Modeling what matters by accounting for measurement error when estimating meaningful score differences: application of longitudinal item response theory to an emulated oncology trial dataset

PURPOSE: Patient-reported outcomes (PROs) are important for understanding patient experiences in clinical trials, but meaningful interpretation of changes in PRO scores remains challenging. This study compared meaningful score difference (MSD) estimates obtained using a conventional anchor-based method (ABM) with those obtained using longitudinal item response theory (LIRT), which explicitly models measurement error. METHODS: An analytic dataset (N = 153) was generated using correlation-driven regeneration based on item-level data from a Phase 3 oncology trial. The primary outcome was the physical functioning (PF) scale of the European Organisation for Research and Treatment of Cancer Quality of Life Questionnaire - Core 30 (EORTC QLQ-C30). Four candidate anchors were evaluated based on their association with change in PF, of which three met the prespecified correlation threshold (|Spearman's r| > 0.30). MSDs for improvement and deterioration were estimated using the conventional ABM and a longitudinal graded response model incorporating the anchor item directly. Ninety-five percent confidence intervals (CIs) were estimated using nonparametric bootstrap resampling for ABM and subject-level empirical bootstrap resampling for LIRT. RESULTS: LIRT generally produced larger absolute MSD estimates than ABM, with the largest differences observed for deterioration. For deterioration of two or more anchor categories, LIRT MSD estimates were - 47.22, - 39.89, and - 42.17 points for difference in Patient Global Impression of Severity (ΔPGIS), difference in item 29 of EORTC QLQ-C30 (ΔI29), and difference in item 30 of EORTC QLQ-C30 (ΔI30), respectively, compared with ABM estimates of - 8.89, - 12.73, and - 20.74. Differences between methods were generally smaller for improvement. LIRT bootstrap CIs were wider for deterioration than for improvement, particularly for the two-category deterioration thresholds, reflecting greater sampling uncertainty in sparsely represented deterioration categories. CONCLUSIONS: Accounting for measurement error using a longitudinal latent-variable framework can materially affect the estimated magnitude of meaningful change, particularly for deterioration. LIRT may provide a useful complement to conventional anchor-based approaches, although estimates from sparsely represented anchor-change categories should be interpreted cautiously. Further empirical and simulation research is needed to evaluate the robustness and generalizability of LIRT-based MSD estimation across clinical settings.