Re-scaling social preference data: implications for modelling

Abstract

As applied in cost-utility analysis, generic health status indexes require that full health and dead are valued as 1 and 0, respectively. When social preference weights for health states are obtained using a visual analogue scale (VAS), their raw scores often lie on a scale with different endpoints (such as "best" and "worst" health). Re-scaling individual raw scores to a 0-1 scale leads to the exclusion of respondents who fail to value dead or full health. This study examined alternative approaches that do not impose such strict exclusion criteria. The impact of a different timing of re-scaling (before or after aggregation) and a different measure of central tendency (median or mean) is measured. Data from a postal valuation survey (n=722) conducted in Belgium are used. The following models are considered: (a) re-scaling values for EQ-5D health states on a within-respondent basis and using mean re-scaled values as proxies for social preference values, (b) using median re-scaled values as proxies for social preference values, (c) computing the median raw VAS values and then re-scale, and (e) re-scaling mean raw VAS values. Exclusion rates, health state rankings and valuations and incremental value differences between pairs of states are computed for each model. Models that use a different timing of re-scaling, are compared ceteris paribus to evaluate the importance of timing of re-scaling and models that use a different measure of central tendency are compared ceteris paribus to evaluate the importance of the measure of central tendency. The exclusion rates are above 20% in the models that re-scale valuations before aggregation and less than 5% in the models that re-scale after aggregation. Health state valuations are found to be different in all two by two comparisons. Although in some comparisons the incremental values are statistically significantly different between models, they are never clinically significantly different. Differences in health state rankings were larger between the models that use a different measure of central tendency than between the models that re-scale at a different time. This study shows that, for the data sample used, the choice of the measure of central tendency is more important for the social health status preference values than the timing of re-scaling. Using median original valuations in the analysis of EQ-5D valuation data is theoretically appealing because it builds on the "median voter" model and has the advantage of a lower respondent exclusion rate. Further analysis, on other data samples, will have to confirm the current findings.