The Effects of Initially Misclassified Data on the Effectiveness of Discriminant Function Analysis and Finite Mixture Modeling

Classification procedures are common and useful in behavioral, educational, social, and managerial research. Supervised classification techniques such as discriminant function analysis assume training data are perfectly classified when estimating parameters or classifying. In contrast, unsupervised classification techniques such as finite mixture models (FMM) do not require, or even use if available, knowledge of group status to estimate parameters or classifying. This study investigates the impact of two types of misclassification errors on the classification accuracy of discriminant function analysis (both linear [LDA] and quadratic [QDA]) and FMM for two groups with a single predictor. Analytic and Monte Carlo results are provided for a variety of misclassification scenarios to investigate the performance of the two methods. Discriminant function techniques recovered the highest overall percentages of correctly classified data, whereas FMM captured higher percentages of the smaller group when group sizes are unequal. LDA marginally outperformed QDA under misclassified conditions.