Assessing the external validity of algorithms to estimate EQ-5D-3L from the WOMAC.

BACKGROUND: The use of mapping algorithms have been suggested as a solution to predict health utilities when no preference-based measure is included in the study. However, validity and predictive performance of these algorithms are highly variable and hence assessing the accuracy and validity of algorithms before use them in a new setting is of importance. The aim of the current study was to assess the predictive accuracy of three mapping algorithms to estimate the EQ-5D-3L from the Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) among Swedish people with knee disorders. Two of these algorithms developed using ordinary least squares (OLS) models and one developed using mixture model.

METHODS: The data from 1078 subjects mean (SD) age 69.4 (7.2) years with frequent knee pain and/or knee osteoarthritis from the Malmö Osteoarthritis study in Sweden were used. The algorithms' performance was assessed using mean error, mean absolute error, and root mean squared error. Two types of prediction were estimated for mixture model: weighted average (WA), and conditional on estimated component (CEC).

RESULTS: The overall mean was overpredicted by an OLS model and underpredicted by two other algorithms (P < 0.001). All predictions but the CEC predictions of mixture model had a narrower range than the observed scores (22 to 90 %). All algorithms suffered from overprediction for severe health states and underprediction for mild health states with lesser extent for mixture model. While the mixture model outperformed OLS models at the extremes of the EQ-5D-3D distribution, it underperformed around the center of the distribution.

CONCLUSIONS: While algorithm based on mixture model reflected the distribution of EQ-5D-3L data more accurately compared with OLS models, all algorithms suffered from systematic bias. This calls for caution in applying these mapping algorithms in a new setting particularly in samples with milder knee problems than original sample. Assessing the impact of the choice of these algorithms on cost-effectiveness studies through sensitivity analysis is recommended.