Utilizing Causal Models to Discover Flaws in Experiments

The burden of this paper is to show that causal models and path analysis may be useful in the analysis of the results of experiments, even though that is not the usual context in which they have been employed. This paper demonstrates that it is feasible to check the tenability of a set of assumptions about the causal structure by which experimental manipulations have their observed effects, provided there are at least two measures (in addition to the manipulation) of the independent variable and at least two measures of the dependent variable. This check proceeds by assessing the goodness of between the model and observed correlations, poor fit indicating that the theoretical causal structure assumed in setting up the experiment is untenable. Thus, while causal models do not specify modes of operationally manipulating independent variables in experiments, they may be useful in discovering when suck manipulations have had artifactual effects.