Mastering 'Metrics: The Path from Cause to Effect

Abstract

Angrist, Joshua D. & Pischke, Jorn-Steffen (2015). Mastering 'Metrics: The path from cause to effect. Princeton, Oxford: Princeton University Press, 304 p., 35 USD, ISBN 978-0-691-15284-4Around five years ago, Joshua D. Angrist and Jorn-Steffen Pischke published their first joint book on econometrics tools for causal inference: Mostly harmless econometrics (2009). Although this book is excellent in many regards (e.g., more than 5000 quotes on Google Scholar), it was not as harmless as the title might suggest. Mastering 'Metrics: The path from cause to effect now fills this gap, as it is a truly nontechnical introduction.Angrist is Ford professor of economics at the Massachusetts Institute of Technology and Pischke is professor of economics at the London School of Economics and Political Science. Both teach applied econometrics and they have published a variety of their own applications of the presented methods. The book is useful in many areas of educational research, because it illustrates the logic behind causal inference when randomized trials are not feasible - this is a standard issue for many educational research questions due to financial, ethical, legal, or other reasons. International large-scale assessments are an example of this: They provide rich information concerning diverse research questions, but are observational cross-sectional designs by nature. The book discusses the underlying logic and assumptions of causal inference and the related methods in a non-technical way, rather than focusing on the actual estimation of statistical models and mathematical properties ("It won't surprise you to learn that there's a formula for IV standard errors and that your econometric software knows it. Problem solved!", p. 110).In the chapters, the authors' five favorite elements in the econometric toolkit are presented methodologically and illustrated in detail using actual applications. Only the most important statistical formulas are presented and thoroughly explained; appendixes to each chapter provide some more technical details. Embedded comics and amusing dialog between fictitious characters make reading the book a fluent and joyful experience; the generally informal language they use is also a benefit in this regard (e.g., "randomized social experiments are expensive to field and may be slow to bear fruit, while research funds are scarce and life is short", p. xiv). The low-threshold and explanatory nature of the book is further underlined by the fact that a supplementary website (httpi//masteringmetrics.com/) provides the datasets used in the examples, as well as further information for instructors. Many examples are educational in nature and the basic ideas of the different methods are well illustrated, meaning that transferring them to the reader's own research questions is straightforward. Each of the first five chapters captures a different approach to causal inference, while the sixth chapter makes a connection specifically to the educational sector.The first chapter Randomized Trials outlines basic experimental concepts like treatment, outcome, control and treatment group, the fundamental problem that we can always only observe one reality in one person, and the idea that randomization makes "other things equal" (p. xii). It also points out why perfect randomization is difficult to achieve in real life. Furthermore, the issue of statistical significance in the interpretation of results is discussed, as analyses are usually only based on samples drawn from populations.As already introduced in the first chapter, treatment and control groups are not necessarily equal in all other aspects, especially under non-randomized conditions. Therefore, the idea of "Regression" is discussed in the next chapter. Regression is presented as a conditioning technique that only delivers credible results if all variables that introduce group differences apart from the treatment are observed. Such variables are then computationally made equal across the groups, so that causal inference can be made. …