Why You Should Construct Primary Endpoints Using Bayesian Methods: Lessons from COVID-19
One of the revelations of the COVID-19 pandemic is that the flexibility and potential of Bayesian designs goes far beyond the benefits connected to informed priors. Rather a number of other benefits to Bayesian designs are sometimes under-appreciated. The importance of using Bayesian methods to choose composite endpoints, for example, particularly in longitudinal studies, can be overlooked when considering Bayesian and Frequentist options.
Cytel statisticians reflected on these benefits during a recent panel discussion.
During a recent panel on COVID-19 drug discovery, led by Cytel VP of Strategic Consulting Natalia Muhlemann, biostatisticians and former regulators reflected on certain misconceptions about Bayesian methods. First and foremost was the fact that the primary advantage of Bayesian designs is that they cut short trial timelines by employing informed priors, that is information about a new therapy gleaned from previous trials and related data.
While this is undoubtedly an important benefit, Bayesian methods can be used to optimize efficient trial design even when samples are small, data is missing or ‘fragmented’, and timelines are constrained.
One of the significant advantages of employing Bayesian methods is that they clear guidance on the choice of which primary endpoint to pursue during a trial. A COVID-19 study, for example, might want to reveal mortality rates as well as days not spent on a ventilator. Even when these endpoints are chosen, there are still several decisions to be made on how to construct them.
Should all composite outcomes be given equal weight or should there be a weighted hierarchy for regulatory submission? (A hierarchy in this case occurs where in a composite endpoint, some endpoints like mortality are considered first and subsequent endpoints considered only if readings of mortality meet a certain state or criterion.)
What does such a hierarchy achieve and which clinical information should inform the simulations that explain these benefits?
How many days should patients be studied and how can information on patients be maximized for a hierarchical construction of endpoints without extending the study longer than necessary?
Constructing these hierarchical models is possible within a frequentist framework, yet far easier when using Bayesian methods. Determining whether or not to adopt hierarchies into composite endpoints, while using all relevant clinical information for decision-making, is itself a simpler operation within a Bayesian framework.
Such decision-making about endpoint selection, using Bayesian methods, is even more important when employing longitudinal studies. Here both the temporal endpoint and the specific outcomes might change during the course of a trial. For example observations of an outcome at fourteen days or twenty-eight days may be affected by the severity of disease, and changes to disease severity can make statistical calculations complicated. Therefore when constructing an endpoint in a longitudinal study and trying to make the most use of clinical information to do so, a framework that allows easy integration and assessment of all relevant clinical information is an invaluable asset.
While Frequentist methods can perform technically complex calculations to undertake these options, there is a case to be made that Bayesian methods are better suited for such decisions. The following webinar, led by Cytel’s Natalia Muhlemann, featuring leading Bayesian industry leaders, offers a lengthy discussion of the uses of Bayesian methods for endpoint selection, in addition to a number of other topics such as flexibility for early stopping, small sample sizes, etc.