| This research investigates the extent to which linguistic features of standardized test antonym items can be used to predict the empirical difficulty of these items. The study was motivated by both theoretical and practical considerations. Theoretically, exploring content-feature correlates of difficulty is a form of construct validation as such inquiry attends to the question of why an item is easy or difficult. The pursuit of evidential links between content and difficulty is a scientific, falsifiable approach to validation. In contrast, conventional response-based analyses do not address the central role played by item content in controlling difficulty. From a practical standpoint, detecting content-based predictors allows for the scaling of similarly constructed items without having to establish difficulty measurements via pretesting.;Twelve variables associated with stem and keyed answer words of 975 published SAT antonym items are considered. Results reveal that the stem word's standard frequency index (SFI) is the most robust predictor of empirical difficulty (as measured by Educational Testing Service's equated delta statistic), distantly followed by the keyed answer word's SFI and part of speech. Across all items, one-third of difficulty variance was explained by this predictor set, confirming the frequency-difficulty relationship reported in studies employing more common words and smaller samples. Prediction was attenuated by the relatively high proportion of words having low frequencies. Presumably, a larger representation of familiar words and more differentiated low frequency values would significantly improve prediction.;These findings suggest that feature-based scaling is an efficient, cost effective means for advancing local development of adaptive assessment systems. A computerized adaptive testing (CAT) system is suggested as one practical, psychometric application of the study's findings. The proposed system would extract verbal stimuli from an electronic word domain and employ the aforementioned word variables as proxies for vocabulary item difficulty. Among the design issues discussed are the automatic construction of vocabulary items, adaptive branching, sequential decision-making, iterative refinement of item calibrations, and the provision of timely output for formative direction of learning and instruction. |
