A Changepoint Rule for Determining Bicluster Retention Cutoffs in Cheating Detection

Abstract

In test cheating detection, biclustering can be used to identify localized groups of examinees who share unusual response patterns, but an unresolved practical issue is determining how many extracted biclusters should be retained and how many examinees should be flagged. Retaining too many biclusters may increase false positives by capturing weak or noise-driven patterns, whereas retaining too few may miss meaningful cheating structures. This study proposes a changepoint-based retention rule that uses the ordered sequence of bicluster p values to identify where strong cheating-related evidence begins to give way to weaker residual patterns, thereby providing a data-driven cutoff for bicluster retention. The method was evaluated using two operational test forms with known cheating labels and a simulation study that varied cheating type, test length, and the proportion of compromised items. Label-based benchmark cutoffs were defined using the F1-score, balanced accuracy, and Youden’s index, with the F1-score treated as the primary benchmark. In the empirical analysis, the estimated changepoints closely aligned with the F1-based benchmark, yielding the same sensitivity and slightly lower specificity across both forms. In the simulation study, the estimated changepoint generally approximated the F1-based benchmark but tended to select more conservative retention cutoffs, resulting in lower sensitivity and small specificity differences across conditions. These findings suggest that the proposed rule can provide a useful basis for determining how many biclusters to flag for further review in operational cheating-detection settings.