Testing alternative regression models to predict utilities: mapping the QLQ-C30 onto the EQ-5D-5L and the SF-6D
Purpose: The purpose of the study was to compare alternative statistical techniques to find the best approach for converting QLQ-C30 scores onto EQ-5D-5L and SF-6D utilities, and to estimate the mapping algorithms that best predict these health state utilities.
Methods: 772 cancer patients described their health along the cancer-specific instrument (QLQ-C30) and two generic preference-based instruments (EQ-5D-5L and SF-6D). Seven alternative regression models were applied: ordinary least squares, generalized linear model, extended estimating equations (EEE), fractional regression model, beta binomial (BB) regression, logistic quantile regression and censored least absolute deviation. Normalized mean absolute error (NMAE), normalized root mean square error (NRMSE), r-squared (r2) and concordance correlation coefficient (CCC) were used as model performance criteria. Cross-validation was conducted by randomly splitting internal dataset into two equally sized groups to test the generalizability of each model.
Results: In predicting EQ-5D-5L utilities, the BB regression performed best. It gave better predictive accuracy in terms of all criteria in the full sample, as well as in the validation sample. In predicting SF-6D, the EEE performed best. It outperformed in all criteria: NRMSE = 0.1004, NMAE = 0.0798, CCC = 0.842 and r2 = 72.7% in the full sample, and NRMSE = 0.1037, NMAE = 0.0821, CCC = 0.8345 and r2 = 71.4% in cross-validation.
Conclusions: When only QLQ-C30 data are available, mapping provides an alternative approach to obtain health state utility data for use in cost-effectiveness analyses. Among seven alternative regression models, the BB and the EEE gave the most accurate predictions for EQ-5D-5L and SF-6D, respectively.