Finally, I decided to research on Bayesian adaptive clinical trials. Unlike conventional clinical trials have substantial uncertainty in intervention arms (e.g. uncertainty in optimal dose/duration), adaptive clinical trials allow us to modify key trial parameters as we are acquiring more information during the experiments, thus reducing the uncertainty and speeding up the evaluation of interventions. Also, adaptive clinical trials allow for dropping arms that do not work well during the experiments based on the interim analysis. Researchers can reallocate people to other arms, thus making it more efficient. In the past decades, statisticians start to use Bayesian in adaptive trials. People may not be familiar with Bayesian, but they actually use Bayesian everyday. If you do have following experience, you have encountered Bayesian. One day, you go to Yelp to see the rating of a restaurant near your house. You found it is 4.9 out of 5, so you think it must be a good restaurant and decide to try. After you try, you find they have the worst food you have ever had, so you pull it to your blacklist and change your belief. As I described in this example, instead of solely based on data to make decisions, the Bayesian approach requires an initial individual belief about the parameter we try to estimate and combine the evidence from the data to reach a Bayesian posterior …show more content…
al (2006) recommends eliciting the prior distributions from a group of experts. Instead of using personal beliefs as the prior information, the prior distribution can be elicited from a group of experts in a particular subject. Each expert will use their expertise to specify the parameters for the prior distribution during a Q&A session with statisticians. After that, statisticians will integrate all experts’ outputs and form a single prior distribution for the parameters. Several studies have been done on eliciting single parameter. However, In clinical trials, researchers need to elicit a joint prior distribution for at least two parameters (e.g. outcome rate with standard care and with improved care). Previous work had also been done by Clemen and Reilly (1999). They considered a joint prior distribution for several parameters. They use a copula to form the joint distribution from an expert 's subjective judgments of marginal distributions and correlations. This method has three potential difficulties. First, assessing the correlations is non-intuitive for experts even for statisticians. I am worried the feasibility of this copula model. Second, in the sense that experts’ elicitation for one parameter may affect by another parameter (i.e. when they try to elicit marginal distribution for improved care, their decision might be potentially affected by the performance of the standard care), asking for marginal distribution may not be a