Collusion Detection Using Joint Response Time Models

Abstract

One method for detecting test collusion, or large-scale answer sharing, is the divergence framework (Belov, Journal of Educational Measurement 50: 141–163, 2013). It uses Kullback–Leibler divergence and a psychometric model to identify groups of test-takers with unusual person-fit distributions. A second phase examines individuals within anomalous groups. Another line of research considers collusion detection methods that depend on the identification of aberrant response times. These methods can be integrated for greater power, using joint statistical models for item response and response time. Here, we explore the value added when collusion detection is conducted under the divergence framework, using two joint models of responses and response times: the lognormal model within a hierarchical framework (van der Linden, Journal of Educational and Behavioral Statistics 31:181–204, 2006; van der Linden, Psychometrika 72:287–308, 2007), and a model extended from the diffusion family of models for choice reaction time (Ratcliff, Psychological Review 85:59–108, 1978; Ratcliff et al., Psychological Review 106:261–300, 1999).