BACKGROUND: Item response theory (IRT), particularly the graded response model, provides a latent trait (θ) that reflects balance ability and offers a complementary framework for psychometric evaluation. Although IRT is recommended for estimating clinical thresholds, its application to cutoff determination for stroke populations remains limited. We hypothesized that IRT-based approaches will yield cutoff values that differ from those derived using classical test theory and provide additional methodological insight for cutoff determination. PURPOSE: To estimate and compare cutoff values for walking independence in inpatients with stroke using classical test theory and IRT. METHODS: Data from 165 inpatients with stroke were analyzed. Balance was assessed using the Brief Balance Evaluation Systems Test (Brief-BESTest). Cutoff scores were calculated using receiver operating characteristic (ROC), IRT combined with ROC (IRT+ROC), prevalence-based IRT method (old IRT), and embedded state-item IRT method (new IRT). Discriminative performance was evaluated using area under the ROC curve (AUC), sensitivity, specificity, likelihood ratios, and accuracy. RESULTS: The cutoff values for walking independence were 12.37 (ROC), 11.46 (IRT+ROC), 10.54 (old IRT), and 7.90 (new IRT). All methods demonstrated comparable performance, with AUC values ranging from 0.769 to 0.794, indicating good discriminative ability of the Brief-BESTest for identifying walking independence. The new IRT method achieved the highest AUC and sensitivity; ROC and IRT+ROC methods achieved similar AUCs. Although cutoff values and classification metrics differed slightly across methods, their confidence intervals overlapped. CONCLUSION: To our knowledge, this is the first study to directly compare cutoff estimation methods incorporating IRT in stroke rehabilitation, suggesting that IRT-based approaches can yield cutoff values with stability comparable to that of classical test theory and providing methodological insight for future research and clinical assessment scale development.
Shinohara et al. (Mon,) studied this question.