Book Reviews

Click to increase image sizeClick to decrease image size

Applied Statistical Inference: Likelihood and Bayes.

Leonhard Held and Daniel Sabanés Bové. Heidelberg, Germany: Springer, 2013, xiii + 376 pp., $59.99 (P), ISBN: 978-3-642-37886-7.

This book provides a high-level, yet intuitive overview of statistical inference from frequentist and Bayesian points of view. It is very well written and enjoyable to read, especially the first six chapters that succinctly cover its fundamental concepts.

The book is structured differently from more traditional books on statistical inference. Chapter 1 introduces several motivating examples. Chapter 2 covers the usual topics related to likelihood-based methods, but also includes more advanced topics such as the expectation–maximization (EM) algorithm and generalized likelihood for survival analysis. I found it quite interesting and refreshing to see how the authors navigate through different topics with ease and present them in an intuitive and coherent framework. In fact, an underlying theme is that each chapter includes all related methods even if they are not discussed together traditionally. Therefore, in Chapter 3 where the authors discuss elements of frequentist inference based on hypothetical repetition of sampling experiments, they include the bootstrap approach along with more traditional methods to find standard error and confidence interval. Chapters 4 and 5 are related to frequentist properties of the likelihood for single parameter and multiparameter models, respectively.

Chapter 6 covers introductory topics on Bayesian inference. Bayesian computation is discussed in Chapter 8. Although in general I found the authors’ decision to break away from the norm and put related methods together in the same chapter quite reasonable and one of the strong points of this book, I think many people will find their decision to include empirical Bayes in a chapter on Bayesian inference questionable. The remaining two chapters, 7 and 9, discuss model selection and prediction for frequentist and Bayesian methods side-by-side.

Each chapter includes many interesting examples. Additionally, Chapters 2 to 9 include a set of carefully designed exercises, which could help readers assess their understanding of the materials.

As the title suggest, the authors mainly focus on application of frequentist and Bayesian methods to solve common inferential problems. For the most part, they avoid philosophical comments that might be perceived controversial. This is of course in line with the statistics community as a whole, which is moving beyond the bitter arguments of past decades and is becoming more pragmatic. I do, however, believe that the authors could point out some relevant facts without being controversial. For example, in Section 3.3, when they discuss significance tests and p-values, they could point out how this approach violates the strong likelihood principle presented earlier in Section 2.5.2. Also, when discussing empirical Bayes in Section 6.7, they could mention the James–Stein estimator and the corresponding proof that maximum likelihood estimations (MLEs) for Gaussian models are inadmissible for dimensions higher than 2. That brings me to a suggestion that I hope the authors consider in their next edition: the book could benefit from a chapter on decision theory and another chapter on large-scale inference. Then, the section on empirical Bayes could move to one of these chapters.

In summary, this is a great book that could be used as a textbook in introductory to intermediate level graduate courses on statistical inference. Its intuitive style would also make it a valuable companion to any of the standard textbooks used in advance-level courses.

Babak Shahbaba

UC Irvine

Bayesian Networks: With Examples in R.

Marco Scutari and Jean-Baptiste Denis. Boca-Raton, FL: CRC Press, 2014, xv + 225 pp., $89.95 (H), ISBN: 978-1-4822-2558-7.

This is a compact book that provides an overview of Bayesian networks using examples in the R statistical software package. It covers the basics of Bayesian networks when all the variables are categorical and are modeled using multinomial distributions, when all variables are continuous and follow Gaussian distributions, and also the case of hybrid Bayesian networks that include both categorical and continuous variables. The examples in R illustrate statistical and computational aspects of learning Bayesian networks from data. They also illustrate the use of Bayesian networks as a visualization tool of complex multivariate distributions and as an inference tool using probabilistic reasoning algorithms. The book would be suitable for a short course on Bayesian networks targeting master’s students in Statistics, or Ph.D. students in other quantitative disciplines. Substantial background in probability theory, statistical inference, and Bayesian statistics and some proficiency with R is needed to fully appreciate the book.

The book is divided into six chapters. Chapter 1 introduces Bayesian networks when all the variables are categorical and uses examples analyzed with the R package blearn to define directed acyclic graphs and the associated conditional probability distributions. The chapter also describes some frequentist and Bayesian methods for estimating the conditional probability distributions and for learning network structure using a multinomial likelihood. Display of the network and examples of algorithms for exact and approximate inference are also described. Chapter 2 has essentially the same structure but focuses on Bayesian networks with continuous variables that follow Gaussian distributions and describes an additional R package for Gaussian networks (rbmm). The third chapter describes Bayesian networks that combine both categorical and continuous variables and shows examples of parameter learning using the BUGS software. Unfortunately the chapter does not address the challenge of learning the structure of hybrid networks from data. Chapter 4 provides some of the theoretical and statistical foundations of Bayesian networks. The chapter introduces a general definition of Bayesian networks and the Markov properties that are described by the directed acyclic graph. The chapter also describes the Bayesian approach to parameter and structure learning, and provides a comprehensive overview of search algorithms for learning Bayesian networks from data. Chapter 5 describes some additional R packages for Bayesian analysis of networks, in addition to providing an overview of the BUGS software and stand-alone software for learning and reasoning with Bayesian networks. Finally, Chapter 6 discusses two applications with real data. The book includes an Appendix on graph theory and on probability distributions. The companion website (www.bnlearn.com/book-crc) provides a brief list of corrections to typos, R code, and the datasets used in the book.

Several excellent books about learning and reasoning with Bayesian networks are available (Pearl 1988; Cowell et al. 1999; Neapolitan 2003; Pourret, Naim, and Marcot 2008; Koller and Friedman 2009) and Bayesian Networks with Examples in R provides a useful addition to this list. The book is usually easy to read, rich in examples that are described in great detail, and also provides several exercises with solutions that can be valuable to students. The book also provides an introduction to topics that are not covered in detail in existing books, for example, it addresses the challenge of working with hybrid networks and includes example of BUGS code that can be used for learning parameters of a hybrid network. It also provides a good list of search algorithms for learning Bayesian network structures. But the major strength of the book is the simplicity that makes it particularly suitable to students with sufficient background in probability and statistical theory, particularly Bayesian statistics. However, the book would not be suitable to Ph.D. students without a statistics background.

Readers who are interested in a gentle introductions to Bayesian networks will find Bayesian Networks with Examples in R useful if they have the right statistics background, while readers who are interested in the foundations of Bayesian networks will find parts of the book too elementary. However, excellent books with substantial focus on the foundations of this topic do exist and can be used to integrate some of the topics that are not comprehensively addressed in Bayesian Networks with Examples in R. A more complete reference list in a future edition of the book could facilitate this integration and provide readers with important contributions in the theory and practice of Bayesian networks, as well as links to additional software for learning and reasoning with Bayesian networks.

Paola Sebastiani

Boston University

Bayesian Phylogenetics: Methods, Algorithms, and Applications.

Ming-Hui Chen, Lynn Kuo, and Paul O. Lewis (eds). Boca Raton, FL: Chapman & Hall/CRC Press, 2014, xxx + 365 pp., $99.95 (H), ISBN: 978-1-4665-0079-2.

Phylogenetics is a field of evolutionary biology in which the goal is to use observable, genetically transmissible information to estimate historical evolutionary relationships among a collection of living organisms. The problem of inferring a phylogenetic tree (a binary tree in which nodes represent either observed or hypothetical organisms and branches represent ancestry-descent relationships) can be viewed as an estimation problem in a statistical sense once a model for the evolutionary process is assumed. Maximum likelihood (ML) estimation was introduced to empirical phylogenetics in the 1960s and 1970s, with the first formal ML method for DNA sequence data described by Felsenstein in 1981. In the mid-1990s, Bayesian methodology was applied to phylogenetics independently by several groups, and revolutionized the field in that analysis based on Markov chain Monte Carlo (MCMC) allowed inference under more complex evolutionary models than could be handled by traditional ML methods. Today, Bayesian methodology is widely used in applied phylogenetics, and numerous software packages exist for addressing a variety of evolutionary problems in an MCMC-based Bayesian framework.

Despite the widespread use of Bayesian methods in empirical phylogenetic practice, the edited volume Bayesian Phylogenetics: Methods, Algorithms, and Applications is the first book devoted solely to Bayesian methods in this field, though several other texts contain sections covering Bayesian methods. The affiliations of the editors of the volume—two (Chen and Kuo) are statisticians while the other (Lewis) is a biologist—reflect the strong interdisciplinary flavor of the book, and indeed the editors state in their preface that they hope that the volume will stimulate future cross-disciplinary research by “bringing state-of-the-art phylogenetics to the attention of the Bayesian statistical community, and state-of-the-art Bayesian statistics to the attention of the phylogenetics community.” The diversity of contributing authors further reinforces this focus, with authors spanning the range from statisticians to phylogeneticists.

The volume consists of 13 chapters, including a short introductory chapter written by the editors that outlines topics contained in the book. Though the chapters are not classified into sections, several themes by which the chapters are organized can be identified. Chapter 2, written by Y. Wang and Z. Yang, provides a careful discussion of the choice of prior distribution for the various parameters involved in a Bayesian phylogenetic analysis. The organization of the chapter into sections describing each type of parameter (e.g., tree topology parameters vs. model parameters, such as branch lengths and parameters of the mutation process) separately is very effective in displaying the wide array of possible choices of priors and in evaluating the possible impacts of such choices, both from computational and practical viewpoints. Further, as the book contains no general introduction to either the Bayesian framework or to phylogeny estimation, this chapter serves the additional purpose of providing an introduction to these topics for readers who may not be entirely familiar with both fields.

Chapters 3–6 collectively address one of the themes that emerge from the volume, namely, estimation of the marginal likelihood. This topic is of particular interest in phylogenetics, because of the importance of Bayesian model averaging and Bayesian model selection for applications in this field. Several of the methods in current use for marginal likelihood estimation are covered in one or more of these four chapters, including harmonic mean estimation, inflated density ratio (IDR) methods, path sampling methods, and generalized stepping-stone methods. Chapters 3 (Arima and Tardella) and 4 (Baele and Lemey) focus on situations in which the tree topology is fixed, while chapters 5 (Holder, Lewis, Swofford, and Bryant) and 6 (Wu, Chen, Kuo, and Lewis) consider modifications of existing methods that allow the topology to vary. Chapters 3 and 4 include nice empirical examples that demonstrate the need for precise estimation of the marginal likelihood in a practical setting, while chapter 5 includes a “brute force” example that makes the nuances involved in considering topological variation concrete. Chapter 6 is largely a collection of theorems and proofs concerning consistency of marginal likelihood estimators in the presence of topological variation, with a short example and discussion following these results. Though these chapters are written independently, they reach similar conclusions concerning the utility of various methods of marginal likelihood estimation. In particular, they all find that harmonic mean estimation performs poorly in comparison to other methods, and that selection of appropriate priors is important, results that are in agreement with current understanding in the statistical literature on this topic (e.g., Raftery et al. 2007).

The next three chapters (chapters 7–9) address the issue of computational efficiency in MCMC methods for inference of phylogenetic quantities. Chapter 7 (Cheon and Liang) begins with an overview of methods for Bayesian phylogenetic inference, before discussing “advanced” MCMC phylogenetic algorithms. The most interesting portion of this chapter is the discussion of stochastic approximation Monte Carlo and sequential stochastic approximation Monte Carlo. Chapter 8 (Bouchard–Côtè) stands out as a highlight of the book, providing an exceptionally clear description of sequential Monte Carlo (SMC) methods for Bayesian phylogenetics, with a view toward limitations of current methods and potential for future innovation. Chapter 9 (Palczewski and Beerli) considers computational challenges for the specific problem of multi-locus analysis at the population level, and presents results that enable efficient inference in this setting.

The remaining chapters (chapters 10–13) describe a selection of important current research topics in Bayesian phylogenetic inference. Chapter 10 (Kuhner) describes ancestral recombination graphs (ARGs), which represent the result of the biological process of recombination, and the use of Bayesian methods for inference in this setting. Chapter 11 (Palacios, Gill, Suchard, and Minin) presents Bayesian methodology for phylodynamic inference, which refers to the setting in which inference of parameters such as effective population size (and, in particular, changes in effective population size over the time) is of primary interest. Examples of this include inference of viral dynamics in an evolutionary framework and inference for samples that are collected at varying times (rather than contemporaneously, as is assumed elsewhere in the book). Chapter 12 (Hobolth and Thorne) considers an inferential framework that makes use of Markov properties for endpoint conditioned processes. These results are applied in the phylogenetic setting by conditioning on the nucleotides at the ends of the branches in a phylogeny. This framework is advantageous because it allows more complex models than are typically possible for MCMC-based inferential methods to be fit, such as models that incorporate dependency in the mutation process for neighboring sites in a DNA sequence. Chapter 13 (Heath and Moore) provides a careful description of the problem of estimation of speciation times for a collection of contemporaneous species, with attention to the problem of “calibration” of a phylogeny (assignment of events along a phylogeny to actual geological time to allow estimation of absolute, rather than relative, times of speciation events). Calibration is arguably the most challenging step in divergence time estimation, and this chapter provides a concise description of methods and issues, including use of fossil information in the calibration process. Taken together, these four chapters provide an effective overview of some of the most critical issues in statistical phylogenetics in which future methodological developments are needed.

In their preface to the book, the editors state that the book is suitable for researchers at or above the graduate student level in either statistics or biology. While this is true for many parts of the book, several of the chapters assume a more intensive background in one area or the other. In particular, several chapters (e.g., chapters 3, 6, and 12) contain fairly technical mathematical/statistical material, and all but the most statistically savvy biologically oriented readers will struggle with portions of these chapters. Likewise, the volume does not include a chapter that provides a basic definition and description of parameters in a standard phylogenetic analysis; instead, some chapters provide a short introduction, while others assume that the reader is already familiar with these concepts. This has the advantage that the chapters are largely self-contained, and thus can be read independently of one another. Also, for readers with some background in phylogenetics or evolutionary genetics more generally, chapters 2 and 7 might provide a sufficient introduction to the area. However, a reader who is completely new to one or the other field would benefit from reading some introductory material before tackling any of the chapters in the volume. On the other hand, researchers from either statistics or biology working in methodology development in genetics will benefit immediately from the clear presentation and level of detail throughout the book.

Overall, the volume does achieve the editors’ primary goal of bringing state-of-the-art developments at the intersection of Bayesian methodology and phylogenetic inference to the forefront in a manner in which they can be appreciated by researchers in both fields. I share the editors’ hope that this will spur further methodological development in these areas. The topics included in the book, centered around the themes of marginal likelihood estimation, computational challenges in MCMC-based inference in phylogenetics, and high-impact topics, seem to have been specifically selected to encourage this, as they represent the most timely and pressing problems in a field that will continue to grow in complexity and importance as large-scale genetic data become more readily available.

Laura S. Kubatko

The Ohio State University

Clinical Trial Design: Bayesian and Frequentist Adaptive Methods.

Guosheng Yin. Hoboken, NJ: Wiley, 2012, xvii + 336 pp., $128.00 (H), ISBN: 978-0-470-58171-1.

In recent years, there has been much interest in adaptive clinical trial designs (FDA 2010). Although there has been a great deal of methodological development appearing in the recent statistical literature, the implementation of new methods has lagged. It is generally accepted that a better understanding of adaptive clinical trial designs, when they are useful, and when they should be avoided, is needed to achieve the perceived benefits of adaptation. Correspondingly, several recent books have been published to provide a general overview and better understanding of different types of adaptive clinical trial design (Chow and Chang 2007; Berry et al. 2008; Chang 2008; Ting 2010). The book from Guoshen Yin, Clinical Trial Design: Bayesian and Frequentist Adaptive Methods, is one more recent addition to this growing literature. As with the other books on adaptive clinical trial design, statisticians are the intended audience. However, this book differs in two important ways. First, in addition to describing adaptive clinical trial designs, the book provides a general overview of clinical trial design. Second, it attempts to address adaptations from both Bayesian and Frequentist perspectives, whereas many other texts focus only on one or the other. To accomplish these objectives, the author provides a basic overview of clinical trial design in the first two chapters. Chapter 3 is devoted to a comparison of the Frequentist and Bayesian approaches. Chapters 4–6 provide descriptions of standard and adaptive trial designs in phase I, phase II, and phase III trials, respectively. The final four chapters provide overviews of more advanced adaptive methods that have been proposed—adaptive randomization, late-onset toxicity designs, drug-combination trials, and targeted therapy designs.

Initially, I found the aim of addressing both Bayesian and frequentist approaches an attractive feature of the book. In particular, Chapter 3 provides a very good comparison of the two approaches. The material provided in this chapter could be extremely useful in many statistics and biostatistics graduate programs. However, at other points, the promise of combining the two approaches falls somewhat short of its potential. Although the book provides a discussion of both approaches when introducing methodology, almost all of the provided examples involve only Bayesian designs. This likely results from the author's experience, which appears to have predominantly involved Bayesian trials. As a result, the presentation of the two approaches is not unbiased in that there is a subtle slant in favor of Bayesian designs. While this may appeal to readers with a more Bayesian background, others might find the examples lacking. Given the ostensive emphasis on both approaches, a more balanced set of examples would be preferred.

Another concern is that all of the examples are focused on the design aspects of the trials. It has been noted by several authors that operational complexities are an important limiting factor (Quinlan and Krams, 2006; Gaydos et al. 2009; Quinlan et al. 2010; Coffey et al. 2012). Some discussion of practical aspects of implementing the designs and the problems that may be encountered in practice would have been welcome additions.

Finally, the attempt to include a general introduction of clinical trials and adaptive trial design in the same book may have resulted in a small target audience. Those interested in learning more about the statistical properties of adaptive trial designs are likely to already have a strong understanding of the fundamental concepts of clinical trials. They are unlikely to find the introductory text on clinical trials very useful. On the other hand, someone with little basic understanding of clinical trials may find these first two chapters useful but may struggle to understand the more advanced statistical concepts presented with the adaptive methods. As a result, the market for this type of book may be somewhat limited. For someone with a background in clinical trials, and an interest to learn more about adaptive designs, this book does not seem to provide anything substantial that would make it stand out from similar books on the market. Thus, if a reader wants to choose a single book on adaptive designs, I would not recommend this one over some of the other books already on the market (e.g., Chow and Chang 2007; Berry et al. 2008; Chang 2008; Ting 2010). Similarly, this book probably does not offer sufficient depth for someone wanting to learn about the basics of clinical trial design; a better choice might be Friedman, Furburg, and DeMets (2003) or Piantadosi (2005). Overall, this book is best suited to the reader that is looking to assemble a library of many different approaches presenting the main concepts of adaptive clinical trial design.

Christopher Coffey

University of Iowa

Fighting for Reliable Evidence.

Judith M. Gueron and Howard Rolston. New York: Russell Sage Foundation, 2013, xvii + 575 pp., $49.95 (P), ISBN: 978-0-87154-493-3.

The Royal Statistical Society's 1887 Charter, issued by “the grace of God,” was graced also with Queen Victoria's declaration that RSS members had “…seriously pursued…the discussion of legislative and other public measures from the statistical point of view” (p. 1). Gueron and Rolston's Fighting for Reliable Evidence embodies the spirit of the Charter. Their book recounts efforts to find acceptance for the idea of randomized trials in the welfare and labor sectors, to mount trials so to learn what works and what does not and for whom, and to meet the challenges of doing so in often volatile political environments. It tells the reader how and by whom the challenges were met.

Histories of randomized trials in medicine come readily to hand. Marks’ (1998) Progress of Experiment, Paul Meier's (1972) articles on the Salk trials, Silverman's (1980) monograph on retrolental fibroplasia, and the James Lind Library (2014) collection are invaluable to those interested in public health. The history of conducting randomized experiments in the social, behavioral, and education sciences is not so well documented. Fienberg et al's (1985) chapter on large scale experiments can be construed as an exception though it focuses attention on statistical rather than operational issues.

Gueron and Rolston's Fighting serves well on this account. Its authors cover 40 years of experience in mounting randomized trials on federally sponsored social welfare and employment programs in the United States. Not for the faint of heart, the volume runs to almost 600 pages. Footnotes are abundant. It is studded with informative references.

The book's authors have been leaders in getting trials off the ground in their ambit. Gueron is a former president of the Manpower Demonstration Research Corporation (MDRC), an organization that has played a remarkable role in conducting the field trials. Rolston was a Senior Executive at the U.S. Department of Health and Human Services’ Administration for Children and Families and in other turf. He served through several presidential administrations in agencies that fostered a taste for better evidence, especially randomized trials, and generated the funding for them.

Their book is autobiographical—a contemporary history. Some readers may be put off by this approach, but there is good precedent. Otto Larsen's (1992) account of the social sciences division at the National Science Foundation is a fine portrayal of early investments in social research in the public interest under his leadership. More to the point, Fighting's authors’ contrapuntal styles help to sustain reader interest. In Gueron's chapters of the book, the prose is tilted toward admirable feistiness. Rolston's prose embodies remarkable perseverance and negotiation ability. Rather more important, the authors’ complementary expertise, reflected in the books’ contents, were essential to making things happen.

The programs that have been put to the test in randomized trials, and not without serious difficulties, have been remarkable. The scale is a bit staggering. Supported Work Demonstrations, a precedent, involved nearly 7000 individuals in 10 sites, who were followed up repeatedly over 20 months. The Work Incentive (WIN) trials involved four sites, and were preceded by 7 months of negotiation to figure how to design the interventions and the trial itself. The Work/Welfare Demonstration project was a major turning point in that it entailed mandatory, rather than voluntary, program participation. In this, random assignment procedures were integrated into normal welfare operations in eight states and with 32,000 individuals so as to learn about how to improve the programs and for whom the putative improvements worked well. The Teen Parent Demonstration and its successors Learnfare and other waiver-based experiments are covered. The 1996 Personal Responsibility and Opportunity Reconciliation Act (PWORA) trials were deployed in almost 45 states. Florida's Project Independence, lasting 6 years, is in the mix, as are California's GAIN trials (8 years).

Statisticians who try mightily to introduce ideas about experimentation into various fields learn from this book about battle, of political, bureaucratic, and rhetorical varieties. Lest some readers regard this language as mere hyperbole, recall Claude Bernard's (1865/1957) declaration in his book on medical experimentation that “…man is naturally metaphysical and arrogant…(F)rom this it follows the experimental methods is not really natural to him.”

High stakes variations on social programs, put to the test in randomized trials in the welfare and employment sectors, engendered controversy. “Insults were hurled and law suits threatened.” (p. 264). Being called a Nazi and a party to Tuskegee ethics cannot have been pleasant, even if the appellations were off the mark, and despite the authors’ tolerance for obloquy and their tough hides. Gueron and Rolston report on external struggles, involving state bureaucracies and political stakeholders, and internal struggles, which involving regime shifts at the federal level and wild variations across States in their taste for dependable evidence on public policy. The workload, we are told, was enormous, and threats to staff morale periodically high. As Gueron observes, “From the current vantage point, the value and feasibility of random assignment may seem so obvious as to be uncontroversial. My point…is to remind readers that gaining this reputation took relentless determination over a long time” (p. 264). In 2004 a Washington DC pundit put the matter pungently in related economic contexts: “If you can't take the dysfunction, get out of the capital.”

In Fighting's ambit, academic opposition to randomized trials, from a Nobel Laureate in economics among others, gets attention. Read the book to identify the miscreant. Of course, high skepticism if not outright opposition to trials was not uncommon during the 1970s and 1980s in other sectors. An elder statesman in education research, Lee Cronbach (1982), for instance, viewed the aspiration to deploy scientifically respectable field experiments as naïve and perhaps misguided. Prior to the 1980s, randomized trials in criminology were thin on the ground (Farrington 2003). Gueron and Rolston aver that support for their trials came from “only a handful” of scholars. They identify a subset of remarkable welfare or labor economists and acknowledge the latter's contributions often. That is to the good. Readers who understand the challenges will appreciate the mutual interests of some statisticians and economists. A few able statisticians such as Light, Singer, and Willett are mentioned, albeit rarely.

The book's inattention to statisticians who have toiled in related vineyards may puzzle, or annoy, some JASA readers. Consider for instance, Fred Mosteller, who in serving on the President's Science Advisory Committee (PSAC) among other duties, boosted randomized trials consistently over 50 years. Mosteller, in fact, collaborated with Senator Daniel Patrick Moynihan who is recognized often in Fighting on account of Moynihan's appetite for dependable evidence in education and welfare. Mosteller and fellow statisticians William Kruskal, John Pratt, and Bernard Greenberg contributed, alongside economists such as Harold Watts and Al Rees, to the Social Science Research Council's Committee on Social Experimentation in the early 1970s; see the preface to Riecken et al (1974). These statistical saints in the pantheon are not recognized in Fighting. The book's authors do, however, recognize reports by the National Academy of Sciences and various groups convened by the Department of Labor and others, each of which included statisticians. These scientific reports added to the governmental support of the trials.

The idea of a fair comparison is, of course, not the property of statisticians. It has a long history. But the idea of such a comparison based on randomized allocation (though not its probabilistic underpinnings) figures prominently in Gueron and Rolston's descriptions of quarrels over whether a randomized trial should be run. The absence of a need for “fancy statistical footwork” is counted as a virtue in selling the idea. If one reads Chapter 7 carefully, the reader will find out about political attempts to dodge a randomized trial and to depend instead on alternative approaches to estimating effects of programs. This is in the context of distrust of relatively complex models based on passive observational data. “Small difference in a great many assumptions (and little evidence) yielded wide variations in estimates…” (p. 229). It was also in the context of related distrust of “politicized economic research” (Schultz 1982, p. 122). In fostering the idea of trials, Gueron and Rolston reiterate another kind of rhetoric: “the Supported Work Demonstrations represent a social initiative, not a research experiment… and… would enable the country to learn some useful lessons from the demonstrated results” (p. 30).

Achieving acceptance for randomized trials as the approach of choice in estimating effects of welfare innovations took virtue well beyond the statistical variety. Developing incentives and justifications, resolving political-institutional disputes, meeting ethical standards, and managing field operations have also been challenges in experiments in crime and justice (Strang 2012), education (Mosteller and Boruch 2002), and elsewhere (e.g., see the entire 2005 issue, Volume 599, of the Annals of the American Academy of Political and Social Sciences). In earlier times, the experience in some commercial/industrial sectors has been no less demanding; see Lide (2001), for instance, on the battery additive controversy and the National Bureau of Standards (now NIST) in the 1960s, and the engagement of statisticians such as Eisenhart.

Fighter's readers get a bit of instruction about constraints on designing fair trials. In some, individuals are the units of random allocation. In others, it is entities such as counties of States or Offices within counties of states that were randomly allocated to different regimens. In still other cases, the jurisdictional units were matched, then randomized (p. 235). Nowadays, some of the studies would be called “cluster randomized trials” in medicine, or called “place randomized trials” in the crime and justice sectors. Repetition, in the sense of full or partial replications of trial appears to have counted heavily. There was appreciation in legislative venues for cumulative evidence, as opposed to one-off studies and their results.

The civil servants, senior executives, and political appointees, in various federal and state agencies, who have encouraged the idea of dependable comparisons has been important. The people, including legislative staffers and their bosses, are named. So too are members of opposing camps. The federal organizations, such as Administration for Children and Families, the Office of Management and Budget, and the White House's Office of Policy Development and relevant agencies at the State level are recognized.

Funding was an incentive for research organizations, just as it was and is for academics. Challenge grants from the Ford Foundation figure into the earliest trials. Funding was awarded more frequently from multiple federal sources, sometimes simultaneously and sometimes sequentially. Conditional waivers from federal regulations were important, to judge from Fighting. A state, for example, might be given permission to depart from certain federal rules in deploying a federal program so long as the program was tied to the condition of producing reliable estimates of the innovation's effect. The relevant section 1115(a) of the Social Security Act allows the relevant agency to waive compliance to ordinary regulations in the context of “experiment, pilot, or demonstration project.” The feds were in the position of deciding what “reliable” means. On account of OMB and others, the tilt was toward randomized experiments.

Statisticians who have not dipped into these waters may be taken aback or even disgruntled, by Fighting's references to MDRC, WIN, TPA, and others. PWORA does not slip easily from the tongue. That nearly 50 abbreviations are defined near the book's end is a comfort. That other more familiar abbreviations, such as ITT, TOT, and SUTVA are absent should not surprise readers. Fighting is not a statistical tract. Its authors succeeded in mounting randomized trials because they understood the benefits of the method and put them plainly and in the substantive context. Recall the declarations of Deming and others who advised us to learn the substantive vernacular to do good statistical work.

A bit more editing might have helped Fighting's readers. Though the book's story lines are engaging, the language is a bit fevered at times. Because the chapters depend on the authors’ experience, the use of personal pronouns is, at times, incontinent. On one stretch of seven pages, for instance, over 50 “I's” and “mys” appear. The timelines and events that drove the development of trials are portrayed in a host of charts. But still more charts would have helped readers to get a good grip on happenings.

The bottom line is this: From Fighting, readers may learn how courage, stamina, and intelligence have been employed in getting randomized controlled trials off the ground over four decades, and in the interest of learning what works better in welfare and training programs in the United States. Queen Victoria, or some of her colleagues at the Royal Statistical Society's inauguration, might have been pleased. More importantly, those who now and in future want to meet the challenges of deploying field trials can better anticipate what may befall them on account of Fighting.

Robert Boruch

University of Pennsylvania

Generalized Linear Models for Categorical and Continuous Limited Dependent Variables.

Michael Smithson and Edgar C. Merkle. Boca Raton, FL: CRC Press, 2013, xxiii + 284 pp., $94.95 (H), ISBN: 978-1-4665-5173-2.

There is no shortage of textbook treatments of generalized linear models (GLMs). What new perspective or emphasis could be brought to this crowded market? This contribution from Smithson and Merkle is distinctive in several respects. It is geared toward a social-science audience, where fitting regression-like models to cross-sectional, individual-level data is an everyday task. Micro-level social data are often discrete, perhaps more than students and practitioners initially think. Examples includes modeling choice from a small set of options (sometimes ordered, sometimes not), or counts of particular behaviors. Social data are often censored or truncated; in addition to survival times, examples include spending or earnings data, hours worked, many scales used in psychological and educational assessments, or response times.

The authors cover an impressive amount of ground in under 300 pages, so much so that the book’s title is perhaps somewhat misleading (as the authors concede on p. 4), but not unpleasantly. The book certainly does encompass GLMs. Chapter 1 introduces GLMs and provides a quick summary of parameter estimation, inference, and model comparison for GLMs. Logistic regression for binomial data is covered in Chapter 2 and models for count data in Chapter 5. But there is much more here. Chapters 3 and 4 cover models for unordered and ordered dependent variables; the authors carefully distinguish the sometimes subtle differences between various models for these types of data, although the absence of even a description of the multinomial probit model was a curious omission. The treatment of models for counts in Chapter 5 includes an extensive discussion of zero-inflated and hurdle models. Beta GLMs for doubly bounded data are the subject of Chapter 6 and models for censored and truncated data appear in Chapter 7. Exercises accompany each chapter. Bayesian extensions are covered in Chapter 8; for instance, hierarchical models are presented that introduce varying intercepts and regression slopes to binomial, Poisson, and Beta GLMs. Thus, in one book, the authors cover many different models for a wide array of discrete and limited dependent variables (LDVs).

Each chapter is accompanied by worked examples in R and Stata, typically drawing on examples from psychology, the authors’ home disciplines. Given the broad reach of the book, it is unsurprising that a large number of R packages are used. This guide to fitting many standard and not-so-standard models for categorical and LDVs in R is a valuable contribution in itself.

At just under 300 pages, a book that is this comprehensive—and devotes considerable space to worked examples with R and Stata code—must move rather expeditiously. The authors acknowledge that more technical and more theoretically elaborate treatments of particular models appear elsewhere. Derivations of the models—from either statistical foundations, or psychological or economic theories about decision making—are compact. The emphasis here is clearly on model fitting and comparisons among particular models for a given type of discrete response or LDV. This is appropriate, given the aims of the book. Yet, I suspect that many readers will be consulting the references for details on how various models follow from first principles, underscoring how particular models embody particular assumptions about decision making (e.g., the development of the logistic regression model and its extensions from utility maximization).

The authors also concede that almost all the models they consider are linear and additive with respect to the predictors, as is typical in most other treatments of GLMs and models for LDVs. Faraway (2005), Hastie and Tibshirani (1990), and Wood (2006) considered nonparametric and semiparametric GLMs, but cover a far smaller set of models than those considered here. Another interesting issue is whether and how a text at this level might cover regularization for GLMs (e.g., Park and Hastie 2007) or emphasize cross-validation as a model evaluation criterion. In much of the social sciences (and many other applied domains), the prevailing approach seems to be to defer a consideration of ideas from machine learning to a separate class and a dedicated text; this approach is adopted by Smithson and Merkle too.

More fundamentally, a book at this level might want to repeatedly stress the shortcomings of Gaussian GLMs for discrete data or LDVs. At least in economics and political science, it is uncontroversial to fit linear regressions to binary data via ordinary least square (OLS), with standard errors estimated with a Eicker–Huber–White covariance matrix (one of many such examples is Snyder and Strömberg 2010, Table 4). This practice might be reasonable, say when the goal of the regression analysis is not to generate a predictive model for the binary outcomes, but to test for differences in proportions across groups with large sample sizes and a few control variables. Many target readers of this book will ask why Gaussian GLMs cannot be used in the settings considered in the book, especially if the alternatives are perceived as overly elaborate and harder to fit and interpret. (In my experience, this is not an uncommon reaction when introducing students to this material.) The authors provide a brief summary of the shortcomings of Gaussian GLMs with binary responses (pp. 17–18), but it is a point that I wish had been stressed in many other settings.

This book will be a valuable resource for graduate students and practitioners in psychology, political science, international relations, anthropology, sociology, education, and communication, branches of the social sciences where the data and models outlined here are routinely encountered. The Beta GLM and some of the models for censored and truncated responses are much newer and more exotic than, say, logistic regression or count models, making Smithson and Merkle’s book a welcome addition even for seasoned practitioners, who are likely to be familiar with the material appearing in the book’s earlier chapters.

Simon D. Jackman

Stanford University

Implementing Reproducible Research.

Victoria Stodden, Friedrich Leisch, and Roger D. Peng. Boca Raton, FL: CRC Press, 2014, xix + 428 pp., $79.95 (H), ISBN: 978-1-4665-6159-5.

Since John Ioannidis published his article entitled, “Why most published research findings are false” (2005), the reproducibility of scientific research—and the lack thereof—has become a research topic in and of itself. The intervening decade has seen the development of helpful software tools, the establishment of more thoughtful protocols, and the revision of academic publishing practices, all aimed at reducing irreproducibility. In Implementing Reproducible Research, three statisticians have compiled in one edited volume an encapsulation of the best research practices that have emerged across the sciences. This collection brings together the expertise and experience of numerous authors and is likely to be valuable to scientists and statisticians alike.

Very usefully, the book begins by distinguishing the reproduction of an analysis from the replication of a study. Reproducibility is defined as, “the calculation of quantitative scientific results by independent scientists using the original datasets and methods” (p. vii). Replication, then, goes the additional step of collecting new data. If replication is one of the foundational principles of science, ensuring a reproducible data analysis is its necessary precursor.

In the preface, the editors specify that their objective is not to convince the reader why reproducibility is important, but rather how it can be achieved. They organize the book into three parts, Tools, Practices and Guidelines, and Platforms, which mirror the three directions of research in reproducibility. One contributor then usefully divides tools for reproducible research into three general categories: tools for literate programming (Chapter 1: knitr in R), workflow management systems (Chapter 2: VisTrails), and tools for environment capture (Chapter 3: Sumatra, Chapter 4: CDE). Although I was already familiar with knitr, an R package used to integrate text and computing, I still learned a lot from the authors’ discussion of the design elements that specifically enable reproducible analyses. The remaining tools were new to me and address the added challenges when faced with a more complex workflow and software that involves many dependencies. These chapters are not full tutorials; the reader will not emerge as a well-versed user, but hopefully will emerge with a sense of which tools they would like to invest time in learning.

The remainder of the book is a mix of more general and technical pieces, some of which are quite domain specific. Chapters 5, 7, and 13 discuss methods from the physical sciences, biology, and machine learning, respectively, though their approach to computation may be of interest to those in other fields. Chapters 6, 9, and 15 discuss arguments and techniques for practicing open science, an important consideration if the goal is to allow anyone to reproduce a study without barriers. Chapter 8 provides an industry perspective of a large-scale data analysis project. Chapter 10 describes how cloud-based virtual machines (VM) can enable anyone to recreate a computational experience in an environment identical to that of the original analysts. Chapter 12 reviews the tenets of traditional intellectual property law and discusses the difficulty this poses for reproducibility. Chapter 14 introduces the reader to runmycode.org, a user-friendly web-based platform that allows one to tweak-the-knobs, so to say, on an analysis.

This book should have broad appeal; the more general chapters should be of interest to any empirical scientist. Most contributors assume some familiarity with scientific computing, and for the most technical chapters, familiarity with UNIX or Linux-based platforms will be useful. I can imagine this book serving as a guide to improved reproducibility within a research lab or as terrific material to kickstart discussion for a university working group on reproducibility. As a compendium of best practices, it would also be a valuable reference for new graduate students who are in the process of forming their research habits.

As a cover-to-cover read, the book can be repetitive (literate programming, version control, and topics in open science are covered in multiple chapters). It can also be disorganized, with sections varying greatly in level of technical background (e.g., Chapter 4 assumes familiarity with Linux while Chapter 8 reverts to a primer on the basics of computing). Additionally, some readers might be surprised to find that, for a book published as part of The R Series, Python gets as much coverage as R, and much of the book is language-agnostic. Considering the many arguments that this book makes for open science, it was reassuring to find that full pdfs of every chapter (including a missing one by Malik et al.) can be downloaded cost-free through the Open Science Framework (implementingrr.org). One cannot help but feel this to be the more natural format for this material: dynamic, modular, and with no barriers to access.

Despite some weaknesses of the book format, Implementing Reproducible Research still introduces some extremely useful tools and practices from leaders in the field. On top of that, it also contains an exciting vision for the future of scientific research. This includes the idea of learning from a research processes instead of just the end result, which becomes feasible when work is reproducible and transparent. Part of this will be a shift away from publishing static papers to publishing full research environments. Another aspect is the promise of collaborative science, from the distributed replication studies of The Reproducibility Project (Chapter 11) to the reevaluation of traditional peer-review practices in light of the paradigm epitomized by GitHub’s pull-request.

The challenge of reproducibility in the computational era is being confronted across the sciences, with each field developing its own tools and best practices. This book is an important step in bringing together a broad group of scientists to share what has been learned.

Andrew Bray

University of Massachusetts, Amherst

Measurement Uncertainty and Probability.

Robin Willink. Cambridge, UK: Cambridge University Press, 2013, xvii + 276 pp., $109.99 (H), ISBN: 978-1-107-02193-8.

Assessment of uncertainty in measurement is of particular importance for the physical sciences, but most undergraduate and even graduate-level courses in physical science address the issue only perfunctorily, with reference to a limited set of standard techniques and little discussion of the methodological and philosophical foundations of these methods. This book, written by a self-described “physicist-turned-statistician, not an expert in measurement” therefore fills an important gap between the relatively superficial treatment of uncertainty quantification in applied science and the well-developed but less accessible statistical literature. The approach is frequentist rather than Bayesian, although one of the strengths of the discussion presented is the willingness to subject both approaches to criticism, so that the limitations of each paradigm are clear, especially with regard to practical experimental situations.

The book is subdivided into three parts. The Introduction and initial chapters, labeled “Principles” and comprising Part I, lay out a manifesto for the clear treatment of measurement uncertainty, introducing a variety of ideas themed around the definition and construction of a 95% confidence interval. There is an emphasis on the real-world objectives and the practice of measurement, with extended examples demonstrating the use of terminology and concepts. The lack of consistent definitions in the literature is clearly a major bugbear of the author; his efforts to clarify and suggest improvements here are well thought out and used consistently. Commendably, this is achieved without becoming over-pedantic or preventing easy reading.

Part II, “Evaluation of uncertainty,” takes the principles outlined in Part I (usefully summarized again at the start of Part II, for reference) and continues to describe particular methods for the evaluation of uncertainty, using the same notation, accompanied by remarks about the pros, cons, and implications of each methodological approach. Evaluation using Monte Carlo methods with and without a linear approximation is treated, as is the propagation of errors under different practical circumstances.

Part III is titled simply “Related topics,” which sounds discouragingly like a collection of semirelated items that the author wished to include but could not fit into a narrative. Upon reading, however, there is much to be gained from each of these more self-contained chapters, and they do link back to the body of Parts I and II. Incorporating some of this material into the first two Parts and titling the remainder something like “Further methodological concerns” would have helped this reviewer to follow the structure, but, as it is all clearly signposted, this is a minor quibble.

Included in Part III is a whole chapter discussing the classic Guide to the Expression of Uncertainty in Measurement (BIPM, IEC, IFCC, et al. 1995). The discussion around the limitations of the Guide, its ambiguities, and notational infelicities as well as the philosophical underpinnings, will be of great use to a certain segment of readers and would be a good starting point (before or in parallel with Part I) for someone familiar with the recommendations of the Guide but less familiar with the deeper conceptual issues in measurement uncertainty.

There are many passing references to the philosophical status of model parameters, but very little discussion of the epistemic difficulties of model parameter estimation (although by the author’s definition “estimation” is certainly the wrong word here) given inevitable model inadequacy. This is a significant omission and in my opinion would have made a worthwhile additional chapter in Part III, where it would be a natural accompaniment to the remarks in Chapter 13 on the limitations of the Bayesian approach.

The text is clearly presented and written in a concise but engaging manner, with clear diagrams, useful cross-referencing, and a welcome preference for written explanations rather than dense mathematics. Informative subheadings make it easy to browse or to skip occasional sections without losing the flow, which is particularly helpful since the book proceeds rapidly through a variety of topics, which will be of interest to different readers. This reviewer found Parts I and III most interesting on a conceptual/philosophical level, but Part II will no doubt be of most interest to the researcher seeking practical methods for uncertainty quantification in this framework.

Part I could potentially form a basis for a short graduate-level course, using examples from Parts II and III, but it would not be an obvious choice unless it were found to be a useful complement to a program of other courses. As a pedagogical reference, it would be more at home in further reading lists and I would be likely to recommend it to advanced students with a particular interest in the principles and foundations of uncertainty quantification methods. No exercises are provided in the text although there are extended examples. In general, this book appears most suited to the interested reader already working on related topics, a reader perhaps frustrated by inconsistencies of notation and philosophical approach and looking for a firmer foundation on which to develop their conceptual understanding of the debates in the literature. For that audience, it is a stimulating read covering a wide range of topics, and provides many points of departure for further consideration.

Erica L. Thompson

London School of Economics and Political Science

The Science of Web Surveys.

Roger Tourangeau, Frederick G. Conrad, and Mick P. Couper. New York: Oxford University Press, 2013, viii + 198 pp., $72.00 (P), ISBN: 978-0-19-974704-7.

Web surveys have emerged over the last 10 years as a popular and attractive way to measure public opinion, driven by a global rise in Internet penetration, a relative low cost of distribution, and the ease of administration compared to other modes of data collection. Along the way, an expansive and disparate body of research on web data collection has developed that is overwhelming and often inconsistent. Tourangeau, Conrad, and Couper—three of today’s most influential authorities in survey methodology—combine their mastery of the topic to offer a comprehensive review of the literature in a single book: The Science of Web Surveys. It is a practical and useful reference source that belongs to the bookshelves of anyone interested in web survey methods—expert and novice alike.

The book’s eight chapters are organized around the total error framework. It covers major topics related to nonobservations errors (sampling and coverage, nonresponse), observation errors (response formats, visual designs, interactive features), and using the web as part of a mixed mode data collection strategy. The authors provide a curated perspective on empirical findings, which includes original meta-analyses on important topics and relevant findings from unpublished works. More importantly, they anchor their discussion of findings in established theory so that readers are left with more than just a sense of “best practices,” but instead learn to make informed methodological decisions consistent research purposes.

Tourangeau et al. begin by reviewing sampling and coverage challenges faced by web surveys. The authors consider the theoretical and empirical evidence of the accuracy of estimates obtained by web samples, and decompose differences into coverage, self-selection, and nonresponse biases. They demonstrate that while Internet penetration has grown considerably over time, demographic and attitudinal differences persist between people based on their use of the Internet. A thorough and easily understandable description of common statistical correction methods—relied upon to remove coverage and selection biases—is provided (post-stratification, raking, GREG weighting, propensity scoring), along with a comparison of their effectiveness in general and on web surveys in particular. An included meta-analysis of relevant studies indicates that regardless of method used, statistical corrections typically remove less than half the bias in estimates from web surveys, and in some instances increase bias.

Some readers may be disappointed that the meta-analysis in Chapter 2 does not (1) differentiate between probability and nonprobability web sample sources in its description of remaining bias, even when some studies in the meta-analysis include such comparisons (e.g., Yeager et al. 2011), or (2) address special cases from electoral polling where nonprobability web panels have been shown to be as accurate or more accurate than surveys from alternative modes using probability samples (e.g., Twyman 2008; Vavreck and Rivers 2008). Tourangeau et al. avoid this contentious and evolving debate. Instead, they give readers a valuable and informed framework to think about such issues through an extensive description of the types of probability and nonprobability web samples that are possible and statistical models for the consequences of nonprobability sampling on the size and direction of bias within an estimate.

Chapter 3 describes how web survey response rates are substantially lower than with other survey modes and the authors touch on topics that impact participation, including mode of invitation, number of contact attempts, identifying sponsorship of the study, and using incentives. Taken together, Chapters 1–3 are succinct and provide a thorough foundation for understanding the complex sampling and nonresponse issues impacting web surveys.

The second half of the book focuses on basic visual design decisions for web surveys. Readers are introduced to the many attractive features that give web surveys an advantage over other modes of data collection and warned of the risk that poor design will lead to higher levels of measurement error. Topics such as pagination versus scrolling, the use of grid questions, selective emphasis of important elements of text, choices of response format, and the optimal placement of images and text on screens are thoroughly covered with a synthesis of empirical evidence and insight into the psychology of respondents. The authors’ broad discussion of features includes an important reminder that designers have little control over the browser and hardware the respondent uses to access and complete surveys, and thereby have only partial control over the look and feel of the survey itself.

In the chapter on interactive features, Tourangeau et al. dispel the popular notion that interactive features are always positive for data quality and respondent experience. Their review of the literature points to mixed results: interactive definitions, auto sum tallies, and responsive prompts can improve data quality, while visual analog scales can sometimes harm data quality. The discussion of progress indicators is nuanced and focused on the signal an indicator provides to respondents rather than a blanket recommendation. Given how software vendors and market research firms often trumpet novel interactive features as a way to differentiate themselves from their competition, this chapter is particularly helpful for consumers of survey tools.

The authors use the last two chapters to situate web surveys within the larger field of survey research and other modes of data collection. The authors’ meta-analysis of mode effects shows that, despite a lower response rate, web surveys eliminate interviewer effects and can reduce cognitive burden. Importantly, the evidence shows that web surveys are as good as other self-administered forms of data collection for eliciting sensitive information and better than interviewer-administered forms.

The discussion of mode effects is tied to mixed-mode surveys and leads to one of the most important contributions of the book—a framework for practitioners to use when deciding whether to use a unimode design approach to minimize differences across modes or to use a “best practices” design approach in each mode. If the goal is to make overall estimates, the “best practices” approach is typically better; however, the unimode approach may be more suitable for comparisons of ratings. This conclusion is supported through easily understood examples and explanations of relevant statistical formulas. Such guidance is immensely important even if one only conducts web surveys, given the rise of unintentional mobile web surveys due to the increasing move to consuming web content on mobile devices. The authors conclude the book with a clear summary and offer a helpful list of recommendations, solidifying this book as a useful reference source.

Tourangeau, Conrad, and Couper’s The Science of Web Surveys provides a well-organized and careful overview of the science of web surveys. They ground the relevant empirical work in both the psychology of the respondent and statistical formulas that help readers understand the underlying principles of web surveys. The content is presented in such an easy-to-read manner that it is suitable as an introduction for those who are new to the field; while those more familiar with web surveys will appreciate the volume of topics and citations brought together in a single source. Every survey methodologist should have two copies on his/her bookshelf—one to keep for reference and one to give to those interested in web surveys.

Curtiss L. Cobb III

Facebook

Statistical Analysis of Network Data With R.

Eric D. Kolaczyk and Gábor Csárdi. New York: Springer, 2014, xiii + 207 pp., $59.99 (P), ISBN: 978-1-4939-0982-7.

I have been working on network data (Internet and power grid) for long enough that my colleagues expect that I am some sort of expert. Lest they discover truth, I decided to review Statistical Analysis of Network Data with R (SANDR) by Eric Kolaczyk and Gábor Csárdi with the hope of actually learning something. I am pleased to say that I am now much closer to meeting my colleagues’ expectations.

In Chapter 1, SANDR explains the basic idea of a network and the sorts of analyses one might apply to one. The authors explain that the key idea of a network is “the notion of elements in a system and their interconnectedness.” A graph (vertices and edges, where edges are pairs of vertices) is a mathematical object frequently used to represent a network. The chapter lays out three broad categories of analysis one might want to use with networks. First, there is visualization and summarization. Second, there are analyses of network structure and what mechanism might have generated the network. Third and finally, there are analyses of the processes that describe how neighboring nodes are similar or how traffic, in some sense, moves between them. The book covers all three in roughly this order. This chapter also briefly discusses the use of R. Better yet, the book provides details on performing all of the covered analyses in R and code is available through the R package sand. This transforms the book from being just an introduction to the concepts and models into being a hands-on tutorial that allows the reader to quickly begin applying the ideas to new data.

Chapter 2 describes the manipulation of network data using the igraph package, which is the workhorse for much of the book’s material. This covers the creation of graph objects, basic graph operations, and the use of edge and vertex attributes. Chapter 3 presents visualization, starting with basic plots and eventually covering the use of color, size, and labeling. Chapter 4 presents a wide variety of descriptive and summary analyses for networks; it includes both the concepts and the necessary R commands. The order of this early material is one of the things I like about the book. First, it reinforces the basic applied statistics approach of plotting and summarizing data before moving on to more sophisticated analyses. At the same time, it provides a gentle introduction to lots of basic graph concepts and terminology.

Chapters 5 and 6 begin the meatier material by covering mathematical (Chapter 5) and statistical (Chapter 6) models for networks and graphs. Chapter 5 covers a number of random graph models, starting with the basic Erdös-Rényi graph model and moving on to extensions and other random graph generating mechanisms. The chapter ends with a discussion of using these models for comparison when assessing the features of real-world graphs. Chapter 6 introduces statistical models for graphs and covers exponential random graph models, network blocks models, and latent network models. Each of these is helpfully analogized with well-known nongraph statistical models. All three cases begin with a general discussion of the model followed by model specification and fitting using whichever R package is appropriate. Each case ends with a discussion of goodness-of-fit, again emphasizing good applied statistical practice. Chapter 7 also considers network structure, but is more narrowly focused on predicting or making inference about links between nodes. It also considers the special case of estimating tree-shaped networks from data flowing across the network.

Chapters 8 and 9 take up the problems of modeling node processes and flows. Chapter 8 considers how node attributes are distributed or evolve based on the structure of the graph. It presents models for both static and dynamic processes on the nodes. Chapter 9 considers models for traffic or flows and focuses on the idea of the traffic matrix, which describes how flows move between different origin and destination nodes. The chapter considers both cases where this matrix is observed and used in modeling and cases where this matrix is to be predicted. Chapter 10 contains a discussion of networks that evolve over time and considers these issues in the context of an example involving contacts between people in a nursing home.

My feeling is that Chapters 1 through 6, the first part of Chapter 7, and Chapter 8 are very strong and would serve well as a text on the material for advanced undergrads, beginning grad students, or other researchers in fields where such data arise. Most of the important basics of the field are covered. If students mastered this material, they would be well positioned to begin working on data and making further progress on their own. Chapter 10 also raises some interesting ideas. The discussion of tree topologies in Chapter 7 and the flow modeling in Chapter 9 are not necessarily bad, but they do not feel as fundamental as the other material. These sections actually include the material that I know best from my own work, specifically aspects dealing with network tomography. These problems are presented clearly, but a detailed discussion of solutions is beyond the scope of the book. As a result, these sections are not as complete as others.

SANDR is reminiscent of the excellent text Linear Models with R by Julian Faraway. Like that book, SANDR covers a lot of basic and important material while teaching the reader how to work with data and models in R. The reader is assumed to have some basic statistical and R programming knowledge, but need not have much background on the book’s topic. If you like the style and level of detail in the Faraway book, you will also like SANDR.

SANDR is a good introduction to the material. Although topics are not deeply covered, the range is quite broad. The R code allows the reader to quickly get their hands dirty with real computing. The book appears to be the only one available that covers the material at an introductory and practical level. A more thorough treatment of the material can be found in Kolaczyk’s Statistical Analysis of Network Data: Methods and Models. SANDR will prepare the reader for moving on to the more in-depth book or can be used to obtain the background necessary to read articles in the field. On the whole, I am happy to recommend it.

Earl C. Lawrence

Los Alamos National Laboratory

In the Public Domain

Statistical Methods for QTL Mapping.

Zehua Chen. Boca Raton, FL: CRC Press, 2013, xix + 288 pp., $94.95 (H), ISBN: 978-1-43-986830-0.

Statistical Methods for QTL Mapping introduces concepts and techniques used in mapping quantitative trait loci (QTL) with a major focus on experimental crosses. The book is unlike a standard textbook in that it does not provide practical exercises. Nonetheless, it is a good reference for advanced graduate students and researchers working on statistical genetics research. It provides the basic techniques and advanced tools necessary for conducting basic genetic data analysis and advanced research in the field. It also includes the theoretical derivations that underlie the statistical analysis of genetic data. Senior graduate students and researchers will find this book extremely useful in developing a systematic understanding of the statistical foundations of QTL mapping, and in identifying the rich research topics in this field. The book can be used as an advanced reference tool or to supplement a textbook such as Wu, Ma, and Casella (2007) or Siegmund and Yahir (2007) in a statistical genetics course.

The book is organized as eight chapters. From the mapping perspective, the topics can be categorized into four parts. The first, containing chapters 1–3, covers the basic terms and concepts in genetics, statistics, and QTL mapping. Specifically, chapter 1 covers basic topics of genetics and chapter 2 covers statistics foundations necessary to understand QTL mapping. Finally, chapter 3 introduces aspects of quantitative genetics related to QTL mapping. These chapters are essential to understand the more advanced analysis of genetic data that is covered in the rest of the book.

The second part (chapters 4–6) presents detailed methods for QTL mapping from the maximum likelihood perspective. The topics range from one-dimensional QTL mapping to multiple interval mapping and QTL mapping with dense markers. The scope is quite broad, covering theoretical development as well as practical implementation of the methods. The contextual and technical details are nicely organized in a way that is both engaging and easy to read. Some of the topics such as selective genotyping, threshold determination, QTL mapping of traits with a spike distribution, and mixture generalized linear models for nonnormal traits are rarely covered in other books but clearly illustrated here. In addition to showing the mapping principles under different scenarios, the author presents detailed theoretical derivations along with relevant algorithms. These will help readers gain insight into the methods.

The third part (chapter 7) focuses mainly on the development of Bayesian methods. Unlike maximum likelihood, Bayesian methods are known for their advantage in multiple QTL mapping since they can deal with a large parameter space and can increase the mapping precision when multiple QTLs are present. The genetic basis of many complex traits often involves multiple QTLs and Bayesian methods have shown their practical merits in many studies (Yandell et al. 2007). Researchers interested in Bayesian approaches should find this chapter useful.

The last part (chapter 8) summarizes the methods for multi-trait QTL mapping and eQTL mapping. Some of the methods discussed are at an advanced level, such as multivariate sparse partial least-square regression and multivariate sequential methods. Multi-trait mapping is a useful tool in practice, especially when distinguishing pleiotropic versus close linkage effect. eQTL mapping is a relatively new topic in the literature (Rockman and Kruglyak 2006), which aims to aid understanding of the genetic basis of gene expression and further infer different modes of gene regulation (Li, Williams, and Cui 2011). While there are review articles on eQTL mapping (e.g., Kendziorski and Wang 2006), there is a lack of textbooks on the topic and I hope the author will expand on eQTL mapping in a future edition.

Although this is an exciting book written by a well-known expert, I have a few suggestions for future editions. First, the book does not have exercises that would help students digest and understand the material. This is a critical omission that could prevent the book from being a useful textbook. Second, longitudinal traits measured over time (e.g., body weight, or drug response) are routinely collected in many studies. These traits can help scientists understand the dynamics of gene function, but the topic is not covered in the book. Finally, I hope the author can expand upon the topics of gene–gene and gene–environment interactions in future editions.

In summary, I found the book’s discussion of various techniques and analytical tools in QTL mapping particularly appealing. The book is unique in terms of the breadth and depth of its statistical treatment in QTL mapping. Besides covering basic strategies in QTL mapping, advanced statistical methods such as penalized likelihood methods are introduced with rigorous statistical derivations followed by practical implementations. The author also provides R code. As more and more genetic data are routinely collected, there is a growing demand for techniques that can handle such data. Researchers working in statistical genetics with a particular focus on QTL mapping will find this book valuable. It can also help researchers working in other fields to identify potential research topics and begin working in this exciting field.

Yuehua Cui

Michigan State University