Question: How to include prior knowledge in data generation (and not in data analysis)?

[jcwie]

Dear package developers,

thank you very much for providing the R package bayesCT for the simulation of Bayesian adaptive trials! This is much welcome in these times with still grand persisting unmet needs for such software!

My question: Is there a way to apply BayesCT to simulate the study characteristics of a Bayesian clinical trial with two-arms binomial outcomes and assuming beta-binomially distributed event probabilities for the outcomes of the simulated trials?

In other words, I would like to plan a Bayesian analysis with uninformative prior but assume some (beta-binomial distribution) knowledge about "p_treatment" and "p_control" for the simulation of trial outcome. The function "binomial_outcome()" seems to allow only specific fixed probability values which I do not find intuitive when thinking Bayesian and facing only limited knowledge about model parameters. I would like to include this uncertainty information in the adaptive design modelling reflecting my subjective believe that I developed on the basis of pilot data. Reading through your documentation I only found a way to include prior information into the planned analysis of the study (via "bayesDP"). Contrarily, my plan is to analyze the study with uninformative prior to make it fit for performance validation and make it as convincing as possible for others. I may summarize this as follows: I would like to include my limited prior knowledge about study-arm specific probabilities into the data generation process but not into study analysis.

Maybe I am just not sufficiently understanding your package bayesCT and my sought opportunity is already available in your package?

Best regards,
Jan Wiemer

[graemeleehickey]

I might be misunderstanding the question, but binomial_outcome() is just simulating the data under the null or alternative hypothesis. The prior is specified via beta_prior(). If you have information on what the proportions would be from pilot data, then you can use those estimates to inform binomial_outcome(), with the Beta prior kept as alpha = beta = 1 (or some other "flat" parameter choices), which reflects the analysis model -- not the simulation model.

…

On Mon, 4 Oct 2021 at 14:37, Jan C Wiemer ***@***.***> wrote: Dear package developers, thank you very much for providing the R package bayesCT for the simulation of Bayesian adaptive trials! This is much welcome in these times with still grand persisting unmet needs for such software! My *question*: Is there a way to apply BayesCT to simulate the study characteristics of a Bayesian clinical trial with two-arms binomial outcomes and *assuming beta-binomially distributed event probabilities for the outcomes of the simulated trials*? In other words, I would like to plan a Bayesian analysis with uninformative prior but assume some (beta-binomial distribution) knowledge about "p_treatment" and "p_control" for the simulation of trial outcome. The function "binomial_outcome()" seems to allow only specific fixed probability values which I do not find intuitive when thinking Bayesian and facing only limited knowledge about model parameters. I would like to include this uncertainty information in the adaptive design modelling reflecting my subjective believe that I developed on the basis of pilot data. Reading through your documentation I only found a way to include prior information into the planned analysis of the study (via "bayesDP"). Contrarily, my plan is to analyze the study with uninformative prior to make it fit for performance validation and make it as convincing as possible for others. I may summarize this as follows: *I would like to include my limited prior knowledge about study-arm specific probabilities into the data generation process but not into study analysis.* Maybe I am just not sufficiently understanding your package bayesCT and my sought opportunity is already available in your package? Best regards, Jan Wiemer — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#45>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABIG3QMGELKHNGWUFB6WJZLUFGU3NANCNFSM5FJMTDSQ> .