Customer Reviews

Data Analysis Using Regression and Multilevel/Hierarchical Models

5 star:		(18)
4 star:		(1)
3 star:		(1)
2 star:		(1)
1 star:		(0)

 

 

Gelman and Hill have put together a fabulously well-integrated look at general modeling with a focus on hierarchical structures. The book starts with simple modeling principles and continues well into material that would satisfy a third semester course in many social science grad programs. This book does something that is extremely hard: presenting serious technical ideas without overwhelming language and detail, making the chapters unusally easy to read and digest. They also do a very nice job of balancing Bayesian and traditional approaches without denigrating or over-promoting either. This should considerably broaden the appeal. Furthermore, the emphasis on R and WinBugs means that readers can immediately (and for free) run through the techniques.

I see this book as primarily a teaching tool, although many will use it as a reference. In this light, it is without peer right now in terms of coverage (basically all of the standard/basic regression models that get taught to social science grad students), price/page ratio (0.15366), and accessibility. Many of us have used econometric texts for such purposes over the years, living with a slightly mismatched set of criteria to rely on the quality of these works (Greene, Mittlehammer et al., etc.), but now there is a competitor that fits much more nicely with non-economic methods training (less of a fixation with asymptotics, no need for 200 named flavors of each model, and so on). Finally, the practical advice and admonitations that accompany the model descriptions will be immensely helpful to practitioners.

I came to this text with a very pragmatic need: I needed power calculations of a multi-level model, and I needed them fast. I skipped directly to Chapter 20, which is the most accessible treatment of multi-level power-calculations I have ever read. A few hours later, I had the calculations I needed done. (Take home point: this book has a wonderfully practical side.)
To my surprise, I also really understood what I had done, why I had done it, and other approaches that I might have taken. That is, the text very effectively provides the broader theoretical overview, gives a concise real-statistics treatment, and pragmatically teaches you how to actually do the analyses you need to do. Gelman & Hill have that rare ability to both teach the abstract and directly help you do the practical. (Fans of Paul Allison's books will love this one, too.) This is a must-have for the shelf, and I am sure I will come back to it repeatedly.

Andrew Gelman is a top researcher in Bayesian statistics as well as an excellent writer. He has written an excellent text on Bayesian data analysis that uses the Markov Chain Monte Carlo methods for dealing with hierarchical linear models. This book starts out on an introductory level covering a wide variety of statistical modeling problems including logistic regression and multilevel logistic regression, generalized linear models and causal inference. The MCMC methods are taught using BUGS and R. This book is not exclusively Bayesian as both likelihood and Bayesian procedures are presented. The topics are general but the emphasis is on social science applications. It is very comprehensive and has received enthusiastic reviews from well known statisticians including Dick Deveaux, Brad Carlin and Jeff Gill. Jeff's review is here on amazon. Jeff is a colleague of mine and he has written one of the finest introductory texts on Bayesian methods including the hierarchical models. His text is now out in its second edition. Jeff also wrote his book with the social scientists in mind.

Jeff's review has been the most looked at and voted the most helpful on this site. As this topic is a specialty area for him more than it is for me, I recommend that if you are interested in the material in this book that his review is very much worth reading.

I have to qualify this review by saying that I proceeded from the 11th chapter since the first ten were more or less review. Also, I am not a statistician by any stretch of the imagination. My math background is pure math and economics degrees with some too-practical econometrics. In spite of that, I understood this book quite well. Hence my positive review. Compared to other comprehensive treatments of HLM, such as Singer and Willett or Hox, this book is in a universe all its own. I actually took Hox's course from him and still barely understood HLM, yet got the highest marks in the class. That's not a good thing. I felt like I wasted my time.

I actually learned a great deal from this book, and more than practical method (which I have since used), I actually understood what it was I was doing. The few R examples I did were worth it, and I would try them out if you can. In the past I have made two abortive runs at learning MLM/HLM, but this time it stuck. This book is extraordinarily well-written, as if it has been taught to non-statisticians a number of times. This is perhaps due to the presence of Hill as coauthor. Her public affairs students are not likely to value the math for its own sake. I alotted myself a month to master the latter chapters, some of which were completely new to me and it took me less than a week.

Drawbacks:
Typos: None of these were in substantive portions of the text such as equations and data print-outs. Still, a few in the wording were present. Mine is a first printing, however, so these might not be in your copies.
Program use: I think that they should also have offered SAS, SPSS, or Stata excercises. I only incidentally learned R, but would prefer to use a more standard software package for the excercises.

I found this book after reading up on the weaknesses of traditional psychological statistics and methods. I read through 3 or 4 classic texts on classical regression (Casella and Berger, etc) and a couple more texts on Bayesian analysis (including Gelman's own Bayesian Data Analysis). When this book came out, I was quite excited and pre-ordered it.

The book far exceeded my expectations. It is written in a crisp clear style and presented very concisely not only how to write but how to interpret multilevel models and especially Bayesian models. I feel I would have been just almost as well off having just read this book and not taken the time reading the other text books.

The book is especially refreshing to me because it explains the basics of data analysis in ways that many psychologists -- who typically don't have a background in probability or linear algebra -- can understand. For this reason alone, this is the book I recommend to any colleague who wants to learn more about statistics. The fact that it might inspire them to create more realistic models or become a Bayesian is just an added benefit. In fact, I have purchased several copies and keep most of them in continuous circulation among coworkers who ask to borrow them.

That said, the following might be noted:

(1) There are typos and errors in some of the graphs in the first edition. This is unfortunate, but not a huge deal. The book is cheap for a textbook already, and purchasing an ultra-cheap used first printing with all the typos won't really affect how much you learn. So if you're shopping around for texts and are worried about cost, that's the way I would go.

(2) I don't have a math or computer science background, but I do understand the concepts underlying most of the math that goes into the book. I don't know how understandable the book would be without a basic familiarity with matrix notation or classical regression. I imagine the book would be less interesting for people with more of a math background, in which case Bayesian Data Analysis may be more your speed. With that in mind, many social scientists have taken classes in regression and should be fine to read the book.

(3) Another reviewer mentioned that contacting the authors was a waste of time. I have had the opposite experience contacting the authors. They have been extremely helpful when I've contacted them. In addition I find that the web site contains a lot of useful information. Professor Gelman posts questions he receives about the book and their answers to his blog, which can be found from his main website. I suggest searching there for answers to common questions.

I am reading this book for two reasons: improving my understanding of some statistical issues and becoming more proficient with modern statistical techniques. The book has been helpful on both fronts, often providing new (to me) points of view for looking at a problem and giving very accessible entry points to more advanced techniques. I have enjoyed very much reading the book and am looking forward to the opportunity to test some of the techniques.

In my opinion, the authors have chosen a good set of examples and have managed to keep me hooked on them. Initially I was a bit reluctant to buy a 'statistics for social sciences' type of book (I come from natural resources/genetics), but the material can be easily transferred to other settings.

The book requires some previous knowledge of statistics (it is no 'linear models for dummies'), targeting readers that already have some experience working with linear regression. Some previous experience with hierarchical models would not hurt either.

Concerning the use of R as the statistical language for the book I think it is a great choice. R is becoming the lingua franca of statistical computing, it is free and the authors do a good job at introducing the language. Even if you are using other language (SAS or SPSS for example) the book will still provide good theoretical explanations and useful comments on data analysis.

I have seen some comments on typos, I have seen some, but are not that egregious as to distract from reading the book.

Andrew Gelman has written an excellent book about regression models, with examples solved in the R language. He provides enlightning views of even complex subjects, such as mixed-effects models. A reader not familiar with R, should probably acquire some knowledge of R before he/she can fully benefit from the book, but this in itself is a worthwhile investment. (R is freely available; see [...]). Although it is an introductory book, the author manages to convey valuable new insights to more advanced readers. This is a book that after you read it once you will pick up time and again to enjoy the presentation of the topics and to benefit your own work. Highly recommended, in particular to those getting started with R (or Splus for that matter).

I read this book looking for an accessible and comprehensive treatment of multilevel models. The topic of social science appealed because this area offers different examples yet has used multilevel techniques widely. See Bryk and Raudenbush for example. The issues I have with this book is that it is over long. It could easily have been made shorter. The writing is often terse and there are clearer ways to put the point across.
The real weakness with the book is the website and support materials. The website is very poorly organized and although the authors provide some R examples and code their instructions for getting up and running with these packages to work through the examples is far from clear. Contacting the authors is similarly a waste of time.
If you want to understand how to run regression models with R the best book is John Fox's The R and S-Plus Companion to Applied Regression. Whose website and customer facing skills are also markedly superior. For getting into multilevel models fox's appendix on his website is well woth reading. To understand Longitudinal models Bryk and Raudenbush or Singer book is better written.
I like the idea of using R and Winbugs but the authors just haven't packaged the practicalities well.

Pros:
They tackle a complex topic from many different angles. They present enough code and theory to get people up and running with the techniques, assuming some prior familiarity with likelihood based inference and R. Otherwise you might need to dig through some of the references to understand everything. Regardless, this book is a valuable reference to keep in your library.

They use matrix notation sparingly and this helps the reader focus on the important concepts of multilevel modeling. I am not even remotely a statistician so my attention would have been lost if I had to sort through a bunch of matrix transpositions and inversions in addition to all of the multilevel notation.

The authors provide many useful references that help reinforce difficult ideas/concepts and that elaborate on topics that are not explored in depth.

I had no prior experience using WinBUGS and the authors provided enough information for me to successfully execute some models that integrate R and WinBUGS. That is no small feat and the authors should be commended because somehow I understood what was going on.

Cons:
The organization of the book seems scattered and could be a little more consistent. On pp 245-246, the authors go on a diatribe about "fixed" and "random" effects terminology, claim that much of the literature that applies these terms does so inconsistently, disown these terms by saying they will avoid using them entirely, and then continue using these terms throughout the book.

The website needs some work. You need to already know how to use R to open different types of files (and maybe some basics of variable assignment)in order to reproduce all of their examples. This book will not hold your hand through the steps like many R books.

This book is full of examples and very well written, contains everything one needs for deep insight into multi level analysis