A unified multi-level model approach to assessing patient responsiveness including; return to normal, minimally important differences and minimal clinically important improvement for patient reported outcome measures

OBJECTIVE This article reviews and compares four commonly used approaches to assess patient responsiveness with a treatment or therapy (return to normal (RTN), minimal important difference (MID), minimal clinically important improvement (MCII), OMERACT-OARSI [Outcome Measures in Rheumatology-Osteoarthris Reseach Society International] (OO)) and demonstrates how each of the methods can be formulated in a multilevel modelling (MLM) framework. DESIGN Cohort study. SETTING A cohort of patients undergoing total hip and knee replacement were recruited from a single UK National Health Service hospital. POPULATION 400 patients from the Arthroplasty Pain Experience cohort study undergoing total hip (n=210) and knee (n=190) replacement who completed the Intermittent and Constant Osteoarthritis Pain questionnaire prior to surgery and then at 3, 6 and 12 months after surgery. PRIMARY OUTCOMES The primary outcome was defined as a response to treatment following total hip or knee replacement. We compared baseline scores, change scores and proportion of individuals defined as 'responders' using traditional and MLM approaches with patient responsiveness. RESULTS Using existing approaches, baseline and change scores are underestimated, and the variance of baseline and change scores overestimated in comparison with MLM approaches. MLM increases the proportion of individuals defined as responding in RTN, MID and OO criteria compared with existing approaches. Using MLM with the MCII criteria reduces the number of individuals identified as responders. CONCLUSION MLM improves the estimation of the SD of baseline and change scores by explicitly incorporating measurement error into the model and avoiding regression to the mean when making individual predictions. Using refined definitions of responsiveness may lead to a reduction in misclassification when attempting to predict who does and does not respond to an intervention and clarifies the similarities between existing methods.