Causal Inferences in Nonexperimental Research

When sociologists were just barely getting into a causal analysis in a serious way, Hubert Blalock published Causal Analysis for Non-Experimental Research. Recall that in 1964 even OLS regression analysis was beyond the ken of most sociologists, and cross-tabs were still the staple statistical tool used when eexamining relationships. Books such as Rosenburg's The Logic of Survey Analysis and Davis' vastly over-rated Elementary Survey Analysis taught social scientists to do causal analysis with one or two controls, but only with cumbersome, case-intensive, hard to interpret, tabular elaborations. While instructive, this primitive approach promised little in the way of causal modeling that might be properly specified. Davis took the position that an elaboration with as many as four variables was quite uncommon and probably poorly informed. In fact, specification, spherical disturbances, multicollinearity, autocorrelation,and other commonplace terms that are part and parcel of the process of understanding regression-based causal modeling were entirey new to most social scientists. With publication of Blalock's 1964 book, things began to change. Yes, change was slow. It was commonplace to report statistically significant paths without the coefficients. Perhaps this was done because the paths' meaning was not understood, especially when they involved indirect effects. So just let signficicance/non-significance speak crudely for itself. Nevertheless, models became more informative, and more and better modeling was soon in the literature, almost all of it relying heavily on Blalock's work. We are, indeed, fortunate that Blalock's book on causal modeling was published well before Duncan's. Blalock's book was a practical instructional tool that actaully enabled analysts to teach themselves and teach others. Duncan's later (1975) book was more advanced, but served best to show how much he knew and how little his students did not. Non-experimental research has become extremely sophisticated, though there are still those who argue that causal analysis cannot be done with utmost accuracy without random assignment. That, I think, is the gross over-statement of a misguided purist, and even today Blalock is worth reading.