Question format is the best predictor of item discrimination: a multivariable analysis

Item discrimination, the point-biserial correlation between performance on a question and total exam score, was analyzed in relation to question format, difficulty, and Bloom's taxonomy, which are rarely studied together in a joint framework. To estimate their independent contributions in undergraduate biology examinations, simultaneous multivariable linear and logistic models were conducted on an upper year biology item bank (n = 754). Predictors were question format (multiple-choice question [MCQ], short answer [SA], true/false [TF]), difficulty (easy >80%, moderate 60%-80%, hard <60%), and Bloom (recall, understanding, application, analysis). Question format was the strongest independent predictor. Relative to MCQs, SA items showed higher discrimination and greater odds of meeting the ≥0.35 threshold (odds ratio [OR] = 4.18), while TF items were less discriminating (OR = 0.58). Moderate and hard items exceeded easy questions (ORs = 2.75 and 2.45, respectively). For Bloom, higher-order items outperformed recall questions overall; both understanding and analysis items discriminated better than recall items (ORs = 1.4 and 3.56, respectively). A post hoc grouping of moderate difficulty questions showed that SA + higher-order items (0.53) exceeded MCQ + recall items (0.35), with 89% vs 53% of these items, respectively, meeting a discrimination threshold of ≥0.35. In simultaneous models, question format was the strongest independent predictor of item discrimination, moderate difficulty optimized discrimination, and higher-order objectives exceeded recall. These results provide exploratory insights from a single-institution case study and suggest that adding some higher-order short-answer questions of moderate difficulty may represent a pragmatic strategy for improving assessment quality.