Toward a Causal Model Approach for Adjusting for Preexisting Differences in the Nonequivalent Control Group Situation

Abstract

The traditional approach for partialing out the effects of confounding factors in a non equivalent control group situation is to calculate a partial correlation coefficient or a (partial) regression coefficient controlling for one or more "covariates." Since the covariates generally are imperfect measures of the factors, this procedure will typically yield biased estimates of effect. An alternative approach, which allows for the presence of measurement error, is discussed and applied to some data from the original Head Start evaluation. Contrary to the original analysis, which found the Head Start program to be totally ineffective, the alternative approach yields small positive estimates of effect. The resulting model is conceptually feasible and fits the data well. However, it is only one of many possible causal models which could have been formulated. Guidelines are needed for building causal models in the evaluation research framework when little is known a priori regarding the causal relationships.