Causal Models, Unobserved Variables, and Experimental Data

While techniques of causal modeling have been employed in the analysis of nonexperimental data, they have been largely ignored by social scientists who deal with experimental data. This paper discusses ways in which causal models may be fruitfully used in the analysis of experimental data, and illustrates their uses. Central to the paper's exposition is the representation of social science experiments in terms of causal processes involving unobserved variables. Within this framework, two key experimental issues are addressed: the evaluation of manipulated independent variables and the evaluation of indicators of dependent variables. Costner's (1971) suggestions regarding techniques for examining rival interpretations of experimental results are also discussed. This type of analysis of experimental data permits us to take explicit account of error in the induction and measurement of independent and dependent variables, thereby yielding better estimates of true causal parameters. Data from an experimental investigation of clients' reactions to initial interviews are used to illustrate this approach.