Identification Problems in the Social Sciences,   ISBN:9780674442849

Customer Rating: 

Identification Problems in the Social Sciences is a landmark book in social science methodology. No sociologist who takes statistical methods seriously can afford to ignore it. While many other statistical books teach readers how to apply advanced techniques, this book was written with a different objective: it pushes the reader to think hard about what can be learned from observed data alone and in doing so sets the upper limits of statistical reasoning when no or very weak assumptions are invoked.

To empirical researchers like myself who have heavily relied on statistical models for inference, reading this book can be both educational and humbling. It is educational in that the book teaches new ways of setting nonparametric bounds for quantities of interest. Yet it is humbling because Manski shows that the more powerful and sharper statistical results we have become accustomed to seeing and producing essentially depend on stronger assumptions. Thus, Manski forcefully shows the classical tradeoff between robustness and statistical power, although the focus here is not on efficiency but on identifying power. When observed data are thin, it takes strong assumptions to yield sharp results. There is no free information in statistics. Either you collect it, or you assume it.

Many sociologists are already familiar with Manski's earlier methodological contributions, a few of which appeared in sociological journals. An accomplished econometrician who often proves theorems, Manski has stayed a close friend of sociology and sets out to orient this book to a sociological audience. The book is written in plain and lucid English, augmented with relatively simple formulas based on properties of conditional probabilities and inequalities. No high level mathematics is required to read the book. A few simple but real examples in the areas of sociology, demography, criminology, education, and economics greatly facilitate the reading.

By "identification problems" Manski means something broader than the difficulty of identifying simultaneous structural equations. Identification problems are those problems that would not disappear even if the researcher could eliminate sampling error by increasing the sample size to infinity. Thus, Manski isolates a class of problems that are difficult to deal with not because of inadequate data points but because observed data reveal weak information. Besides the classical case of simultaneity, Manski discusses the problems of extrapolation, selection, mixing, response-based sampling, prediction-from-intention, and reflection. Each of these topics is covered in a separate chapter. Although the chapters are closely related and should be read as a whole, some readers will be tempted to read only the chapters most relevant to their interest. If you do so, keep in mind that Chapter 2 on the selection problem and Chapter 7 on the reflection problem are worth reading for all sociology readers.

Manski has worked on the identification problems for some time, and the book draws results from his earlier publications. Sophisticated readers will want to read his other publications that have appeared in journals. However, it would be a mistake to assume that this book is merely a collection of Manski's articles published elsewhere. The book is tightly integrated with a common concern and a common approach. Throughout the book, Manski takes a distinct philosophical stand: he begins with a discussion of what can be learned from the worst-case scenario and then gradually introduces weak assumptions to improve identification. This overtly conservative approach is highly commendable and should be adopted by others in both the teaching and the application of statistical methods.

Manski's solutions to the identification problems are innovative. He breaks away from the conventional wisdom that only point estimates are of interest. Instead, he utilizes available information to furnish nonparametric bounds. Although the bounds may appear too wide to some readers at first sight, I have found them to be quite informative. One needs to keep in mind that Manski is able to bound estimates with no or very weak assumptions. Moreover, he demonstrates that he can substantially tighten bounds at the expense of additional assumptions.

One message that the book clearly conveys is the importance of nonparametric statistics. I concur that sociologists should welcome the new development in nonparametric statistics and make more use of nonparametric methods in their empirical work. However, I am afraid that sociology's road to nonparametric statistics will be bumpy for the time being, as at least two obstacles need to be removed. First, empirical researchers need to make a cultural shift toward presenting results in unconventional forms such as graphics, bounds, and conditional estimates. Second, methodologists (econometricians and statisticians included) need to make nonparametric methods more practical and more routinized. I have in mind the difficulties of running out of cases or making presentations unduly complicated as one adds more dimensions of control.

Fortunately, Manski's book gives us good directions to follow to solve various practical problems in the future. The book is not a technical manual, but a new philosophy of doing statistics in the social sciences. Regardless of your predispositions, Identification Problems in the Social Sciences is a great book to read. Even if you disagree with Manski's solutions, you will still admire the book's elegance and logical clarity. I highly recommend the book.

Customer Rating: 

(From the back cover) "This book provides a language and a set of tools for finding bounds on the predictions that social and behavioural scientists can logically make from nonexperimental and experimental data. (...) draws on examples from criminology, demography, epidemiology, social psychology, and sociology as well economics to illustrate this language and to demonstrate the broad usefulness of the tools".

This is a little brilliant book on the relation between data, the use of assumptions, and identification of parameters or distributions that social scientists may be interested in. In the words of the author, "this book examines the conditional predictions that can and cannot be made given specified assumptions and empirical evidence". The beauty of this book is indeed its emphasis on the link between maintained assumptions and features of a population that can be identified. Manski has spent a large part of his career emphasizing that results derived from strong arbitrary assumptions do not have much scientific value. In some cases, calculating BOUNDS for parameters of interest can be much better than having a point estimate obtained by "denying" lack of knowledge of important aspects of reality. One lesson you will derive from this book is "we need to develop a greater tolerance for ambiguity".

Here is an example of the many empirical problems dealt with in the book (which is mostly methodological, and hence technical enough to be not suited for a lay reader). Suppose you have a sample of homeless individuals and you want to study the reasons why they may still be homeless some months afterwards. However, months later you only have information on a subset of your initial sample. Without making any assumptions about what causes individuals to exit your sample, what can you learn? Can you learn more if you make assumptions on the causes that lead individuals out of your sample?

Another reviewer has described this book as an introduction to "nonparametric estimation". This is susprising and totally misleading, as there is close to NOTHING in this book about estimation, parametric or otherwise. This book is about identification. That is, the question addressed here is always something like the following: suppose that you have perfect knowledge of certain features of the data, and suppose that you are interested in certain other features of the data. What kind of assumptions do you need to make in order to be able to learn about these other features? How does your ability to learn change when you change the assumptions and/or the initial information? This book will NOT tell you how to do the estimation, or how to estimate variances and confidence intervals. Identification will only tell you if you CAN estimate something, but it does not tell you HOW you can actually estimate it. If you are looking for an introduction to nonparametric ESTIMATION, you should probably look at Pagan and Ullah, which is an excellent introduction.

The book is not too technical, but it does require some prior knowledge of math and especially probability.

Customer Rating: 

This book is a nonparametric introduction written for social sciences students. This implies that it's light on the mathmetical side. But the nonparametrics are easy to understand anyway. This is a highly recommended book for those who want to formalize their thoughts and want to test that model with real life data but find it difficult to apply classical parametric methods.