Casual Models Do Not Support Scientific Conclusions: A Comment in Support of Freedman

Overview A critical distinction in methodological work is between (a) building (and applying) statistical for the processes that generate the social science data and (b) tossing the data at available statistical methods. In my own work I strive for (a) and discourage others from settling for (b). Regrettably, expositions and applications of the popular causal modeling methods (under the various names path analysis, structural equation models, LISREL, etc.) contain much of (b) and little of (a). In fact, my favorite typographical error casual models (which I've suffered in print) is enjoyable in large part because of its accidental accuracy. And an argument can be made that the methodological proselytizing for and dominance of causal has retarded the much more useful methodological work of (a). A similar theme is present throughout Freedman's paper, as in the last paragraph of his conclusion which begins My opinion is that investigators need to think more about the underlying social processes... . Earlier in the paper Freedman requires that the as-if-by-experiment conclusions must depend on a theory of how the data came to be generated. The translation of substantive theory into methods for data collection and analysis is where I think the fertile interaction between statisticians and social scientists lies (rather than in arguing a thumbs up or thumbs down on path analysis).