Abstract
It is well known that answering causal questions in the absence of randomized experiments is considerably challenging. So, empirical sociologists should always welcome ideas about how they could use conventional statistical techniques to make causal statements about the issues they study. Mostly Harmless Econometrics makes an important contribution to the understanding of quantitative techniques applied to nonexperimental settings by placing these techniques in the potential outcome (counterfactual) framework of causality. The authors are not throwing out totally new techniques here but rather offering a remarkably detailed yet practically simplified account of the methods, along with the assumptions necessary for a causal interpretation of the models. The label econometrics in the title should not give the wrong impression about the relevance of this book for sociologists. It is a great applied statistics text from which mainstream sociology scholarship would benefit a great deal. Needless to say, the issue of causality deserves more attention in the quantitative methods texts in sociology.
The opening chapters of the book (chapters 1 and 2) introduce general conceptual issues related to the research process and the potential outcome model of causality. Specifically, in chapter 1, the authors discuss key questions empiricists should ask in every research project, including what the causal question is, the research design required to set the stage for answering the question, the analytical strategy that will eventually be used, and the mode of inference. Empirical research questions of causal nature essentially boil down to these: Does variable A cause variable B? If so, to what extent is this so? In chapter 2, the authors emphasize that the process by which individuals (often with different characteristics) are exposed to the cause of interest is paramount from a causal analytic point of view. In terms of research design, randomized experimental designs provide the most convincing setup. Using the potential outcome framework, they illustrate how random assignment eliminates the baseline or selection bias in estimating the average causal effect. The next best designs are quasi-experimental designs, which cash in on naturally occurring phenomena or circumstances that assign people into treatment and control groups like intentional random assignment. Given that social scientists do not always work with data generated from such designs, the authors systematically delineate suitable analytical methods and the assumptions that allow for a causal interpretation of research results.
The remainder of the book (parts II and III) contains topical chapters of real interest to quantitatively oriented social scientists. Part II contains three long chapters on regression techniques for adjusting for variables considered to be correlated with both the causal and outcome variables, instrumental variable estimators of causal effects, and the differences-in-differences method (fixed effects) for controlling for unobserved but constant confounders in a panel data setting. The chapter on instrumental variables was particularly instructive for me because it clarified for me the connection between the traditional instrumental variable estimators and the local average treatment effects parameter derived from a potential outcome perspective. In part III of the book, the authors tackle some less widely applied methods in sociology, including regression discontinuity designs and quantile regression. The last chapter, on standard errors, is very insightful. The authors caution that reliance on the “robust” standard error (as opposed to the conventional standard error estimator) may be misleading if the assumptions justifying its use are violated, a situation that is not uncommon. They show robust standard error estimates to be biased under a range of conditions that they spell out.
Generally, the authors use a fair amount of equations, theorems, and proofs to illustrate their points, but they faithfully include intuitive descriptions of what the equations substantively mean or imply. This way, people who want to dig a bit deeper can consider the mathematical details, while those who are happy to avoid them can do so. However, for the most part, this book is written for those who can follow equations and have good grounding in probability and statistics. Mostly Harmless Econometrics is effectively for graduate-level scholarship or more specifically for those interested in quantitative methodology, as it would be occasionally frustrating for those with little background in advanced statistical methods.
One appealing feature of Mostly Harmless Econometrics is its clear presentation and focus on the ideas, using very straightforward language and relevant examples. The authors deliver their points with good humor, with amusing introductory quotations and pictures after the chapter titles. This book is a lot more readable than most econometrics textbooks, and its lightness does not compromise accuracy.
The authors wrote the book primarily for PhD economics students and others: empirical social scientists who need to understand the rationale for various techniques used in causal analysis. Sociology instructors who want to put a causal spin on conventional topics such as regression will find this book useful. This book is not likely to appeal to graduate sociology students who are not quantitatively oriented, but for those seeking an advanced background in applied statistics, Mostly Harmless Econometrics goes way beyond “harmless” to a value-added experience.