Abstract
Keywords
Generic preference-based measures, such as the EQ-5D 1 and SF-6D,2,3 can be used in the calculation of Quality Adjusted Life Years (QALYs) to inform the economic evaluation of health interventions. The ‘quality’ adjustment is based on the preferences of the general population for health states described by the measure, and is anchored at 1 for full health and 0 for dead. Traditionally, preferences have been derived using ‘iterative’ choice-based techniques such as the Time Trade Off (TTO), 4 which has been widely used to produce utility values for the EQ-5D-3L, 5 and the Standard Gamble (SG), which has been used to value the SF-6D. 3
More recently, studies have explored the use of Discrete Choice Experiments incorporating duration (DCETTO) to estimate utility values anchored on the full health–dead scale. DCETTO has been shown to produce logically consistent value sets for a range of descriptive systems in the UK, Canada and Australia, including the EQ-5D-3L,6,7 the EQ-5D-5L8-10 and the SF-6D. 11 For example, a methodological study based in the UK valued EQ-5D-5L online using DCETTO with 1,799 members of the general population completing 15 DCETTO choice sets each (the PRET-AS study).9,10 The results demonstrated generally logically ordered utility decrements within each dimension. The overall health state values ranged from 1 (for state 11111, i.e., no problems on any dimension) to -0.845 (for state 55555, i.e., worst health state with extreme problems on all dimensions) with approximately one-third of the states modeled as worse than being dead. The distribution of values was unimodal, and the difference in health state value between the best (11111) and the next best (12111) state was small in comparison to, for example, the UK EQ-5D-3L TTO value set. 5
Although there is evidence for the feasibility of DCETTO, important methodological and design issues remain. One issue relates to the duration attribute. DCETTO data are modeled by examining the interaction between the level of each dimension with duration (see Bansback and others 6 for details), and therefore the accuracy of this parameter affects the overall precision of the anchored value set. The standard error of the duration coefficient observed in the PRET-AS study above was relatively large. Since only three levels of duration (1, 5, and 10 years) were used (generating six unique ratios combining different levels of duration), and with only 15% of pairs having different durations across the options, there is scope to explore this dimension further. Increasing the number of levels and combinations of duration may allow for a wider range of variation across scenarios.
A second issue relates to the efficiency of the DCE design. Previous designs have assumed no prior knowledge regarding the value of the parameters (i.e., it used ‘zero prior’ values). However, these previous studies provide some information about the parameter values, and therefore could be used to improve the statistical efficiency of the experimental designs, requiring fewer respondents or responses for each respondent. 12 This could be approached in two ways. One approach would be to include the coefficients from an existing study as ‘non-zero prior values’ to select the health profile pairs (i.e., pairs of scenarios described in terms of EQ-5D-5L states with specified durations) in an efficient design. The expectation associated with the incorporation of prior information in the design concerns the efficiency of the design (i.e., the precision of the coefficient estimates), rather than the estimated coefficients themselves. At the same time, it should be noted that a more statistically efficient design may mean that each choice is more difficult for the respondent, which may lower the respondent efficiency (see Johnson and others 12 for a discussion of statistical versus respondent efficiency). The other approach would be to generate the design in two stages: first to select pairs of EQ-5D-5L states using zero-priors in an efficient design, and then to use predicted utility values from previous studies to combine the states with duration values to form the health state profile. The duration values are chosen so that the expected distribution of responses to each health state profile pair will fall around a given proportion.12,13
Based on these issues, this paper reports a study that seeks to improve the precision of future DCETTO designs by addressing three aims:
To increase the number of levels of the duration attribute of the DCETTO design and, hence, the proportion of choice sets where duration varies (in comparison to the earlier UK EQ-5D-5L study, PRET-AS).
To compare the impact of using a design with zero prior values and a design with non-zero prior values on the preciseness of the coefficient estimates and the extent to which they are logically ordered.
To examine an alternative, two-stage approach to DCETTO design (allocating duration based on the estimated value of the states), and compare this with the (one-stage) designs that use zero and non-zero prior values.
Methods
Choice Set Design
The DCETTO choice sets used in this study are based on EQ-5D-5L, which was developed from the EQ-5D-3L instrument to improve the sensitivity of the descriptive system and standardize the wording across the dimensions. 14 EQ-5D classifies health states across five dimensions (mobility, self-care, usual activities, pain/discomfort and anxiety/depression). In contrast to the three-level version, the 5D-5L adds two intermediate levels of severity (none, slight, moderate, severe, extreme/unable). The DCETTO profiles used in this study consist of “you” living in a particular EQ-5D-5L ‘health scenario’ for one of six levels of duration T (where T = 6 months, 1, 2, 4, 7 and 10 years) followed by death. This generates 20 unique ratios of duration that allow for trading off across a wider range of durations that is more subtle than the six ratios possible in PRET-AS. Respondents were asked which health profile they prefer. Figure 1 displays a screenshot of a choice set.
Screenshot of a DCETTO choice set used in this study.
Experimental Design
The EQ-5D-5L describes 3,125 possible health states, and combining these with six levels of duration, amounts to 18,750 possible health profiles and therefore over 350m possible profile pairs. The general guidance for pairwise DCE choice design is that the minimum number of pairs required is the number of parameters to be estimated. As described by Bansback and others, 9 DCETTO models the pairwise choice data in terms of interactions between the health state levels (categorical) and duration (continuous), and therefore the number of parameters is 21 (interactions between each of the EQ-5D-5L level dummies and continuous duration ((5-1) × 5 × 1= 20), plus continuous duration). However, further analyses involving EQ-5D-5L main effects and quadratic duration (the above 21, plus EQ-5D-5L main effects 5 × (4-1) = 20, interactions between these and duration squared 20 × 1=20, and duration squared) would require 62 parameters.
Three designs are used in this paper. The first two are based on selecting pairs of health profiles in a single step. Type Ia was designed to address aim 1, using zero prior values for the parameter values. Type Ib used non-zero prior values taken from the PRET-AS study, and was compared to Type Ia to address aim 2. Both designs were based on the D-efficiency criterion, which is a summary measure of the precision that would be achieved from a given design given the prior values. The Types Ia and Ib designs include 69 (58%) and 86 (72%) pairs, respectively, where duration differs across the profiles.
The Type II design addressed aim 3, and involved firstly generating 120 EQ-5D-5L health state pairs and, secondly, selecting the duration (either 6 months, 1 year, 4 years, 7 years or 10 years) for each profile. This was done using coefficients generated from the PRET-AS study in an optimization procedure to select the durations that resulted in an expected split of 70% v. 30% in the choice of health profiles. These values were chosen because they lie within the range identified by Kanninen 13 for the optimal probability split for DCE designs (see also Fowkes and Wardman 16 ). In total, 111 (93%) of the choice sets achieved a predicted probability between 66% and 74%. The remaining nine choice sets had EQ-5D-5L scenarios that were too different to achieve a 70% to 30% split, and were allocated either a matched duration of 10 years, or different durations of 10 years and 6 months. In total, 100 (80%) of the choice sets were assigned different durations. Again, no restrictions were applied.
For the Type I designs, the experimental design programme Ngene 15 was used to select 120 pairs and allocate each pair to one of 12 blocks of 10 for each survey version. Stata was used to select the 120 pairs for the Type II design.
There are concerns that some EQ-5D-5L states are more difficult to imagine than others and some states may even appear to be “implausible”. 5 However, since there is no agreed measure of how difficult a given health state is to imagine or any threshold along this measure beyond which states become unimaginable, no health state combinations were excluded from any of the designs.
Survey Design, Recruitment and the Sample
Respondents were recruited from an existing commercial internet panel (IPSOS Observer), and were selected according to quotas based on the UK general population for age (across five age groupings) and gender. First, potential respondents were invited by e-mail and accessed the survey webpage, where they read detailed project information and consented to take part. Those consenting to participate then completed questions on demographic background and self-reported health status, the Office of National Statistics (ONS) wellbeing questions, 17 and EQ-5D-5L for their own health. They were then presented with information about the DCETTO tasks including details about the EQ-5D-5L health dimensions, and instructions to imagine: that they would experience each health state for the period shown without relief or treatment; that death would be very swift and completely painless; and that they would have no other health problems besides what was indicated. This was followed by ten DCETTO choice sets. Respondents were screened out if they completed the survey in less than the minimum time of 2 minutes.
Analysis
DCETTO Modeling
We followed the analysis described in detail by Bansback and others.
6
Conditional logit regression
18
was used to estimate the coefficients of a utility function μ defined by a vector of four dummy variables for each EQ-5D-5L attribute
The data are a series of binary outcomes indicating respondent i’s choices between profiles j = 1, 2. The coefficient β represents the value of living in full health for one year and is expected to be positive; λ represents the disutility of living with the specified set of EQ-5D-5L health problems (
As is shown by Bansback and others,
6
the value for each health state
Thus, for full health, this value is 1 (since when
The results are reported in terms of the ‘unanchored’ coefficients (β and λ), which are on a latent scale and therefore their magnitudes are not directly comparable across models, and the ‘anchored’ coefficients (λ/β), which are on the scale with 1 for full health and 0 for dead, and are therefore comparable across models.
Assessing the Duration Coefficient and Comparing the Designs
To assess the duration attribute (aim 1), we examined the sign and ordering of the coefficients using the Type Ia results. The sign of the duration coefficient should be positive (as utility increases with the time spent living in full health), whereas the sign of the interaction coefficients on the levels of each dimension are expected to be negative, since they are all worse than the baseline (which is level 1, no problems). Furthermore, the levels in each dimension should have a logical ordering, whereby more severe levels should have larger decrements from the baseline. This was compared to the model from the PRET-AS study.9,10
To assess the impact of non-zero prior values (aim 2) and the two-stage design (aim 3), we first compared the models produced by the three Types of design, by comparing differences in the sign, ordering and significance of the coefficients across the models, and the standard errors (of the anchored coefficients only). We next examined differences in the anchored scales by assessing the predicted values of six select EQ-5D-5L states (five very mild states and the very worst state).
To gauge response efficiency, the difficulty of choice sets within each design was considered. The difficulty of a choice set was proxied by the distribution of respondents across the two options within the pair: an easy choice set is one where one option emerges as the clear majority choice by a wide margin (e.g., 90% of respondents choose the majority choice), while a difficult choice set is one where the majority choice has a much narrower margin (e.g., 55% of respondents choose it). The distribution of the 120 choice sets across different levels of difficulty was compared across the three Types.
The three designs were compared in two further analyses. First, we used the approach proposed by Swait and Louviere 19 to test the null hypothesis that preferences are heterogeneous across the three sub-samples. The likelihood-ratio test statistic is given by LR = −2(LLR– LLU), where LLR is the log-likelihood of a model estimated on the pooled sample which allows for scale differences but assumes that the value of living in full health for a specified duration (β) and the disutility of an EQ-5D-5L health state for a certain duration (λ) do not vary across the sub-samples (this is defined as the restricted model). LLU is the sum of the log likelihoods of three conditional logit models estimated on the sub-samples, which, together, form the unrestricted model (this allows for variation in preferences across the three sub-samples). The restricted model is estimated using the clogithet Stata module.20,21
Furthermore, we examined non-trading with respect to duration, defined as respondents consistently choosing the health profiles with the longest duration. While such lexicographic behavior may reflect “genuine” preferences, 22 it may also indicate a simplifying heuristic is being used, which results in non-trading across dimensions. By Type, we examined the number of times the profile with the longer duration was chosen, and for each Type and survey version, the number of pairs where duration differs, and the number of respondents always selecting the option with a longer duration when available to them in the version they completed. Respondents who never chose the profile with the shorter duration were defined as “duration-based non-traders” and their demographic characteristics were compared to the rest of the sample. DCETTO models excluding non-traders were compared to the full sample models.
Results
Response Rate and Demographics
For Type Ia, approximately 12,000 panel members were invited to take part by e-mail, with 1,618 (13%) accessing the survey. Of these, 340 (21%) were screened out due to full quotas, 433 (27%) did not pass the information and consent pages, 41 (3%) dropped out during the survey, 2 (0.001%) completed in less than 2 minutes, with 802 (50%) fully completing the survey. Each of the 12 survey versions, including ten choice sets, was completed by between 48 (6%) and 81 (10%) respondents.
Similar figures apply for Type Ib. Approximately 11,000 members of the online panel were invited to take part, and 1,567 (14%) accessed the survey. Of these, 340 (22%) dropped out due to full quotas, 377 (24%) did not consent to take part (or did not pass the information page), 50 (3%) dropped out during the survey, and 800 (51%) respondents fully completed the survey (none were excluded for taking less than 2 minutes). Between 56 (7%) and 77 (10%) completed each of the 12 survey versions.
For Type II, approximately 3,800 members of the panel were invited to take part by e-mail, and 643 (17%) accessed the survey. Of these, 222 (35%) did not consent to take part (or did not pass the information page), 21 (3%) dropped out during the survey, and 400 (62%) respondents fully completed the survey (none were excluded on the basis of taking less than 2 minutes). Between 30 (8%) and 40 (10%) completed the 12 survey versions.
The demographics of the samples are reported in Table 1, and are similar across all survey versions. Type Ib has significantly more respondents in the best EQ-5D-5L health state (11111) than Type Ia but self-reported health is not significantly different. There are no other significant differences in demographic characteristics across the groups. The characteristics of the PRET-AS sample are included for comparison where available. 10
Demographics of the Sample for Each Design and the PRET-AS Data
Null hypothesis: the values for Types Ia, Ib and II are not different
Source: Mulhern and others. 10
DCETTO Models
Table 2 reports the unanchored coefficients for the Types Ia, Ib, II and the PRET-AS samples, and Figure 2 displays the anchored coefficients for the same designs. The significance levels of the coefficients in Table 2 are in comparison to level 1 (reference), and to the level directly before (one level milder). For the Type Ia design with zero prior values, there are small non-significant inconsistencies between levels 2 and 3 of the pain/discomfort and anxiety/depression dimensions. This indicates that slight and moderate problems in these dimensions are valued on average as being no different in terms of severity.
Unanchored Model Results with Selected Health State Values from the Anchored Models
Null hypothesis is that the coefficient is no different from the coefficient for the adjacent milder level
Comparison of the anchored coefficients from the four designs compared in this study (Type Ia, Type Ib, Type II and PRET-AS).
The Type Ib design with non-zero prior values has more non-significant inconsistencies between mobility levels 4 and 5; and usual activities for levels 1, 2 and 3 (where the coefficients for levels 2 and 3 are both positive leading to a non-significant increase in utility as the health level decreases). Levels 1, 2 and 3 of the pain/discomfort dimension are also inconsistent, where level 2 has a non-significant positive coefficient. Level 3 has a non-significant negative coefficient, as expected, but the difference between the two coefficients is significant.
For Type II, mobility level 2 has the ‘wrong’ sign but is non-significant. Regarding the earlier PRET-AS study, the design includes non-significant coefficients for mobility level 2 and self-care level 2. All the remaining coefficients in PRET-AS are logically ordered and significant.
The standard errors of the anchored model coefficients differ across the designs. The Type Ia design standard errors (range, 0.021 to 0.025) are approximately half the size of the Type Ib standard errors (0.039 to 0.049), indicating increased precision. The Type II standard errors are in the same range as that of the Type Ia design (0.021 to 0.026), with approximately half the sample size, and the PRET-AS errors range from 0.011 to 0.015 (with a substantially larger number of observations).
The bottom rows in Table 2 display the predicted anchored utility values for six health states for the three designs. The Type Ib and Type II model produced utility values above 1 (which are not significantly different to 1), but this is not the case for Type Ia. PRET-AS also produced an estimate above 1 (not significant). Regarding the relative ranking of the five mildest states, there is no clear pattern across the four models. The value for the worst state (55555) ranges from -0.852 (Type Ib) to -0.706 (Type Ia).
Figure 2 compares the anchored utility decrements produced for each design. The similarity of the decrements across the designs varies across the dimensions.
Objective (1) – Investigating the Duration Attribute
Table 2 illustrates that, for the Type Ia design, increasing the number of duration levels and the pairs where duration varies (in comparison to the PRET-AS design) still produces a model with generally logically ordered coefficients. There are differences with the PRET-AS model, where the difference between the slight and moderate severity levels is smaller and the anchored coefficients have smaller standard errors. However, note that the PRET-AS model is based on more than three times the number of observations.
Objectives (2) and (3) – Comparing the Different DCE Design Approaches
Table 2 shows that the Type Ib design has more evidence of disordering than Type Ia. For example, there is a disordering between the coefficients for levels 1 and 2 of the pain/discomfort dimension, where level 2 has a positive coefficient. This would mean that an increase from no pain/discomfort to slight pain/discomfort would lead to an increase in utility. However, none of the disordering is statistically significant. The standard errors of the anchored coefficients (which are directly comparable) are larger for Type Ib than Type Ia. The Type II model displays a high level of ordering, with half the number of observations of Types Ia and Ib.
Figure 3 is a set of histograms illustrating the distribution of the 120 choice sets across five classes of difficulty, ranging from where 90% to 100% of respondents chose the majority choice to where 50% to 59% of respondents made the same choice. For example, in the Type Ia dataset, 98.6% of respondents preferred the health state 32131 for 4 years over 55354 for 4 years. This was defined as easier than the choice between 25111 for 1 year and 33451 for 7 years, where 52% and 48% preferred each profile, respectively. Panel (a) is for Type Ia, with zero prior values, panel (b) is for Type Ib, with non-zero prior values, and panel (c) for Type II. The patterns across the three Types are clearly different: Type Ia has a relatively uniform distribution of choice sets across the five classes of difficulty; Type Ib has a distinctly upward sloping pattern, with 40% of the choice sets in the most difficult class; and Type II is clearly unimodal, but the mode at 80% and above is less difficult than the intended 70% to 30% split.
Comparison of the distribution of the 120 choice sets across different difficulty indicator groups for the Type Ia, Type Ib and Type II designs.
Further Analyses: Heterogeneous Preferences across Sub-samples
Table 3 reports the model estimated on the pooled (Types Ia, Ib and II) dataset, which allows for scale differences but assumes that the value of living in full health for a specified duration (β) and the disutility of an EQ-5D-5L health state for a certain duration (λ) do not vary across the samples. The likelihood ratio statistic is 134.85, which implies that the null hypothesis—that the parameters are equal across the groups—is rejected. The results suggest that the scale is lowest for Type Ib, followed by Type Ia, and then Type II. This indicates that the choices are more “noisy” for Type Ib, with Type II the least noisy.
Pooled Model including All Three Designs
p values for a test for the scale parameter being equal to 1.
Further Analyses: Duration-based Non-trading
The Type Ia design includes 68 (56.7%) choice sets where duration varies resulting in 4,567 observations. Of these, 2,847 (62.3%) result in the respondent choosing the profile with the longer duration. The equivalent numbers for Type Ib are 86 (71.7%) choice sets with 5,785 observations, leading to 2,987 (51.6%) with the longer duration chosen. For Type II, 101 (84.2%) choice sets differ, with 3,326 observations, and 2,126 (63.9%) with the longer duration chosen. Table 4 reports, for each Type and survey version, the number of pairs where the duration differs; and across each block and overall, the number of duration-based non-traders; i.e., respondents never selecting the option with a shorter duration when available. At the individual respondent level, for Type Ia, 130 (16.2%) are non-traders, i.e., they always chose the option with the longer duration; for Type Ib, this was 68 (8.5%), and for Type II, 38 (9.5%). The proportions of non-traders differed significantly between Type Ia and Type Ib (P < 0.000) but not between Type Ib and Type II (P = 0.648).
Proportion of Respondents Non-trading based on the Duration Attribute
Number of pairs in each block (out of 10) where duration differs across the health profiles.
Number of respondents completing each block.
Number of respondents never choosing the shorter duration for the profiles in their block where duration differs.
The demographic characteristics of the duration-based non-traders and the rest do not differ. The analysis comparing the full sample models with those excluding non-traders indicates limited differences, with the same pattern of coefficient ordering (and disordering) within each dimension demonstrated. There are minor differences in the anchored value ranges (with the models excluding non-traders having a value for 55555 between 0.1 and 0.2 lower for each of the designs). (Details are available from the corresponding author.)
Discussion
This paper presents the results of a study addressing three design issues relating to the use of DCETTO to value health states from descriptive systems such as the EQ-5D-5L: the effect of the duration attribute (aim 1); the effect of using non-zero prior values in the design (aim 2); and the effect of a two-stage design (aim 3).
Regarding aim 1, the results demonstrate that increasing the number of duration levels and pairs where duration varies (in comparison to an earlier DCETTO study with EQ-5D-5L in the UK) is feasible, and produces a model with coefficients that are generally logically ordered. Other DCETTO studies have also used more than three levels and found generally consistent models.6–8,11 It is possible that using more levels will increase the validity of responses, as it allows for smaller differences between life years in choice sets. Using more levels of duration that have closer ratios may make it less likely that respondents will choose on the basis of duration alone (i.e., lexicographic decision making). Even with the closer ratios between duration levels designed in this study (including 10 years v. 7 years; and 7 years v. 4 years with the smallest ratios), the results suggest that there is a minority of respondents who choose based on duration only throughout, and this varies across the designs.
Although the results were compared with the earlier UK EQ-5D-5L PRET-AS study, it should be noted that the duration attribute is not the only difference across the two studies. Other differences include the DCE design (for PRET-AS we used the modified Fedorov algorithm implemented in Stata, while the Type I designs in the current study used the Ngene swapping algorithm), the sample size (1,799 for a single design in PRET-AS; 2,002 across three designs for the current study), and the internet panel used to provide the sample. Respondents from different internet panels may exhibit different behaviors based on the criteria for panel membership and recruitment strategies. The number of DCETTO choice sets per respondent was also different. Since the results of PRET-AS suggested that models using 10 pairs were more consistent than those using 15,9,10 the current study used 10. Other studies in this area have used similar numbers of choice sets and produced logically consistent models.6-8,11
Regarding aim 2, the model estimated with the non-zero prior value design data was more inconsistent than the model estimated using the zero prior value data, with less precise coefficient estimates. To our knowledge, this is the first comparison of zero and non-zero prior value designs using DCETTO, and the findings are unexpected, since the introduction of the non-zero prior values should, in theory, improve the precision of the parameter estimates. While the use of prior values that differ substantially from the true parameter values would lead to an inefficient design, our prior values are taken from an earlier published study (PRET-AS),9,10 which also used DCETTO to estimate values for the EQ-5D-5L in the UK. On the other hand, it seems to be that the non-zero design resulted in a larger number of choice sets, where the proportions of respondents across the options is closer. It is reasonable to assume that such choice sets are more difficult and challenging for the respondent to complete, resulting in larger estimation errors. The inclusion of some choice sets that contribute less towards improving the precision of the parameter estimates but allow the respondents to remain focused by not overburdening them, may be the reason that the Type 1a design is found to be the most efficient overall. Nevertheless, it should also be noted that the drop-out rate from the survey amongst those who started was no higher for Type Ib across the two types (both approximately 3%).
Regarding aim 3, the results demonstrate that the two-stage design of allocating duration based on the estimated health state severity is feasible, as it produces a generally logical and ordered model using the same number of pairs as Type Ia and Ib, but with half the sample size. This two-stage design is, in effect, very similar to what is proposed by Kanninen, 13 where the levels of a key attribute, in our case, duration, is manipulated to achieve the required proportions based on the results from an earlier round of data collection. The positive results that we obtained mean that the two-stage design approach has potential and could be tested further in comparison with other more established design methods.
Non-trading will impact the DCETTO process and therefore the validity of the models produced. The analysis of duration-based non-trading suggests that some respondents may choose selections that are based on duration alone. Overall, 236 (12%) of the 2,002 respondents always choose the option with the longer duration; although, the proportion differs across the designs indicating that the preferences for duration in comparison to the health dimensions may differ across the blocks. However, removing duration-based non-traders does not change the characteristics of the models. It is important to note that lexicographic behavior may reflect genuine preferences. For example, always choosing the profile with a longer duration could be a genuine preference based on religion or maximizing time with family, as suggested by qualitative work. 10 This could also apply to the EQ-5D-5L dimensions (where respondents always choose the option with the least severe level on one dimension, and only consider a second dimension if there is a tie): these may be genuine preferences for that particular aspect of health, which helps the respondent to decide between the profiles presented.
This study is not without drawbacks, which could be areas for future research. We cannot fully gauge respondent behavior whilst completing the choice sets. Further work to try to understand how respondents complete choice sets, and the impact of heuristics and response strategies, such as those found in other DCE studies,10,23,24 would prove useful. Regarding the study design process, we designed one set of states for each of the two Type I designs, but we could have developed a range of designs and selected those with more similar numbers of pairs that displayed different durations (which the lexicographic choice analysis suggests is an important aspect of the process). We also do not know the extent to which the non-zero prior values used were indicative of the preferences of the tested population. This may limit the inferences that can be drawn about the use of prior values in DCETTO studies per se. However, the prior values were taken from an earlier valuation study of EQ-5D-5L using DCETTO, with a similar sample in terms of observable characteristics, using the same mode of administration, all of which would suggest high comparability. Testing other prior values from different studies, in different countries, and for different descriptive systems, may help to establish the impact of the use of non-zero prior values on the consistency of the models produced and whether they are useful in the design of health state valuation studies using DCETTO.
Footnotes
Acknowledgements
This study was funded by the EuroQol Research Foundation. The views expressed do not necessarily reflect the views of the EuroQol Research Foundation. Ethics approval was obtained from the University of Sheffield Research Ethics Committee. We are grateful to Ben van Hout, Julie Ratcliffe, Elly Stolk, Richard DeAbreu Lourenco, and the three anonymous referees for Medical Decision Making for their comments, and all the respondents who took part. The usual disclaimers apply.
Earlier versions of this paper were presented at the UK Health Economists’ Study Group conference, June 2014, Glasgow; the EuroQol Group Plenary, Sept 2014, Stockholm; and the International Academy of Health Preference Research, Amsterdam, November 2014.
Financial support for this study was provided entirely by a grant from the EuroQol Group. The funding agreement ensured the authors’ independence in designing the study, interpreting the data, writing, and publishing the report. The following authors are members of the EuroQol Group: Brendan Mulhern; Nick Bansback; Aki Tsuchiya.