Building a composite score for patient self-report of flare in osteoarthritis: a comparison of methods with the Flare-OA-16 questionnaire

Abstract

ObjectiveThis study aims to compare methods of constructing a composite score for the Flare-OA-16 self-reported questionnaire. MethodsParticipants with knee and hip osteoarthritis (OA) completed a validated 16-item questionnaire assessing five domains of flare. Three estimation methods were compared: i) second-order confirmatory factor analysis (CFA); ii) logistic regression, according to the participant’s self-report of flare (yes/no); and iii) Rasch method, with weighted scores in each dimension. The distribution (floor effect (FF) and ceiling effect (CF)) were described and the known-group validity (by self-reported flare) tested by Wilcoxon rank-sum test. Similarity between the scores was analyzed by intraclass correlation coefficient (ICC) and their performance against self-report compared by areas under ROC curves (AUC). Intra-score test-retest reliability at 14 days was assessed by ICC. ResultsIn a sample of 381 participants, 247 reported having a flare. CFA showed fit indices (CFI=.95; RMSEA=.08) and estimated composite mean score=4.33(SD=2.85)[FF=14.9%, CF=0%]. For the logistic regression estimation, mean composite score was 6.48 (SD=3.13)[FF=0%, CF=0%]. With Rasch model, mean composite score was 4.35 (SD=2.60)[FF=14.9%, CF=0%]. Similarity analysis indicated a greater concordance between CFA and Rasch scores (ICC=.98) than between logistic regression score and the two others (ICC=.88 with Rasch score and .90 with CFA score). The AUC indicated similar performance of all methods: logistic model (AUC=.89 [.85-.92]), CFA and Rasch model (AUC=.86 [.82-.90]). The difference between groups was significant (p<.05) for scores estimated by CFA (3.98), Rasch model (4.95) and logistic regression (4.30). The reproducibility was ICC=.84[.75-.90] for Rasch and CFA scores and ICC=.78[.66-86] for logistic model. ConclusionsThree alternatives explored to build a composite score showed similar construct validity. Some metric superiority (better score distribution and reproducibility) of the Rasch model is promising for the detection of occurrence and assessment of severity of a flare in OA.