Factorial Modeling in Support of Causal Inference

Abstract

This paper describes and justifies a new method for analyzing correlations in support of causal inference. Named Factorial Modeling (FaM), its motivations are (1) Social scientists have an obligation to hypothesize the probable causes of the phenomena they seek to explain, and (2) In the interests of discipline and parsimony, causes should be operationalized as uncorrelated variables. Applying a simple algebra to the correlation matrix, FaM produces a structural equation for each variate in the research, thus analyzing all the variances and covariances. The advantage of the FaM method is that the natural language of domains of measurement which are known to be relevant can be respected in the hypothesizing of causes. Since the algorithm does not attempt to maximize or minimize anything, a loose fit to the data will be obtained; but it is suggested that such loose-fitting models may travel well to other situations to which generalization is attempted. The FaM method is illustrated on a small example, for which path analysis, LISREL-type analysis, canonical correlation, and commonality analysis results are also given to provide comparisons with other methods of modeling. Predictions of the impacts of policy manipulations under a model obtained from FaM are demonstrated.