Comments on Professor Freedman’s Paper

Abstract

Professor Freedman makes a number of useful comments on path analysis. Perhaps their reiteration by an eminent statistician will have more of an effect than similar statements by others have had. For example, Keynes (1939, 1940), Baumrind (1983), and Cliff (1983) seem to have had little effect on the volume of such naive applications of path analysis. I am somewhat concerned, though, that careless reading of his will serve to reinforce the naive interpretations of path analysis that he is trying to counteract. A superficial reading of its preliminaries leads to the impression that causality is established by drawing straight lines from one thing to another. This may lead the whimsical or literal-minded reader to establish some surprising causal relations with the aid of a crayon. More fundamental to possible misunderstanding is the formulation of the theorem itself. This seems to prove that the correlation between variables is the sum of products of path coefficients and correlations with proximate causes. This will seem to the path analyst to reflect exactly what he or she has been doing. I believe that Professor Freedman's purpose in the paper is to discourage the practice of treating correlations in this way, but in the context he seems to be accepting it. That is, correlational data are rarely sufficient to demonstrate causal relations, but the theorem seems to condone the use of correlations to establish causality. A point that he emphasizes that may not have received due attention in earlier statements of this kind is some of the fallacies in the interventionist interpretation of a path model. What many users of path analysis would like to conclude, or perhaps would like others to conclude, is that if there is a path between X and Y as revealed by a path analysis, then intervening to change X will result in the change in Y that is shown by the coefficient in the model. Such hopeful interpretations are unsophisticated for the following reasons: 1. It is hard to tell whether there will be a change in other variables correlated with X or not. (A promotion of the father of a particular child will change the father's place on the occupational scale. Will it also change the numerous other variables that are correlated with occupation?) 2. Close examination of almost all applications will reveal that the variables that furnish the numbers actually used in a path analysis are surro-