Frederick Wolfe1, Kaleb Michaud, Gene Wallenstein. 1. National Data Bank for Rheumatic Diseases, and University of Kansas School of Medicine, Wichita, Kansas 67214, USA. fwolfe@arthritis-research.org
Abstract
OBJECTIVE: To compare the US EQ-5D with the UK EQ-5D and the SF-6D in patients with rheumatoid arthritis (RA). To provide mappings for each of the scales based on clinical variables. METHODS: We studied 12,424 patients with RA with 66,958 longitudinal observations using linear regression. In our mapping models we used the Health Assessment Questionnaire (HAQ) as a continuous predictor variable and as individual items. More complex models included the addition of a visual analog pain scale, the mood scale from the SF-36, and demographic and comorbidity covariates. We compared various models using root mean squared error (RMSE), in-sample and out-of-sample mean absolute error (MAE), and other measures of prediction accuracy and model fit. RESULTS: At any level of clinical severity, the US EQ-5D always had a higher utility score than the UK EQ-5D; and overall, the US scores were 0.094 units higher. The best models explained 64% to 72% of variance in utility scores, with RMSE values of 0.07 (SF-6D), 0.11 (EQ-5D US), and 0.17 (UK EQ-5D). There was a substantial increase in predictive accuracy by using pain and mood as predictor variables in the mapping. CONCLUSION: The US EQ-5D differs from the UK version and from the SF-6D in mean scores and ranges. When determined by mapping, the US EQ-5D has a much lower prediction error than the UK EQ-5D. Simple mapping models that use HAQ and pain have acceptable error rates, although more complex models that include mood scores and individual HAQ items substantially improve predictive accuracy.
OBJECTIVE: To compare the US EQ-5D with the UK EQ-5D and the SF-6D in patients with rheumatoid arthritis (RA). To provide mappings for each of the scales based on clinical variables. METHODS: We studied 12,424 patients with RA with 66,958 longitudinal observations using linear regression. In our mapping models we used the Health Assessment Questionnaire (HAQ) as a continuous predictor variable and as individual items. More complex models included the addition of a visual analog pain scale, the mood scale from the SF-36, and demographic and comorbidity covariates. We compared various models using root mean squared error (RMSE), in-sample and out-of-sample mean absolute error (MAE), and other measures of prediction accuracy and model fit. RESULTS: At any level of clinical severity, the US EQ-5D always had a higher utility score than the UK EQ-5D; and overall, the US scores were 0.094 units higher. The best models explained 64% to 72% of variance in utility scores, with RMSE values of 0.07 (SF-6D), 0.11 (EQ-5D US), and 0.17 (UK EQ-5D). There was a substantial increase in predictive accuracy by using pain and mood as predictor variables in the mapping. CONCLUSION: The US EQ-5D differs from the UK version and from the SF-6D in mean scores and ranges. When determined by mapping, the US EQ-5D has a much lower prediction error than the UK EQ-5D. Simple mapping models that use HAQ and pain have acceptable error rates, although more complex models that include mood scores and individual HAQ items substantially improve predictive accuracy.