FormalPara Key Points for Decision Makers

This review supports the use of both the 3L and the 5L in a broad range of patients, populations and countries.

The 5L showed better or at least similar measurement properties when compared to the 3L.

Evidence on responsiveness is inconclusive and requires further research.

1 Introduction

The EQ-5D is one of the most widely used instruments to describe and value health [1, 2]. It is a generic, self-completion questionnaire consisting of two parts: a 5-item descriptive system and a thermometer-like visual analogue scale ranging from 0 to 100 (the EQ-VAS). It comprises five items, each describing one dimension [Mobility (MO), Self-Care (SC), Usual Activities (UA), Pain/Discomfort (PD), and Anxiety/Depression (AD)]. The original questionnaire, introduced in 1990 [3], allows respondents to choose between three options; level 1, representing no problems; level 2, reflecting small or moderate problems; and level 3, indicating extreme problems (or ‘unable to’). Self-ratings on the three levels in the five dimensions (items) can be summarized to produce 243 health states, also known as a health profile. Health profiles can be assigned index values derived from econometric techniques to elicit societal preference weights. These index values can then be used in economic evaluation of health programs [1, 4].

The EQ-5D was conceptualized to capture deviations from ‘normal’ health, thereby focusing on self-reported health and health-related quality-of-life problems while not attempting to capture aspects beyond health. Internationally, it is currently one of the most widely used preference-based quality-of-life questionnaires. A large body of literature demonstrates that the instrument is valid and reliable [5,6,7]. However, although the EQ-5D was developed to supplement other instruments, this simple and short measure has been increasingly used as a ‘stand-alone tool’ [8, 9]. The increase in use of the EQ-5D in the field of health technology assessment raises concerns about methodological measurement issues [10]. The first is the EQ-5D’s ceiling effect, or a high proportion of participants reporting ‘no problems’ on one or all dimensions [11,12,13]. Second, some studies found the EQ-5D to be less responsive to changes in health compared with other preference-based measures [e.g. Health Utility Index (HUI), Short-Form 6-Dimension (SF-6D), Quality of Well-Being Scale–Self Administered (QWB-SA)] [14,15,16,17,18,19,20,21]. To address these concerns, paired with the inherent aspiration of the ever-expanding research community to continually improve the instrument, a new version of the EQ-5D was developed by the EuroQol group [22,23,24]. The new version expanded the response choices from three to five levels and changed the wording of some of the response categories (Table 1). The new version is called the EQ-5D-5L [25], and can describe 3125 (= 55) health conditions.

Table 1 Response levels of the EQ-5D-3L and EQ-5D-5L
Full size table

Since introducing the EQ-5D-5L in 2011, many studies have comparatively investigated the measurement properties of the original EQ-5D (now interchangeably referred to as EQ-5D-3L or 3L) and the newer EQ-5D-5L (now interchangeably referred to as EQ-5D-5L or 5L). In this review we summarize the consolidated findings from these studies.

2 Methods

2.1 Data Sources, Search Strategy, Study Selection, and Inclusion and Exclusion Criteria

We conducted a systematic literature search to identify all studies in the English and German languages comparing the 3L and the 5L published between January 2007 and May 2016 using the following keywords: ‘EQ-5D-5L’, ‘EQ-5D 5L’, ‘EuroQol AND 5L’, ‘EuroQoL AND 5 level’. Electronic searches were performed in the PubMed, EMBASE, and PsycINFO (EBSCO) databases, in addition to the EuroQol Research Foundation website, for relevant publications. The inclusion criteria were primary study or conference paper comparing the final versions of the 3L and the 5L(studies using experimental versions were excluded). Articles were further excluded if they did not assess the EQ-5D, were of another publication type, it was not an empirical study in adults, were not in English or German, or were not available in full text. The review was updated during the process of manuscript revision using the same search algorithms, and inclusion and exclusion/eligibility criteria as detailed above. The search was conducted in articles published between May 2016 and January 2018. The process of study selection is shown in Fig. 1.

Fig. 1
figure 1

Literature search and selection process

Full size image

2.2 Screening and Data Extraction

Two researchers independently reviewed the title and abstract of all identified studies, while a third reviewer (TK or MFJ) was consulted in case of variance. After removing duplicates, full-text articles were reviewed by one reviewer (IB) and doubled-checked by the second reviewer (YSF) for missing extractions. For cases of papers that used the same data, those with more information on the indicators of interest were always included. When publications addressed different information based on the same data, both papers were included. For each article, the following information was extracted: authors, title and year of publication, sample characteristics (e.g. sample size, percentage of females, mean age), country, outcome measures used, aims of the study, study design, and parameters describing relevant measure properties. The measurement properties were distributional properties, informativity, inconsistencies, responsiveness and test–retest reliability. All of these properties were assessed in terms of results related to the descriptive systems of the 3L and 5L. For responsiveness and test–retest reliability, results on index values were also included.

2.3 Quality Assessment of Studies

The quality of the full-text articles included for review was assessed using a 9-item critical appraisal tool (see the electronic supplementary material [ESM]). The items were defined based on the ‘Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies’ from the National Heart, Lung, and Blood Institute (NHLBI) [26], and contained (1) objective/research questions; (2) study population; (3) groups recruited/eligibility criteria; (4) study design; (5) sample size; (6) response rate; (7) data collection; (8) outcome parameter; and (9) statistical tests/analysis. Study quality was assessed as excellent, good, fair or poor, with the corresponding number of criteria fulfilled being 8–9, 5–7, 3–4 or 0–2.

2.4 Distributional Properties

We compared the 3L and 5L on their classical distribution characteristics, such as the number and proportion of missing values, the number and percentage reporting the best (ceiling; ‘no problems’) and worst (floor) level of health state in each dimension and across all dimensions (e.g. ‘11111’ and ‘33333’ for the 3L or ‘55555’ for the 5L, respectively). The results are presented as ranges of percentages or numbers of studies in which the 5L performed equal to, worse than, or better than the 3L (e.g. How often did more than 15% [as suggested by Terwee et al. [27]] of the study sample report ‘no problems’ when using the 5L compared with the 3L?) We used random effects logit transformation to calculate pooled proportions from single proportions using R’s ‘meta’ package specifically for proportion reporting ‘no problems’ across all dimensions (‘11111’). Pooled proportions give an idea of the overall ceiling effect when taking into account the sample sizes across included studies.

2.5 Informativity

Shannon’s index is based on information theory and allows an assessment on the informational and discriminatory power of each descriptive system.

According to Shannon’s indices, an item is most efficiently used when all responses are evenly distributed across all response options [28], with a higher index indicating more information captured by the instrument. While H′ represents the extent to which the information is evenly distributed across all categories, Shannon’s J′ additionally takes into account the number of response options or descriptive categories of the measurement system. J′ can take values between 0 and 1, with a J′ of 0 representing that all responses are concentrated in one response level (most uneven distribution; worst discriminatory power) and 1 representing that all response levels are evenly distributed (even distribution; best discriminatory power). There is no straightforward interpretation for H′. Since H′max depends on the number of levels, within our context H′ can take values between 0 (no informational richness/discriminatory power) and 1.58 (log2L, with the number of levels L = 3) for the 3L and 2.32 (log2L, with L = 5) for the 5L (which corresponds to the highest informational richness/discriminatory power).

Within this review, H′ and J′ were extracted from the studies or calculated using these formulas, where p i is the proportion of responses in the ith response option:

$$H^{\prime} = - \mathop \sum \limits_{i = 1}^{L} \left( {p_{i} \log_{2} p_{i} } \right)$$
$$J^{\prime} = \frac{H'}{H'\hbox{max} }$$

Both indices are reported for each EQ-5D dimension. We aggregated the mean information gain by the 5L, which was calculated through dividing Shannon’s H′ of the 5L by Shannon’s H′ of the 3L (H′5L/H′3L) and Shannon’s J′ for the 5L by Shannon’s J′ for the 3L (J′5L/J′3L), respectively, with H′/J′ ≥1 showing the 5L descriptive classification system to be more informative than the 3L.

2.6 Inconsistencies

Due to the two additional response levels, we expect a redistribution that can be quantified with the help of the parameters already described (i.e. classical distribution properties on the one hand and Shannon’s indices on the other). In order to assess whether this redistribution of responses is conclusive in terms of content, we also considered inconsistent responses, as defined by Janssen et al. [8], as a qualitative distribution parameter, or if, and to what degree, 3L and 5L response pairs differ from each other. Operationally, we (1) transformed the 3L response levels 1, 2 and 3 to 5L response levels 1, 3 and 5 to calculate (2) the size of difference of corresponding responses. Paired responses differing more than one level were defined as ‘inconsistent’, with a size of inconsistency ranging from 1 to 3. All studies included in the review used the methods of Janssen et al. [8] to calculate inconsistencies.

We report and compare the percentage of inconsistencies by dimension, the range of percentage of inconsistencies by dimension, and the total number and average of inconsistencies. Notwithstanding the fact that the mere presence of inconsistent responses does not provide any information about the underlying causes, their consideration is of particular interest when they occur systematically, e.g. only in certain patient groups, which could affect validity, responsiveness and reliability.

2.7 Responsiveness

To evaluate how the instruments capture changes in health over time, we collected all reported distribution-based effect sizes (ESs), such as the standardized ES and the standardized response mean (SRM), and non-parametric test statistics, such as the Wilcoxon signed-rank order test or the probability of superiority (PS) as defined by Grissom and Kim [29]. The ES is the mean change divided by the standard deviation of the baseline measurement. It disregards the variation in change which is considered by the SRM (the ratio of the mean change to the standard deviation of the change). The Wilcoxon test is the non-parametric equivalent of the t-test for dependent samples and is applied when the prerequisites for a parametric procedure are not met.

2.8 Test–Retest Reliability

Several methods can determine whether a measurement tool consistently produces the same results if the attribute of interest remains stable [30, 31]. We extracted and summarized any reported information regarding the magnitude of agreement of data collected at two points in time: intraclass correlation coefficients (ICCs), Cohen’s Kappa (κ), weighted Kappa (wκ), and percentage of agreement (POA).

An ICC quantifies the dependency of interval-scaled data pairs if the order of measurement is negligible. Values range from -1 to 1 with values less than 0 indicating a reliability of 0 and values higher or equal to 0.70 indicating good reliability [32].

κ is the most widely used measure to assess the agreement for categorical data [33]; it measures the random corrected degree of agreement between two ratings. In contrast to the simple percentage of agreement of two ratings, it considers that ratings will sometimes agree or disagree by chance. When additionally taking into account the size of the deviation (one vs. several categories) within ordinal-scaled data (such as the EQ-5D responses), calculating wκ is indicated [34]. Kappa is 1 if two ratings perfectly match, and 0 when agreement equals chance. Kappa is negative if the match is poorer than chance [35]. Note that a wκ using quadratic weights is one type of ICC. Based on the Guidelines for Reporting Reliability and Agreement Studies (GRRAS), a κ > 0.40 and ICCs >0.6 were considered acceptable [30].

3 Results

Of the 497 studies identified from the search, 215 were selected for full-text review based on title and abstract screening. Of those, 190 did not meet the inclusion criteria and were excluded. The remaining 20 articles that compared methodical properties of the official versions of the 3L and 5L were included in the review (Fig. 1 [36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59]). An update carried out in the course of the manuscript revision resulted in a further four hits, therefore the final review is based on a total of 24 articles. All papers were of good to excellent quality (see the ESM).

The sample size of the included articles ranged from 50 to 7294 for the 3L, and 50 to 6800 for the 5L (Table 2). Data were collected in 18 different countries in the following settings: general population (8 studies) and patient populations (16 studies). All but two studies directly compared the 3L and the 5L (head-to-head, i.e. the same respondents completed both the 3L and 5L questionnaires). In head-to-head comparison studies, the 5L was administered before the 3L (Table 3). Two of the crossover studies (i.e. studies that randomized the administration order of the 3L and 5L) reported that order of administration had no influence on response trends [37, 43].

Table 2 Characteristics of the studies included in this systematic review
Full size table
Table 3 Study design and type of questionnaire administration of the studies included in this systematic review
Full size table

3.1 Missing Values and Distributional Properties

Fifteen studies reported missing values below 5% for both 3L (range for the dimensions: 0–1.9%; range for the profile: 0–6.6%) and 5L (range for the dimensions: 0–1.6%; range of the profile: 0–4.0%). One study found 8.5% left the 5L blank and 0.8% left the 3L blank entirely, which is probably due to the methodology of how the 3L is first presented in this study [52]. Floor effects by dimension were reported in 19 studies and were almost always below 5% (3L: 0–26.1%; 5L: 0–6.5%) [Table 4]. Mean absolute reduction in floor effects ranged from 0.16 percentage points (Usual Activities) to 4.18 percentage points (Pain/Discomfort). For the profile, floor effects ranged from 0 to 2.7% for the 3L and 0 to 1.8% for the 5L (five studies).

Table 4 Results of the floor and ceiling effects
Full size table

All studies reported information on the number or proportion reporting ‘no problems’ in any dimension or for the profile (‘11111’) [Table 4]. The percentage reporting ‘no problems’ ranged from 2.3 to 99.5% for the 3L and from 0.6 to 99.5% for the 5L. Using the 5L could reduce ceiling effects up to 16.9 percentage points (Mobility) to 30.0 percentage points (Self-Care). The highest absolute reduction of ceiling effects was found for Self-Care (−1.3 to 30.0 percentage points), followed by Pain/Discomfort (1.5 to 20.0 percentage points), and Anxiety/Depression (−3.4 to 19.7 percentage points). Regarding the profile, full health state profiles were reported for 0.6 to 68.0% of the samples studied with the 3L, compared with 0 to 55.0% of the samples studied with the 5L (Fig. 2).

Fig. 2
figure 2

Ceiling for the profile (‘11111’) compared with the 3L and the 5L. f.-up follow-up

Full size image

Figure 3 presents the pooled ceiling effects (proportion reporting ‘11111’) for studies using patient (16 studies) and population (8 studies) samples. The pooled proportion of ceiling in the patient population was 0.23 [confidence interval (CI) 0.170–0.296] for 3L and 0.18 (CI 0.131–0.238) for 5L. Furthermore, the pooled proportion of ceiling in population-based studies was 0.53 (CI 0.474–0.593) for 3L, compared with 0.43 (CI 0.369–0.492) for 5L. The pooled proportions did not change substantially when excluding the two studies that did not use direct head-to-head comparisons (3L = 0.55, CI 0.472–0.623; 5L = 0.44, CI 0.367–0.517).

Fig. 3
figure 3

Ceiling for the profile by sample type: forest plot with study proportions, pooled proportions, and 95% CI of reporting ‘11111’ of the EQ-5D-3L against the EQ-5D-5L. CI confidence interval, P proportion, N sample size, THA total hip arthroplasty, UK United Kingdom, US United States

Full size image

3.2 Informativity

Fourteen studies provided information on Shannon’s H′ and Shannon’s J′. In general, Shannon’s H′ was always higher in the 5L than in the 3L, and Shannon’s J′ was higher for the 5L than the 3L, in all but five studies. Across all studies and dimensions, mean Shannon’s H′ ranged from 0.72 to 1.43 for the 5L and from 0.47 to 0.98 for the 3L (Fig. 4). Mean information gain for Shannon’s H′ (H′5L/H′3L) ranged from 1.44 for Anxiety/Depression to 1.65 for Mobility. Shannon’s J′ differences between the 3L and the 5L were marginal (Fig. 4), with a mean information gain (J′5L/J′3L) ranging from 1.02 for Self-Care to 1.16 for Mobility.

Fig. 4
figure 4

Shannon’s H′ and J′ for the 3L and the 5L

Full size image

3.3 Inconsistencies

Eighteen studies provided information on inconsistencies. The total number and proportion of inconsistencies were, with four exceptions, well below 5%, ranging from 0 to 10.6% across 18 studies (Fig. 5). The most inconsistencies were reported for Usual Activities (mean percentage 4.1%), whereas the lowest number of inconsistencies was found for Mobility (2.5%). The total proportion of inconsistencies was lowest (range 0–5.4%) in healthy and chronic populations and highest (range 6–10.6%) in orthopedic settings (Fig. 5).

Fig. 5
figure 5

Percentage of inconsistencies by dimension and overall. THR total hip replacement

Full size image

3.4 Responsiveness

Of the three studies analyzing responsiveness, two studies examined the index-level utility scores (using preference-based weights) [42, 45], whereas one study assessed responsiveness on the dimensional-level using percentage of improved, stable and deteriorated patients, and PS, a measure defined by Grissom and Kim [29, 37]. Distribution-based ES measures were only included in one of these studies [42]. In this longitudinal cohort, stroke patients were classified into three groups of improved, stable and deteriorated patients based on two external criteria: the Barthel Index and the modified Rankin Scale. Although both the 3L and the 5L were responsive, showing moderate ES and SRM, the 5L appeared to be (slightly) less responsive than the 3L but more responsive than the EQ-VAS. The other two studies overall found better responsiveness for the 5L compared with the 3L when using non-parametric test statistics in populations of liver disease patients and inpatient rehabilitation patients (Table 5) [37, 45]. Importantly, the two studies [42, 45] that analyzed the EQ-5D-5L on the index-level estimated index values using the crosswalk method, which maps 3L preference weights onto the 5L responses, which should be considered when interpreting these results.

Table 5 Evidence of studies reporting on responsiveness for the indices or on dimension level
Full size table

3.5 Test–Retest Reliability

Six articles studied the reproducibility of the EQ-5D measure, with all but one specifying two or more measures of agreement. ICC was used in all six studies—wκ and POA in three studies, and Kappa in two studies. The time interval between repeated measurements varied from 1 to 3 weeks (Table 6). When using ICC, the studies reported moderate to excellent reproducibility for both 3L and 5L index scores, with ICC ranging from 0.52 to 0.83 for the 3L and from 0.69 to 0.93 for the 5L. When using unweighted Kappa, studies reported good to very good agreement (κ3L = 0.39–0.93, κ5L = 0.36–0.98, mean κ3L = 0.692, mean κ5L = 0.678), while studies using wκ statistics found mostly fair to moderate agreement (wκ3L = 0.31–0.70, wκ5L = 0.33–0.69, mean wκ3L = 0.527, mean wκ5L = 0.541). There is no clear pattern of better reliability for either the 3L or the 5L. POA was always the same or higher for the 3L when compared with the 5L (POA3L = 0.78–0.97, POA5L = 0.64–0.97, mean POA3L = 0.877, mean POA5L = 0.773).

Table 6 Evidence of studies reporting on test–retest reliability (listed by year of publication)
Full size table

4 Discussion

The EQ-5D-5L was developed to improve the discriminative and evaluative properties of the EQ-5D-3L. Since publication of the 5L, a body of evidence has emerged that allows us to determine whether it has improved upon those properties. This review systematically summarizes the evidence of studies comparing the methodological properties of the EQ-5D-3L and EQ-5D-5L, with a special focus on redistribution of responses, including ceiling effects, floor effects, inconsistent responses, reliability and responsiveness. In the face of the reviewed results, both instruments demonstrated appropriateness for use in a wide range of study populations, addressing a variety of research questions and using different study designs. They show (1) the 5L responses logically distribute from the 3L, and (2) the 5L has advantages in terms of ceiling, (re-)distribution/distributional properties and how the descriptive system is used, but there are (3) some areas, such as responsiveness, in which the evidence is mixed and further research is needed. Furthermore, other aspects beyond the reviewed methodological parameters are important when choosing between 3L and 5L.

The low percentage of inconsistencies found in head-to-head studies demonstrates that the 3L redistributes logically to the 5L and that results of the 5L and 3L are comparable. The 5L is successful in reducing ceiling effects; a smaller proportion of respondents reported ‘11111’ on the 5L than on the 3L, especially in healthier samples. Thus, the 5L is suggested if the main goal is to discriminate among patients with milder health states. Moreover, the 5L outperformed the 3L when considering Shannon’s H′, with H′ being approximately 1.5-fold higher for the 5L compared with the 3L, without a relevant decrease of J′.

Missing values are negligible for both instruments demonstrating acceptance by respondents. Floor effects are also negligible for both instruments, meaning few respondents reported having the third or fifth levels of function in EQ-5D dimensions (e.g. ‘unable to wash or dress myself’). Most value sets assign negative weights to poor EQ-5D health states, meaning respondents valued many of these health states as worse than death (death is anchored at zero).

There is mixed evidence for better reliability on dimensional level, while evidence on index values shows better performance of 5L. Evidence on comparative responsiveness of the 3L and 5L is mixed [37, 42, 45]. This is surprising since adding levels to the 3L was intended to improve the responsiveness of the 3L. While two studies found the 5L to be slightly more responsive than the 3L when using non-parametric test statistics, Golicki et al. [42] found the crosswalk-derived 5L index to be less responsive than the 3L index when using several distribution-based approaches. There could be an explanation for why Golicki et al. conflicted with the other two studies. Crosswalk-derived utility scores tend to underdetect health gains [60,61,62]. For a preference-based instrument, it may be more appropriate to assess how changes in 5L versus 3L index scores are reflected in incremental cost-effectiveness ratios (ICERs) or quality-adjusted life-years (QALYs) [63]. Furthermore, differences with how participants value 3L versus 5L health states must be more closely examined [64, 65]. More research into sensitivity to change of the 5L and 3L is needed.

4.1 Limitations

This review has several limitations. Although all but two studies directly compared the 3L and the 5L, there are several reasons that the results of this review are difficult to generalize. The data have been derived from (1) different studies, (2) sampled from different population or patient samples, (3) use different language versions or values sets of the EQ-5D, and (4) use varying research designs (e.g. order of 3L vs. 5L, placing other, and how many, questionnaires in between 3L and 5L). Due to the differing methods, designs, analyses and potential cross-cultural differences in EQ-5D response patterns [66], it was difficult to summarize results. There are no guidelines for preference-based measures or established guidelines and standards (such as, for example, COSMIN). The EuroQol Group could create a task force to develop reporting standards in order to ensure future studies are well-defined and use more homogenous methods.

However, choosing between using the 3L and 5L requires consideration of aspects beyond methodological characteristics (which were specifically addressed in the scope of this review), such as setting and respondents, purpose of use, and availability of instruments and value sets. For all self-assessment instruments, and for preference-based instruments in particular, the choice of instrument should always take into account the perspective of those who complete the instrument, i.e. patients or respondents. There is evidence that patients prefer the 5L to the 3L, although the reason is not clear [8, 45, 67]. Fewer patients reported problems filling in the EQ-5D-5L questionnaire, and more patients deem the 5L to be easier to answer than the 3L and can find statements to describe their own health state on the 5L.

Another crucial aspect is the available language version, and, related to that, the availability of a value set to calculate the index score for the target population. Currently, both the 3L and 5L are available in more than 120 languages (3L: >170; 5L: >130) and for various administration modes (www.euroqol.org). To calculate an index score, the availability of a value set for the target population is necessary. The number of value sets available for the 3L (at least 27) is much larger than for the 5L (at least 8), with the crosswalk serving as the interim scoring method, while population-specific 5L value sets are being developed. There are also some cases where a 5L value set is available but a 3L is not; for those situations, population-specific 3L scores cannot be calculated.

5 Conclusions

This review supports the use of both the 3L and the 5L in a broad range of patients, populations, and countries. The 5L performs slightly better in terms of reducing ‘ceiling’ effects, and similarly in many other distributional properties. More research must be conducted to clarify both instruments’ performance on sensitivity to change and reliability, for which our review found mixed results from a few studies. The EuroQol group considering guiding end users with the decision to use the 5L or 3L as the choice of instrument would be based on aspects beyond measurement properties. The evidence presented in this paper can benefit the development of new EQ-5D versions, such as a 5L version of the child-friendly EQ-5D-Y [68], or exploring additional dimensions to the current five-dimension format (‘bolt-ons’) [69, 70].