Head-to-head comparison between the EQ-5D-5L and the EQ-5D-3L in general population health surveys

Health-related quality of life has been gaining importance in research, clinical practice and health planning [1, 2] by providing complementary information to health indicators based on morbidity and mortality. This is especially relevant to describe health in developed countries, where life expectancy has been increasing steadily after their epidemiological transition. Evaluating the general population’s health is one of the specific applications proposed for health-related quality of life instruments [3].

The EQ-5D has been frequently selected for national health surveys [4‐10], given its low respondent burden and its consistently proven metric properties [6, 11, 12] . However, the high percentage of individuals with the best health state in the EQ-5D [13, 14] has been repeatedly highlighted as a limitation, since this may reduce its capacity to discriminate within good health [6, 15, 16], and its responsiveness in some health areas [17‐19]. The traditional EQ-5D descriptive system, composed of five dimensions (mobility, self-care, usual activities, pain/discomfort, and anxiety/depression) with three levels of severity, defines 243 distinct health states [20] resulting from all the possible combinations (i.e., 3⁵). This is a very low number compared with other instruments, such as the Health Utilities Index [21] with 972,000 or the SF-6D [22] with 18,000 possible combinations.

To improve its discriminative capacity and sensitivity to change, and to reduce ceiling effects, the EuroQol Research Foundation decided to develop a new EQ-5D version increasing the number of response options from three (EQ-5D-3L) to five levels (EQ-5D-5L), resulting in 3125 health states (i.e., 5⁵). Face and content validity of the new EQ-5D-5L were demonstrated for both the English and Spanish versions through focus group research [23]. Studies performed in cancer [24, 25], hepatitis B [26], or hip arthroplasty [27] patients showed improvements for discriminative capacity [24, 26], construct validity [24‐26], and responsiveness, without diminishing its reliability [25], as well as a large decrease in the percentage of individuals with the best health state.

Given the recent development of the EQ-5D-5L, there are still few head-to-head studies in general population comparing its metric properties with the traditional 3 L version. Studies carried out in South Korea [28], Alberta (Canada) [29], England [30] and Lombardy (Italy) [31], mainly based on national health surveys, examined both versions of EQ-5D in general population. Yet the South Korean study published in 2013 [28] was performed only in a small sample (n = 600), neither the Canadian [29] nor the English health surveys [30] administered both versions together, while the Italian survey did, but without comparing them. The decrease in the percentage of individuals with the best health state varied in these studies, from 42.1 to 32.3% in Alberta [29], from 56.2 to 47.6% in England [30], from 43.9 to 38% in Lombardy [31], and from 65.7 to 61.2% in South Korea [28]. The aim of this study is a head-to-head comparison to assess to what extent expanding the number of levels in the EQ-5D from three to five has improved its distribution, discriminatory power, and validity in the general population.

Of the 9658 participants in the CHIS between January 2011 and December 2012, 7554 individuals aged 18 to 102 years old were analyzed after excluding 2104 people younger than 18. Mean age of participants was 47.1 (SD = 18.9), and 50.9% were female (Table 1). More than half had completed secondary studies, 40% belonged to social class IV, and 48.5% suffered three or more chronic conditions. Only 15% of the individuals reported some limitation of activities in the previous 6 months, and 34.3% claimed to have either excellent or very good perceived health (Table 1).

Cross tabulations of responses to both EQ-5D versions (Table 2) showed that most of the participants reporting no problems in the 3L version remained at Level 1 in the 5L version, and only 1–2% moved to slight problems. In contrast, a large proportion of individuals reporting extreme problems in the 3L version had moved to severe problems (Level 4) in the 5L version. This proportion was particularly marked for pain/discomfort (75.5%) and anxiety/depression (66.4%). Grey cells show the pairs previously defined as inconsistencies. The number of inconsistencies was highest in the pain/discomfort domain (n = 189; 2.4%) and lowest in the self-care one (n = 54; 0.6%). The average proportion of inconsistencies by dimension was 0.9%.

Regarding the validity of the redistribution between three and five levels, the mean of the perceived health VAS was over 75 in the subgroup of individuals reporting no problems in both versions for all dimensions (range 75.4–79.7). Confirming the hypothesis established a priori, the perceived health VAS mean in subgroups of individuals selecting an EQ-5D-5L category of more severe problems is worse than in those remaining in the same category as in the EQ-5D-3L. Similarly, those moving to milder problems in the EQ-5D-5L presented better perceived health. For example, in the last row of Table 2 (extremely anxious or depressed in the EQ-5D-3L), the 66.4% who moved to a milder level in the 5 L (severe problems) presented better perceived health than those who remained at the extreme level (11.5%): mean VAS of 41.7 vs. 29.5.

Figure 1 shows Shannon indices of EQ-5D-3L and EQ-5D-5L. The maximum information captured by the system (H’max in light bars), and also the absolute informativity (H′ in dark bars) is higher in 5 L than in 3 L version. However, when H′ is compared with the H’max, the relative information area captured (J’) is significantly lower in EQ-5D-5L than in 3 L for all dimensions except self-care. This difference is especially marked in pain/discomfort (J’ = 0.59 vs. 0.68) and anxiety/depression (J’ = 0.42 vs. 0.50).

Figure 2 shows the plot between EQ-5D-3L and EQ-5D-5L indices. The cloud of points and the biggest clusters of individuals were concentrated around the perfect agreement diagonal, indicating a high correlation between both indices. A slight deviation to higher values with the EQ-5D-5L than the EQ-5D-3L is also observable.

Table 3 shows the statistics describing distribution of the EQ-5D indices. Ranges observed in our sample matched exactly with the theoretical ranges (from −0.65 to 1). The proportion of individuals with the worst health state was negligible (< 0.15%), while the proportion with the best health state was 61.8% with EQ-5D-3L and 60.8% with EQ-5D-5L. Means (SD) were 0.87 (0.25) and 0.89 (0.22) for EQ-5D-3L and EQ-5D-5L. Sensitivity analysis performed with the EQ-5D-5L index constructed with the newly developed Spanish value set [41] (see Additional file 1) showed consistent results: mean 0.90 (SD = 0.19) in the entire sample, and mean 0.92 (SD = 0.14) after excluding the 249 subjects with negative values. Differences between EQ-5D-3L and EQ-5D-5L indices remained quite stable regardless the value set used.

Figure 3 shows results by each specific chronic condition: the proportion of individuals with the best health state (11111) in the EQ-5D-3L (blue bars) and EQ-5D-5L (green bars), and also the mean (95% CI) of perceived health VAS in subgroups with and without the best health state. In both indices, chronic allergies presented the highest proportion of subjects with the best health state (50.6 and 50.1%), and urinary incontinence the lowest (13.1 and 12.0%). Regardless of the index used, the proportion of individuals with the best health state was ≤ 30% in half of the chronic conditions from the checklist (cervical pain, tumors, arthrosis, arthritis or rheumatism, peptic ulcer, poor circulation, other health illnesses, cataracts, myocardial infarction, chronic constipation, anxiety or depression, other mental disorders, stroke, osteoporosis, and urinary incontinence). The mean of the VAS for the subgroup with the best possible health state defined by EQ-5D-3L and EQ-5D-5L (in dark blue and green lines, respectively) was over 70 within all specific chronic condition groups, ranging 71.3–79.8 and 72.6–81.3, respectively. Perceived health VAS means in the subgroups defined by the EQ-5D-3L were very similar to those obtained in the subgroups defined by EQ-5D-5L. For the subgroup with some health problem (not 11111), mean of VAS was always lower than 60 (light blue and green lines).

This head-to-head comparison of EQ-5D-5L with EQ-5D-3L in the general population of Catalonia shows that redistribution of levels is mostly in individuals reporting extreme problems on the EQ-5D-3L, which moved to level 4 on the EQ-5D-5L, but not for those reporting no problem, who remained at the top. This explains the very small reduction in the percentage of individuals with the best health state, from 61.8% with EQ-5D-3L to 60.8% with EQ-5D-5L, and the increment of the index mean (from 0.87 to 0.89) in our sample.

One of the original contributions of this study is that, as far as we are aware, this is the first time that distribution and validity of the EQ-5D-5L have been compared head-to-head to those of the EQ-5D-3L in a health survey on general population. In the Lombardy study both versions were also administered, but they were not compared since the publication was focused on reference norms [31].

Our study has some limitations. Firstly, the two versions of the EQ-5D were always administered in the same order: first the 3 L and then the 5 L. This proximity might have affected the EQ-5D-5L, which was always administered second. However, the comparison with the 6th wave (see Additional file 2), where only the EQ-5D-5L was administered, showed no differences in EQ-5D-5L dimensions, except for pain/discomfort (72.4% versus 67.6% of individuals reporting no problem, p = 0.003). This finding indicates that the fact of administering the two versions together did not modify the response to the EQ-5D-5L when administered alone (as in the 6th wave). Furthermore, results from the 2011 National Health Survey of Spain (62.4% of individuals with the best health state) where only the 5 L version was administered [42] also support our EQ-5D-5 L findings in Catalonia. Secondly, an interviewer bias may have played a role, and this could be differential for those dimensions where the wording of the response option had been modified in the EQ-5D-5L. For example, in the extreme of mobility (“confined to bed” for the EQ-5D-3L versus “unable to walk about” for the EQ-5D-5L), interviewers might have attenuated the severity. Finally, our sample is only representative of Catalonia. However, given the similarities in national indicators such as life expectancy or healthy life years in the general population of Catalonia, Spain, and other European regions [43], it is likely that our results will be generalizable to similar developed countries.

The small reduction observed in the percentage of individuals with the best health state, from 61.8% with EQ-5D-3L to 60.8% with EQ-5D-5L, is due to the negligible movement from level 1 out of 3 (“no problem”) to level 2 out of 5 (“slight problems”) in all dimensions. This absolute reduction of 1% (relative reduction of 1.6%) in the proportion of individuals with the best health state was lower than that reported in the population of South Korea and Lombardy (absolute reductions of 4.5, and 5.9%, respectively) [28, 31]. The Canadian and English studies [29, 30] reported greater differences of 9.8 and 8.6%; but as previously remarked, they were not head-to-head comparisons, so this could be explained by other reasons related to the study design, rather than to differences between EQ-5D versions.

This is the first time that redistribution of a large proportion of individuals from extreme to severe problems has been reported in a general population. Depending on the dimension, between 18.3 and 75.7% of individuals reporting extreme problems in the 3 L version moved to level 4 (severe problems) in the 5 L one. The better perceived health in this latter subgroup (VAS mean over 40 in most domains), compared with the subgroup remaining in extreme problems (VAS mean ranging from 29.5 in anxiety/depression to 36.5 in self-care), supports the validity of the redistribution phenomenon observed in the side of the EQ-5D descriptive system indicating poor health. This may indicate that the EQ-5D-5L can measure the health state of individuals with severe (but not extreme) health problems in the Catalan general population better than the EQ-5D-3L. This partly explains why the index mean of the new version was higher (0.89) than that obtained with the traditional version with three levels (0.87). Due to its small sample size (N = 600), the South Korean study could not observe this redistribution because there were too few participants on level 3 of EQ-5D-3L (0–6 individuals) [28], while the Italian study did not assess the redistribution [31]. It is important to highlight the low average proportion of inconsistencies between both EQ-5D versions in our study (0.9%), which was comparable to the South Korean general population (1.1%) [28], and lower than that reported among patients with cancer (3.5%) [25] or with chronic conditions (2.9%) [39].

As expected, extending the EQ–5D descriptive system from three to five levels resulted in significantly higher absolute, but slightly lower relative (evenness) discriminatory power. J’ values have also been found slightly lower in some dimensions of EQ-5D-5L in previous comparative studies [37, 39, 40]. The absolute and relative informativity of both EQ-5D versions in our study (0.36–1.37 and 0.21–0.68, respectively) were similar to those reported by Pattanaphesaj et al. [40] (0.12–1.40 and 0.08–0.63), but lower than those observed in others [37, 39]. The relatively good health of people from the Catalan general population could partly explain the lower absolute informativity observed in our study.

The difference observed between EQ-5D-3L and EQ-5D-5L indices for medians and means (SD) merits a comment. The EQ-5D-5L index presented a slightly higher median and mean, but a reduced SD compared with the EQ-5D-3L index. Since the crosswalk 3 L–5 L value set applied to calculate the EQ-5D-5L index had been derived from that originally developed for the 3 L version, these differences may be mainly explained by the increment in the number of levels. For this reason, it is recommendable that national health surveys using the EQ-5D-3L that decide to replace it with the EQ-5D-5L maintain both versions, at least in a random subsample, for a temporary period. Results in these subsamples will allow anchoring results of the two versions, in order to take into account the version effect and correctly monitor the evolution of health along time. Otherwise, changes observed when monitoring populations could be mistakenly attributed to health worsening/improvement instead of measurement differences between versions.

The most prevalent chronic conditions in this sample were low back pain (30%), arthrosis, arthritis or rheumatism (27.8%), and high blood pressure (25.6%), while stroke was the least prevalent with a rate of 2.4% (data not shown). Contrarily to the a priori hypothesis, both EQ-5D versions had an almost identical validity measuring health in individuals who self-reported chronic conditions and with the best health state. This unexpected result is probably explained by the very similar percentage of individuals with the best health state within each specific chronic condition, regardless of the EQ-5D version. Although larger reductions in this percentage were reported in studies of specific conditions such as hepatitis B (21.6 to 16.7%) [26] and surgery patients (30 to 18%) [27], the decline observed in the groups with specific chronic conditions within our sample was ≤3% in all cases. This difference could be due to self-reporting instead of clinical diagnoses.

The increase of levels provided by the EQ-5D-5L contributed to the validity and discriminatory power of this new version. The group of individuals with poor health was redistributed into different severity levels, while in the EQ-5D-3L they were stuck in the category of extreme problems. The proportion of individuals with the best health state is still very high in the EQ-5D-5L. Nonetheless, results of perceived health VAS support validity of the observed redistribution. Furthermore, consistency between both EQ-5D versions and with results from the 2011 Spanish National Health Survey enhance the reliability of responses from this subset of general population in good health.

Our findings support the validity and discriminatory power of the new EQ-5D-5L for health measurement of the general population. However, it would be advisable to maintain both versions in parallel for a temporary period when introducing the new EQ-5D-5L to a national health survey currently using the EQ-5D-3L version in order to establish an anchor.