A New Mixture Model for Misclassification With Applications for Survey Data

Abstract

Social scientists often rely on survey data to examine group differences. A problem with survey data is the potential misclassification of group membership due to poorly trained interviewers, inconsistent responses, or errors in marking questions. In data containing unequal subsample sizes, the consequences of misclassification can be considerable, especially for groups with small sample sizes. In this study, the authors develop a new mixture model that allows researchers to address the problem using the data they have. By supplying additional information from the data, this two-stage model is estimated using a Bayesian method. The method is illustrated with the Early Childhood Longitudinal Study data. As anticipated, the more information supplied to adjust for group membership, the better the model performs. Even when small amounts of information are supplied, the model produces reasonably robust estimates and improves the fit compared to the no-adjustment model. Sensitivity analyses are conducted on choices of priors.