Bayesian Statistics in Medical Device Clinical Trials
This guidance provides direction on the use of Bayesian statistical methods in medical device clinical trials. It covers key statistical aspects of trial design and analysis using Bayesian approaches, including prior information incorporation, sample size determination, interim analyses, and decision-making criteria.
What You Need to Know? 👇
What are the key benefits of using Bayesian statistics in medical device clinical trials?
Bayesian methods offer several advantages including incorporation of prior information to potentially reduce sample sizes, flexible adaptive trial designs, better handling of missing data, and natural accommodation of interim analyses without traditional multiplicity concerns.
When should I meet with FDA to discuss my Bayesian trial design?
FDA recommends scheduling meetings before study initiation to discuss prior information and experimental design. For trials requiring an IDE, meet with FDA before IDE submission. Early collaboration is crucial for Bayesian trials.
What constitutes appropriate prior information for a Bayesian medical device trial?
Good prior information typically comes from clinical trials conducted overseas, patient registries, data on similar products, or pilot studies. Prior information based on empirical evidence is generally less controversial than expert opinion alone.
How do I determine if my prior distribution is too informative?
A prior may be too informative if the prior probability of study success approaches the posterior success criterion, or if the effective sample size borrowed exceeds the actual trial size. FDA recommends sensitivity analyses to assess robustness.
What operating characteristics must I provide for a Bayesian trial design?
You must provide Type I and II error rates, power calculations, sample size distributions, prior probability of study claims, and if applicable, stopping probabilities at interim analyses through simulation studies across various scenarios.
Can I switch from frequentist to Bayesian analysis after seeing trial results?
No, FDA strongly recommends against switching analytical approaches after trial initiation. Such post-hoc changes are not scientifically sound and weaken submission validity. The analysis method should be pre-specified and agreed upon with FDA.
What You Need to Do 👇
Recommended Actions
- Schedule early meetings with FDA to discuss:
- Prior information selection and justification
- Trial design and analysis plans
- Operating characteristics evaluation
- Develop comprehensive protocol including:
- Prior information details
- Success criteria
- Sample size justification
- Operating characteristics
- Simulation results
- Implement proper trial conduct measures:
- Randomization
- Masking where applicable
- Data quality control
- Interim analysis procedures if planned
- Prepare detailed documentation of:
- Statistical models and assumptions
- Software code and calculations
- Sensitivity analyses
- Model checking results
- Plan for clear presentation of results:
- Understandable labeling language
- Appropriate summaries of posterior distributions
- Credible intervals reporting
- Comprehensive final report
- Consider post-market surveillance planning using Bayesian methods to update safety and effectiveness information
Key Considerations
Clinical testing
- Clinical trial planning and rigorous conduct are important regardless of statistical approach
- Basic tenets of good trial design remain the same for both Bayesian and frequentist trials
- Randomization and masking principles should still be followed
- Pre-specification of analysis method (Bayesian vs frequentist) is required
Software
- Software programs capable of Bayesian calculations should be used (e.g., WinBUGS)
- Program code and data used for simulations should be submitted
- Electronic submission of calculations is recommended for FDA review
Labelling
- Results from Bayesian trials need to be expressed clearly in device labeling
- Bayesian terminology should be understandable
- Credible intervals should be reported
Safety
- Minimum sample size should be determined based on safety endpoints
- Safety endpoints may require larger sample sizes than effectiveness endpoints
- Post-market surveillance can use Bayesian approaches to update safety information
Other considerations
- Prior information selection and incorporation must be carefully justified
- Operating characteristics (type I/II errors) must be evaluated
- Interim analyses and adaptive designs need pre-specification
- Model assumptions should be checked and sensitivity analyses performed
- Hierarchical models can be used to combine data across studies
- Missing data handling requires appropriate methods and sensitivity analyses
Relevant Guidances 🔗
Related references and norms 📂
- MCMC: Markov Chain Monte Carlo techniques for Bayesian computation
- DIC: Deviance Information Criterion for model selection
- CODA: Convergence Diagnosis and Output Analysis for MCMC