The idea behind most if not all statistical experiments (clinical trials, observational and real-world studies, animal trials, agriculture, laboratory experiments, etc.) is to estimate some population parameters of interest. Unlike the traditional approach where these paramters are assumed to be fixed (but unknown), in the Bayesian framework these are assumed to be random variables. Thus a probability distribution needs to be defined for such a parameter prior to the experimentation. This is called the prior distribution (\(\pi(\theta)\)) for a paramater \(\theta\) of interest. Once the data (\(\mathcal{D}\)) is obtained, using Bayes theorem, the prior is updated to get the posterior (\(\pi(\theta|\mathcal{D})\)). Inference for \(\theta\) such as hypothesis testing and interval estimation is then based on this posterior distribution.
Since the Bayesian approach does not necessarily depend on large-sample theory, it is a popular choice for data analysis in early phase trials or experiments that are conducted with limited sample sizes. In addition, Bayesian trial designs also offer flexibility in terms of mid-course adaptive (data-driven) changes to the study design. This makes the framework very attractive for conducting trials in rare diseases. Bayesian designs are also common for registrational trials for medical devices. In recent times they have also been applied to vaccine trials, for example, the Pfizer-Biotech RNA COVID vaccine trial.
In general, and especially for confirmatory trials, the use of non-informative or weakly-informative priors is advised. These weak priors mostly represent an almost flat distribution for the parameter of interest and as data accumulates we let the data dictate the shape of the parameter distribution. In high unmet need inidcations there might be a case for using informative priors constructed from historical trial data, however proper borrow-control methods need to be used in order to protect from false-discovery rate inflation arising from prior-data conflicts. Historical trial data or real world data can however be used to construct informative priors to be used for Bayesian interim decision making via the use of predictive power.
Bayesian trial designs do require extensive simulations to set design parameters that are to be pre-specified in protocols. See Simulation Guided Designs for more details.
Our statisticians with expertise in both Bayesian and traditional statistics and Bayesian clinical trial designs can help with the rigorous planning and extensive simulations needed for setting up and optimising trial designs within the Bayesian framework and to defend such design with the regulatory bodies. The following are a list of related experiences in this area:
- Bayesian confirmatory trial in CVD device trial using dynamic borrowing from historical trial data.
- Bayesian designs for dose-escalation studies incorporating toxicity and early efficacy signals such as biomarkers related to the mechanism of action.
- Use of a Bayesian clinical utility score for dose selection for a vaccine Phase-2 trial.
- Bayesian adaptive and group sequential designs for late phase trials in rare diseases incorporating prior evidence.
- Bayesian adult-pediatric bridging trials.
- Bayesian geographical-bridging device trials.
- Bayesian sequential design for Ph-2 or PoC vaccine trials.
- Bayesian seamless dose finding and efficacy trials using step-wedged and other within-subject designs for rare indications.
- Hybrid designs in Oncology with interim adaptation based on Bayesian computation of predictive power incorporating prior real world data.
- Bayesian Phase-2 design in Oncology borrowing from literature control.
- Hybrid design in CVD using Bayesian predictive power in presence of possible delayed treatment effect.
- Single arm trial in a rare indication with historical literature control.
- Bayesian confirmatory veterinary trial.
- Training in Bayesian statistics and designs for clinicians and statisticians.
- Bayesian design tools and software.