Conditions for the Presence or Absence of a Bias in Treatment Effect : Some Statistical Models for Head Start Evaluation

This paper examines the possibilities of carrying out unbiased evaluations for compensatory education programs when group selection is not random and/or when pretests are unavailable. Four models are presented which demonstrate that bias can be avoided if particular selection procedures are used. The models are then considered for their usefulness in evaluating Head Start by using the data collected for the Westinghouse Learning Corporation-Ohio University study of that program. Suggestions are made on how the appropriate model could be selected and how the presence of bias could be determined. In the final section of the paper the importance of correct specification of the model and knowledge of the selection procedures are shown to be crucial for unbiased evaluation of nonrandom experiments and ex post facto analyses.