Do country-specific preference weights matter in the choice of mapping algorithms? The case of mapping the Diabetes-39 onto eight country-specific EQ-5D-5L value sets

Abstract

<p><i>Purpose</i>: To develop mapping algorithms that transform Diabetes-39 (D-39) scores onto EQ-5D-5L utility values for each of eight recently published country-specific EQ-5D-5L value sets, and to compare mapping functions across the EQ-5D-5L value sets.</p>

<p><i>Methods</i>: Data include 924 individuals with self-reported diabetes from six countries. The D-39 dimensions, age and gender were used as potential predictors for EQ-5D-5L utilities, which were scored using value sets from eight countries (England, Netherland, Spain, Canada, Uruguay, China, Japan and Korea). Ordinary least squares, generalised linear model, beta binomial regression, fractional regression, MM estimation and censored least absolute deviation were used to estimate the mapping algorithms. The optimal algorithm for each country-specific value set was primarily selected based on normalised root mean square error (NRMSE), normalised mean absolute error (NMAE) and adjusted-r2. Cross-validation with fivefold approach was conducted to test the generalizability of each model.</p>

<p><i>Results</i>: The fractional regression model with loglog as a link function consistently performed best in all country-specific value sets. For instance, the NRMSE (0.1282) and NMAE (0.0914) were the lowest, while adjusted-r2 was the highest (52.5%) when the English value set was considered. Among D-39 dimensions, the energy and mobility was the only one that was consistently significant for all models.</p>

<p><i>Conclusions</i>: The D-39 can be mapped onto the EQ-5D-5L utilities with good predictive accuracy. The fractional regression model, which is appropriate for handling bounded outcomes, outperformed other candidate methods in all country-specific value sets. However, the regression coefficients differed reflecting preference heterogeneity across countries.