Counterfactuals and Causal Inference: Methods and Principles for Social Research (review)

This volume is a major contribution to our understanding of causality in observational social science. Morgan and Winship use a sophisticated counterfactual understanding of causality as a framework to integrate three major sets of methods for casual inference: statistical methods involving regression and matching analysis, instrumental variable techniques, and the specification of causal mechanisms. The analysis proceeds in a comprehensive and logical fashion, illustrating its arguments with intuitively appealing graphs and with a running set of empirical examples, such as the debates stimulated by the work of James Coleman and his associates on the impact of Catholic schooling on student achievement. Counterfactuals and Causal Inference is an important work that is likely to become required reading in courses on research design and causal inference in sociology and political science. It is written so clearly, with its major points explained in straightforward prose, that it will be of great value to students and faculty whose work is principally qualitative, as well as to the quantitatively oriented audience to which it is principally directed.

The authors build on a venerable tradition in social science, expressed by Samuel A. Stouffer in 1950 when he argued that, when testing alternative ideas, “it is essential that we always keep in mind the model of a controlled experiment, even if in practice we may have to deviate from an ideal model.”(7) The central problem is what Paul W. Holland called the Fundamental Problem of Causal Inference: one cannot simultaneously observe the treatment and control conditions for the same unit. In developing their quasi-experimental, counterfactual approach, Morgan and Winship specify three distinct and complementary strategies for causal inference: (1. conditioning on other potential variables that could affect the outcome, as in regression and matching analysis; (2. using appropriate exogenous variables as instrumental variables; and (3. establishing an “isolated and exhaustive” mechanism that links the outcome variable to the causal variable of interest.

Part 2 of the volume discusses conditioning strategies. Morgan and Winship emphasize that regression analysis, although important, can have the lamentable effect of short-circuiting careful thinking about theory or about the non-random ways by which observable data were generated. The “Sherlock Holmes inference,” in which the investigator constructs theories after having studied the data, requires specification assumptions that make it vulnerable to a wide variety of sources of bias. Matching techniques, while often valuable, carry with them their own limitations. The authors conclude that matching and regression should, whenever possible, be used together; but that even jointly, they have limitations that require going beyond conditioning strategies.

In Part 3 the authors address issues for which conditioning strategies are not sufficient. Chapters 6 and 7 discuss identification strategies, employing instrumental variables. Morgan and Winship argue that when a valid instrumental variable can be found for a narrowly-defined causal inference problem, this strategy can be a very powerful complement to regression techniques. However, it is more difficult than it may seem to find valid instrumental variables. When such variables affect the outcome variable through pathways other than through the causal variable of interest, or when they only weakly affect the causal variable, they may generate highly misleading results.

Chapter 8 discusses causal mechanisms. Here Morgan and Winship follow closely the analysis of Judea Pearl. They emphasize that the causal mechanism must be “isolated and exhaustive.” Identifying and specifying isolated and exhaustive causal mechanisms of course requires explicit theory, which Morgan and Winship regard as indispensable, in the end, to valid and reliable causal inference. A causal effect identified by conditioning and instrumental variable strategies is, they say, “best explained when it is also identified by an isolated and exhaustive mechanism. This is the gold standard for an explanatory causal analysis.”(242)

Counterfactuals and Causal Inference is a valuable guide to the pitfalls of efforts at causal inference, and various strategies to cope with these threats to inference. The authors repeatedly demonstrate that there is no...