BOOK REVIEWS 1365

Kpotufe, S. (2011), “k-NN Regression Adapts to Local Intrinsic Dimen- approach are also useful. This book should serve as the pri-
sion,” in Advances in Neural Information Processing Systems 24, eds. J. mary text for anyone who wishes to study the potential outcomes
Shawe-Taylor, R. S. Zemel, P. L. Bartlett, F. Pereira, and K. Q. Wein- approach.
berger, Cambridge, MA: MIT Press, pp. 729–737. [1364]
LaLonde, R. J. (1986), “Evaluating the Econometric Evaluations of Train-
The focus on the potential outcomes approach also means
ing Programs with Experimental Data,” American Economic Review, 76, that other alternative approaches, such as directed acyclic graphs
604–620. [1364] (DAGs) and structural equation models (SEMs), are omitted
from this book. In the end, it is impossible for any single book
to cover every topic within this growing field of causal inference.
Given that the authors of this book are the major developers and
Cosma Rohilla Shalizi
proponents of the potential outcomes framework, the scientific
Carnegie Mellon University/Santa Fe Institute
community benefits most from the book that elucidates their
advocated approach. For the approach based on the DAGs and
SEMs, researchers can consult the authoritative book by Pearl
Causal Inference for Statistics, Social, and Biomedical (2009).
One limitation of the book is that it does not cover a number
Sciences: An Introduction. Guido W. Imbens and
of important topics. The authors spend the first 11 chapters
Donald B. Rubin. New York: Cambridge University Press, discussing the potential outcomes framework and classical
2015, xix + 625 pp., $60.00 (H), ISBN: 978-0-52-188588-1. randomized experiments in detail. As a result, the book does
not discuss such topics as cluster randomized experiments,
Guido Imbens and Donald Rubin have written an authoritative interference between units, longitudinal data, mediation analy-
textbook on causal inference that is expected to have a lasting sis, nonbinary treatments, regression discontinuity designs, and
impact on social and biomedical scientists as well as statisticians. difference-in-differences designs. Applied researchers should
Researchers have been waiting for the publication of this book, consult other work to learn about these additional topics for
which is a welcome addition to the growing list of textbooks causal inference that are relevant in many fields.
and monographs on causality. Causal inference plays an essen- The authors write that their target audience is “researchers in
tial role in empirical studies that test scientific theories because applied fields.” In my view, this book is best suited for applied
scholars are concerned about causal relationships rather than researchers who have a solid understanding of basic probabil-
associations. Similarly, policy makers, who are evaluating the ity and statistics. Those seeking a nontechnical introduction to
efficacy of public policy programs, and marketing researchers, statistics may have a hard time following the detailed mathe-
who are assessing the effects of advertisements, must build an matical derivation presented in the book. These researchers may
effective causal model rather than a good predictive model. In find other textbooks such as Morgan and Winship (2007) and
the age of “big data,” this distinction between causality and asso- Angrist and Pischke (2009) more accessible. In addition, while
ciation is becoming more important than ever. The book should the book includes many real-world applications, it does not pro-
be of interest to many researchers and practitioners in a variety vide tips on how to implement the methods. Online materials
of fields. including the computer code for the results presented in the
Although the study of causality goes back to Neyman and book would be a useful supplement to the text.
Fisher in the 1920s, for a long time only limited progress was The authors’ focus on the fundamental causal inference prob-
made in statistics regarding this important topic. A critical lems is apparent in their treatment of causal inference with
turning point may have been the publication of an article by observational studies. For example, Chapter 13 discusses the
Holland (1986), which highlighted the utilities of the potential estimation of propensity scores through an iterative process of
outcomes framework developed by Neyman and Rubin. Since covariate balance checking and model respecification. Although
then, statisticians and methodologists from various disciplines several methods have been recently proposed to estimate the
have made numerous contributions to improve causal inference propensity score by directly optimizing covariate balance, these
in both experimental and observational studies. Over the last methods are not introduced. Similarly, Chapter 15 only dis-
three decades, causal inference transformed itself from an cusses classic matching methods while newer matching meth-
obscure topic to a major research theme in statistics. Imbens ods are not included (see also Chapters 17 and 18). It may be
and Rubin played a significant role in this development and that a separate book is required for a comprehensive treatment
their perspective from numerous leading publications are now of methods for observational studies.
elegantly summarized in this book. The materials in Part IV (Chapters 17–20) provide a succinct
The most important feature of the book is that it is written introduction to the series of recent articles on matching meth-
exclusively from the potential outcomes perspective. After an ods written by the authors. While applied researchers often
accessible and informative discussion of the potential outcomes compute the uncertainty estimates conditional on the matched
framework in Part I, the authors apply this approach to random- sample, the authors show how to incorporate the uncertainty
ized experiments and then extend it to observational studies. regarding the process of matching given a particular matching
The discussion is thorough with an effort to build everything method. In practice, however, as illustrated in Chapters 13–15,
from the first principles. In particular, the detailed derivation applied researchers must choose among different matching
of variances under the Neyman’s framework and posterior dis- methods and/or alternative propensity score models. Future
tributions under the Bayesian approach are insightful. Explicit research should address the issue of how to quantify additional
efforts to connect the Neyman’s approach to the model-based uncertainty arising from the selection of matching methods.
1366 BOOK REVIEWS

In conclusion, the authors should be congratulated for the location and scale problem, and nonparametric regression. The
publication of this impressive volume. The book provides a book offers critiques of standard statistical methods through-
unified introduction to the potential outcomes approach with out, with the final chapter targeting the likelihood principle,
the focus on the basic causal inference problems that arise in stopping rules, asymptotics, maximum likelihood, and model
randomized experiments and observational studies. The book selection. Davies’s work, though likely contentious, is both
is most useful for researchers in statistics, social, and biomed- thoughtful and novel. I would primarily recommend the book
ical sciences who have a solid background in probability and to those interested in the foundations of statistics. Applied
statistics and are looking for a rigorous, but not overly techni- researchers with a strong grounding in probability theory may
cal, introduction to causal inference. While the book does not also be interested in exploring the methods.
cover some important topics that may be of interest (and yet the All opinions and conclusions expressed in this review are
book is more than 600 page long!), it provides readers with the those of the reviewer and do not necessarily reflect the views
foundation to explore the exciting and growing methodological of the Federal Deposit Insurance Corporation.
research on causal inference.
Alicia A. Lloro
Federal Deposit Insurance Corporation
References In the Public Domain
Angrist, J. D., and Pischke, J. S. (2009), Mostly Harmless Econometrics,
Princeton, NJ: Princeton University Press. [1365]
Holland, P. W. (1986), “Statistics and Causal Inference” (with dis-
cussion), Journal of the American Statistical Association, 81, 945–
Data Analysis With Competing Risks and Intermediate
960. [1365] States. Ronald B. Geskus. Boca Raton, FL: Chapman &
Morgan, S. L., and Winship, C. (2007), Counterfactuals and Causal Infer- Hall/CRC Press, 2015, xxix + 247 pp., $79.95 (H),
ence: Methods and Principles for Social Research, Cambridge, UK: Cam-
bridge University Press. [1365] ISBN: 978-1-46-657035-1.
Pearl, J. (2009), Causality: Models, Reasoning, and Inference, Cambridge,
UK: Cambridge University Press. [1365] Competing risks studies are increasingly common in the era of
data science and big biomedical data. These studies essentially
Kosuke Imai involve dependent or informative censoring mechanisms. One
Princeton University example of competing risks is in breast cancer where a patient
can have progression of the cancer after receiving treatment,
or the patient may die without progression. It continues to be
debated in the research community whether treatment benefits
Data Analysis and Approximate Models: Model should focus on prolonged time to progression or time to death.
Choice, Location-Scale, Analysis of Variance, Some breast cancer patients may also experience progression
Nonparametric Regression and Image Analysis. Laurie before death, and progression in this case is an intermediate state
Davies. Boca Raton, FL: Chapman & Hall/CRC Press, 2014, between the initial state of diagnosis or beginning of treatment
and the final state of death. There are many other examples of
xvi + 304 pp., $104.95 (H), ISBN: 978-1-48-221586-1.
competing risks, not only in medical science but also in engi-
neering and the social sciences.
Laurie Davies’s Data Analysis and Approximate Models presents The book targets readers such as epidemiologists and medi-
an alternative approach to statistical analysis that explicitly treats cal statisticians. In contrast to the previously published Compet-
probability models as approximations. The approach is neither ing Risks and Multistate Models with R, by Beyersmann, Allig-
frequentist nor Bayesian: the data are taken as given, there are nol, and Schumacher (2012), this book is less heavy on counting
no true but unknown parameters, and there is no assumption of process notation and theory (often used in survival analysis) and
a true model. instead emphasizes practical examples (using R) and interpreta-
The basic idea is that “a model is an adequate approximation tions.
to a dataset if ‘typical’ data generated under the model ‘look like’ Some of the terminology in the book deviates from that used
the real data. The words ‘look like’ mean that certain features in classical survival analysis and epidemiology practice in the
of the data are of importance and must be exhibited by ‘typical’ U.S. Putting these relatively minor differences aside, the book
sets generated under the model... The word ‘typical’ is opera- reads relatively smoothly. The first chapter “Basic Concepts” is
tionalized by specifying a percentage such that this percentage quite extensive, covering concepts such as censoring complete
of samples generated under the model exhibit the features defin- data (the usual noninformative right-censoring in survival data
ing ‘look like’ ” (p. xiv). The difficulty lies in specifying which can bring substantial complications to the otherwise “censor-
features define “look like.” Davies notes that selection of such ing complete data”). This chapter also introduces left trunca-
features will depend on the purpose for analyzing the data and tion, focusing on conditional inference with no discussion of the
eschews the notion that this process be dictated by purely statis- unconditional approach that has recently appeared in the sur-
tical concepts. In practice, selection of these features could prove vival analysis literature. To my knowledge, little work has been
tricky, and will likely require careful consideration. done using unconditional inference for competing risks.
After formalizing the concepts of approximation, Davies Chapters 2 and 3 consider nonparametric estimation in the
discusses approximation in the context of discrete data, the one-sample setting, that is, without any regressors or predictors.