In January, the Food and Drugs Agency (FDA) has evolved into acceptance and even in some cases a recommendation (!) of the use of Bayesian statistics in the evaluation of devices and medical treatments (see the press release for more information: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/use-bayesian-methodology-clinical-trials-drug-and-biological-products).
Historically, clinical trials were based mainly on frequentist statistics, based on tests of classic hypotheses and significance thresholds (in general a p-value less than .05). However, these methods have certain limitations, especially when it comes to integrating pre-existing information or working with small samples. In addition, many fields and more particularly in psychology, from which I come, have suffered and still suffer today from inappropriate uses of classical methods (p-hacking, lack of statistical power, etc.) which have led to publication biases and publication pressure (publish or Perish) producing a reproducibility crisis. It is in this context that Bayesian approaches have gained in importance.
In its documentation, the FDA explicitly recognizes the benefits of the Bayesian framework, including its ability to incorporate previous data (from previous studies, registers or clinical expertise) in the form of a priori distributions. This helps optimize the design of the tests, reduce the size of the samples needed and accelerate the evaluation process, while maintaining a high level of scientific rigor. These advantages directly follow the scientific philosophy aimed at building brick-by-brick knowledge based on existing knowledge.
The position of the FDA is formalized in several guidance documents, particularly in the field of medical devices. For example, his guide to the use of Bayesian statistics for clinical trials on medical devices explicitly encourages the use of Bayesian methods when appropriate. The FDA stresses that these approaches can improve the effectiveness of trials while providing more interpretable results for decision-makers. An emblematic case is that of implantable devices or rare diseases, where available data is limited. In these situations, the Bayesian approach makes it possible to exploit reliable historical data to strengthen the analysis, which would be difficult to achieve with classic frequentist methods. Beyond the advantages mentioned, the Bayesian methods also make it possible to take into account unmeasured variables which is also more complicated with the classic methods (see the example of the Bayesian billiard table).
However, the FDA also insists on the need for transparency and rigor in the implementation of Bayesian models. The choices of a priori distributions must be justified, sensitivity analyses are strongly recommended, and the models must be clearly documented in order to guarantee their reproducibility and credibility. For more details on Bayesian workflows and good practices, see https://doi.org/10.48550/arXiv.2011.01808, https://doi.org/10.1038/s43586-020-00001-2.
Today, the Bayesian approach is no longer perceived as a marginal alternative, but as a powerful and legitimate tool in the statistical arsenal of clinical trials. Its growing adoption by the FDA reflects a broader shift towards more flexible methods, able to adapt to the increasing complexity of biomedical data. These advantages would also be particularly interesting in other areas, economics, politics, psychology… Let’s hope that this evolution also inspires other scientific communities.