Should Bayesian tools be used in phase 3 clinical trials?
Our department just had an excellent journal club about the use of Bayesian methods in clinical trials. The ensuring discussion raised some questions about the use of Bayesian and Frequentist methods in clinical trials.
One professor raised the point that a phase 3 trial should represent stand-alone evidence that a treatment works (or that we can’t conclude that it works), and therefore should not include any prior information in the tools used to analyze the data (regardless of whether the tools are Bayesian or Frequentist). I am assuming that the Bayesian argument (which was not as vocal) would be that scientists should use all prior information available to more efficiently estimate treatment effects in a clinical trial (regardless of the phase of the trial). This improved efficiency would hopefully result in exposing as few patients as possible to potentially worthless drugs that have side effects, or drugs that are harmful, as well as decreasing the amount of time that a good drug needs to get to market.
One very helpful point that a professor raised was that there were two topics really being discussed: (1) taking prior information into account (which can be done with Bayesian or Frequentist tools); and (2) are Bayesian or Frequentist tools better for taking prior information into account, in the context of determining treatment effects of drugs.
Questions that need answering:
After thinking about these topics, I have come up with several questions whose answers might inform the use of Bayesian methods in clinical trials:
- Is previous information implicitly used in all clinical trials? (The FDA usually requires two independent studies that establish a significant effect of treatment to approve a drug. The fact that the first trial [or second trial] was even conducted is a result of previous information that indicated a significant effect of the drug could potentially be found.)
- If the answer is “yes,” then what kind of effect does that answer have on estimates/intervals/hypothesis testing in Phase 3 trials analyzed with Frequentist tools, and the final decision about whether a drug has an effect after both trials have been conducted?
- If there is an effect of implicitly assuming information about previous results, then would Bayesian or Frequentist tools be better to take this prior information into account?
- Perhaps a question that should be asked first is: does the use of methods that take into account implicitly assumed information (i.e., making implicit information explicit), regardless of the tools used to do so (Frequentist or Bayesian), affect the inference/estimation of a treatment effect compared to not taking into account the assumed information?
- Is there a difference in how well doctors and/or patients interpret the results of a Phase 3 trial analyzed using Bayesian tools versus results of a Phase 3 trial analyzed using Frequentist tools?
- This information would let us know whether the argument that “Bayesian tools use probability in a more intuitive way” is tenable.
I think the answers to these questions would clear up a lot of the debate about using Bayesian methods in clinical trials, especially in phase 3 trials. I look forward to comments on this post.