Donald Berry, PhD

This chapter supplements and complements Chapter 32, “Theory and Practice of Clinical Trials,” by Marvin Zelen. In particular, understanding the basic principles espoused in Chapter 32 is a prerequisite for the present chapter. The two chapters support each other to a substantial degree, but are different in attitude. The purpose of the present chapter is to describe statistical approaches to cancer research that allow for building new designs and incorporating new analyses. Some of the methods described here have been introduced into research practice and others are still being developed. The goals of the innovations presented here are (1) to more effectively use patient resources while treating patients in clinical trials more effectively, and (2) to identify better drugs and other therapies more rapidly, moving drugs more quickly through the development process. The methods exploit available evidence and place information gleaned from an ongoing clinical trial into the context of what is already known. These new methods tend to be intuitively appealing. But like most innovations, some are controversial. Although they presage the future of clinical trial research, in oncology and more generally, not every method described here is destined to become standard in medical research.

This chapter addresses two overall categories of innovations. One represents a fine-tuning of the traditional practice of statistics. The other is based on an alternative view of the foundations of statistics. Separating the two categories is not possible. I will set each method in the context of statistical approach, but I will present the various methods in an integrated fashion.

The “alternative view” of the foundations of statistics is the Bayesian approach. Since not all readers will be familiar with this approach, I will describe it and relate it to the more traditional frequentist approach. Readers who are familiar with Bayesian ideas may wish to skip “Bayesian Updating” below. An important distinction between the two approaches is one of attitude. The Bayesian approach is ideal for on-line learning (as data accrue), and the frequentist approach is tied to a particular experimental design. But the two approaches support each other. For example, much of this chapter's development of clinical trial design employs the Bayesian approach as a tool for finding designs that tend to treat patients in the clinical trial more effectively and that identify better drugs more rapidly. But the design thus derived is checked for its frequentist properties (such as false-positive rate and power). Ensuring that a design has prespecified frequentist properties means that the design is frequentist and that the Bayesian approach is a tool for finding good frequentist designs.

Expanding the horizons of statistical designs and analyses to the extent described here relies on the availability of high-speed computers and sophisticated computational methods. In the past ten years there has been an explosion of Bayesian computational procedures that can be used to derive efficient designs. In addition, high-speed computers can be used to simulate trials having these designs to evaluate and compare their properties, such as power and false positive rate.

© 2003 BC Decker Inc