Duke 2008 | ASHE | Equity and Efficiency in Health and Healthcare

Objectives: Health states defined by classification systems like the EQ-5D can be valued using methods that aim to elicit cardinal measures directly, such as the time trade-off (TTO) and visual analogue scale (VAS), but also by ranking the health states. In this study, we estimate international value sets for the 243 EQ-5D states based on rank, TTO and VAS responses and test their equivalence. Background: Aggregate-level ordinal rankings can be used to impute cardinal measures, but valuation results and evidence of their validity have only recently begun to emerge and therefore remains limited. For instance at the previous EuroQoL Group Scientific Meeting evidence was presented that ranking can replicate TTO valuations of health-states collected for a subset of EQ-5D states in the same study; in other work, regression models have been fit to ranked EQ-5D states in specific country datasets, and the predicted valuations from the rank model corresponded closely to the empirical TTO valuations for the same states. Thus far, however, only single-country applications have been reported. This paper contributes EQ-5D valuation sets useful for international health policy and further tests the equivalence of values based on ordinal and cardinal responses. If the values of the direct and ranking procedure prove sufficiently similar, the evidence would further support the expanded use of simpler health state valuation techniques, which would be especially advantageous when illiteracy and innumeracy hinder practical measurement of societal preferences over health states, as in the developing world. Research Design: We estimated the coefficients of a conditional logit and a linear probability model of rank responses as well as the coefficients of two linear models of TTO and VAS responses using pooled data from eight countries: Slovenia, Argentina, Denmark, Japan, Netherlands, Spain, United Kingdom and the United States which gave us 179,431 responses from 11,483 subjects. The main difference between the two rank models is that one models utility in ln(odds), as suggested by McFadden, and the second in probability. To allow comparison across methods, we normalize all predicted health state values so that dead has a value of 0.00, and the state with no problems on any dimension (11111) has a value of 1.00. These values were permitted to vary according to seventeen state variables: two level indicators for each of the five dimensions, five indicators for the seven possible combinations of any 1s, 2s and 3s, a count of 2s squared, and a count of 3s squared. This specification has previous models estimated in the United States and United Kingdom nested within its framework. Furthermore, we compare predicted values for the 242 EQ-5D states, excluding 11111, across methods in terms of correlation and concordance. Results: The non-optimal gap reduces when utilities are linearly modeled in probability instead of log-odds, as suggested by McFadden. The rank-based values are highly correlated with both TTO and VAS values. Compared to the log-odds model, the linear probability model produces rank-based values with greater concordance to TTO/VAS values. Conclusions: In former investigations we tested if ranking in the large pool of data produce values of health states comparable with direct measures as TTO and VAS. In this investigation we provide further evidence by showing convergent validity between TTO/VAS and alternative rank-based models for the whole valuation space. This evidence emphasizes the promise of a valuation technique that incorporates ordinal responses. Overall, rank-based values merits closer investigation.