All Times EDT
Keywords: Probability of success, historical control data, Bayesian meta analysis
Sample size and power for a Ph.3 study depend on the null hypothesis, expected treatment effect, and standard deviation (or control rate for binary endpoint). They do not take account of the uncertainty about true treatment effect and standard deviation/control rate. As the power is usually set at 80% to 90%, the gain in power is small if the true treatment effect is better than the expected but the loss in power could be high if the true treatment effect is worse than the expected. Therefore, the real success rate of Ph.3 is in general much lower than the nominal power in the sample size calculation. In order to increase the success of Ph.3 studies, it’s better to predict the success for Ph.3 study considering uncertainty about the treatment effect and standard deviation/control rate. This practice may lead to re-assess sample size in the benefit of higher probability of success. Ph.2 study is often conducted using unequal randomization between new treatment and control in order to get more data about the new treatment. In order to overcome the smaller control data from Ph.2 study, historical control data may be used with Ph.2 control data to predict the success of Ph.3 study. Bayesian meta analysis could be used to discount the historical control data. Due to huge historical control data, the prior effective sample size from the Bayesian meta analysis may be much larger than the Ph.2 control data. We will discuss (a) test for prior conflict, (b) include Ph.2 control data in Bayesian meta analysis, and (c) limit the prior effective sample size.