Comparing Approaches to Link SF-36 PF-10 Scores to PROMIS Physical Function: A Validation Study in Three Clinical Samples

Background: Physical function (PF) is a central patient-reported outcome (PRO) in many clinical conditions. However, the variety of existing PRO measures (PROMs) yield scores on different scales, limiting the score comparability and interpretability. To overcome this gap, the Patient-Reported Outcomes Measurement Information System (PROMIS®) established a standardized T-score metric using item response theory (IRT). As such, different PROMs measuring PF can be linked to this common metric, allowing for efficient harmonization of scores. Linking algorithms allow conversion of SF-36 PF-10 scores to the PROMIS-PF metric, but these methods have not been validated in independent clinical samples. Objective: To validate and compare two established linking methods for the translation of SF-36 PF-10 scores to the PROMIS-PF metric in clinical populations. Design: Two previously proposed linking approaches were applied to estimate PROMIS-PF T-scores based on the SF-36 PF-10: Item-level linking, Cross-walk tables. The directly observed T-scores from the 20-item PROMIS-PF short form (PROMIS-PF20a) served as a benchmark against which the linked T-scores from the SF-36 PF-10 were compared. Results were compared to a newly estimated IRT-model based on the study’s dataset. Participants: Patients from cardiology (n = 185), rheumatology (n = 172), and psychosomatic medicine (n = 262), who completed both the PROMIS-PF20a and the SF-36 PF-10. Main Measures: PROMIS-PF20a, SF-36 PF-10. Key Results: All linking approaches demonstrated high association with observed PROMIS-PF20a T-scores (Pearson correlation ≥ 0.84) and indicated negligible practical differences at the group level (standardized mean difference < 0.2). Conclusions: Two currently available linking approaches can reliably translate SF-36 PF-10 scores to standardized PROMIS-PF T-scores across different clinical samples, eliminating the need for re-estimating models in new datasets. As all linking algorithms ultimately presented highly comparable results, cross-walk tables may be preferred as the most practicable approach, allowing for score conversion without complex statistical modeling.