The true significance of ‘high’ correlations between EQ-5D value sets

Abstract

High correlation coefficients for EQ-5D value sets derived from different samples, e.g. across countries, are conventionally interpreted as evidence that the people in the respective samples have similar health-related quality of life preferences. However, EQ-5D value sets contain many inherent rankings of health state values by design. By calculating coefficients for value sets created from random data, we demonstrate that ‘high’ correlation coefficients are an artefact of these inherent rankings; e.g. 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 ‘true’ associations. After calculating significance levels based on our simulations – available as a 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.