A recent study found that negative utility values elicited using composite time trade-off (TTO) were barely associated with the severity of EQ-5D-5L health states, suggesting poor discriminative ability. Assuming negative values provide limited information, this study aimed to explore the usefulness of censoring negative TTO values at 0 in modeling EQ-5D-5L valuation data.
We analyzed EQ-5D-5L valuation data from China, The Netherlands, Canada, Singapore, and Thailand. For each data set, we estimated value sets using 2 Tobit models, one left-censored at −1 (current practice) and one left-censored at 0 (our proposed method), and compared the model performances. We hypothesized that censoring at 0 and censoring at −1 would produce similar values, though on slightly different scales.
When censoring at 0, logical inconsistencies and statistical significance were improved but the value range was compressed. In the cross-attribute level effects model, the 3-level parameters were similar between the models censored at 0 and −1, but the rank order of some dimension parameters was altered. Health state values predicted by the 2 censoring models approximated a perfect agreement after rescaling.
Censoring TTO values at 0 improved model estimation and fit but produced higher utility values than models censoring at −1. Investigators of future EQ-5D value set studies using the composite TTO method are advised to examine the validity of negative TTO values before choosing modeling strategies.