Comparing EQ-5D scores for comorbid health conditions estimated using five different methods

BACKGROUND: While health state utility values (HSUVs) for many single health conditions are now in the public domain, due to the large number of possible combinations of comorbid health conditions (CHC) the HSUVs for these are not readily available. As a consequence, HSUVs for CHCs are frequently estimated using data obtained from cohorts with single health conditions. With researchers presenting conflicting results, there is currently no consensus on the most appropriate method to estimate HSUVs for CHCs.

OBJECTIVE: The objective of the study was to assess the accuracy of five different methods in the same dataset.

METHODS: EQ-5D data (n=41,174) from the Health Survey for England was used to compare HSUVs generated using the following techniques: the additive, multiplicative and minimum methods, the adjusted decrement estimator (ADE), and a linear regression model.

RESULTS: The additive and multiplicative methods under estimated the majority of HSUVs and the magnitude of the errors increased as the actual HSUV increased. Conversely, the minimum and DAE methods over estimated the majority of HSUVs and the magnitude of errors increased as the actual HSUV decreased. Although the simple linear model produced more accurate results than the others, there was a tendency to under predict higher HSUVs and over predict lower HSUVs and 20% of the errors were greater than the MID (|0.074|) for the EQ-5D. We found the magnitude and direction of mean errors in the estimated scores could be driven by the actual scores being estimated in addition to the technique used and in general the HSUVs estimated using an adjusted baseline were more accurate.

We found the additive and minimum methods performed very poorly in our data. While the simple linear model gave the most accurate results, the model requires validating in external data and additional research exploring alternative model specification is warranted. Our comparison of errors in subgroups of actual EQ-5D scores highlights the need to present additional data when reporting results of analyses in this area as conclusions using average errors in truncated ranges could be misleading.