The item-focussed tree (IFT) model, a model-based recursive partitioning, combines a classification tree with logistic regression to detect differential item functioning (DIF). DIF is a measurement bias that occurs when patients with the same health status do not interpret patient-reported outcome measures (PROMs) items similarly. Our study purpose was to compare the generalizability of DIF analyses for the IFT model in k-fold and holdout cross-validation (CV) methods using real-world clinical and simulated data. Real-world data included 247 patients. Simulation parameters were sample size, number of items, and DIF effect sizes in three items induced by five binary variables. In clinical data 5-fold CV identified DIF for five of seven items, while both holdout and 2-fold CV methods identified two DIF items, but on different variables. In simulations, holdout CV (0.24) had lower misclassification error rates than k-fold CV (0.33), suggesting its suitability for large samples (>100).