copyfair: An R Package for Item and Ability Conditioned Answer-Copying Detection in Multiple-Choice Examinations

Answer-copying indices for multiple-choice examinations (e.g., the K-index, S2, ω, and the generalized binomial test) typically compare an examinee pair's error similarity against a single global or ability-marginal baseline-an approach that can conflate genuine collusion with population-wide "attractive distractor" effects, in which unrelated examinees independently select the same wrong option. We introduce copyfair, an R package that extends the ability-conditioned diagnostic logic of condfair to answer-copying detection: it models the expected number of matching incorrect responses jointly as a function of examinee ability and item-specific distractor popularity, tests localized disparity with a wild-bootstrap omnibus procedure, and returns Benjamini-Hochberg-corrected pairwise flags. A simulation study confirms the method recovers an injected, ability-concentrated collusion signal (p = .0005). Applied to the PIRLS 2011 Indonesia reading assessment (N = 4,788), an ability-only baseline over-flagged cross-school pairs sharing a common misconception (85.7% of flagged pairs), while the item-and-ability-conditioned model concentrated flags on same-classroom pairs (57.1%, a 114-fold enrichment over the 0.5% baseline rate), consistent with the physical/social proximity expected of genuine collusion.