Guidance Recap Podcast | Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products
Thank you for joining us for another episode of the Guidance Recap Podcast. The Guidance Recap Podcast provides highlights for FDA guidance documents straight from the authors. My name is Kylie Haskins, and I am the host for today’s podcast. In today’s episode, I am excited to be talking with Dr. James Travis. He is a Master Mathematical Statistician with the Office of Biostatistics in the Office of Translational Sciences at CDER. He will be sharing some thoughts with us on the newly published draft guidance titled “Use of Bayesian Methodology in Clinical Trials of Drug and Biological Products.” Welcome, Dr. Travis. Thank you for speaking with us today.
What are Bayesian statistical methods?
Bayesian statistical methods operate using a different framework based on Bayes theorem, a more than 250-year-old mathematical rule. The Bayesian approach uses previous data to construct what is known as a prior distribution, which is combined with the current data to form the posterior distribution, which is used to make inferences about the questions in the trial. This use of the previous data in the prior distribution is the main feature distinguishing Bayesian from traditional statistical approaches.
Bayesian methods can be used in place of traditional statistical methods in any situation, but the explicit use of previous data makes Bayesian methods ideally suited for certain situations. For example, when designing pediatric studies, we can consider extrapolation approaches where we rely on information from adult studies in the pediatric assessment. We can use Bayesian methods to explicitly incorporate the adult information into our analyses to reduce the size of the pediatric study. Bayesian approaches may also help facilitate adaptive trials where modifications to the design may occur based on interim assessments. Bayesian approaches have been discussed in other guidance documents related to these settings such as the ICH E11A and E20 guidances.
Why is this guidance important?
We are seeing more proposals for trials that use Bayesian approaches in some way, and this guidance is important to ensure that FDA’s needs and expectations are clear for sponsors as they propose and implement these approaches. This need for transparency was recognized in the commitment under the seventh iteration of the Prescription Drug User Fee Act.
Could you summarize the recommendations in this draft guidance?
Sure. In this guidance, FDA recommends how to appropriately use Bayesian statistical methods in clinical trials that evaluate the effectiveness and safety of new drugs.
Drug clinical trials can use Bayesian methods in various ways. For example, sponsors can use Bayesian methods to govern the timing and adaptation rules for an interim analysis in an adaptive design; to inform design elements, such as dose selection, for subsequent clinical trials, or as the key analysis to support primary inference in a trial.
The draft guidance addresses relevant considerations on how to design clinical trials using Bayesian approaches so that the trial meets the regulatory requirements for drug and biological product development programs. The guidance provides recommendations on how to specify decision criteria, which are usually used to measure trial success. It also discusses important operating characteristics which are measures of how a trial is likely to perform and succeed and are key to informing factors such as the sample size. Construction of the prior distribution based on previous trials is also a key topic as it is unique to Bayesian approaches.
Finally, the guidance provides recommendations on the use of software for Bayesian inference, considerations for missing data when using previous data to construct the prior, and detailed recommendations on how to document Bayesian approaches when designing and planning a study and reporting results following completion of the study. Also, the guidance includes examples from drug development programs where Bayesian methods were used. I can provide examples during our discussion.
What should sponsors consider when using available information to construct the prior distribution?
In general, the process for determining a prior should begin with identification and review of all relevant external information that is available. This should be systematic as we need to ensure that all relevant sources are identified, and we need to understand the reason why any relevant source is included or excluded. Sources of information can potentially include relevant pharmacokinetic, pharmacodynamic, and clinical data from previous trials, nonclinical data, systematic reviews of trials, real-world data, and professional or expert guidelines or consensus documents on the given disease topic.
To avoid bias and to ensure that results are interpretable, consideration should be given in advance on how to evaluate whether to leverage a particular source, how much information to leverage, and how to specifically use the information in constructing the prior distribution. In the evaluation, sponsors should consider several factors. Some of these factors are data quality and reliability, prespecification, and relevance. Proper planning and evaluation ensure that conclusions that rely on such information are reliable and interpretable.
Any information leveraged should be relevant to the applicable regulatory question of the clinical trial. The quality and reliability of the information should be adequate for the type of regulatory decision informed by the analysis. For example, if the information is intended to support approval, it should ideally be of comparable quality to phase 3 trial data. To avoid bias, information that may suggest skepticism of the existence or magnitude of a treatment effect should not be excluded.
When using Bayesian methodology, what should be considered when defining the success criteria for a clinical trial?
When designing a clinical trial, we prespecify criteria for determining whether the primary and secondary objectives of the trial have been met. The need to specify criteria is no different for Bayesian methods than for other traditional methods.
The typical “default” success criterion based on the familywise Type I error rate may not be applicable or appropriate for Bayesian approaches, such as when borrowing information. Specification of a success criterion is most often based on the posterior probability that the true treatment effect size exceeds some threshold. For example, if we use zero as the threshold, then this will be the direct probability that there is a benefit to the drug.
When using Bayesian methodology, what should be considered when defining the operating characteristics of a clinical trial?
In designing clinical trials, sponsors compute key operating characteristics such as the power to understand how the trial is likely to perform in terms of supporting correct conclusions and the reliable estimation of treatment effects. In trials with Bayesian approaches, the inference is based on the posterior distribution, and so the relevant operating characteristics are also based on the posterior which is based on both the prior distribution and observed data. Characteristics such as the probability of study success given a range of possible treatment effect sizes and the expected bias and variability of the treatment effect estimates are important for design decisions and can be estimated using statistical trial simulations.
Could you please share a few examples of how Bayesian methods have been used in drug development?
I will share a few examples described in the guidance, and these and other examples in the guidance come from applications submitted to FDA.
The first example is about borrowing information from previous clinical trials. A randomized, double-blind, placebo-controlled phase 3 study borrowed from a previous trial to evaluate REBYOTA, which is a fecal microbiota transplant product. The program evaluated REBYOTA for the prevention of recurrent Clostridioides difficile infection in individuals. The primary analysis of the phase 3 study evaluated the effectiveness of REBYOTA by using a Bayesian model to formally incorporate data from a previous phase 2 placebo-controlled study of the drug. The results supported the effectiveness of REBYOTA, which FDA approved for marketing in 2022.
Sponsors can also consider borrowing information across similar diseases or disease subtypes. Through the Complex Innovation Trial Design Meeting Program, sponsors proposed a randomized, double-blind, placebo-controlled study in patients with epilepsy with myoclonic-atonic seizures. The study used a Bayesian primary analysis to borrow information from previously conducted trials that evaluated the effect of the drug in different types of epilepsy.
A third example explores borrowing information between subgroups of a patient population, also called a subgroup analysis. Sponsors used a Bayesian hierarchical model to estimate treatment effects across regions in the trial Liraglutide Effect and Action in Diabetes: Evaluation of Cardiovascular Outcome Results. The trial compared liraglutide to placebo in patients with type 2 diabetes mellitus and at high risk for cardiovascular disease. The analyses helped clarify potential differences in drug effects in Asia, Europe, North America, and the rest of the world.
The last example is related to dose-finding trials in oncology. Sponsors proposed dose-escalation designs using Bayesian methods, with goals such as improving efficiency, reaching the maximum tolerated dose sooner; optimizing dose selection, that is minimizing toxicity and/or improving efficiency; and evaluating flexibility in terms of cohort sizes and timing of assessments. In the literature, Bayesian designs aiming to identify maximum tolerated dosages have been proposed for early phase trials in oncology.
For regulatory submissions, are there special considerations for documenting and reporting Bayesian analysis plans?
Many considerations are similar, for example, sponsors should specify and justify the study’s design, estimands, and analyses in a protocol and consider applicable guidances. Where Bayesian designs will differ is how the proposed prior is justified. For example, the protocol will need to include detailed information to support the proposed prior distribution and any external information borrowing, likelihood function, success criteria, and trial operating characteristics. Bayesian methods also often need to more heavily rely on statistical simulations, and it is important to provide extensive documentation on these to enable FDA review. Sponsors should submit all relevant information to FDA during the design stage and as early as possible to ensure that there is sufficient time for FDA feedback prior to the initiation of the trial.
For completed clinical trials, sponsors should describe in a clinical study report the design, analyses, and results.
For our final question, what are a couple of key items that you especially want listeners to remember?
I’d like to highlight that:
- The guidance provides recommendations to facilitate the appropriate use of Bayesian statistical methods in making primary inference from clinical trials that evaluate the effectiveness and safety of new drugs.
- In a Bayesian analysis, data collected in a study are combined with a prior distribution that captures the prestudy information about a parameter of interest to form a posterior distribution that expresses the updated, post-study information about the parameter of interest.
- In general, the process for determining a prior should begin with an identification and review of all relevant external information that is available.
- The guidance offers examples of how drug development programs have used Bayesian approaches in clinical trials that test the safety and effectiveness of a new drug.
- Sponsors should discuss with FDA their plans for using Bayesian methods as early as possible and before beginning their study.
Dr. Travis, thank you for taking the time to share your thoughts on the Bayesian methodology draft guidance. We have learned so much from your insights on this document. We would also like to thank the guidance working group for writing and publishing this draft guidance.
To the listeners, we hope you found this podcast useful. We encourage you to look at the snapshot and to read the guidance.