New Books in Reveew

Users of structural equation modeling (SEM), a method that has also been known as causal modeling, have learned to tread lightly around the subject of causality. After all, the empirical raw material of SEM is typically a covariance matrix derived from nonexperimental data, and research dogma indicates that this is insufficient for making statements about cause-and-effect relations. At least as far back as Robert Ling's (1982) scathing review of David A. Kenny's (1979) book. Correlation and Causality (a classic text that is still well worth reading), users of SEM methods have found themselves on the defensive, careful not to claim too much. This, however, has produced something of a paradox. The models estimated with SEM clearly depict variable A as having an effect on variable B and distinguish between covariance relations and directional paths—that is, causal effects. Thus, SEM users propose structures that are causal but tend to disavow the causal element when they evaluate their results. Especially for the practitioner, the causal component is likely to be the point of the whole exercise: What a manager wants to know is, "If I do X, how will that changeY?" Furthermore, reasonable people use causal language and reach causal conclusions all the time. The government releases economic statistics, the stock market subsequently moves, and the observer concludes that the new information moved the market. Rain falls, water drips through a hole in the roof, and an observer makes the connection. People do this without the aid of either experiments or sophisticated data analysis. True, sometimes the observers are wrong—for example, at one point medical researchers thought exposure to aluminum was a cause ofAlzheimer's syndrome, but now they consider the linkage spurious. Nevertheless, people proceed through life making causal inferences from nonexperimental data. Pearl's Causality and Spirtes, Glymour, and Scheines's (SGS's) Causation, Prediction, and Search (2d ed.) urge researchers to resolve the paradox by dropping the pretense and acknowledging the causal content of their models. A —^ B means more than that A is correlated with B. It means that B changes in response to changes in A, and the lack of an arrow in the opposite direction means that A does not change in response to changes in B. Returning to the path analysis roots of SEM, these authors argue that justification for causal interpretation of model parameters follows from satisfying basic criteria. These criteria have as much to do with proper sampling design as they do with modeling. However, these authors do not argue that their approach to inferred causation reveals "truth." Instead, as Pearl (pp. 47-48) writes, "It identifies the mechanisms we can plausibly infer from nonexperimental data; moreover, it guarantees that any alternative mechanism will be less trustworthy than the one inferred because the alternative would require more contrived, hindsighted adjustment of parameters (i.e., functions) to fit the data." These books summarize intertwined research programs. Within the data mining and artificial intelligence community, Pearl and SGS are associated with machine learning and "Bayes networks," which look like structural equation models but are encountered in the context of discovery. The overall focus of the research has been on understanding how causal inferences are made and how they can be made reliably. One aim is to determine just how much analysts may discover from a data set using nothing more than a set of algorithms and common sense. In the data mining and knowledge discovery in databases literature, the issue arises because the organization has an excess of data, an adequate supply of computing power, and a shortage of analyst time. Many SEM users have faced a similar dilemma. The researcher gains access to an attractive secondary data set, develops a theoretical model, finds variables in the data set that can serve tolerably as indicators of the latent variables, runs the analysis—and comes nowhere close to an acceptable fit. The analyst resorts to ex post modifications and