The book by Proschan, Lan and Wittes is a
welcome addition to the recently published books on sequential
clinical trials. The books in this area tend to fall in the theoretical
or practical/case study categories. The books by Whitehead (1997),
Jennison and Turnbull (2000) are examples of theoretical books
with a textbook flavor. On the other hand, Ellenberg, Fleming
and DeMets (2002) and DeMets, Furberg and Friedman (2005) focus
on operating principles/case studies/lessons learned in clinical
trials with interim assessments. Proschan, Lan and Wittes (2006)
addresses both methodological and practical aspects of sequential
monitoring of clinical studies by including a detailed description
of relevant statistical methods, numerous case studies, numerical
examples and advice from experts.
The book begins with a description of the underlying mathematical
framework. This framework is built around Brownian motion and
helps the authors develop a unified approach to virtually all
types of clinical trials, including trials with continuous, binary
and survival outcome variables, repeated measures, etc. The authors
describe "classical" group-sequential methods in Chapters 3-7,
related topics in Chapters 8-10 and a more advanced topic (sample
size modification) in Chapter 11. In what follows, I will provide
a short outline of each chapter with emphasis on points that I
found particularly interesting.
Chapters 3, 4 and 5 form the backbone of the book and introduce
key concepts such as conditional power, stopping rules for benefit,
harm and futility, and error spending. The theory and applications
of conditional power are covered in Chapter 3. The authors discuss
the standard frequentist definition and a Bayesian-type extension
that relies on averaging the conditional power over the posterior
distribution of the treatment difference (predictive power approach).
Chapter 4 deals with traditional group-sequential designs based
on Pocock and O'Brien-Fleming boundaries and Chapter 5 describes
the error spending function approach introduced in the seminal
paper by Lan and DeMets (1983). It is important to note that the
authors utilize error spending functions to both design and execute
group-sequential trials. This approach enables a seamless transition
between the design and monitoring stages and is more natural/efficient
than an alternative approach that begins with a traditional group-sequential
design and then relies on an approximate error spending function
during the monitoring stage. Chapter 5 also contains an interesting
discussion of benefit and harm/futility stopping rules in sequential
trials. The authors discuss the inherent asymmetry of benefit
and harm stopping boundaries (a negative treatment difference
does not have to be statistically significant to justify an early
termination of the clinical trial) and the relationship between
benefit and futility stopping rules. They recommend computing
the benefit and futility boundaries in an independent manner to
give data monitoring committees (DMCs) more flexibility with respect
to futility stopping. One exception to this rule is the comparison
of two active treatments. Another exception that comes to mind
is data monitoring in mortality clinical trials (in this case,
it seems prudent to make futility stopping mandatory to avoid
exposing patients to an inefficient treatment).
Chapter 6 discusses sequential monitoring in clinical trials with
time-to-event outcomes, for example, survival trials. It includes
a good discussion of issues specific to survival monitoring such
as the computation of information fractions and analysis of non-proportional
hazards. Chapter 7 describes adjusted inferences at the last look
(bias-corrected point estimates, confidence intervals and p-values)
that account for the sequential nature of the trial. Chapter 8
covers statistical methods that can be considered when the Brownian
motion-based framework becomes unreliable (for instance, permutation
methods in small samples).
Chapter 9 touches upon a topic which is often downplayed in statistical
papers on trials with sequential designs --- monitoring of safety
data. The chapter describes more formal safety decision rules
that are similar to decision rules for efficacy variables and,
in addition, provides a summary of recommendations that come in
handy when a DMC reviews adverse events (for example, approaches
to adverse event classification, including hierarchical classification
schemes).
Chapter 10 briefly reviews Bayesian approaches to sequential data
monitoring, including a Bayesian formulation of the frequentist
decision rules described earlier in the book. The chapter also
discusses the selection of prior distributions in Bayesian stopping
rules and explores connections between frequentist and Bayesian
monitoring in sequential trials.
Chapter 11 deals with a topic that has attracted much attention
in recent years --- adaptive clinical trials. Specifically, it
discusses sample size reassessment at an interim analysis based
on updated estimates of nuisance parameters (for example, the
variance of a continuous outcome variable) or an updated estimate
of the treatment effect. The former case is quite straightforward
whereas the latter one has caused some controversy. The authors
should be commended for giving a detailed and objective assessment
of available options in this important area of research and providing
practical advice.
The last chapter (Chapter 12) quickly reviews topics not covered
in the book. This includes, among other things, multiplicity-adjusted
confidence intervals for the true treatment difference computed
at each interim look (known as repeated confidence intervals).
The authors argue that repeated confidence intervals are not generally
informative and I agree that these confidence intervals are inferior
to bias-adjusted confidence intervals derived at the last look.
However, DMC members may find repeated confidence intervals useful
in benefit/risk assessments when it is important to understand
the likely magnitude of clinical benefit given the interim results.
Finally, the two appendices give a crash course in survival analysis
and information on group-sequential software.
One of the attractive features of the book is an impressive collection
of case studies. Examples of real clinical trial are used throughout
the book to illustrate properties of statistical approaches discussed
in each chapter and get important points across to the reader.
To give an example, the CAST I study is first mentioned in the
introduction and then reappears in the discussion of futility
rules based on conditional power, benefit and harm error rates
in sequential clinical trials and safety monitoring. Several case
studies in this book are really thought-provoking and help the
reader appreciate how much work goes into the design and monitoring
of sequential trials (CAST I, CAST II, RALES and other trials).
These examples make it clear that there are no cookie-cutter solutions
in the world of sequential clinical trials.
Numerical examples are included in almost every section. They
range from fairly basic examples such as the computation of conditional
power (Chapter 3) to more complex ones arising in challenging
statistical problems such as the derivation of bias-adjusted estimates
of the treatment effect and confidence limits following a group
sequential trial (Chapter 7). All of these examples include detailed
instructions and will be appreciated by practitioners. It is worth
noting that most of numerical examples involve straightforward
calculations and reinforce an important point: sequential monitoring
of clinical studies does not always require specialized software
and in many cases calculations can be done on a pocket calculator.
The book also features software-driven numerical examples; however,
all of these examples rely on open-source software freely available
on the Internet. This includes the R language and software developed
by David Reboussin, David DeMets, KyungMann Kim and Gordon Lan
(
http://www.biostat.wisc.edu/landemets/).
My last comment pertains to the use of graphical displays in this
book. Effective graphical displays are known to be helpful when
one is dealing with complex concepts and Figures 3.3 and 3.6 serve
as excellent examples of plots that are worth a thousand formulas.
The figures give an elegant geometric interpretation of conditional
power. It literally takes a few seconds to understand key properties
of conditional power, for example, the relationship between conditional
power and assumptions about the distribution of future data in
a clinical trial (conditional power under the alternative hypothesis
of a positive treatment effect is higher than that based on the
null hypothesis). The reader will find quite a few other helpful
plots and diagrams throughout the book.
The new book gives an excellent overview of issues related to
the design and conduct of sequential clinical trials. Researchers
working in this area will find this comprehensive book very useful.
DeMets DL, Furberg CD, Friedman LM (editors).
Data Monitoring in Clinical Trials: A Case Studies Approach. Springer,
New York, 2005.
Ellenberg SS, Fleming TR, DeMets DL. Data Monitoring Committees
in Clinical Trials: A Practical Perspective. Wiley, New York,
2002.
Jennison C, Turnbull BW. Group Sequential Methods with Applications
to Clinical Trials. Chapman and Hall/CRC Press, London/Boca Raton,
FL, 2000.
Lan KKG, DeMets DL. Discrete sequential boundaries for clinical
trials. Biometrika. 1983; 70:659-63.
Whitehead J. The Design and Analysis of Sequential Clinical Trials
(Second edition). Wiley, London, 1997.
