Do health preferences contradict ordering of EQ-5D labels?

Scale and rate heterogeneity in the EQ-5D-5L valuation

OBJECTIVES

To estimate values on a quality-adjusted life year (QALY) scale using individual preference evidence, choice analyses typically include ancillary parameters, such as scale factors and discount rates. These parameters potentially differ among respondents. In this study, we investigated how allowing heterogeneity in scale and rate affects the estimation of EQ-5D-5L values.

METHODS

Using the first wave of the 2016 EQ-5D-5L valuation study (N = 1017), we estimated a conditional logit (CL) model and three mixed logit models: random scale, random rate, and bivariate. Prior to the exploratory study, we hypothesized that scale and rate are correlated and that allowing heterogeneity in both parameters decreases the number of insignificant incremental effects. We confirmed the exploratory findings by re-estimating these models using paired comparison responses from a second wave (N = 1229).

RESULTS

Scale and rate exhibited significant heterogeneity and were positively correlated. As hypothesized, allowing this heterogeneity improved the face validity of the EQ-5D-5L value set by reducing the number of insignificant incremental effects (from 6 to 2 p-values > 0.05; out of 20). Nevertheless, the CL and bivariate mixed logit estimates are highly correlated and concordant (Pearson correlation coefficient of 0.897, Spearman correlation coefficient of 0.888, Lin's concordance coefficient of 0.763).

CONCLUSIONS

Allowing this heterogeneity adds three parameters to the estimation (two variances and a correlation) and improves the face validity of the EQ-5D-5L values. This finding may influence experimental design and choice analysis in health valuation more generally.

Different Frameworks, Similar Results? Head-to-Head Comparison of the Generic Preference-Based Health-Outcome Measures CS-Base and EQ-5D-5L

OBJECTIVE

We compared two generic, preference-based health-outcome measures: the novel patient-centered Château-Santé Base (CS-Base), entailing a multi-attribute preference response framework, and the widely used EQ-5D-5L, regarding effects of different measurement frameworks and different descriptive systems.

METHODS

We conducted a cross-sectional study using a random sample of patients (3019 reached, 1988 included) in the USA with various health conditions. The CS-Base (12 attributes, each with four levels), EQ-5D-5L and the 5D-4L (an ad hoc, multi-attribute preference response-based measure that includes five attributes similar to the EQ-5D-5L, but with four levels) were used as health-outcome measures. We compared the proportions of problems reported on health attributes, statistical robustness and face validity of coefficients, attribute importance, differentiation between health states based on health-state values obtained with these measures, and user experience.

RESULTS

All the CS-Base and 5D-4L coefficients had logical orders and significant differences from the reference level (p < 0.001). Substantial differences were observed in the CS-Base and 5D-4L coefficients between all levels on all attributes, while subtle differences were seen in those of the EQ-5D-5L. Attribute importance of usual (daily) activities were lowest or second lowest in all the three health-outcome measures. Attributes with the highest importance in the CS-Base, 5D-4L, and EQ-5D-5L were respectively mobility, anxiety/depression, and pain/discomfort. Four attributes are similar between the CS-Base and EQ-5D-5L, eight are exclusive to CS-Base. Of the eight, vision and hearing had the highest importance. Health-state values showed a smoother distribution with minimal discontinuity in the CS-Base and EQ-5D-5L than in the 5D-4L. In user experience evaluation, both CS-Base and the 5D-4L showed mean scores above 50 (indicating positive evaluation) in terms of the description of health and ease of understanding.

CONCLUSIONS

This study demonstrated that CS-Base and 5D-4L, which are grounded in the multi-attribute preference response framework, produced statistically robust coefficients, with better face validity than those for the EQ-5D-5L. CS-Base and the EQ-5D-5L outperformed the 5D-4L in differentiating between health states. Notwithstanding differences in content, measurement frameworks, and estimated coefficients, the computed health-state values were similar between CS-Base and EQ-5D-5L.

Estimating the Relationship Between EQ-5D-5L and EQ-5D-3L: Results from a UK Population Study

OBJECTIVES

The aim of this study was to estimate the relationship between EQ-5D-3L and EQ-5D-5L, in both directions, using a single model.

METHODS

An online survey containing both variants of EQ-5D, with randomised ordering, was administered to a large UK sample in 2020. A joint statistical model of the ten EQ-5D responses (five at 5L, five at 3L), using a multi-equation ordinal regression framework was estimated. The joint model ensures mappings in either direction are fully consistent with the information in the sample and satisfy Bayes' rule. Three extensions enhance model flexibility: a copula specification allows differing degrees of correlation between the 3L and 5L responses at the upper and lower extremes of health; a normal mixture residual distribution gives flexibility in the distributional form of responses; and a common factor captures correlations in responses across the five dimensions.

RESULTS

Almost 50,000 responses were received. Thirty-five percent of respondents reported an existing medical condition. Ninety percent of possible 3L and 43% of possible 5L health states were observed. The preferred model specification includes age, sex and the responses to the EQ-5D instrument. Close alignment to the observed data was observed both in within-sample and out-of-sample comparisons.

CONCLUSION

The results from this study provide a means of translating evidence to or from EQ-5D-3L to or from 5L based on a large-scale UK population survey with randomised ordering. Mapping can be performed either using descriptive system responses, individual utility scores or summary statistics.

Psychometric properties of the EQ-5D-5L compared with EQ-5D-3L in cancer patients in Iran

Background and Objective

Psychometric evidence to support the validity and reliability of the EuroQol-5 Dimensions (EQ-5D) in cancer patients is limited. This study aimed to test the validity and reliability of the EQ-5D-5L (5L) in comparison with EQ-5D-3L (3L) in cancer patients.

Methods

Data of 650 cancer patients were collected through consecutive sampling method from three largest governmental cancer centers in Iran between June 2021 and January 2022. The data were gathered using the 3L, 5L, and the European Organization for Research and Treatment of Cancer quality of life questionnaire (QLQ-C30) instruments. The 3L and 5L were compared in terms of ceiling effect, discriminatory power, convergent and known-groups validity, relative efficiency, inconsistency, agreement, and reliability.

Results

Compared with the 3L, ceiling effect decreased by 27.86%. Absolute and relative informativity of discriminatory power improved by 45.93% and 22.92% in the 5L, respectively. All convergent validity coefficients with 5L were stronger than with 3L. Both 3L and 5L demonstrated good known-groups validity, and the relative efficiency was higher for 5L in 4 out of 7 patients' characteristics. The two instruments showed low overall inconsistency (1.45%) and 92.57% of the differences of observations between the 3L and 5L were within the 95% limit of agreement. The interclass correlation coefficient (ICC) for 3L and 5L indexes were 0.88 and 0.85, respectively, and kappa coefficients in the 3L dimensions (range=0.66-0.92) were higher than the 5L(range=0.64-0.79).

Conclusions

The 5L demonstrated to be better than the 3L in terms of ceiling effect, inconsistency, discriminatory power, convergent validity, relative efficiency.

Preference Paths and Their Kaizen Tasks for Small Samples

BACKGROUND

Stated preference research currently lacks a form of evidence that is well suited for small samples. A preference path is a sequence of two or more choices showing the evolution of an object following an adaptive process.

OBJECTIVES

The aims were to introduce preference paths and their kaizen tasks and to demonstrate how to analyze their evidence using a small sample.

METHODS

Twenty respondents were assigned the same 16 profiles generated from an orthogonal array based on the five attributes of the EQ-5D-5L descriptive system. Each kaizen task began with an opt-out paired comparison (i.e., choosing between the initial 10-year profile and the opt-out "dying immediately"), followed by choosing three changes, and ended with a second paired comparison (final profile versus opt-out) if the respondent chose opt-out initially. By maximum likelihood with respondent clusters, we estimated the 20 main effects using conditional logit and Zermelo-Bradley-Terry (ZBT) specifications.

RESULTS

Apart from demonstrating heterogeneity and profile effects, all main effect estimates were non-negative, and most were significant (15 for logit and all 20 for ZBT; p value < 0.05). Under the logit and ZBT specifications, the value of the worst EQ-5D-5L profile (55555) is - 0.920 quality-adjusted life years (QALYs) or - 1.478 QALYs, respectively. Furthermore, the findings illustrate a log-linear relationship between the logit and ZBT main effects.

CONCLUSION

This paper demonstrates the feasibility of a stated-preference study that estimates 20 main effects using path evidence from 20 respondents (16 kaizen tasks, 15-min interviews). This approach shows promise for future application in stated-preference research, particularly in small samples.

A New Approach to Assessing Children's Interpretation of Severity Qualifiers in a Multi-Attribute Utility Instrument-The EQ-5D-Y-5L: Development and Testing

Introduction

The beta EQ-5D-Y-5L is a new patient-reported outcome measure (PROM) for children aged 8–15 years that is currently under development by the EuroQol Group. The EQ-5D-Y-5L is similar to the EQ-5D-Y but has five levels of severity per dimension rather than three. The increased number of levels increases the granularity of the responses but possibly has also increased the difficulty of distinguishing between levels. The EuroQoL’s Version Management Committee (VMC) required a robust method to determine how well children distinguish between the five EQ-5D-Y-5L ordinal severity qualifiers (i.e. ‘no problems’ through to ‘extreme problems’), which are a critical aspect of both health measurement and the valuation of health states.

Objective

This paper describes the development, testing, selection, and piloting of such a method.

Methods

Following a literature review and consultation with the wider VMC and a Language Support Services agency, a range of exercises were developed to assess the ordering and comprehension of the five severity qualifiers. Three exercises were pre-tested with children in Spain and New Zealand. One exercise, preferred and understood by children, was then piloted.

Results

Five children in Spain and 11 in New Zealand tested the three exercises. In both countries, all children found the three exercises easy to understand and complete. Of the 12 children who expressed a preference, nine said they preferred the card ranking. Card ranking also allowed the interviewer to observe difficult choices being made as the children physically rearranged the card order until they settled on their final order. Following rigorous assessment of translatability and cultural portability by an independent Language Support Service, card ranking was piloted in South Africa (n = 9) and in Indonesia (n = 10), where it highlighted severity qualifier order inversions that would otherwise not have been detected.

Conclusion

The card ranking exercise was found to be a preferred and acceptable means of testing the ordering of translations of severity qualifiers among children. Additional formal testing of the exercise in other countries and languages is now underway. The approach developed and tested by the VMC for cognitive debriefing of beta EQ-5D-Y-5L language/country versions may also be useful in determining the adequacy of translated qualifiers in debriefing of adult EQ-5D-5L versions and other PROMs.

Parallel Valuation of the EQ-5D-3L and EQ-5D-5L by Time Trade-Off in Hungary

OBJECTIVES

The wording of the Hungarian EQ-5D-3L and EQ-5D-5L descriptive systems differ a great deal. This study aimed to (1) develop EQ-5D-3L and EQ-5D-5L value sets for Hungary from a common sample, and (2) compare how level wording affected valuations.

METHODS

In 2018 to 2019, 1000 respondents, representative of the Hungarian general population, completed composite time trade-off tasks. Pooled heteroscedastic Tobit models were used to estimate value sets. Value set characteristics, single-level transition utilities from adjacent corner health states, and mean transition utilities for all possible health states were compared between the EQ-5D-3L and EQ-5D-5L.

RESULTS

Health utilities ranged from -0.865 to 1 for the EQ-5D-3L and -0.848 to 1 for the EQ-5D-5L. The relative importance of the 5 EQ-5D-5L dimensions was as follows: mobility, pain/discomfort, self-care, anxiety/depression, and usual activities. A similar preference ranking was observed for the EQ-5D-3L with self-care being more important than pain/discomfort. The EQ-5D-5L demonstrated lower ceiling effects (range of utilities for the mildest states: 0.900-0.958 [3L] vs 0.955-0.965 [5L]) and better consistency of mean transition utilities across the range of scale. Changing "confined to bed" (3L) to "unable to walk" (5L) had a large positive impact on utilities. Smaller changes with more negative wording in the other dimensions (eg, "very much anxious/feeling down a lot" [3L] vs "extremely anxious/depressed" [5L]) had a modest negative impact on utilities.

CONCLUSION

This study developed value sets of the EQ-5D-3L and EQ-5D-5L for Hungary. Our findings contribute to the understanding of how the wording of descriptive systems affects the estimates of utilities.

Does Device or Connection Type Affect Health Preferences in Online Surveys?

BACKGROUND AND OBJECTIVE

Recent evidence has shown that online surveys can reliably collect preference data, which markedly decrease the cost of health preference studies and expand their representativeness. As the use of mobile technology continues to grow, we wanted to examine its potential impact on health preferences.

METHODS

Two recently completed discrete choice experiments using members of the US general population (n = 15,292) included information on respondent device (cell phone, tablet, Mac, PC) and internet connection (business, cellular, college, government, residential). In this analysis, we tested for differences in respondent characteristics, participation, response quality, and utility values for the 5-level EQ-5D (EQ-5D-5L) by device and connection.

RESULTS

Compared to Mac and PC users, respondents using a cell phone or tablet had longer completion times and were significantly more likely to drop out during the surveys (p < 0.001). Tablet users also demonstrated more logical inconsistencies (p = 0.05). Likewise, respondents using a cellular internet connection exhibit significantly less consistency in their health preferences. However, matched samples for tablets and cell phones produced similar EQ-5D-5L utility values (mean differences < 0.06 on a quality-adjusted life-year [QALY] scale for all potential health states).

CONCLUSION

Allowing respondents to complete online surveys using a cell phone or tablet or over a cellular connection substantially increases the diversity of respondents and the likelihood of obtaining a representative sample, as many individuals have cell phones but not a computer. While the results showed systematic variability in participation and response quality by device and connection type, this study did not show any meaningful changes in utility values.

Head-to-Head Comparison of EQ-5D-3L and EQ-5D-5L Health Values

BACKGROUND

The EQ-5D is a widely used preference-based instrument to measure health-related quality of life. Some methodological drawbacks of its three-level version (EQ-5D-3L) prompted development of a new format (EQ-5D-5L). There is no clear evidence that the new format outperforms the standard version.

OBJECTIVE

The objective of this study was to make a head-to-head comparison of the EQ-5D-3L and EQ-5D-5L in a discrete choice model setting giving special attention to the consistency and logical ordering of coefficients for the attribute levels and to the differences in health-state values.

METHODS

Using efficient designs, 240 pairs of EQ-5D-3L health states and 240 pairs of EQ-5D-5L health states were generated in a pairwise choice format. The study included 3698 Dutch general population respondents, analyzed their responses using a conditional logit model, and compared the values elicited by EQ-5D-3L and EQ-5D-5L for different health states.

RESULTS

No inconsistencies or illogical ordering of level coefficients were observed in either version. The proportion of severe health states with low values was higher in the EQ-5D-5L than in the EQ-5D-3L, and the proportion of mild/moderate states was lower in the EQ-5D-5L than in the EQ-5D-3L. Moreover, differences were observed in the relative weights of the attributes.

CONCLUSION

Overall distribution of health-state values derived from a large representative sample using the same measurement framework for both versions showed differences between the EQ-5D-3L and EQ-5D-5L. However, even small differences in the phrasing (language) of the descriptive system or in the valuation protocol can produce differences in values between these two versions.

Valuing EQ-5D-5L health states 'in context' using a discrete choice experiment

BACKGROUND

In health state valuation studies, health states are typically presented as a series of sentences, each describing a health dimension and severity 'level'. Differences in the severity levels can be subtle, and confusion about which is 'worse' can lead to logically inconsistent valuation data. A solution could be to mimic the way patients self-report health, where the ordinal structure of levels is clear. We develop and test the feasibility of presenting EQ-5D-5L health states in the 'context' of the entire EQ-5D-5L descriptive system.

METHODS

An online two-arm discrete choice experiment was conducted in the UK (n = 993). Respondents were randomly allocated to a control (standard presentation) or 'context' arm. Each respondent completed 16 paired comparison tasks and feedback questions about the tasks. Differences across arms were assessed using regression analyses.

RESULTS

Presenting health states 'in context' can significantly reduce the selection of logically dominated health states, particularly for labels 'severe' and 'extreme' (χ² = 46.02, p < 0.001). Preferences differ significantly between arms (likelihood ratio statistic = 42.00, p < 0.05). Comparing conditional logit modeling results, coefficients are ordered as expected for both arms, but the magnitude of decrements between levels is larger for the context arm.

CONCLUSIONS

Health state presentation is a key consideration in the design of valuation studies. Presenting health states 'in context' affects valuation data and reduces logical inconsistencies. Our results could have implications for other valuation tasks such as time trade-off, and for the valuation of other preference-based measures.

Further evidence on EQ-5D-5L preference inversion: a Brazil/U.S. collaboration

PURPOSE

A preference inversion occurs when "worse" health (instead of "better" health) along a scale or score is preferred. Our aim was to confirm past findings of EQ-5D-5L preference inversions among English-speaking respondents and to explore inversions among Portuguese-speaking respondents. Anecdotal evidence suggests that inversions may be more common in the Portuguese version, where the translation of the fourth level "severely" (gravemente) means "gravely."

METHODS

Through an infusion clinic in Tampa, Florida, United States and a cardiology clinic in Rio de Janeiro, Brazil, 740 respondents completed a tablet-based survey, which included the EQ-5D-3L and -5L followed by paired comparisons designed to assess preference inversions between the fourth and fifth levels of each of the five domains: Mobility (MO), Self-Care (SC), Usual Activity (UA), Pain/Discomfort (PD), and Anxiety/Depression (AD). An example from the AD dimension would be: "Which do you prefer? Starting today, 30 days with health problems: Severely anxious or depressed or Extremely anxious or depressed" (i.e., Level 4 AD vs. Level 5 AD).

RESULTS

In the English-speaking respondents, preference inversion was only observed to a substantial extent in the AD dimension (U.S. N = 470; 7% MO, 14% SC, 14% UA, 20% PD, and 45% AD). Inversions were more common among the Portuguese-speaking respondents (Brazil N = 270; 11% MO, 32% SC, 35% UA, 49% PD, and 65% AD). Specifically, 44 out of 68 Brazilian respondents (65%) preferred "extremamente" (Level 5 AD) over "gravemente ansioso (a) ou deprimido (a)" (Level 4 AD).

CONCLUSIONS

This evidence confirms previous findings for the U.S. English version of the EQ-5D-5L and led to a relabeling on the Portuguese version. It demonstrates the usefulness of collaboration between psychometric, econometric, and linguistic experts in developing the wording for and translating preference-based measures of health-related quality of life. Further research may explore inversions in other translations. The authors recommend that preference inversion tests should be included in the development and translation process.

Is Dimension Order Important when Valuing Health States Using Discrete Choice Experiments Including Duration?

BACKGROUND

Discrete choice experiments with duration (DCE_TTO) can be used to estimate utility values for preference-based measures, such as the EQ-5D-5L. For self-completion, the health dimensions are presented in a standard order. However, for valuation, this may result in order effects. Thus, it is important to understand whether health state dimension ordering affects values. The aim of this study was to examine the importance of dimension ordering on DCE values using EQ-5D-5L.

METHODS

A choice experiment presenting two health profiles and a third immediate death option was developed. A three-arm study was used, with the same 120 choice sets presented online across each arm (n = 360 per arm). Arm 1 presented the standard EQ-5D-5L dimension order, arm 2 randomised order between respondents, and arm 3 randomised within respondents. Conditional logit regression was used to assess model consistency, and scale parameter testing was used to assess model poolability.

RESULTS

There were minor inconsistencies across each arm, but the magnitudes of the coefficients produced were generally consistent. Arm 3 produced the largest range of utility values (1 to -0.980). Scale parameter testing suggested that the models did not differ, and the data could be pooled. Follow-up questions did not suggest variation in terms of difficulty.

CONCLUSIONS

The results suggest that the level of randomisation used in DCE health state valuation studies does not significantly impact values, and dimension order may not be as important as other study design issues. The results support past valuation studies that use the standard order of dimensions.