Over the last few years, adaptive designs were a focus area of biostatistical research. The topic received a broader attention since it was felt that adaptive clinical trials can provide means to improve the efficiency of drug development, which was facing a high attrition rate of compounds during its late and most costly phases. In contrast to the development of group sequential designs, where a state of the art summary in book format took decades to appear, two companion books on adaptive designs have appeared this and the previous year. The more recent one by Mark Chang (who is also the co-author of the first one on “Adaptive Design Methods in Clinical Trials”)1 undertook the effort to provide implementations of adaptive methods as SAS and R programs to ease application of the theory.

The book is divided into 18 chapters, the first of which constitutes a roadmap for the contents of the following parts of the book. It starts with a definition of adaptive designs as clinical study designs using accumulating data to decide about modifications to aspects of the trial without undermining the validity and integrity. Adaptive trials have a potential to reduce the number of patients and the corresponding amount of data, however, they may need more time for upfront planning. Whether they reduce the amount of study drug in all cases, as pointed out in the book, may be questionable. In any case, they may require the change of processes in drug supply management and other operational aspects and increase the number of analyses and the need for independent boards for data review in order to preserve trial integrity.

Chapter 2 reviews sample size calculations for classic non-adaptive clinical studies aiming at demonstrating superiority, non-inferiority, or dose-response. Chapter 3 provides a general theory of adaptive designs that covers stopping boundaries, point estimates, and confidence intervals. From group sequential trials, one is familiar with the fact that without any adjustment, trials that stop early with rejecting the hypothesis will deliver biased estimates of treatment effects, p-values and, confidence intervals. However, since only compounds with significant definitive trials will be submitted to health authorities for approval, the author concludes that the estimates from non-adaptive designs have a bias issue as well. The discussion of this interesting aspect could have been more clearly elaborated and the table in the respective section fully explained in the text.

Chapter 4 describes examples of adaptive methods using a combination of p-values and supplies the corresponding SAS macros. The value of printing these macros without structuring and commenting the statements raises some concerns. Some programming lines contain as much code as could be fitted in, may be to save space. This leaves the reader with being faced with kind of a black box despite that he has the full program text at hands. The methods are adequately explained; only the sample size adjustment in some examples is not very well founded and could have been left for Chapter 9.

Chapter 5 on methods with inverse-normal p-values discusses the Lehmacher–Wassmer approach and classic group sequential designs including spending functions. The presentation of these methods is not fully successful in explaining the nice relationship between these approaches. It sounds a bit odd in a book introducing software to read that, to determine the stopping boundaries, prefix some input and vary the other until the result is close enough to the desired result. The following Chapter 6 discusses implementation of K-stage methods.

Chapter 7 looks at adaptive designs from the conditional error function point of view. The results by Müller and Schäfer provide likely the most flexible approach to adaptive designs. Chapter 8 on recursive adaptive designs is a mathematical tour de force adding an additional point of view to adaptive designs. The sample size re-estimation designs discussed in Chapter 9 rely on the adjustment based on effect size ratio or conditional power at the time of the interim analysis. These methods have implications for estimates and confidence intervals and require unblinding of the data. If the main purpose of a sample size re-assessment is the degree of uncertainty about variability (or other nuisance parameters) when the trial needs to start, methods to re-assess variability under blinded conditions may be considered first. It would be worthwhile to mention those in this chapter. Chapter 10 provides some thoughts on multiple endpoint adaptive designs.

Chapter 11 deals with seamless phase II/III designs, which were most controversially discussed over the past years. These designs intend to combine a (dose-)selection stage (comparable to a separate phase II study) and a confirmation stage (similar to a classic phase III studies) by using the data from the selected treatment arms of the first stage together with additional data from the second phase for a confirmative trial. The gain compared with separate phase II and phase III studies is the re-use of data from stage one and the absence of “white space” between the trials. However, there may be additional patients to be recruited on the arms that are going to be dropped until the results of the interim analysis are available. To account for the many treatment options in stage one, a type I error control is required. Methods exist that control the type I error in the weak sense (i.e., for the hypothesis of equal effects in all treatment arms) and in the family-wise sense. This chapter could have been elaborated a bit more because of the relevance of these designs in a regulatory context, for example in regard their acceptability as well-controlled studies. It should also be emphasized more clearly that a type I error adjustment may be required even if the treatment selection does not depend on p-values from the first stage of the study, but for example on safety aspects. This adjustment limits the number of arms among which to select during stage one.

Biomarker adaptive designs as treated in Chapter 12 can become one of the promising areas for the use of adaptive designs. Most of this chapter is concerned with classifying biomarkers. Adaptive designs can be applied to decide whether it is worthwhile to proceed with the total or with the biomarker-positive population. Chapter 13 on treatment switching and crossover is interesting but mainly theoretical. The response-adaptive allocation designs of Chapter 13 are clearly a bit off the main field of adaptive designs. The adaptive dose-finding design Chapter is restricted to oncology dose-escalation trials, an important area but clearly not exhaustive for dose-response studies. Chapter 16 on Bayesian adaptive designs in general is clearly too short. To do justice to that area, maybe other material in the book would have to be shortened. Chapter 17 on planning and execution is reasonably short and touches on some practical aspects. Chapter 18 discusses adaptive designs in the context of efficiency and sufficiency and adds personal viewpoints of the author.

In the preface it is said that the aims of the book are to provide a unified and concise presentation of adaptive design theories, furnish the reader with computer programs in SAS and R for the design and simulation of adaptive trials, and to offer a quick way to master the different adaptive designs through examples.

In the view of this reviewer, the book fulfills its preset targets only partially. Though the author manages to start with a very general framework, the presentation of the approaches is somewhat patchy. This criticism may not be attributable exclusively to the author; it rather reflects the state of the art of this statistical field, which is still more like a collection of different approaches than a uniform theory. In this sense, the book may have aimed for too much at this point in time.

Some chapters, like the one on response-adaptive designs, could have been omitted for their limited use and acceptability; others like the one on Bayesian methods would deserve a more thorough elaboration. The thoughts of the last chapter are relevant for the theoretical appreciation of adaptive designs. They may be less relevant for the practitioner, since many times inefficient means just a bit less efficient than optimally possible. The book is not clearly taking positions in favor of theory or practice, which is probably the reason for the variability of chapters in that regard.

The SAS and R programs are available on the web. This is good news since it avoids painful re-typing and error checking. One could even ask whether it was necessary to reproduce the code in the book. For statisticians in the pharmaceutical industry, it may be helpful to know to what extent the programs are validated.

The book can provide all statisticians interested in the adaptive design field with an overview of the topic and software to run practical examples. However, the flow of the presentation of the material is inhomogenous in style and requirements in regard to the mathematical background.


  1. Top of page
  2. Reference
  • 1
    Chow SC, Chang M. Adaptive Design Methods in Clinical Trials. Boca Raton: Chapman & Hall/CRC, 2007.