How Significant Are “High” Correlations Between EQ-5D Value Sets?

Abstract

The calculation of quality-adjusted life years, as used for cost–utility analysis, depends on the availability of value sets representing people’s preferences with respect to health-related quality of life (HRQoL). A value set consists of HRQoL index values for all health states representable by the particular descriptive system used, of which the EQ-5D (EuroQoL, 5 Dimensions) is by far the most widely used. High correlation coefficients for EQ-5D value sets derived from different samples—across countries and/or using different valuation techniques—are conventionally interpreted as evidence that the people in the respective samples have similar HRQoL preferences. However, EQ-5D value sets—for both versions of the system (EQ-5D-3L and EQ-5D-5L)—contain many inherent rankings of health state values by design. By calculating correlation coefficients for value sets created from random data, we demonstrate that “high” coefficients are artifacts of these inherent rankings, such as median Pearson’s r = 0.783 for the EQ-5D-3L and 0.850 for the EQ-5D-5L instead of zero. Therefore, high correlation coefficients do not necessarily constitute evidence of meaningful associations in terms of similar HRQoL preferences. After calculating significance levels based on our simulations—available as an online resource for other researchers—we find that many high coefficients are not as significant as conventionally interpreted, whereas other coefficients are not significant. These “high” but insignificant correlations are in fact spurious.