Valuation of EQ-5D Health States in Poland: First TTO-Based Social Value Set in Central and Eastern Europe
ABSTRACT
Objective: Currently, there is no EQ-5D value set for Poland. The primary objective of this study was to elicit EQ-5D Polish values using the time trade-off (TTO) method.
Methods: Face-to-face interviews with visitors of inpatients in eight medical centers in Warsaw, Skierniewice, and Puławy were carried out by trained interviewers. Quota sampling was used to achieve a representative sample of the Polish population with regard to age and sex. Modified protocol from the Measurement and Value of Health study was used. Each respondent ranked 10 health states and valued 4 health states using the visual analog scale and 23 using the TTO. Mean and variance stability tests were performed to determine whether using a larger number of health states per respondent would yield credible results. Modeling included random effects and random parameters models.
Results: Between February and May 2008, 321 interviews were performed. Modeling based on 6777 valuations resulted in an additive model with all coefficients statistically significant, R2 equal to 0.45, and value −0.523 for the worst possible health state. Means and variance did not differ significantly for states valued in the middle and at the end of the TTO exercise.
Conclusions: This is the first EQ-5D value set based on TTO in Central and Eastern Europe so far. Because the values differ considerably from those elicited in Western European countries, its use should be recommended for studies in Poland. Increasing the number of health states that each respondent is asked to value using TTO seems feasible and justifiable.
Introduction
Economic evaluation of health technologies in the cost–utility analysis framework aims at providing maximal utility—as perceived by a given society—within a limited budget, and thus should be based on social preferences that are to be satisfied. The EQ-5D questionnaire is a widely known tool that can be used to elicit social preferences [1].
The EQ-5D was translated into Polish in 1997, after the EuroQol Group guidelines and in interaction with EuroQol translation review members [2]. This version is currently used in clinical research conducted in Poland. The main limitation against the wider application of the Polish EQ-5D in clinical and pharmacoeconomic studies in Poland is the lack of either population norms or a national EQ-5D value set. As a result, the Agency for Health Technology Assessment in Poland (AHTAPol) has been recommending the use of the EQ-5D European value set [3]. This value set was derived using the visual analog scale (VAS) methodology developed during the EuroQol BIOMED Research Programme funded by the European Union (1998–2001) [4]. This might not be an optimal choice, as in health economics, the preferred outcome measure is quality-adjusted life-years (QALYs) and the VAS scale is less associated with the QALY paradigm than choice-based valuation methods, like for instance time trade-off (TTO) [5]. Moreover, the EQ-5D European value set was based only on data collected in Western European countries (Finland, Germany, The Netherlands, Spain, Sweden, and the United Kingdom), lacking data from Poland or any other Central European country. The primary objective of this study was therefore to establish a Polish EQ-5D value set using TTO. The secondary objectives included comparison with the EQ-5D European VAS value set and other potentially useful value sets, as well as assessing the possible bias resulting from expanding the TTO experiment to 23 states per respondent.
Methods
EQ-5D
EQ-5D essentially consists of two pages—the EQ-5D descriptive system (page 2) and the EQ visual analog scale (EQ VAS) (page 3) [1]. The EQ-5D descriptive system comprises the following five dimensions: mobility, self-care, usual activities, pain/discomfort, and anxiety/depression. Each dimension has three levels: no problems, some problems, and severe problems. The respondent is asked to indicate his or her health state by ticking (or placing a cross) in the box against the most appropriate statement in each of the five dimensions. This decision results in a one-digit number expressing the level selected for that dimension. A health state is defined by combining one level from each of the five dimensions. A total of 243 possible health states are defined in this way. Each state is referred to in terms of a five-digit code. For example, state 11111 indicates no problems on any of the five dimensions, while state 11223 indicates no problems with mobility and self-care, some problems with performing usual activities, moderate pain or discomfort, and extreme anxiety or depression.
EQ-5D health states, defined by the EQ-5D descriptive system, may be converted into a single summary index by applying a formula that essentially attaches values (also called weights) to each of the levels in each dimension [6]. The index can be calculated by deducting the appropriate weights from 1, the value for full health (i.e., state 11111). Information in this format is useful, for example, in cost–utility analysis. Value sets have been derived for EQ-5D in several countries using the EQ-5D VAS or the TTO valuation techniques [20].
Sample
Between February and May 2008, 10 trained undergraduated medical students surveyed a representative sample of the Polish adult population. Two training workshops for the interviewers were conducted by two study investigators (JB, DG). Each interviewer had to conduct at least one simulated interview. Survey quotas with respect to age and sex were prepared based on demographic data from the Central Statistical Office in Poland [8]. Face-to-face interviews were conducted among visitors of inpatients in eight medical centers in Warsaw (Mazowieckie voivodship), Skierniewice (Łódzkie voivodship), and Puławy (Lubelskie voivodship). Respondents were not promised any compensation but were given an unexpected gift of limited value only after the interview had taken place. The study was approved by the Medical University of Warsaw ethics committee (KB/24/2008) and all respondents gave informed consent.
Study Design and Pilot Tests
This study was based on the most frequently cited EQ-5D valuation study, the Measurement and Value of Health (MVH), conducted in the United Kingdom in 1993 [9]. The MVH study was a large exercise, in which each of 3395 respondents valued 13 different health states. The MVH research group collected values for 43 (out of 245 potential EQ-5D health states [including “unconscious”]), frequently eliciting more than 500 observations per health state.
Because of budget limitations of the current study, it was necessary to modify the MVH protocol. Based on the findings of Lamers et al. [10] who studied the efficiency of the MVH design, the current study proposed to 1) carry out approximately 300 interviews; 2) collect approximately 150 valuations per health state; and 3) increase the number of health states valued per respondent to more than the 17 health states (as in the EQ-5D Dutch [10] and Japanese [11] studies).
Five pilot interviews were conducted by one of the study investigators (DG) and one of the student interviewers. Respondents ranked a set of 19 health states and valued the ranked health states using the EQ VAS and the TTO technique. These pilot interviews showed that the ranking and VAS valuation exercise may be more time-consuming than the TTO exercise alone. Given that TTO was the primary aim of this study, it was decided to 1) reduce the number of health states to be valued in the ranking exercise to 10; 2) reduce the number of ranked health states to be valued using the EQ VAS to 3 or 4 (the best, the worst, the closest to the average, “immediate death”); and 3) increase the number of health states ranked in the TTO exercise to 23. The above changes were the most substantial deviations from the MVH protocol. Possible problems resulting from the increase in the number of states to be valued using TTO, e.g., order effects, were addressed using statistical tests (see “MVH deviations verification” section). The data from pilot TTO exercises were included into the final study data set.
Interview Procedure
Each respondent was asked to perform the following tasks: 1) indicate his or her own health status using the EQ-5D descriptive system; 2) perform the ranking exercise; 3) value the ranked health states using the EQ-5D VAS; 4) rate his or her own health on the EQ VAS; 5) perform the TTO exercise; and 6) answer some socioeconomic background questions. Two sets of 25 cards describing health states according to the EQ-5D descriptive system were used alternatingly by the interviewers (Table 1). Dead and 11111 are not valued in the TTO exercise by its design. All health states used in the MVH study were used, except “unconscious.” Two additional health states (23333 and 32333) were chosen from those proposed by Kind [12]. Cards describing health states were divided into two fixed sets in such a way that 1) there was equal representation of “very mild,”“mild,”“moderate,” and “severe” health states in both sets; and 2) the largest number of logical comparisons was allowed (e.g., health states “11112” and “11113” were included in set 1, while health states “11121” and “11131” were included in set 2). This procedure was intended to facilitate ranking by the respondent, to shorten the survey time and to control data quality. Health states applied in the ranking exercise (Table 1, states marked with *) were also selected in the same way to facilitate the survey and shorten the duration of the preliminary part of the study. Equalization of the value of health states in both sets used in the ranking exercise was not a priority. Cards used in the ranking exercise were marked in the corner on the reverse side by a black dot to allow interviewers to make a quick selection among the cards in each set. Health state cards were shuffled before the TTO exercise and were presented in random order. Interviewers were asked not to present the first three cards describing health states worse than death.
| Health state category | Set 1 | Set 2 |
|---|---|---|
| Very mild | 11112* | 11121* |
| 11211 | 12111 | |
| 21111 | ||
| Mild | 11122* | 12121 |
| 22112 | 12211 | |
| 22121* | ||
| Moderate | 11113* | 11131* |
| 11133 | 11312 | |
| 12222 | 12223 | |
| 13332 | 13212 | |
| 21133 | 13311 | |
| 21232 | 21222* | |
| 21312 | 21323 | |
| 22122* | 22222* | |
| 22222* | 23313 | |
| 22331 | 32211 | |
| 23321 | 33212 | |
| 32313 | 33321* | |
| 32331 | ||
| Severe | 22323* | 22233 |
| 23333 | 23232 | |
| 32232 | 32223 | |
| 32333* | 33232 | |
| 33333* | 33323* | |
| 33333* | ||
| Anchoring states | 11111* | 11111* |
| DE* | DE* |
- States severity was defined following Kind [12]: very mild—one level 2 problem; mild state—with no level 3 problems and up to three level 2 problems; severe—no level 1 problems and at least two level 3 problems; moderate—neither mild nor severe.
- * Cards used in ranking exercise.
- DE, death.
The TTO exercise used the same visual probe as used in the United Kingdom and in the United States (where researchers also used a protocol similar to the UK MVH protocol) [9,13]. This probe, often called a “time board,” allows for both positive and negative TTO values. The interview book used by Shaw et al. [14] was translated into Polish and used for training the interviewers and in the pilot interviews. During these pilot tests, it emerged that the book was rather complex. For instance, the description of a single TTO exercise occupied three pages in English (US) and four pages in the Polish version. Because we had planned to elicit values for as many as 23 states using TTO, the TTO valuation task would have taken more than 90 pages. Moreover, it was considered likely that a 90-page-long protocol would obstruct the flow of the interview. Based on the interviewers' suggestions and subsequent pilot tests, the instruction and documentation of the TTO exercise in the protocol book was reduced to a graphic system, which allowed the registering of results of five TTO exercises on a single page (http://www.ispor.org/Publications/value/ViHsupplementary/ViH13i2_Golicki.asp). Every interviewer was trained and had continuous access to a separate instruction standardized on the TTO methodology as carried out in the US study [13].
Respondents were allowed to trade time in months and weeks instead of years when no valuation changes were noted for a period of 9 years on side 1 of the time board (positive values). This was a modification introduced during a TTO valuation study by German EuroQol Group members [15]. Results of the TTO exercise were read out from the scale in the protocol book with an accuracy of 0.25 of a year. States regarded as better than dead were anchored on a scale ranging from full health to dead: X/10. States regarded as worse than dead were calculated as X/10 − 1, so scores were bounded by 0 and −1 [9].
Exclusion Criteria
To ensure “rational” trade-offs, respondents who misunderstand the task were removed [16]. These respondents were identified according to the following exclusion criteria: fewer than three states valued, all states valued worse than dead, all states valued the same, and “serious logical inconsistencies.” These respondents were distinguished from those who provided “irrational” values resulting from “normal cognitive imperfections.” A logical inconsistency was defined as being an instance where one health state could be clearly seen to be better than another but the respondent ranked it as worse. A logical inconsistency was called “serious” if the difference in valuation was greater than 0.5. It was considered a clear sign that the respondent had misunderstood the task when he or she had 10 or more serious inconsistencies. In these cases, all responses relating to that particular respondent were excluded. Extreme values, defined as values more than 2 SD from the mean, were also excluded.
MVH Deviations Verification
One of the aims of the present study was to evaluate the possible bias resulting from the modifications made to the original MVH study design and by expanding the TTO experiment to 23 states per respondent. We anticipated that respondents might be too fatigued to credibly answer the last TTO questions. At least two issues would arise as a result: the mean valuation of a health state might be different or the variance of this valuation might increase. In the first case, it would pose a problem relating to the credibility of the valuation; in the second case, the overall error of estimation might increase.
To assess the possible bias involved, several tests were conducted. First, we tested whether the mean valuation for each health state differed when it was valued in the middle of the experiment (as 6th–17th state) or at the end of the experiment (as 18th–23rd state). The first five health states were omitted in this comparison, because they might have been perceived as a warm-up task in the TTO experiment. During valuation of first health states, respondents are just learning the rules of TTO exercise and the variance of the valuation may be significantly high. Beside that, the first three states differ from states valued later, because we asked interviewers not to show states worse than dead at the beginning of the TTO exercise. Because we intended to increase the power of the test and minimize the type II error (not finding a difference in means when there was one), a higher than usual significance level of P = 0.1 was used. At the same time, to control for the multiple hypotheses testing for the 44 individual states, we used the Hölm–Bonferroni correction for multiple hypotheses testing. A t test with separate variance estimation was used, but the common variance assumption did not change the results. Tests for equality of variances were performed analogously.
Modeling
The dependent variable was the loss of utility associated with a specific health state, i.e., 1 − u, where u is the utility. The predictor variables included binary variables dk,l(j) equal to 1 or 0 depending on whenever level l of domain k of health state j was 1, 2, or 3. Furthermore, we used derived variables described in earlier EQ-5D valuation studies—N3, D1, I2, I2sq, I3, and I3sq (for definitions, see Table 2) [9,13].
| Variable | Definition |
|---|---|
| MO2 | 1 if mobility is at level 2; 0 otherwise |
| MO3 | 1 if mobility is at level 3; 0 otherwise |
| SC2 | 1 if self-care is at level 2; 0 otherwise |
| SC3 | 1 if self-care is at level 3; 0 otherwise |
| UA2 | 1 if usual activities is at level 2; 0 otherwise |
| UA3 | 1 if usual activities is at level 3; 0 otherwise |
| PD2 | 1 if pain/discomfort is at level 2; 0 otherwise |
| PD3 | 1 if pain/discomfort is at level 3; 0 otherwise |
| AD2 | 1 if anxiety/depression is at level 2; 0 otherwise |
| AD3 | 1 if anxiety/depression is at level 3; 0 otherwise |
| N3 | 1 if any dimension is at level 3; 0 otherwise |
| D1 | Number of movements away from full health beyond the first (ranging from 0 to 4); D1 = max (0; d1,2 + d2,2 + d3,2 + d4,2 + d5,2 + d1,3 + d2,3 + d3,3 + d4,3 + d5,3 − 1) |
| D1sq | The square of the D1 variable |
| I2 | Number of dimensions at level 2 beyond the first |
| I2 = max (0; d1,2 + d2,2 + d3,2 + d4,2 + d5,2 − 1) | |
| I2sq | The square of the I2 variable |
| I3 | Number of dimensions at level 3 beyond the first |
| I3 = max (0; d1,3 + d2,3 + d3,3 + d4,3 + d5,3 − 1) | |
| I3sq | The square of the I3 variable |
The data had the panel structure with one level being the respondent index, and the other the health state being evaluated. Two approaches were used in modeling. In the first approach, a simple random effects model was built in. For example, it was assumed that the loss of utility assigned by i-th respondent to the j-th health state was described as
where α denotes the model parameters, dk,l(j) was defined as above, ηi,j denotes the error term associated with a single TTO experiment, and υi denotes the error term associated with i-th respondent (and fixed for this respondent across all TTO experiments). It was assumed that ηi,j and υi were independent and normally distributed with zero means. Additionally, models with other variables describing the health state j than just dk,l(j) were used. Random effects modeling was performed with GRETL software [17].
As a second approach, a more complex random parameters model was estimated using Bayesian statistics. In this more complex model, it was assumed that the respondents could differ not only in the error terms, but also in the model parameters. This approach allowed for a full incorporation of demographic differences. The specification of a model was as follows:
where α, dk,l(j), and ηi,j denote as above, and εi,0 and εi,k,l denote the random variability of model parameters on the individual level (notice that υi is incorporated in εi,0 term). It was assumed that εi,0 and εi,k,l were independent normally distributed random variables with fixed variance across all respondents. The random parameters model was estimated using the Bayesian approach and Markov Chain Monte Carlo (MCMC) method in WinBUGS software (MRC Biostatistics Unit, Cambridge, UK) [18]. Noninformative priors were used as follows:
During the MCMC simulation, 10,000 initial simulations and 10,000 sample simulations were used. Ninety-five percent confidence intervals were calculated using the percentile method.
The quality of the models was assessed on two levels—on the individual TTO valuation level and on the health state level. On the individual level, standard R2 coefficient and mean absolute difference between theoretical and empirical value (mean absolute error [MAE]) were calculated. For all 44 health states used in the experiment, the mean absolute difference between value predicted by the given model and average valuation (for the data set used in modeling) was calculated as well as number of states, for which this difference was larger than 0.05 or 0.1.
Comparison with Other Countries' Value Sets
EQ-5D model coefficients and health state values estimated in the present Polish valuation study were compared with those estimated in other countries. Two TTO and two VAS value sets were chosen: 1) the United Kingdom TTO value set (MVH A1 [9]) as it is the “original” standard; 2) the German TTO value set [15] as it emanates from the country closest geographically to Poland; 3) the European VAS value set [5] as recommended by the AHTAPol; and 4) the Slovenian VAS [19] as an example of a country that can be defined as Central European with a similar political history (although not necessarily sharing cultural similarities with Poland). We performed plain comparison of model coefficients and calculated the mean absolute difference between health states values, the number of health states (out of 243) with values more than 0.05 (or 0.1) different from Polish value, and the correlation coefficient between value sets.
Results
In total, 321 respondents completed the interview, of which 53% were females. Age and sex distributions of the respondents after exclusions are shown in Table 3. The interviewed sample group was representative of the Polish general population in terms of age and sex, but contained a large proportion of individuals with higher education, employed people, and students as well as a low percentage of widowed individuals. Approximately 62% of survey participants came from Warsaw, 19% from other towns, and 15% from rural areas. Overall, 86% of individuals in the sample group were inhabitants of Mazowieckie voivodship. The majority of problems reported in the EQ-5D descriptive system were pain/discomfort (40.1%) or anxiety/depression (37.8%). The mean health state recorded on the EQ VAS was 81.4 (SD 14.5), and the mean interview time was 42 minutes (SD 13).
| Study sample, after exclusions (n = 305) | Polish general population (%) | |
|---|---|---|
| Male (years) | 46.9% | 49.9 |
| 18–24 | 7.5% | 7.5 |
| 25–34 | 9.8% | 11.1 |
| 35–44 | 8.9% | 8.7 |
| 45–54 | 9.5% | 9.8 |
| 55–64 | 7.5% | 7.8 |
| 65–74 | 3.6% | 4.1 |
| Female (years) | 53.1% | 51.1 |
| 18–24 | 7.5% | 7.2 |
| 25–34 | 10.5% | 10.8 |
| 35–44 | 9.5% | 8.5 |
| 45–54 | 10.5% | 10.1 |
| 55–64 | 9.5% | 8.8 |
| 65–74 | 5.6% | 5.7 |
| Mean age (SD) | 42.8 (15.7) | Not available |
| Educational level | ||
| Low | 5.2% | 23.7 |
| Middle | 52.1% | 58.0 |
| High | 42.6% | 18.3 |
| Marital status | ||
| Single | 18.4% | 20.3 |
| Married/living together | 72.1% | 64.2 |
| Widowed | 4.6% | 10.4 |
| Divorced | 4.9% | 5.1 |
| Work | ||
| Employed | 62.3% | 53.7 |
| Unemployed | 3.0% | 4.4 |
| Pensioner | 3.0% | 4.8 |
| Retired | 14.8% | 14.8 |
| Student | 11.1% | 6.3 |
| Housewife/househusband | 3.9% | Not available |
| Belief in life after death | 63.8% | Not available |
| EQ-5D | ||
| Those reporting problems on | ||
| Mobility | 16.8% | Not available |
| Self-care | 3.3% | |
| Usual activities | 13.8% | |
| Pain/discomfort | 40.1% | |
| Anxiety/depression | 37.8% | |
| EQ VAS own health | ||
| Mean (SD) | 81.4 (14.5) | Not available |
- EQ VAS, EQ visual analog scale.
There was no response with fewer than three states valued, or with all states valued worse than dead, or with all states valued the same. We identified 532 serious logical inconsistencies in 120 (37%) interviews. Sixteen respondents with 10 or more serious logical inconsistencies were excluded from the final analysis. These respondents did not differ in demographic characteristics from the whole sample group. Eleven of the 16 excluded respondents (69%) were interviewed by the same interviewer, suggesting that the surveyor himself might have been the cause of the logical inconsistencies. Additionally, 206 extreme values, deviating from the mean score by more than 2 SD, were considered invalid and were excluded from the analysis. As a result, the number of useable valuations was reduced from 7351 to 6983 by excluding the 16 respondents with a high number of serious logical inconsistencies and, subsequently, to 6777 after excluding the extreme values (Table 4).
| Observed value (data before quality check) | Observed value (data after quality check) | |||||||
|---|---|---|---|---|---|---|---|---|
| Health state | No. of observations | Mean | SD | % of negative | No. of observations | Mean | SD | % of negative |
| 11112 | 171 | 0.896 | 0.212 | 1 | 157 | 0.925 | 0.116 | 0 |
| 11113 | 171 | 0.656 | 0.425 | 9 | 150 | 0.753 | 0.250 | 2 |
| 11121 | 149 | 0.880 | 0.206 | 1 | 137 | 0.912 | 0.132 | 0 |
| 11122 | 173 | 0.826 | 0.287 | 2 | 160 | 0.848 | 0.249 | 1 |
| 11131 | 149 | 0.286 | 0.619 | 28 | 140 | 0.333 | 0.595 | 26 |
| 11133 | 170 | 0.195 | 0.648 | 34 | 163 | 0.232 | 0.630 | 31 |
| 11211 | 170 | 0.900 | 0.168 | 0 | 154 | 0.935 | 0.091 | 0 |
| 11312 | 147 | 0.685 | 0.362 | 5 | 135 | 0.743 | 0.246 | 2 |
| 12111 | 148 | 0.901 | 0.168 | 0 | 133 | 0.934 | 0.100 | 0 |
| 12121 | 150 | 0.853 | 0.203 | 0 | 135 | 0.891 | 0.140 | 0 |
| 12211 | 149 | 0.849 | 0.178 | 0 | 131 | 0.886 | 0.129 | 0 |
| 12222 | 170 | 0.727 | 0.356 | 4 | 157 | 0.781 | 0.237 | 0 |
| 12223 | 149 | 0.527 | 0.462 | 11 | 131 | 0.635 | 0.304 | 4 |
| 13212 | 150 | 0.615 | 0.403 | 7 | 134 | 0.712 | 0.234 | 1 |
| 13311 | 150 | 0.490 | 0.513 | 16 | 132 | 0.605 | 0.373 | 8 |
| 13332 | 170 | −0.071 | 0.655 | 49 | 162 | −0.040 | 0.653 | 48 |
| 21111 | 170 | 0.915 | 0.140 | 0 | 156 | 0.934 | 0.094 | 0 |
| 21133 | 170 | 0.202 | 0.635 | 29 | 162 | 0.232 | 0.623 | 28 |
| 21222 | 149 | 0.760 | 0.259 | 1 | 137 | 0.799 | 0.195 | 0 |
| 21232 | 170 | 0.287 | 0.631 | 26 | 160 | 0.324 | 0.603 | 24 |
| 21312 | 170 | 0.549 | 0.479 | 11 | 151 | 0.673 | 0.287 | 3 |
| 21323 | 149 | 0.417 | 0.554 | 20 | 133 | 0.530 | 0.430 | 13 |
| 22112 | 170 | 0.783 | 0.306 | 3 | 156 | 0.826 | 0.196 | 0 |
| 22121 | 149 | 0.803 | 0.262 | 1 | 140 | 0.825 | 0.195 | 0 |
| 22122 | 170 | 0.754 | 0.311 | 3 | 155 | 0.797 | 0.212 | 0 |
| 22222 | 319 | 0.663 | 0.405 | 7 | 290 | 0.747 | 0.238 | 0 |
| 22233 | 150 | 0.058 | 0.620 | 40 | 142 | 0.081 | 0.619 | 39 |
| 22323 | 172 | 0.296 | 0.595 | 25 | 158 | 0.366 | 0.534 | 21 |
| 22331 | 171 | 0.071 | 0.657 | 39 | 163 | 0.099 | 0.652 | 37 |
| 23232 | 149 | 0.046 | 0.627 | 41 | 141 | 0.061 | 0.626 | 40 |
| 23313 | 149 | 0.129 | 0.616 | 38 | 141 | 0.165 | 0.601 | 36 |
| 23321 | 173 | 0.293 | 0.598 | 25 | 160 | 0.356 | 0.551 | 21 |
| 23333 | 169 | −0.204 | 0.626 | 60 | 161 | −0.213 | 0.627 | 62 |
| 32211 | 149 | 0.464 | 0.559 | 19 | 132 | 0.573 | 0.445 | 13 |
| 32223 | 149 | 0.187 | 0.587 | 34 | 141 | 0.216 | 0.580 | 32 |
| 32232 | 171 | −0.050 | 0.650 | 50 | 163 | −0.027 | 0.645 | 49 |
| 32313 | 171 | 0.024 | 0.653 | 45 | 163 | 0.058 | 0.647 | 43 |
| 32331 | 169 | −0.110 | 0.627 | 53 | 161 | −0.092 | 0.623 | 52 |
| 32333 | 172 | −0.295 | 0.597 | 69 | 152 | −0.384 | 0.508 | 74 |
| 33212 | 149 | 0.278 | 0.600 | 29 | 139 | 0.319 | 0.574 | 27 |
| 33232 | 150 | −0.183 | 0.600 | 60 | 142 | −0.167 | 0.607 | 60 |
| 33321 | 148 | 0.033 | 0.648 | 48 | 140 | 0.068 | 0.640 | 46 |
| 33323 | 150 | −0.150 | 0.606 | 55 | 142 | −0.143 | 0.611 | 56 |
| 33333 | 318 | −0.362 | 0.542 | 70 | 285 | −0.461 | 0.458 | 78 |
| Total or mean | 7351 | 0.383 | 0.474 | 24 | 6777 | 0.424 | 0.411 | 22 |
The two final models resulted from the random effects modeling are presented in Table 5. The first model encompasses all statistically significant variables, including I3sq (tested for significance of individual variables with t test and the whole set by F test). The R2 value of the model amounted to 0.4524. Because this model contains the nonintuitive I3sq variable, it might be considered as less creditable. Therefore, a second model using only dk,l variables was estimated. On the other hand, exclusion of statistically significant variables and the correlation with the other independent variables introduces bias. Nonetheless, the parsimonious model was much more intuitive and had an R2 value of 0.4517, which is only marginally different from the saturated model.
| Basic | I3sq | Bayesian | ||||
|---|---|---|---|---|---|---|
| Coefficient (SD) | P-value | Coefficient (SD) | P-value | Coefficient (SD) | 95% CI | |
| Constant | 0.049 (0.018) | 0.007 | 0.035 (0.018) | 0.053 | 0.054 (0.013) | 0.028–0.080 |
| MO2 | 0.052 (0.011) | 0.048 (0.011) | 0.051 (0.009) | 0.035–0.069 | ||
| MO3 | 0.331 (0.014) | 0.363 (0.016) | 0.325 (0.018) | 0.289–0.361 | ||
| SC2 | 0.054 (0.012) | 0.057 (0.012) | 0.047 (0.010) | 0.028–0.067 | ||
| SC3 | 0.235 (0.015) | 0.269 (0.016) | 0.224 (0.015) | 0.196–0.253 | ||
| UA2 | 0.046 (0.014) | 0.032 (0.014) | 0.023 | 0.048 (0.011) | 0.026–0.069 | |
| UA3 | 0.212 (0.014) | 0.224 (0.014) | 0.212 (0.014) | 0.183–0.239 | ||
| PD2 | 0.057 (0.011) | 0.063 (0.012) | 0.058 (0.009) | 0.042–0.075 | ||
| PD3 | 0.489 (0.012) | 0.513 (0.013) | 0.485 (0.021) | 0.443–0.526 | ||
| AD2 | 0.026 (0.013) | 0.036 | 0.030 (0.013) | 0.018 | 0.027 (0.010) | 0.007–0.046 |
| AD3 | 0.207 (0.012) | 0.235 (0.013) | 0.204 (0.013) | 0.179–0.229 | ||
| I3sq | −0.012 (0.002) | |||||
| R 2 overall | 0.452 | 0.452 | ||||
| MAE | 0.039 | 0.033 | 0.041 | |||
| No. (of 44) > 0.05 | 10 | 12 | 12 | |||
| No. (of 44) > 0.10 | 3 | 3 | 3 | |||
- All coefficients were significant at P < 0.001 unless otherwise stated.
- CI, confidence interval; MAE, mean absolute error.
Ninety-five percent confidence intervals for the Bayesian model excluded zero; thus, all domains on all levels significantly influenced the utility values (Table 5). In a Bayesian estimation, R2 was not meaningful and is not reported here. It is also worth noting that the random parameters modeling results were very similar to the random effects modeling, indicating that the heterogeneity of the surveyed population had generally no impact on the results. Because Bayesian modeling resulted in no extra predictive value (MAE of 0.039 for the parsimonious model vs. 0.041 for the random parameters model), we therefore decided to base the Polish value set on the classical random effects model, with only dk,l variables and without any interaction variables. The final full Polish EQ-5D value set is presented in Table 6.
| State | Utility | State | Utility | State | Utility | State | Utility |
|---|---|---|---|---|---|---|---|
| 11111 | 1.000 | 13132 | 0.201 | 22223 | 0.535 | 31321 | 0.351 |
| 11112 | 0.925 | 13133 | 0.020 | 22231 | 0.310 | 31322 | 0.325 |
| 11113 | 0.744 | 13211 | 0.670 | 22232 | 0.284 | 31323 | 0.144 |
| 11121 | 0.894 | 13212 | 0.644 | 22233 | 0.103 | 31331 | −0.081 |
| 11122 | 0.868 | 13213 | 0.463 | 22311 | 0.633 | 31332 | −0.107 |
| 11123 | 0.687 | 13221 | 0.613 | 22312 | 0.607 | 31333 | −0.288 |
| 11131 | 0.462 | 13222 | 0.587 | 22313 | 0.426 | 32111 | 0.566 |
| 11132 | 0.436 | 13223 | 0.406 | 22321 | 0.576 | 32112 | 0.540 |
| 11133 | 0.255 | 13231 | 0.181 | 22322 | 0.550 | 32113 | 0.359 |
| 11211 | 0.905 | 13232 | 0.155 | 22323 | 0.369 | 32121 | 0.509 |
| 11212 | 0.879 | 13233 | −0.026 | 22331 | 0.144 | 32122 | 0.483 |
| 11213 | 0.698 | 13311 | 0.504 | 22332 | 0.118 | 32123 | 0.302 |
| 11221 | 0.848 | 13312 | 0.478 | 22333 | −0.063 | 32131 | 0.077 |
| 11222 | 0.822 | 13313 | 0.297 | 23111 | 0.664 | 32132 | 0.051 |
| 11223 | 0.641 | 13321 | 0.447 | 23112 | 0.638 | 32133 | −0.130 |
| 11231 | 0.416 | 13322 | 0.421 | 23113 | 0.457 | 32211 | 0.520 |
| 11232 | 0.390 | 13323 | 0.240 | 23121 | 0.607 | 32212 | 0.494 |
| 11233 | 0.209 | 13331 | 0.015 | 23122 | 0.581 | 32213 | 0.312 |
| 11311 | 0.739 | 13332 | −0.011 | 23123 | 0.400 | 32221 | 0.463 |
| 11312 | 0.713 | 13333 | −0.192 | 23131 | 0.175 | 32222 | 0.437 |
| 11313 | 0.532 | 21111 | 0.899 | 23132 | 0.149 | 32223 | 0.256 |
| 11321 | 0.682 | 21112 | 0.873 | 23133 | −0.032 | 32231 | 0.031 |
| 11322 | 0.656 | 21113 | 0.692 | 23211 | 0.618 | 32232 | 0.005 |
| 11323 | 0.475 | 21121 | 0.842 | 23212 | 0.592 | 32233 | −0.176 |
| 11331 | 0.250 | 21122 | 0.816 | 23213 | 0.411 | 32311 | 0.354 |
| 11332 | 0.224 | 21123 | 0.635 | 23221 | 0.561 | 32312 | 0.328 |
| 11333 | 0.043 | 21131 | 0.410 | 23222 | 0.535 | 32313 | 0.147 |
| 12111 | 0.897 | 21132 | 0.384 | 23223 | 0.354 | 32321 | 0.297 |
| 12112 | 0.871 | 21133 | 0.203 | 23231 | 0.129 | 32322 | 0.270 |
| 12113 | 0.690 | 21211 | 0.853 | 23232 | 0.103 | 32323 | 0.090 |
| 12121 | 0.840 | 21212 | 0.827 | 23233 | −0.078 | 32331 | −0.135 |
| 12122 | 0.814 | 21213 | 0.646 | 23311 | 0.452 | 32332 | −0.161 |
| 12123 | 0.633 | 21221 | 0.796 | 23312 | 0.426 | 32333 | −0.342 |
| 12131 | 0.408 | 21222 | 0.770 | 23313 | 0.245 | 33111 | 0.385 |
| 12132 | 0.382 | 21223 | 0.589 | 23321 | 0.395 | 33112 | 0.359 |
| 12133 | 0.201 | 21231 | 0.364 | 23322 | 0.369 | 33113 | 0.178 |
| 12211 | 0.851 | 21232 | 0.338 | 23323 | 0.188 | 33121 | 0.328 |
| 12212 | 0.825 | 21233 | 0.157 | 23331 | −0.037 | 33122 | 0.302 |
| 12213 | 0.644 | 21311 | 0.687 | 23332 | −0.063 | 33123 | 0.121 |
| 12221 | 0.794 | 21312 | 0.661 | 23333 | −0.244 | 33131 | −0.104 |
| 12222 | 0.768 | 21313 | 0.480 | 31111 | 0.620 | 33132 | −0.130 |
| 12223 | 0.587 | 21321 | 0.630 | 31112 | 0.594 | 33133 | −0.311 |
| 12231 | 0.362 | 21322 | 0.604 | 31113 | 0.413 | 33211 | 0.339 |
| 12232 | 0.336 | 21323 | 0.423 | 31121 | 0.563 | 33212 | 0.313 |
| 12233 | 0.155 | 21331 | 0.198 | 31122 | 0.537 | 33213 | 0.132 |
| 12311 | 0.685 | 21332 | 0.172 | 31123 | 0.356 | 33221 | 0.282 |
| 12312 | 0.659 | 21333 | −0.009 | 31131 | 0.131 | 33222 | 0.256 |
| 12313 | 0.478 | 22111 | 0.845 | 31132 | 0.105 | 33223 | 0.075 |
| 12321 | 0.628 | 22112 | 0.819 | 31133 | −0.076 | 33231 | −0.150 |
| 12322 | 0.602 | 22113 | 0.638 | 31211 | 0.574 | 33232 | −0.176 |
| 12323 | 0.421 | 22121 | 0.788 | 31212 | 0.548 | 33233 | −0.357 |
| 12331 | 0.196 | 22122 | 0.762 | 31213 | 0.367 | 33311 | 0.173 |
| 12332 | 0.170 | 22123 | 0.581 | 31221 | 0.517 | 33312 | 0.147 |
| 12333 | −0.011 | 22131 | 0.356 | 31222 | 0.491 | 33313 | −0.034 |
| 13111 | 0.716 | 22132 | 0.330 | 31223 | 0.310 | 33321 | 0.116 |
| 13112 | 0.690 | 22133 | 0.149 | 31231 | 0.085 | 33322 | 0.090 |
| 13113 | 0.509 | 22211 | 0.799 | 31232 | 0.059 | 33323 | −0.091 |
| 13121 | 0.659 | 22212 | 0.773 | 31233 | −0.122 | 33331 | −0.316 |
| 13122 | 0.633 | 22213 | 0.592 | 31311 | 0.408 | 33332 | −0.342 |
| 13123 | 0.452 | 22221 | 0.742 | 31312 | 0.382 | 33333 | −0.523 |
| 13131 | 0.227 | 22222 | 0.716 | 31313 | 0.201 | Dead | 0.000 |
Test of Additional TTO Valuations
The comparison of health state values when assigned during the middle of the experiment (position 6 to 17) or at the end (position 18 to 23) showed no statistically significant differences neither in mean nor in variance using the Hölm–Bonferroni correction (the smallest P-values for means and variances comparison between groups were equal to 0.0161 and 0.006, respectively, with Hölm–Bonferroni threshold of 0.002273). The results are shown in Table 7 (the values have been ordered according to P-values in variance testing). We therefore inferred that additional states were credibly valued (with identical means) and increased the precision of the final estimation (i.e., did not inflate the total variance). Hence, the extension of the number of states was justifiable.
| State | In position 6th–17th | In position 18th–23rd | P-value for equality of variance test | P-value for equality of mean test | ||
|---|---|---|---|---|---|---|
| N | Mean (SD) | N | Mean (SD) | |||
| 12223 | 97 | 0.482 (0.511) | 36 | 0.627 (0.235) | 0.0060 | 0.0278 |
| 22112 | 80 | 0.856 (0.196) | 33 | 0.704 (0.323) | 0.0062 | 0.0161 |
| 22323 | 102 | 0.279 (0.56) | 41 | 0.174 (0.694) | 0.0282 | 0.3911 |
| 11312 | 63 | 0.659 (0.439) | 50 | 0.742 (0.246) | 0.0411 | 0.2057 |
| 13212 | 69 | 0.546 (0.501) | 47 | 0.68 (0.289) | 0.0597 | 0.0702 |
| 23321 | 102 | 0.309 (0.579) | 52 | 0.202 (0.655) | 0.1040 | 0.3206 |
| 21133 | 98 | 0.154 (0.65) | 50 | 0.383 (0.591) | 0.2232 | 0.0337 |
| 32313 | 106 | 0.044 (0.625) | 40 | −0.042 (0.683) | 0.2405 | 0.4937 |
| 23232 | 81 | −0.001 (0.676) | 40 | 0.006 (0.59) | 0.2496 | 0.9547 |
| 12211 | 66 | 0.863 (0.165) | 29 | 0.839 (0.205) | 0.2651 | 0.5883 |
| 11122 | 68 | 0.834 (0.192) | 47 | 0.785 (0.397) | 0.3050 | 0.4419 |
| 33212 | 94 | 0.262 (0.604) | 40 | 0.347 (0.554) | 0.3200 | 0.4287 |
| 21111 | 80 | 0.932 (0.109) | 30 | 0.905 (0.169) | 0.3264 | 0.4206 |
| 33323 | 75 | −0.106 (0.595) | 47 | −0.19 (0.647) | 0.3708 | 0.4718 |
| 22331 | 109 | 0.081 (0.68) | 43 | 0.048 (0.626) | 0.3979 | 0.7761 |
| 22122 | 66 | 0.741 (0.369) | 49 | 0.769 (0.249) | 0.4051 | 0.6307 |
| 32223 | 80 | 0.141 (0.59) | 48 | 0.26 (0.575) | 0.4301 | 0.2632 |
| 21232 | 89 | 0.311 (0.631) | 53 | 0.144 (0.658) | 0.4500 | 0.1400 |
| 22222 | 151 | 0.671 (0.378) | 91 | 0.613 (0.437) | 0.4626 | 0.2922 |
| 32333 | 90 | −0.374 (0.574) | 49 | −0.277 (0.578) | 0.4784 | 0.3435 |
| 32232 | 87 | −0.087 (0.642) | 62 | −0.047 (0.69) | 0.4806 | 0.7222 |
| 11133 | 102 | 0.208 (0.638) | 45 | 0.169 (0.684) | 0.5035 | 0.7473 |
| 13332 | 106 | −0.1 (0.643) | 42 | 0.029 (0.7) | 0.5391 | 0.3080 |
| 11113 | 91 | 0.666 (0.403) | 47 | 0.658 (0.449) | 0.5941 | 0.9236 |
| 33232 | 88 | −0.142 (0.603) | 45 | −0.266 (0.581) | 0.6072 | 0.2538 |
| 11211 | 70 | 0.919 (0.169) | 38 | 0.93 (0.087) | 0.6145 | 0.6455 |
| 13311 | 85 | 0.51 (0.534) | 37 | 0.497 (0.484) | 0.6186 | 0.8958 |
| 21222 | 74 | 0.716 (0.263) | 35 | 0.776 (0.303) | 0.6453 | 0.3140 |
| 21312 | 86 | 0.518 (0.503) | 51 | 0.548 (0.475) | 0.6605 | 0.7319 |
| 11121 | 47 | 0.866 (0.254) | 35 | 0.851 (0.224) | 0.6729 | 0.7844 |
| 11112 | 61 | 0.932 (0.102) | 37 | 0.942 (0.122) | 0.6751 | 0.6629 |
| 12111 | 58 | 0.887 (0.194) | 19 | 0.908 (0.198) | 0.6751 | 0.6946 |
| 11131 | 81 | 0.328 (0.606) | 36 | 0.188 (0.656) | 0.7216 | 0.2815 |
| 33321 | 97 | 0.034 (0.623) | 36 | −0.073 (0.657) | 0.7608 | 0.3984 |
| 33333 | 171 | −0.357 (0.568) | 104 | −0.358 (0.515) | 0.7611 | 0.9873 |
| 21323 | 89 | 0.437 (0.545) | 43 | 0.471 (0.55) | 0.7744 | 0.7393 |
| 22233 | 91 | 0.004 (0.599) | 47 | 0.131 (0.652) | 0.8198 | 0.2675 |
| 22121 | 70 | 0.803 (0.282) | 23 | 0.793 (0.215) | 0.8251 | 0.8522 |
| 32211 | 91 | 0.471 (0.531) | 32 | 0.522 (0.558) | 0.8261 | 0.6504 |
| 23333 | 99 | −0.224 (0.638) | 36 | −0.25 (0.663) | 0.8477 | 0.8411 |
| 12121 | 62 | 0.862 (0.199) | 30 | 0.83 (0.218) | 0.9053 | 0.4989 |
| 12222 | 86 | 0.74 (0.364) | 37 | 0.712 (0.37) | 0.9309 | 0.6982 |
| 23313 | 87 | 0.149 (0.61) | 43 | 0.195 (0.599) | 0.9459 | 0.6814 |
| 32331 | 106 | −0.093 (0.622) | 37 | −0.243 (0.643) | 0.9901 | 0.2228 |
| All | 3851 | 0.326 (0.649) | 1912 | 0.33 (0.651) | 0.9487 | 0.8028 |
- The smallest value in each P-value column has been put in bold. The minimal Hölm–Bonferroni threshold amounts to 0.0023.
Comparison with Other Countries' Value Sets
In a comparison of estimated values, for all health states, it can be seen that the Polish TTO values are higher than the UK values from the MVH survey (Fig. 1a) and were similar to the German TTO values, although estimation of individual states differed (Fig. 1b). Comparison with the European and Slovenian value sets showed the classical pattern of differences between TTO and VAS values, with Polish TTO values higher for the better health states and lower for the worse health states (Fig. 1c,d).
Graphical comparison of Polish EQ-5D TTO value set versus: (a) UK TTO, (b) German TTO, (c) European VAS, and (d) Slovenian VAS value set. TTO, time trade-off; VAS, visual analog scale.
Table 8 presents a statistical summary of cross-country comparisons. Polish health state values correlate significantly with UK values from the MVH A1 value set (R2 = 0.90), but at the same time, mean absolute difference between Polish and UK values is the largest (0.245, compared with 0.117 between Polish and German values).
| Polish TTO | UK TTO (MVH A1) | German TTO | European VAS | Slovenian VAS | |
|---|---|---|---|---|---|
| Constant | 0.049 | 0.081 | 0.001 | 0.1279 | 0.128 |
| MO2 | 0.052 | 0.069 | 0.099 | 0.0659 | 0.206 |
| MO3 | 0.331 | 0.314 | 0.327 | 0.1829 | 0.412 |
| SC2 | 0.054 | 0.104 | 0.087 | 0.1173 | 0.093 |
| SC3 | 0.235 | 0.214 | 0.174 | 0.1559 | 0.186 |
| UA2 | 0.046 | 0.036 | 0.0264 | 0.054 | |
| UA3 | 0.212 | 0.094 | 0.0860 | 0.108 | |
| PD2 | 0.057 | 0.123 | 0.112 | 0.0930 | 0.111 |
| PD3 | 0.489 | 0.386 | 0.315 | 0.1637 | 0.222 |
| AD2 | 0.026 | 0.071 | 0.0891 | 0.093 | |
| AD3 | 0.207 | 0.236 | 0.065 | 0.1290 | 0.186 |
| N3 | 0.269 | 0.323 | 0.2288 | ||
| Mean absolute difference | 0.245 | 0.117 | 0.167 | 0.146 | |
| No. (of 243) > 0.05 vs. Polish | 238 (98%) | 178 (73%) | 210 (86%) | 193 (79%) | |
| No. (of 243) > 0.10 vs. Polish | 219 (90%) | 122 (50%) | 178 (73%) | 147 (60%) | |
| R 2 vs. Polish | 0.90 | 0.81 | 0.74 | 0.73 |
- MVH, Measurement and Value of Health; TTO, time trade-off; VAS, visual analog scale.
Discussion
In this study, we performed 321 face-to-face TTO interviews, directly measuring population values for 44 EQ-5D health states and estimated the Polish EQ-5D value set using random effects modeling.
The Polish EQ-5D valuation study differed in several aspects from the original UK MVH protocol. First, we reduced the ranking exercise to 10 health states and nearly eliminated the VAS valuation of ranked health states (four states) while increasing the number of health states to 23 valued in the TTO exercise. To date, this is the highest number of health states per respondent valued in a national valuation study. Second, similar to valuation studies performed in The Netherlands [10] and in Japan [11], predetermined health state sets were used; yet, in contrast to the before-mentioned studies, there were two sets, not one. Third, we developed a detailed description of the TTO exercise in a separate instruction booklet, leaving the graphic form of recording the TTO results in the protocol book. Additionally, we confirmed the results of random effects modeling in random parameters modeling (using the Bayesian approach), proving that heterogeneity of the surveyed population had no influence on the results.
The limited number of respondents can be regarded as a potential weakness of our study, although the size of the sample group is similar to the sample size in the German [15] and Dutch [10] valuation studies. The sample size proved to be sufficient in obtaining statistically significant model coefficients. Another important weakness of our study may be lack of formal random allocation of responders to our study population. The representativeness of the sample was controlled with respect to two characteristics—age and sex—through quotas. Of course the sample is not representative with respect to other features: education, geography, or having a relative/friend inpatient. It is difficult to state a priori whether this characteristics impact the preferences among health states. On one hand, it can be argued that it can impact the preferences depending on the relative's illness; on the other, it may increase the awareness of the respondent and improve the quality of the results. Although patient relatives may not be an ideal population for valuation exercise, they were interviewed in earlier valuation studies in Spain and Argentina [20,21]. Our highly educated study sample, although not representative for general population, may better comprehend and perform TTO exercise, making the results more consistent and credible, and actually improve goodness of fit. Nevertheless, it may be the case that the highly educated differ from worse educated in the preferences among health states either directly or indirectly via another characteristics (e.g., increased wealth). In this case, our results would be biased. Because there were no previous health surveys using EQ-5D in Poland, it is unlikely to assess the representativeness and compare our study sample with the Polish general population considering health status data. This line of research should be pursued in the following studies.
Finally, by using fewer interviewers, we would probably have improved the quality of data, but for practical reasons, we were unable to reduce their number below 10. Despite these limitations, goodness-of-fit analysis proved that direct utilities from surveys fit the model-derived utilities relatively well with R2 of 0.45, which is close to 0.46 obtained in the MVH study [9].
Moreover, we believe that the present study supports the use of more health states in TTO experiments than previously thought. We found no statistical differences neither in mean nor in variance between valuations when a given health state was valued as 6th–17th in a row or 18th–23rd. Therefore, there is no risk of a bias or efficiency decrease in the estimation. This finding provides evidence for improving the efficiency of valuation protocols and supports the estimation of national value sets in other countries, including Central and Eastern Europe.
In the Polish version of EQ-5D, as in the Dutch and Italian versions, the wording chosen for the third level of “mobility,”“confined to bed,” implies being bedridden. Some of the authors were concerned that this could have caused the Polish values to be lower for health states that included the third level of “mobility”[7]. In comparing dimensions (based on level 3 coefficients and assuming no other problems), individuals in Poland, similar to those in the United Kingdom and Zimbabwe, valued problems associated with “pain/discomfort” as the worst and not the problems associated with “mobility,” as in Argentina, Denmark, Germany, Japan, Spain, the United States, and Spanish-speaking Hispanic US residents [2]. In Poland, similar to Argentina and the Spanish-speaking Hispanic US residents, the “anxiety/depression” domain was ranked as the least important [21,22]. This differs from other countries where the “usual activities” domain was judged the least important.
Conclusions
This is the first EQ-5D value set based on TTO in Central and Eastern Europe so far. Because the values differ considerably from those elicited in Western European countries, its use should be recommended for studies in Poland. Increasing the number of health states that each respondent is asked to value using TTO seems feasible and justifiable.
Acknowledgments
We thank Paul Kind, Aki Tsuchiya, and Benjamin Craig for useful comments and guidance. We also thank Rosalind Rabin for editorial support.
Source of financial support: This study was sponsored by unrestricted grants from AstraZeneca Pharma Poland, GSK Commercial, and Pfizer Poland.
