9 What have you (not) learned?

We summarize some of what you have learned in this book and what else you may want to know. We also suggest areas of future development for the web interface.

You have now learned the basics of designing group sequential trials using the gsDesign web interface. This specifically has related to trials comparing two treatment groups with time-to-event outcomes, binomial outcomes or normal outcomes. However, you have also learned about how to transform the sample size from a fixed design to a group sequential design. Finally, you have learned about information-based designs. We provided background on the use of spending functions and showed a variety of spending functions you might apply. We demonstrated tabular and graphical output available for summarizing the designs you derived. All that was required for any of the above was a web browser that could be on your computer, tablet or phone.

In Chapter 8, we extended what you could do with the web interface by teaching you some basics or applying the gsDesign package in R. We provided examples of how to analyze a group sequential trial and computing repeated confidence intervals. We discussed the concepts and showed how to compute conditional power, predictive power, probability of success, conditional probability of success, and prediction intervals. These are a lot of the basics you may want for applying group sequential designs. However, we also demonstrated two methods of sample size adaptation that can be derived using gsDesign: adaptation based on conditional power and adaptation based on observed statistical information.

There are many topics not covered here that are often of interest to those applying group sequential designs. For instance, when a group sequential design crosses a bound at and interim analysis and the trial is stopped, how do you incorporate any data that was collected between the time of the database cutoff and the final analysis? This is a topic addressed by John Whitehead³², among others. You may be interested in simulation to see how the asymptotic theory used for designs here compares to exact inference; some basic capabilities for simulating trials comparing binomial rates are provided in the function simBinomial(). There is also capability in gsDesign to design trials with an exact binomial outcome for evaluating, say, response rate in a single-arm oncology trial looking at response rates or a vaccine trial comparing rates of rare events between two treatment groups.³³ See help(gsBinomialExact) after loading the gsDesign package in R for more on this; this function is also extended to sequential analysis in the function binomialSPRT().