STATISTICAL MODELS OF EDUCATIONAL PLACEMENT DECISIONS: A COMPARISON OF DISCRIMINANT ANALYSIS, LOGISTIC REGRESSION, AND ORDINARY LEAST SQUARES REGRESSION MODEL

This dissertation contrasted three statistical models of educational placement under conditions of polytomous choice alternatives. The models were OLS regression, discriminant analysis, and logistic regression. They were assessed using two criteria: (1) accuracy of prediction and (2) ability to capture placement policies. Two separate studies were performed.;In Study 1, the models were compared using computer generated placement decisions. Thus, the experimenter knew the underlying placement policy. Process tracing models were constructed for each of three special educators. Each was asked to "think aloud" while making classroom placement recommendations (i.e., regular, learning disabled, or emotionally disturbed) for hypothetical student profiles. Each model was translated into a computer program. Sets of decision data of different sample sizes were then generated by the computer programs. These data were fit to each of the three statistical models. The models were compared using the two criteria. Discriminant analysis satisfied both criteria. Logistic regression yielded the highest proportion of correct classifications but failed to capture the placement policy. OLS regression was the least accurate model. The results suggested that when decisions are made among more than two alternatives, more than one decision strategy may be used by the decision maker. OLS is inadequate in this situation.;In Study 2, five graduate students were asked to make placement recommendations for the hypothetical profiles. Each statistical model was fit to the data obtained from each student. The criterion used to compare models was accuracy of prediction since the true underlying policy was unknown. Discriminant analysis and logistic regression yielded approximately the same proportion of correct classifications. Logistic regression did not yield many significant parameters; thus, its results were difficult to interpret. OLS regression was the least accurate. These results were consistent with the findings of Study 1.