FormalPara Key Points for Decision Makers

A comprehensive systematic review and evaluation was conducted on the psychometric performance of 14 generic childhood multi-attribute utility instruments across 18 psychometric properties.

There is a high degree of variability surrounding psychometric performance across the instruments and the psychometric properties evaluated, with more recently developed instruments targeting preschool children having greater evidence gaps.

Identified gaps in psychometric evidence and common methodological issues in psychometric assessment can inform future research directions and methods in this topic area.

1 Background

Constraints on healthcare expenditure require the comparison of alternative interventions in terms of costs and consequences [1]. Many healthcare decision makers recommend cost-utility analysis (CUA) as a form of economic evaluation enabling such comparisons [2,3,4,5]. The preferred health outcome in CUA is the quality-adjusted life-year (QALY), which combines preference-based health state utility values or health utilities with the length of time in those states [1]. Health utilities derived from generic instruments enable comparison of interventions across health conditions [6]. One approach to measuring health utilities is the use of multi-attribute utility instruments (MAUIs), which describe self- or proxy-reported health status across multiple dimensions according to a prespecified classification system [7]. They are designed to be accompanied by value sets that reflect the stated preferences (typically of a representative sample of the general adult population) for the health states generated by the classification system [8]. The value set application produces health utilities anchored on a scale with 0 = dead and 1 = full health.

Unique challenges arise in measuring health utilities in childhood populations (aged ≤ 18 years). First, biopsychosocial development during childhood means that the relevant dimensions of health status undergo rapid change [9]. This implies that classification systems for adult populations may not be applicable to children, and that the systems should be tailored to different age groups within childhood [10]. Second, the instrument wording and format (e.g., use of pictures, response scale levels) should be tailored to the comprehension level and attention span of the target childhood age group [11, 12]. Proxy respondents such as parents can be used, either when child self-report is not feasible (e.g., for very young preschool children) or when an alternative perspective on the child’s health status and health needs is sought [11]. Accordingly, several reviews have investigated the level of proxy-child agreement for patient-reported outcome measures (PROMs), including MAUIs [13,14,15].

Given these challenges, recent methodological advances have included the development of a range of instruments with childhood-specific classification systems and child-friendly formats and childhood-derived value sets, as well as the use of childhood-compatible instruments [7, 16]. A recent systematic review identified 89 generic multidimensional PROMs designed specifically for or compatible with application in children, 14 of which were MAUIs designed to be accompanied by value sets that generate health utilities [12].

Psychometrics concerns the measurement properties of scales, originating from psychophysics research on subjective judgements about physical stimuli and later applied in healthcare to develop scientifically rigorous PROMs [17,18,19]. Key psychometric properties include content validity, reliability, construct validity, responsiveness, interpretability, robust translation (cross-cultural validity), and patient and investigator burden (acceptability) [18,19,20,21]. Each of these properties requires unique tests and criteria, and contributes to minimum scientific standards for the use of a given PROM in patient-centred outcomes research such that the PROM should demonstrate acceptable performance across all properties included in the minimum standard set [20]. Otherwise, the instrument cannot be considered as providing scientifically credible information [17].

This highlights the need for a comprehensive evaluation of the psychometric evidence of instruments prior to application. However, it is unlikely that a single study can address all psychometric properties of an instrument for all research contexts [20]. Investigators seeking to use an instrument must rely on the weight of evidence across multiple studies, evaluating the consistency of findings while considering the volume and methodological quality of the contributing studies [20]. Therefore, a comprehensive systematic review of the available psychometric evidence for generic childhood MAUIs is warranted to inform their selection based on their meeting the minimum standards of scientific rigour.

Previous reviews have covered only a limited number of instruments [22,23,24,25,26,27,28]. For example, Rowen and colleagues [26] focused on four instruments, namely the CHU9D, EQ-5D-Y (3L or 5L), HUI2 and HUI3; Tan and colleagues [28] similarly focused on five instruments, the above four plus AQoL-6D. Janssens and colleagues [27] identified psychometric evidence (up to 2012) for eight instruments (16D, 17D, AQoL-6D, CHU9D, EQ-5D-Y, HUI2, HUI3 and CHSCS-PS) but not for those developed since 2012. The review by Janssens et al. also only included studies conducted on general childhood populations and excluded psychometric evidence from patient/clinical samples. Moreover, the above reviews excluded key psychometric properties from evaluation, including internal consistency and content validity [26], or cross-cultural validity [27]; the Tan review focused solely on known-group validity, convergent validity, test–retest reliability, and responsiveness [28]. The previous reviews also excluded indirect evidence from studies that did not explicitly aim to conduct psychometric assessment [26,27,28], even though such studies provide relevant psychometric evidence (e.g., evidence for responsiveness generated by clinical trials).

The aim of the current study was to conduct a systematic review that identifies and evaluates the psychometric evidence for all generic childhood MAUIs identified in a recent systematic review [12]. The review is intended to be comprehensive in terms of (1) the instruments covered; (2) the psychometric properties evaluated, drawing on diverse standards in the literature for both properties and evaluation criteria [17,18,19,20,21, 29,30,31]; (3) the coverage of evidence from general and clinical populations; and (4) inclusion of both ‘direct studies’ that explicitly aimed to assess psychometric properties and ‘indirect studies’ that provide evidence relevant to psychometric properties without this as a stated aim. Key review objectives were to:

  1. 1.

    create a catalogue of evaluated psychometric evidence that can aid in the selection of generic childhood MAUIs for application;

  2. 2.

    identify gaps in psychometric evidence to inform future psychometric research; and

  3. 3.

    summarise the commonly used psychometric assessment methods and the relative psychometric performance of instruments by property.

This review does not aim to arrive at conclusions on which MAUI is ‘optimal’. This would depend on the contextual factors for application, including, for example, country of application (determining relevant instrument language version and value set), the childhood population age and type (e.g., general population, clinical), and instrument respondent type. These in turn determine the relative importance of different psychometric properties and the appropriate level of consistency with which the instrument shows acceptable performance for each property in the accumulated evidence. These considerations comprise the minimum standard of scientific rigour for MAUI application. The review provides the evidence base that analysts can use to select instruments according to their application-specific minimum standard and/or conduct further psychometric research to fill relevant evidence gaps.

2 Methods

A prespecified protocol outlining the systematic review methods was developed and registered with the Prospective Register of Systematic Reviews (PROSPERO; CRD42021295959). For reporting purposes, the review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guideline [32] (see the electronic supplementary material [ESM] for the PRISMA checklist). Figure 1 illustrates the systematic review process, terminology, and objectives, which are further discussed in the following sections.

Fig. 1
figure 1

Illustration of the systematic review process, terminology, and objectives. ICC intraclass correlation coefficient

Full size image

2.1 Data Sources and Study Selection

The database searches aimed to identify studies that provide evidence for the psychometric performance of one or more of the following generic childhood MAUIs identified in a recent systematic review [12]: 16-Dimensional Health-Related Measure (16D) [33]; 17-Dimensional Health-Related Measure (17D) [34]; Adolescent Health Utility Measure (AHUM) [35]; Assessment of Quality of Life, 6-Dimensional, Adolescent (AQoL-6D Adolescent) [36]; Child Health—6 Dimensions (CH-6D) [37]; Comprehensive Health Status Classification System—Preschool (CHSCS-PS) [38]; Child Health Utility—9 Dimensions (CHU9D) [39, 40]; EuroQoL 5-Dimensional questionnaire for Youth 3 Levels (EQ-5D-Y-3L) [41]; EQ-5D-Y 5 Levels (EQ-5D-Y-5L) [42]; Health Utilities Index 2 (HUI2) [43]; Health Utilities Index 3 (HUI3) [44]; Infant health-related Quality of life Instrument (IQI) [45]; Quality of Well-Being scale (QWB) [46, 47]; and Toddler and Infant Health Related Quality of Life Instrument (TANDI; recently renamed EuroQoL Toddler and Infant Populations (EQ-TIPS) [48]) [10, 49]. It should be noted that the above MAUIs are designed to be accompanied by value sets with which health utilities can be derived, but not all (e.g., the recently developed TANDI) currently have a value set developed (see the list of value sets published before October 2020 in our previous review [12]). In this case, the psychometric performance of the classification system was evaluated.

The following databases were covered from database inception to 7 October 2021: MEDLINE (OvidSP)[1946-present]; EMBASE(OvidSP) [1974–present]; PsycINFO (OvidSP) [1806–present]; EconLit (Proquest); CINAHL (EBSCOHost) [1982–present]; Scopus (Elsevier); and Science Citation Index (Web of Science Core Collection) [1945–present]. Tables A1–A7 in the ESM provide the search strategy for each database. References of the included studies were also searched. Endnote 20 was used to remove duplicates among the imported references. At least two researchers from a team of seven (JK, SK, TB, MH, EH, SP, and RR) independently reviewed the titles and abstracts and then the full texts on Covidence [50]. At each review stage, if an article received two approvals from any pair of the seven reviewers, it proceeded to the next stage (from title/abstract screening to full-text screening, then to data extraction), with disagreements referred to a third reviewer for the final assessment.

The inclusion criteria were (1) the study provided evidence for at least one psychometric property (see below for list of properties) of at least one of the 14 instruments listed above using any language version; (2) the study obtained data from general or clinical childhood populations (sample mean age ≤ 18 years) or relevant proxy respondents; (3) where the study covered childhood and adult populations, results were reported separately for a childhood subgroup with a mean age ≤ 18 years; and (4) the study was published in English. Studies reporting the original instrument development (as identified previously [12]) were included for content validity evidence. Studies without the explicit aim of assessing psychometric properties (i.e., instrument application studies) but nonetheless contained relevant psychometric evidence (e.g., responsiveness within intervention trials) were included as ‘indirect’ assessment studies. Conference abstracts that met the inclusion criteria were included. Studies that used one or more of the 14 instruments as a criterion standard to test a new, yet-to-be validated instrument were excluded, as were studies that developed and validated value sets for health utility derivation without assessing or providing evidence of the psychometric properties of the health utilities.

2.2 Data Extraction

Data from the included studies were extracted by JK, with 20% of studies double data extracted by the other reviewers (SK, TB, MH, EH, and RR). Disagreements from double extraction were resolved by discussion within the research team and by decision from senior authors (SS, SP) if consensus could not be reached. The solutions reached through the discussions were applied to the other 80% of single-extracted studies.

The following data were extracted according to proformas in Excel (Microsoft Corporation, Redmond, WA, USA): (1) first author name and year of publication; (2) study country(ies); (3) study design—e.g., cross-sectional survey; (4) whether the study explicitly aimed to assess psychometric properties (‘direct’) or not (‘indirect’); (5) psychometric property(ies) assessed or relevant evidence reported; (6) main methodological issues affecting psychometric assessment (described in the next section); (7) instrument(s) assessed; (8) instrument language version(s); (9) instrument component(s) assessed—e.g., index (after value set application), dimension score; (10) country and population (where relevant) where a value set was derived; (11) respondent type(s)—child self-report or proxy report; (12) administration mode(s)—e.g., by self, with interviewer; (13) study population type—healthy, clinical, or general childhood population(s)—and clinical characteristics; (14) target and sample age,—e.g., mean, range—and proportion of females; (15) target and actual sample size; and (16) intervention(s) assessed (if relevant).

2.3 Evaluation and Data Synthesis

Table 1 lists and defines the psychometric properties evaluated. These have been drawn from established standards developed and used by medical industry regulators, previous psychometric research projects, and professional communities who are collectively interested in the psychometric performance of PROMs. The standards included the Consensus-based Standards for the selection of health Measurement Instruments (COSMIN) checklist [29,30,31]; the International Society for Quality of Life Research (ISOQOL) guideline [20]; the Food and Drug Administration (FDA) guideline [18]; the Medical Outcomes Trust (MOT) guidelines [21]; and the taxonomy of psychometric properties previously applied by Smith and colleagues [17]. Table 1 references the related guidelines for each property. There was a broad overlap in the properties covered by the guidelines and also a nontrivial variation, and this variation motivated the referral to multiple guidelines for comprehensiveness. Brazier and Deverill [19] have discussed the relevance of psychometric properties to health utility instruments, and their views were incorporated where relevant. We distinguished between proxy-child agreement and inter-rater agreement between different proxies given the emphasis placed on the former for paediatric PROM application [11].

Table 1 Definitions of psychometric properties assessed by the systematic review and their performance criteria rating
Full size table

To clarify the terminology used (see Fig. 1), a primary study directly or indirectly conducts an assessment of the psychometric performance of instrument(s) by property, which generates primary psychometric evidence. This review in turn conducts an evaluation of this assessment in terms of the quality of its design and methods and the resulting psychometric evidence. This evaluation produces a criteria rating output for each assessment. Assessments for different components of an instrument (e.g., index, dimension) are evaluated separately. As shown in Table 1, a four-part criteria rating for each property (except for interpretability, with a three-part set of criteria) was developed for the evaluation. References are given where the criteria are sourced from specific guidelines. In general, a criteria rating output of ‘+’ indicates psychometric evidence consistent with an a priori hypothesis (formulated by the primary study) in terms of clinical and psychometric expectation; ‘±’ partially consistent; and ‘–’ no evidence or evidence contrary to the a priori hypothesis. A ‘?’ output is given where the poor quality of assessment design and methods prevented a conclusion to be drawn by the evaluation concerning the psychometric performance for a property. General considerations for this quality evaluation across all properties included sample size, missing data level, and appropriateness of the statistical technique used (e.g., suited to continuous, binary, or ordinal nature of instrument outcome) [30]. Thereafter, considerations for the appropriate assessment design and methods differed by psychometric property [30].

It should be noted that the referenced guidelines provide general recommendations for evaluation: the COSMIN checklist, for example, expects the adequate sample size to vary by application context (p. 545) [30]. Further context-specific judgements were thus required for evaluation and these are reported case-by-case in the Excel file in the ESM. A key source of between-context variation was the a priori hypothesis specified by the included study, which introduced between-study variation in the expected correlation, association, or change sought. For example, one study defined a moderate and therefore acceptable correlation as a coefficient above 0.3 [51], while another as above 0.4 [52]. Where criteria and thresholds were specified as part of a priori hypotheses, these were followed by the review. Finally, studies frequently performed multiple assessments for the same psychometric property; i.e., sub-assessments within the assessment for the property. These sub-assessments each received criteria rating outputs, which were then grouped together to produce a summary output for the property. Details on this process are given in footnote b of Table 1.

The criteria rating outputs were synthesised to address the three review objectives:

  1. 1.

    Create a catalogue of evaluated psychometric evidence. The Excel file in the ESM serves as the main catalogue where the criteria rating outputs are tabulated with other extracted variables (e.g., study design, sample size) and the main rationale for each rating. More condensed catalogues are presented in this manuscript.

  2. 2.

    Identify gaps in psychometric evidence. Two metrics were used to identify evidence gaps: (1) the number of cases where no criteria rating output was available to an instrument for a given property; and (2) the number of cases where no criteria rating output was available or where available outputs had no ‘+’ output (no ‘+’ or ‘±’ for interpretability, which had very few ‘+’). The metrics were calculated for the whole evidence base and for a subset of evidence from direct assessment studies. Comparing these two scenarios allows one to understand the impact of including evidence from indirect assessment studies.

  3. 3.

    Summarise the commonly used psychometric assessment methods and the psychometric performance of instruments by property. Common reasons for assessments receiving ‘?’ were noted by property. The relative performance of instruments was compared by property, using the proportion of ‘+’ as the performance metric, while also considering the number of outputs over which the proportion was calculated.

3 Results

3.1 Search Results

Figure 2 presents the PRISMA flow diagram. After screening of titles/abstracts and full texts, 372 studies were included. Studies excluded at full-text screening stage (n = 167) are listed in Table B in the ESM, alongside the main reason for exclusion. Table C in the ESM details co-authors’ contributions to the screening stages.

Fig. 2
figure 2

Preferred Reporting Items for Systematic Reviews and Meta-Analyses flow diagram

Full size image

3.2 Characteristics of the Included Studies

Table D in the ESM reports the characteristics of the included studies and these are summarised in Table 2. The full extracted information on the included studies is available in the Excel file in the ESM. There were 163 (43.8%) studies that were judged to have conducted direct assessments of psychometric properties. Ninety-six (25.8%) studies assessed two or more instruments, 48 of them as direct assessments. Most studies (n = 210; 56.5%) concerned clinical paediatric populations. The largest target age category was that combining pre-adolescents and adolescents spanning the age range of 5–18 years (n = 145, 39.0%).

Table 2 Characteristics of the included studies
Full size table

3.3 Psychometric Evaluation Results

Table 3 summarises the characteristics of the psychometric assessments from which the criteria rating outputs were derived. The full extracted information is contained in the Excel file in the ESM, which serves as the main catalogue that meets the review objective (1).

Table 3 Characteristics of psychometric assessments
Full size table

Table E in the ESM provides a condensed catalogue of the evaluated psychometric evidence by study, instrument, component, and respondent type. There were 2153 criteria rating outputs and 1598 (74.2%) excluding ‘?’ outputs. All outputs and rationale for ratings, alongside the study and assessment characteristics, are reported in the Excel file in the ESM as the main evidence catalogue. Figure A in the ESM displays the output numbers and proportions by instrument, component, and psychometric property, except for AHUM, CH-6D, and IQI, which had four or fewer outputs.

3.4 Psychometric Evidence Gaps

This section addresses the second review objective and identifies psychometric evidence gaps. Table 4 shows the number of criteria rating outputs by instrument and property. Seven properties—internal consistency, inter-modal reliability, and structural, discriminant, empirical, concurrent and predictive validities—had fewer than 30 outputs, while three had more than 300: known-group validity (n = 500), acceptability (n = 352), and hypothesis testing (n = 309). Similarly, there was high variation in output numbers across instruments, with the HUI2 (n = 472) and HUI3 (n = 623) comprising nearly half of all outputs.

Table 4 Evaluation of psychometric evidence gaps by psychometric property and instrument from all included studies (n = 372)
Full size table

Of the 270 cells in Table 4 defined by instrument and property, 116 (43.0%) were empty, indicating absence of evidence, while 146 (54.1%) were either empty or had 0% ‘+’ output. The psychometric evidence gap varied greatly by instrument, with the longer established instruments having smaller gaps: EQ-5D-Y-3L had the least number of cells (n = 2) that were empty or had 0% ‘+’ output, followed by HUI3 (n = 4), HUI2 (n = 5), and CHU9D (n = 5). By contrast, the more recently developed instruments targeting preschool children had high numbers of empty or 0% ‘+’ output cells: CHSCS-PS (n = 12); IQI (n = 17); and TANDI (n = 12).

Another noticeable feature is the frequent evidence gaps for reliability (IC, TR, IR, and IM), which had 32 of 56 (57.1%) cells that were empty or had 0% ‘+’ output. This compares with 24 of 70 (34.3%) for construct validity (KV, HT, CNV, DV, and EV) and 5 of 42 (11.9%) for KV, HT, and CNV combined. Proxy-child agreement is a property specifically relevant to childhood MAUIs, yet its evidence gap was similarly high: 10 of 14 (71.4%) cells were empty or had 0%, and likewise 7 of 11 (63.6%) after excluding CHSCS-PS, IQI, and TANDI, which target preschool children and hence likely rely solely on proxy report.

Table F in the ESM shows the number of criteria rating outputs and proportion of ‘+’ from the subset of 163 direct assessment studies. There were 118 (43.7%) empty cells in Table F, only two more than in Table 4. This suggests that the further inclusion of indirect assessment studies in Table 4 expanded the volume but not the type of psychometric evidence. However, there was a noticeable difference in the number of empty or 0% ‘+’ cells, with 160 (59.3%) in Table F compared with 146 (54.1%) in Table 4. Hence, the inclusion of indirect assessment studies reduced the number of cases where there was no ‘+’ output. This was particularly so for 16D, 17D, and AQoL-6D, for which the number of empty or 0% ‘+’ cells declined from 12, 13, and 16, respectively, in Table F, to 9, 9, and 13, respectively, in Table 4.

3.5 Psychometric Assessment Methods and Performance by Property

This section addresses the third review objective and describes the common psychometric assessment methods and the relative performance of instruments by property, except for predictive validity, which had no criteria rating output.

3.5.1 Internal Consistency

Internal consistency criteria rating outputs (n = 25) were available for eight instruments. Most evidence consisted of Cronbach’s alpha (n = 19, 76.0%), which fell below the acceptable threshold of 0.7 only twice, once each for CHU9D (from seven outputs) [53] and EQ-5D-Y-3L (from three) [54]. The proportion of ‘+’ was highest for TANDI at 100% (from one output), followed by CHU9D at 85.7% (from seven outputs).

3.5.2 Test–Retest Reliability

Test–retest reliability criteria rating outputs (n = 64) were available for 10 instruments. The interval between initial test and retest varied significantly from 24 hours to 1 year. Nineteen outputs from eight studies involved test–retest intervals longer than 4 weeks [55,56,57,58,59,60,61,62]; of these, five studies limited the test–retest sample to those who maintained stable health conditions in the interval [55, 57, 60,61,62]. Statistical methods included intraclass correlation coefficient (ICC) for continuous outcomes (e.g., index, visual analogue scale [VAS]) and the Kappa coefficient and percentage agreement for categorical outcomes (e.g., ordinal item response). Figure B in the ESM shows the criteria rating outputs by instrument. QWB and TANDI had 100% ‘+’ outputs, although the total numbers per instrument were small (one and two, respectively). 16D and 17D had 100% ‘+’ from four outputs. Among others, EQ-5D-Y VAS (from the 3L and 5L versions) had the highest proportion of ‘+’ (75.0%) from 12 outputs.

3.5.3 Inter-Rater Reliability

Inter-rater (proxies) reliability criteria rating outputs (n = 77) were available for five instruments. Proxies involved in comparison included parents, physicians, nurses, and therapists. Statistical methods included ICC, Kappa coefficient, and percentage agreement. Figure C in the ESM shows the criteria rating outputs by instrument. QWB had 100% ‘+’, followed by CHSCS-PS at 66.7%, but both instruments had few outputs (two and three, respectively). HUI2/3 had more outputs than other instruments, but the proportions of ‘+’ were low at 11.8% and 6.7%, respectively.

3.5.4 Inter-Modal Reliability

Inter-modal reliability criteria rating outputs (n = 11) were only available for three instruments and from three studies [63,64,65]. Modal comparisons of interest were between home- and clinic-based administration [63], between paper and online versions [64], and between telephone interview, face-to-face interview, and postal self-administration [65]. EQ-5D-Y-3L dimension and VAS scores had ‘+’ outputs twice each: HUI2 ‘±’ twice, and HUI3 ‘±’ twice and ‘–’ three times.

3.5.5 Proxy-Child Agreement

Proxy-child agreement criteria rating outputs (n = 149) were available for eight instruments. Proxies included caregivers, parents, physicians, nurses, teachers, and therapists. Statistical methods included ICC, Kappa coefficient, and percentage agreement. Figure 3 shows the outputs by instrument. Agreements were generally low, with 24.8% (37 of 149) of outputs being ‘+’. CHU9D had the highest proportion of ‘+’ at 33.3%, followed by HUI2 at 31.3% from a considerably larger number of outputs.

Fig. 3
figure 3

Proxy-child agreement criteria rating outputs by instrument. Note: Absolute numbers of criteria rating outputs are displayed within each bar

Full size image

3.5.6 Content Validity

At least one content validity criteria rating output was available for all instruments (n = 35). Around half of the outputs (n = 18) were reported in the original development of instruments, i.e., whether a conceptual framework for the purpose of measurement was stated, whether qualitative research with children or appropriate proxies were conducted for dimension/item elicitation, and whether cognitive interviews and piloting were conducted. An output of ‘+’ for original development was given to CHSCS-PS [38], CHU9D [39, 40], IQI [45], and TANDI [10, 49]; ‘±’ to 17D [34], AHUM [35], AQoL-6D [36], EQ-5D-Y-3L [41, 66], EQ-5D-Y-5L [42], HUI2 [43], HUI3 [44], and QWB [46]; and ‘–’ to 16D [33] and CH-6D [37].

Among other outputs (n = 17), some concerned content validation of modified versions (n = 7): cognitive bolt-on to EQ-5D-Y-3L [67] and the ‘child-friendly’ version of HUI2/3 [68]; the latter suggests conceptual issues in applying the standard HUI2/3 with children (aged 5–18 years in the study). Although not counted in the rating scale, 69 of 105 studies (65.7%) that directly or indirectly assessed HUI2 removed the fertility dimension. The remainder (n = 10) concerned post-development content validations of CHU9D [69, 70], EQ-5D-Y-3L [70, 71], HUI2 [72,73,74], and HUI3 [75]. These studies conducted surveys and qualitative research with children and proxies to evaluate the instruments’ conceptual relevance to the childhood health construct of interest. Only the more recently developed instruments targeting preschool children—CHSCS-PS, IQI, and TANDI—received 100% ‘+’ outputs. Among instruments for older children, only CHU9D had an above-zero percentage of ‘+’ (40%).

3.5.7 Structural Validity

Only 10 criteria rating outputs were available for structural validity. Exploratory or confirmatory factor analysis was conducted for 16D [76] (n = 1), CHU9D [77, 78] (n = 3), EQ-5D-Y-3L with cognitive bolt-on [67] (n = 1), and TANDI [49] (n = 1). One study conducted multi-trait analyses of the correlations between dimension scores and dimension-relevant item responses for the HUI3 and found hypothesised correlations for all dimensions except cognition [68]. Another study estimated the between-item correlations for the EQ-5D-Y-3L, finding low–moderate correlations [79]. CHU9D and TANDI received 100% ‘+’ from three and one outputs, respectively.

3.5.8 Cross-Cultural Validity

There were 34 cross-cultural validity criteria rating outputs for 16D translations into French [80] and Norwegian [81]; 17D into French [80]; CHU9D into Chinese [51], Danish [82], Dutch [83], and Canadian French [84]; EQ-5D-Y-3L into English, German, Italian, Spanish, and Swedish [41, 79, 85, 86] and Japanese [87]; EQ-5D-Y-5L into Chinese for the mainland [88] and Hong Kong [89]; HUI2 and HUI3 into Dutch [90], Canadian French [72, 91], German [92], Russian [93, 94], Spanish [95, 96], Swedish [97], and Turkish [98]. EQ-5D-Y (3L and 5L) thus reported on the highest number of translated versions, likely aided by the established EuroQoL translation guideline. However, the overall proportion of ‘+’ outputs was 54.5% (6 of 11). Deficits in rating were generally due to a lack of backward translation or pilot testing with children. One study assessed the feasibility of a Spanish version of HUI2 and found that translation errors in emotion items resulted in a high volume of missing parental responses [96].

3.5.9 Known-Group Validity

Known-group validity criteria rating outputs (n = 500) were available for all instruments except AHUM and IQI. Group comparisons included those between patients and healthy/general controls, between disease presence/absence, severities, and diagnosis types, and between sociodemographic groups where expected differences were specified a priori (e.g., by parental education, employment status or socioeconomic status [52, 99]). Statistical methods included the Mann–Whitney U and Kruskal–Wallis tests for continuous outcomes and the Chi-square test for categorical outcomes. Figure 4 shows the outputs by instrument. CH-6D, CHSCS-PS, and TANDI had 100% ‘+’ from one or two outputs, and AQoL-6D, EQ-5D-Y-5L, and QWB likewise had high proportions of ‘+’ outputs (75.0%, 62.5%, and 62.5%, respectively) from less than 10. Among others, CHU9D had the highest proportion of ‘+’ outputs at 52.1%.

Fig. 4
figure 4

Known-group validity criteria rating outputs by instrument. Note: Absolute numbers of criteria rating outputs are displayed within each bar

Full size image

3.5.10 Hypothesis Testing

Hypothesis test criteria rating outputs (n = 309) were available for all instruments except IQI. Statistical methods included multivariate linear regression to estimate the adjusted association between the instrument score and a variable of interest (e.g., disease severity) and comparison against the a priori hypothesis. A high proportion (101 of 309, 32.7%) of outputs were ‘?’ since ad hoc tests without a priori hypotheses (e.g., by age group and sex) were placed under this property. Figure 5 shows the outputs by instrument. Other than AHUM with a single ‘+’ and AQoL-6D with 40% ‘+’ from five outputs, EQ-5D-Y-3L had the highest proportion of ‘+’ outputs at 35.7%, followed by EQ-5D-Y VAS at 34.5%.

Fig. 5
figure 5

Hypothesis testing criteria rating outputs by instrument. Note: Absolute numbers of criteria rating outputs are displayed within each bar

Full size image

3.5.11 Convergent Validity

Convergent validity criteria rating outputs (n = 202) were available for all instruments except AHUM, CHSCS-PS, and IQI. The main statistical method involved correlations between comparable dimensions or between index utilities/total scores that measured similar constructs (e.g., between EQ-5D-Y-3L pain and PedsQL physical functioning dimensions [100], or between CHU9D index and PedsQL total score [101]), and assessing whether they were significantly different from zero and higher than prespecified thresholds of strength (e.g., r > 0.40). Figure 6 shows the outputs by instrument. CH-6D had 100% ‘+’ from only one output, while 16D and 17D had 85.7% ‘+’ from seven combined and EQ-5D-Y VAS had 39.1% ‘+’ from 23.

Fig. 6
figure 6

Convergent validity criteria rating outputs by instrument. Note: Absolute numbers of criteria rating outputs are displayed within each bar

Full size image

3.5.12 Discriminant Validity

Only seven discriminant validity criteria rating outputs were available for four instruments. Studies prespecified hypotheses of negligible correlations between dimensions and index/total scores measuring different constructs, e.g., between HUI3 index and quality-of-life dimensions of being, belonging, and becoming [102]. EQ-5D-Y-5L dimensions and VAS had one ‘+’ each; EQ-5D-Y-3L dimensions had one ‘+’ and one ‘±’, as did the HUI3 index; and CHU9D dimensions had one ‘?’ due to poor reporting.

3.5.13 Empirical Validity

Empirical validity criteria rating outputs (n = 29) were available for five instruments. They all concerned differences in index utility scores across groups defined by self-reported general health status. Most outputs (n = 18) concerned the CHU9D, all but one of which was ‘+’ (94.4%). Other instruments also had high proportions of ‘+’: AQoL-6D, 100% from two; EQ-5D-Y-3L, 100% from one; HUI2, 80% from five; and HUI3, 100% from three. This suggests acceptable empirical validity for the five instruments and particularly for the CHU9D.

3.5.14 Concurrent Validity

Concurrent validity criteria rating outputs (n = 17) were available for six instruments. Studies prespecified ‘gold standard’ measures and expected correlation strength and direction between the gold standard and the given instrument. EQ-5D-Y-3L had the highest number of outputs (n = 5) with 40% ‘+’, while CHSCS-PS (n = 1), CHU9D (n = 3), and HUI3 (n = 2) had 100% ‘+’.

3.5.15 Responsiveness

Responsiveness criteria rating outputs (n = 147) were available for nine instruments. Studies assessed the statistical and clinical significance of changes in instrument score over time and whether these were consistent with a priori hypotheses and changes in the reference measure score (e.g., a disease-specific measure or clinical outcome). A significant proportion of outputs (64 of 147, 43.5%) was ‘?’ due to the non-inclusion of a reference measure. Figure 7 shows the outputs by instrument. EQ-5D-Y-5L had 100% ‘+’ from one output. Among others, HUI2 had the highest proportion of ‘+’ at 46.7%, followed by CHU9D at 45.5%.

Fig. 7
figure 7

Responsiveness criteria rating outputs by instrument. Note: Absolute numbers of criteria rating outputs are displayed within each bar

Full size image

3.5.16 Acceptability

Acceptability criteria rating outputs (n = 352) were available for all instruments except AHUM, CH-6D, and IQI. Assessed features included missing data rates (due to instrument, not survey, design), ceiling effects (no study reported significant floor effects), time for completion, comprehensibility, and number of unique health states per sample respondent. The prespecified criterion that studies evaluating only one acceptability feature receive ‘?’ resulted in a high proportion of outputs being ‘?’ (175 of 352, 49.7%). Figure 8 shows the outputs by instrument. CHSCS-PS had the highest proportion of ‘+’ at 50% but from only two outputs, followed by QWB at 33.3%, similarly from few outputs (n = 6). Among instruments with 10 or more outputs, CHU9D had the highest proportion of ‘+’ at 29.7%.

Fig. 8
figure 8

Acceptability criteria rating outputs by instrument. Note: Absolute numbers of criteria rating outputs are displayed within each bar

Full size image

Synthesising individual acceptability features separately, there were 183 assessments of ceiling effect, 254 of high missing data, 52 of response time, and 78 of other features (e.g., number of unique health states). Concerning the two most assessed features, namely ceiling effect and high missing data (see Table 1 for thresholds used), Tables G and H in the ESM compare their respective occurrences by instrument. In Table G, ceiling effect occurrences were tabulated separately for all childhood populations and samples including patients only. The ceiling effect occurrences did not diminish for patient populations even though a lower ceiling effect would be expected. As for the high missing data occurrence in Table H, AQoL-6D, CHSCS-PS, and TANDI had no occurrence, although only from one or two assessments each. Among others, EQ-5D-Y-3L and -5L had the lowest percentages of samples with high missing data: 11.5% and 0%, respectively.

3.5.17 Interpretability

Interpretability criteria rating outputs (n = 185) were available for nine instruments. Only two studies sought to establish childhood population norm data, one for EQ-5D-Y-3L index in Japan [103] and the other for EQ-5D-Y-5L dimensions and VAS in Hong Kong, China [104]. Two studies calculated the minimal clinically important difference (MID) based on effect size for the EQ-5D-Y-3L index [105] and unweighted sum of item scores [106], while another estimated it as half of baseline standard deviation for EQ-5D-Y-3L unweighted sum and VAS [107].

Other studies used external data from the literature to interpret results, e.g., comparing the instrument scores with those of general/healthy childhood samples (not necessarily population norms), for which the output was evaluated as ‘±’. Studies referencing external MIDs repeatedly used those derived from non-childhood populations, for which the output was ‘?’. For example, studies for HUI2/3 referenced MIDs of 0.03 for index and 0.05 for dimension score from Horsman and colleagues [108], which were not based on primary derivation from childhood population but on expert opinion. Figure 9 shows the outputs by instrument. Comparing the proportion of ‘±’ and ‘+’ combined (given the low incidence of ‘+’), 16D/17D had the highest proportion (85.2%) among instruments with 10 or more outputs.

Fig. 9
figure 9

Interpretability criteria rating outputs by instrument. Note: Absolute numbers of criteria rating outputs are displayed within each bar

Full size image

4 Discussion

This review evaluated the psychometric performance of 14 generic childhood MAUIs [12], drawing on directly or indirectly provided psychometric evidence from 372 primary studies. Psychometric performance criteria were drawn from established standards in the literature to generate criteria rating outputs for 18 psychometric properties. The resulting 2153 outputs constitute a comprehensive psychometric evidence catalogue for generic childhood MAUIs, which analysts can use to inform MAUI selection for application. Further synthesis identified psychometric evidence gaps (e.g., instruments targeting young children below the age of 3 years) and summarised the range of psychometric assessment methods and results by property (e.g., whether a reference measure was included to test the instrument responsiveness against the measured health status change). No one instrument outperformed others across all properties, and the comparisons had to consider both the proportion of ‘+’ and the number of outputs from which the proportion is calculated.

As noted above, the evidence catalogue is necessary but insufficient to finalise the optimal MAUI selection for application. Considering established standards in the literature [17,18,19,20,21, 30], stakeholders involved in instrument selection should establish a minimum standard of scientific rigour for the given research setting. This would involve decisions on (1) which properties are relevant—proxy-child agreement, for example, would not be relevant for applications relying solely on proxy report; (2) the relative importance of different properties—responsiveness, for example, would be most important for application in clinical trials; and (3) the appropriate level of consistent acceptable performance for each relevant property within the accumulated evidence. As noted, the decision on (3) would involve setting thresholds for both the proportion of ‘+’ and the number of outputs from which the proportion is calculated: several instruments had very high proportions of ‘+’ but from a small pool of outputs. We note that the small pool of outputs may be due to recent development of an instrument and/or limited evidence generation relating to a specific psychometric property for that instrument. Moreover, the application will most likely target a specific childhood population, meaning that only a subset of the accumulated evidence (defined, for example, in terms of country, instrument language version, age group, and clinical characteristics) will be relevant. By extracting a wide range of study- and assessment-related variables, the current catalogue should aid analysts in retrieving the necessary psychometric evidence for the given minimum standard. Indeed, without application-specific target populations and minimum standards, reaching a conclusion on the overall performance of each measure from the highly heterogeneous evidence base is difficult [26].

For the decision on (1), a further question is whether the properties regularly assessed for non-preference-based PROMs and health status measures are relevant for preference-based MAUIs [19, 26]. Brazier and Deverill [19] argue that construct validity and responsiveness are not relevant to health utilities (although relevant to responses on the classification system prior to value set application) since they concern preferences over health level/change, not the level/change itself. They instead recommend empirical validity, one test of which is whether utility scores reflect hypothetical preferences over health/disease states expressed as self-rated health status [19, 109]. Moreover, internal consistency for MAUIs should concern the relevance of items to societal preferences, not between-item correlations within the same dimension [19]. Structural validity should concern orthogonality of the items rather than their conformity to hypothesised dimensionality [110]. Likewise, content validity for MAUIs should concern the classification system’s relevance to individuals’ utility function, not its comprehensiveness in reflecting the health construct of interest [19].

In practice however all the above properties are included in psychometric assessments of MAUIs by primary studies and in their evaluations by psychometric reviews. For example, the recent review by Rowen and colleagues [26] did not distinguish between empirical validity and construct validity, while the review by Janssens and colleagues [27] covered internal consistency, structural validity, and content validity. There was likewise a nontrivial volume of evidence for these properties from primary studies, e.g., n = 25 for internal consistency (particularly for CHU9D) and n = 35 for content validity (for all instruments, which contradicts the finding of Tan and colleagues [28] that content validity evidence was insufficient to be synthesised). These suggest that the psychometric properties specified in Table 1 remain relevant to generic childhood MAUIs. Analysts could therefore draw on the Table 1 criteria for the psychometric assessment or evaluation of MAUIs (generic or disease-specific, childhood or adult), particularly regarding the psychometric performance of both the utility index and the underlying health classification system, while noting the methodological ambiguities present in this space intersected by the disciplines of psychometric research and health economics.

Another key strength of the Table 1 criteria is the incorporation of the ‘?’ criteria rating output, given to 25.8% of 2153 psychometric assessments, which identified psychometric assessment designs and methods of poor quality. This quality evaluation was not incorporated in the Rowen review [26]. The Janssens review [27] incorporated the ‘?’ output but at the study level rather than at the more granulated assessment level. It nevertheless noted “significant methodological limitations” in several included studies (p. 342) [27]. The Tan review [28] highlighted the methodological limitations in the assessment of test-retest reliability and responsiveness in particular. In this review, the commonly observed reasons for the ‘?’ output were narrated in Sect. 3.5 by property, e.g., (1) the reliance on external MIDs derived from expert opinion or non-childhood populations [108, 111]; (2) assessment of only one acceptability feature out of many available; (3) lack of a reference measure or MID to help interpret responsiveness evidence; and (4) lack of a clear a priori hypothesis in testing associations, particularly between instrument scores and sociodemographic variables. These methodological gaps, further detailed case-by-case in the Excel catalogue, should help inform the design of future psychometric assessments of existing and new measures.

A significant finding of this review is the high volume of psychometric evidence available from indirect studies (900 of 2153, or 41.8% of all criteria rating outputs) that did not explicitly aim to conduct psychometric assessments. Thirty percent of outputs from indirect studies were ‘?’, which was higher, but not substantially so, than the 22.1% from direct studies. More than half of outputs for known-group validity, hypothesis testing, and responsiveness were drawn from indirect studies. These findings, i.e., the general non-inferiority and the nontrivial quantity of the psychometric evidence from indirect studies, justify the inclusion of both study types. This is a key strength of our review relative to previous reviews that covered direct studies only and which reported the scarcity of responsiveness evidence in particular [26,27,28]. It was found that the inclusion of indirect studies, while expanding the volume of evidence, did not increase the range of psychometric properties for which there was evidence. It did however expand the number of properties with at least one ‘+’ output. Therefore, indirect study evidence supplements that of direct studies, which should enable future primary studies to focus on those properties for which little direct or indirect evidence currently exists, particularly reliability (internal consistency, test-retest, inter-rater, inter-modal) and proxy-child agreement as found in Sect. 3.4. Given that the scarcity of reliability evidence and the uncertainty over proxy-child agreement have already been emphasised in previous reviews [26,27,28], greater research attention on these properties is warranted.

This review has further key strengths. First, it is an up-to-date and comprehensive review on the psychometric performance of generic childhood MAUIs. Previous reviews predate the development of some notable instruments, including the EQ-5D-Y-5L targeting school-aged children, and TANDI and IQI targeting preschool children [22,23,24,25,26,27, 112, 113]. The Rowen and Tan reviews deliberately focused on EQ-5D-Y, HUI2/3, AQoL-6D, and CHU9D [26, 28], thereby excluding instruments targeting young children below the age of 3 years. By contrast, the significant psychometric evidence gaps for these instruments were a key finding in Sect. 3.4, with no evidence available for inter-modal reliability, cross-cultural validity, discriminant validity, empirical validity, responsiveness, and interpretability. Their inclusion in our review will help with planning future studies addressing these evidence gaps. Otherwise, they would face greater difficulty than the longer established ones in meeting a given minimum standard of scientific rigour. The current review is also comprehensive in terms of the range of properties covered: the Rowen review, for example, excluded internal consistency and content validity [26]; the Janssens review excluded cross-cultural validity [27]; while the Tan review focused on known-group and convergent validities, test-retest, and responsiveness, and chose not to extract data on content validity despite this being a prespecified research objective [28].

The review nevertheless has several limitations. First, the review did not quantify (e.g., by using the COSMIN checklist [30]) a methodological/reporting quality score for each study and property. Such quantification for all 372 studies was deemed impractical in terms of time and research resources. This nevertheless neglects variation in methodological quality among assessments that avoided the ‘?’ output. It is likely, for example, that assessments with larger sample sizes (all other factors being equal) provide more robust psychometric evidence than those with smaller sizes, even if both sets meet the minimum level of methodological quality. Second, the criteria rating applied uniform performance thresholds (e.g., p value < 0.05 for statistical significance and ICC >0.7 for proxy-child agreement being ‘+’) unless specified otherwise and justified by primary studies. This created situations where an instrument found to be favourable relative to another within a study nevertheless received a negative rating and vice versa. For example, in the study by Ungar and colleagues [61], the child-parent dyad approach to questionnaire administration produced higher ICCs for proxy-child agreement than independent administrations for the HUI2 index and emotion dimension score and the HUI3 ambulation score. However, the ICC remained below 0.7, yielding a ‘–’ rating for the dyad approach. Third, the presentation of the evidence synthesis results in Sects. 3.4 and 3.5 were not disaggregated to each component of the instrument (e.g., index, dimension, VAS) as done in the Rowen review [26], but the case-by-case evaluation results are available in the Excel catalogue in the ESM. Distinctions could also be drawn between different variable types derived from dimension and index scores, e.g., index utility as a continuous variable versus categorised into disability levels (mild, moderate, and severe).

There were also several caveats with the review methods applied. First, some studies applied instruments to childhood groups outside the instruments’ intended target age ranges (e.g., HUI2/3 on children aged 3–6 years [114]), but assessments from such studies were not assigned a ‘?’ output. The rationale was to compare the rating outputs (not conducted in this manuscript) between instrument applications on intended and unintended age groups, which would not be possible if the latter were uniformly assigned ‘?’. Second, evidence on cross-cultural validity was difficult to obtain via peer-reviewed publications. For example, according to the HUI website, HUI2/3 have at least 30 translated versions [115], but the development of only seven versions were identified by this review. It is plausible that evidence on other psychometric properties might have been similarly missed by limiting the search to published articles and conference abstracts (e.g., instrument development steps in grey literature containing content validity evidence). Third, no study was identified that applied modern psychometric test theories such as Rasch analysis and item response theory [116]. This corroborates the finding of the Janssens review [27]. However, given the increasing use of modern theories for adult instrument development (e.g., EQ-5D-5L [117]), future psychometric reviews of childhood instruments will need to include criteria relevant to these theories.

5 Conclusion

This systematic review provides comprehensive and up-to-date evidence on the psychometric performance of generic childhood MAUIs that are designed to be accompanied by preference-based value sets. The inclusion of indirect evidence from studies without the explicit aim of psychometric assessment increased the comprehensiveness of the review. The catalogue of evaluated psychometric evidence provides a valuable resource for researchers and policymakers, particularly those involved in cost-effectiveness analysis, modelling, and decision-making, in selecting instruments for specific applications. The candidate instruments should meet a minimum standard of scientific rigour defined by the established criteria and consideration of the application setting. The final instrument(s) selected from the candidates should be that (those) with the most consistent performance according to the accumulated evidence in the review. The identified psychometric evidence gaps also motivate future psychometric studies, particularly the gaps on reliability and proxy-child agreement and on MAUIs targeting preschool children. The commonly observed issues in assessment design and methods, such as the statement of the a priori hypothesis for testing associations and changes, should likewise inform future psychometric studies.