A statistical approach to detect cheating interviewers

Abstract

Survey data are potentially affected by cheating interviewers. Even a small number of fabricated interviews might seriously impair the results of further empirical analysis. Besides reinterviews some statistical approaches have been proposed for identifying fabrication of interviews. As a novel tool in this context, cluster and discriminant analysis are used. Several indicators are combined to classify "at risk" interviewers based solely on the collected data. An application to a dataset with known cases of cheating interviewers demonstrates that the methods are able to identify the cheating interviewers with a high probability. The multivariate classification is superior to the application of a single indicator such as Benford's law.