkers of disease that are specific to individual patients. I will give an example (called I-SPY 2) of a Bayesian adaptive biomarker-driven trial in neoadjuvant breast cancer. The goal is to efficiently identify biomarker signatures for a variety of agents and combinations being considered simultaneously. Longitudinal modeling plays a vital role.
Although the Bayesian approach supplies important tools for designing informative and efficient clinical trials, I’ve learned to not try to change things too abruptly. In particular, we can stay rooted in the well-established frequentist tradition by evaluating false-positive rates and statistical power using simulation. The most exciting aspect of this story is the potential for utilizing Bayesian ideas in the future to build ever more efficient study designs and associated processes for developing therapies, based on the existing solid foundation.