Quantitative Decision Making in Pharmaceutical Development via Confidence Distributions
The Bayesian paradigm is often chosen for decision making in clinical development due to its probabilistic statements around parameters. These probability statements are visually depicted through prior and posterior distributions, "distribution estimates" of an unknown quantity of interest, and are powerful tools for visualizing and pooling prior information and expert opinion with current data. Under the frequentist paradigm the analogous distribution estimate is a confidence distribution, a sample-dependent ex-post object supported on the parameter space that depicts all possible p-values and confidence intervals one could construct given the observed data. Confidence distributions are a powerful visual tool and allow for the inclusion of historical data and expert opinion via meta-analysis. We demonstrate the use of hypothesis testing via confidence distributions when defining end-of-study success and displaying study results, as well as in performing inference on power for progression through all phases of pharmaceutical development. Desired inference on phase 3 power is used to reverse engineer the hypothesis, significance level, and sample size required in phase 2. Extrapolation between endpoints is also demonstrated, and a discussion is provided on multiple comparisons.
READ FULL TEXT