Estimating Quality Weights for EQ-5D Health States with the Time Trade-Off Method in South Korea

Introduction

Interest in economic evaluation has increased along with the rapid growth of costs of public health. Among the several approaches used for economic evaluation, cost-utility analysis is useful for making comparisons across a broad set of interventions because it uses quality adjusted life years (QALYs). Calculation of QALYs requires quality weights for each health state.

The EQ-5D is a good preference measure that uses the health state classification system consisting of five dimensions (Mobility, Self-Care, Usual Activities, Pain/Discomfort, and Anxiety/Depression) with three levels of “description,” corresponding to “no problems” (level 1), “some problems” (level 2), and “extreme problems” (level 3), thus generating 243 possible health states [1]. The quality weight for each EQ-5D state can be found from a value set.

Although several studies in South Korea have used QALYs, they have used quality weights from other countries. We found, however, that there are differences between Korea and other countries in observed quality weights for some EQ-5D health states [2]. Many researchers have suggested that each country should develop its own social value set [3–5]. We therefore sought to develop a Korean valuation set for EQ-5D health states using time trade-off (TTO) and a statistical model based on quality weights for some EQ-5D health states in the general population of South Korea. We also compared the estimated values with those in other countries.

Methods

Individuals aged 20 years and over living in two adjacent regions, Seoul and Gyeonggi-do, were eligible for this study. A multistage proportionate quota sample was chosen with a consideration of age and sex distribution of the population. We selected 51 small areas and chose homes in each area with a systematic sampling. We finally interviewed individuals by visiting their homes. The survey was performed from March 24, 2005 to April 21, 2005 by 19 interviewers trained in valuing methods such as visual analog scale (VAS) and TTO.

The valuing processes were similar to those of UK Measurement and Valuation of Health (MVH) study [6]. At first, all respondents evaluated their own health states using EQ-5D. Then, they valued 15 hypothetical EQ-5D states which consisted of 11 EQ-5D health states randomly selected from a subset of 41 states by severity and “11111,”“33333,”“unconscious,” and “dead” states given to all respondents using VAS from 0 to 100. Lastly, 13 hypothetical health states except for “11111” and “dead” states were valued with TTO after dividing those states into “better than death” states and “worse than death” states. For “better than death” states, interviewers tried to find a respondent's point of indifference between a length of time (t) in the perfect health state (11111) and 10 years in the target health state. For “worse than death” states, respondents were asked to choose between dying immediately and living a length of time (10-t) in the target health state followed by t years in the perfect health state. We used health state description cards with visual aids to facilitate a valuation task. After their valuation process, respondents were given an opportunity for revision.

Because the meanings of t in “better than death” and “worse than death” states are different, adjustment is needed. Quality weights for states “better than death” were adjusted using the formula t/10. Quality weights for states “worse than death” were calculated using the formula –t/10, so that all values lie in the range [−1, 1]. The data of respondents that had four or more logical inconsistencies were excluded from modeling. The logical inconsistency was defined as the case of the value being larger than that of a health state that is dominant to that state, according to logical ordering of the health state [3].

The initial model included 5 dimensions of EQ-5D with 10 dummy variables indicating the presence of a level 2 or 3 in each dimension. The additional variables considered as independent variables for model development were N3, D1, I2, I2s, I3, and I3s [6,7]. “N3” represents “whether any of the dimensions is at level 3”[6] and “D1” means “the number of movements away from perfect health”[5]. “I*” and “I*s” represent the number of dimensions at level * beyond the first and the square of I*[5]. During model development, we considered four kinds of dependent variables as alternatives: 1 minus the quality weight for a given health state (q1), TTO value of “33333” state minus a given health state (q2), the logarithm of q1 and q2.

For determining the final model, we compared those models using residual diagnostic plots, Wald-type test and mean absolute error (MAE: the average of the absolute differences between observed and predicted values). The residual plots showed that the variability of standardized residuals decreased as predicted values increased. To overcome this heteroscedasticity, we considered a log linear model for the variance as well as a random effects model for mean response [7]. MAE was considered as an important model selection indicator because the objective of this modeling was to predict the quality weights of EQ-5D health states. In conditions with similar MAE values, the simplest model was selected because of parsimony.

To compare estimated quality weights for EQ-5D health states between this study and other foreign studies, the rank correlation coefficients and mean absolute differences (MAD) between estimated quality weights from the final model and those from the UK [6], USA [5], and Japan [4] were calculated.

Results

Respondents composed of 251 men and 249 women. The mean age was 41.3. The proportion of the subjects with elementary school or below, middle and high school, and university or above was 3.8%, 64.6%, and 31.6%, respectively. The sociodemographic characteristics of respondents are detailed elsewhere [2]. Most of the respondents did not differ from the Korean general population. Observed TTO values for the 42 EQ-5D health states and the “unconscious” state are shown in Table 1.

The number of respondents who had no logical inconsistency was 333 (66.6%) of 500 respondents. The number of respondents who violated over 3 inconsistencies was 12 (2.4%). Therefore, 488 respondent values were included in modeling data set.

Most of the five dimension variables were statistically significant in all alternative models. Predictive performance was improved by logarithmic transformation of q1 model. The q1 log model with basic variables was selected as the final model because all variables were statistically significant, MAE was relatively lower (0.074), and it was simple (Table 2).

The rank correlation coefficient of estimated values between this study and foreign studies showed a strong positive correlation (the UK: 0.759, the USA: 0.747, Japan: 0.721). However, the magnitude of MAD was substantial, 0.455, 0.237, and 0.198, respectively (Fig. 1).

Discussion

TTO values for 42 EQ-5D health states were elicited from 500 respondents in the general population of South Korea. The q1 log model was chosen as the final model for estimating quality weight for all EQ-5D health states. Although most models including five-dimension variables were significant, the q1 log model that included only basic variables, a random effects model, was selected as the prediction model because of goodness of fit using MAE. The magnitude of MAE (0.074) was higher than that of studies in the UK, the USA, and Japan. This may be because of, in part, the different patterns of quality weights for EQ-5D health states among countries. Most EQ-5D health states in this study had higher quality weights than those in other countries. The quality weight of “33333” state in this study was lower than in other studies. These differences may have contributed to the relatively high MAE in this study.

All regression coefficients in the final model were statistically significant by the Wald-type test. At level 3, the dimension receiving the highest coefficient was “self-care” followed by “usual activity,”“anxiety/depression,”“mobility,” and “pain/discomfort.” The order of these coefficients also differed from those in the UK, the USA, and Japan [4–6].

Although the overall value set showed a similar pattern among countries, estimated quality weights of EQ-5D health states in the general population of South Korea differed substantially from those in the UK, USA, and Japan. MAD between the Korean and Japanese value sets was lower than that between the Korean and UK and USA sets. This may be because of, in part, the magnitude of cultural similarities among countries. Even the lowest MAD, 0.198, may change the results or conclusions of studies on burden of disease, economic evaluation, and clinical trials when considering that the maximum possible MAD is 2. These MADs provide a rationale for the development of a Korean EQ-5D value set. TTO values may differ from country to country more than VAS values, as shown among European countries [3], between Japan and the UK [4], and between the UK and Spain [8].

Conclusion

This study successfully determined quality weights for all EQ-5D health states using TTO in a general population of South Korea. Substantial differences in the EQ-5D value set among countries were found. Special caution is needed when a value set from one country is applied to another with a different culture.