Enhancing the multi-dimensional assessment of quality of life: introducing the WHOQOL-Combi

Data were collected through the Comprehensive Longitudinal Assessment of Salford Integrated Care (CLASSIC) survey of health in older patients. Participants were 65 years or over, registered with a general practice in Salford, UK, and diagnosed with one or more long-term health conditions (n = 3686) [21, 22]. This socio-economically deprived population has some of the nation’s poorest health. WHOQOL-Combi and EQ-5D-5L questionnaires were administered during the third follow-up (2016), 18 months after baseline. The EQ-5D-5L was recently modified from the EQ-5D ‘gold standard’ measure [23].

International and national WHOQOL modules (see Table 1), were reviewed for generic facets. In a conceptual review of facet definitions we examined module manuals, WHOQOL publications and internal WHO documents. Item wording in long form modules and procedures for selecting module items were examined. The psychometric performance of each item, domain and module was scrutinised. Measures were excluded if: (i) no short form had been standardised (e.g. children’s QoL [24]); (ii) the module was standardised in one culture (e.g. UK WHOQOL-Pain [25]), not the WHO minimum of three from different Continents; (iii) ‘experts’ had proposed or endorsed new concepts and/or items, without user input (e.g. poverty [26]), so the requirements of patient/person-reported outcome measures (PROMs) were not satisfied [27].

Several issues emerged during item selection that needed attention, for instance, apparent duplication of concepts in more than one module. Retention could increase completion burden. Similar facet themes included perceptions of the future in fear of the future (WHOQOL-HIV [14]), and past, present and future activities (WHOQOL-Old [15]). However recurring themes also signalled that this concept could be pertinent. As one item per facet was needed to maintain parity, both items above were pre-tested with CLASSIC data. Second, contents sometimes overlapped within modules, e.g. autonomy and participation/isolation (WHOQOL-Old [15]). As their psychometric performances did not indicate differences, both were pre-tested. Third, an intimacy facet in the WHOQOL-Old was similar to sex life in the core, so replacement or supplementation was tested. Fourth, a political rights and freedoms dimension had been proposed for an international poverty module [26], a Thai children’s measure by their parents [24], and as a New Zealand national item [32]. As political rights extend beyond health, this was not considered further, but a freedom facet in the WHOQOL-Old was retained.

WHOQOL measures assess QoL in the past 2 weeks. Each item is rated on one of several 5-point Likert response scales [33, 34]. A new item from a module is normally inserted into the core at the end of the relevant response scale block to speed completion. Socio-demographic details on gender, age, educational qualifications, living and work situation were documented.

Module facets had been designated to a domain during instrument development, and their position in the structure was reassessed. Item pool analysis for the preliminary WHOQOL-Combi was conducted without modular boundary constraints. Item reduction followed previous exclusion procedures [4, 5]. Negatively phrased items were reversed. Raw scores were calculated for items in the five provisional domains, then transformed linearly onto a 0–100 scale. Embedded WHOQOL-BREF items were scored as four domains.

Domain scores were not calculated if two or more items were missing. A two-item mean was calculated for the general facet. Cases were deleted where > 20% of items in the scale were missing. When no more than two items were missing, means were imputed from that person’s remaining domain items. During item reduction, item properties were tabulated: normality statistics, ceiling/floor effects and internal consistency reliability (ICR). When mapping the preliminary structure, Pearson Product Moment correlations (r) between items, domains and general QoL were calculated, and item-total correlations corrected as applicable; the acceptance cut-off was > .40 and significance p < .05 (one-tailed). Where two or more results were unacceptable, item exclusion was considered. Where performance was mixed, items were retained for further psychometric testing.

The tenability of the hypothesised factor structure was investigated through confirmatory factor analysis (CFA) and Rasch analysis using non-imputed scores. Informed by the established factor structure of WHOQOL instruments/modules, a higher-order five-factor model (physical, psychological, social, environment and SRPB) was initially tested using CFA with diagonally weighted least-squares and polychoric correlations [35] (LISREL v.8.80 [36]). Since the overall sample size well exceeded recommendations for asymptotically distribution-free methods [35], analyses were conducted on split samples to demonstrate robustness via replication. Consequently, two samples of n = 1000 were randomly extracted from the overall sample: Sample A for the main CFA, and Sample B for replication. Following WHOQOL validation procedures [26], item error co-variances were not correlated, providing a conservative exploration of model fit.

Since Chi-square inflates with sample size [37] , model fit was assessed using the indices and root-mean-square error of approximation (RMSEA) < 0.06, comparative fit index (CFI) > 0.95, and standardized root-mean-square residual (SRMR) < 0.08 [38]. As these cut-offs may be uncertain for asymptotically distribution-free methods [38], the entire pattern of fit indices was interpreted. Following earlier research [32, 39], the hypothesised factor structure was further confirmed using Rasch analysis with an additional randomly selected n = 500 (Sample C), which met Rasch sample size criteria [40]. This was conducted in parallel, to resolve misfit sources in CFA models. This advantageous approach also provided confidence that the robustness of the factor solution could be generalised to a different psychometric method. Partial-credit Rasch analysis (RUMM 2030 software [41]) followed previous research [42] by using a domain-level subtest approach. This permits the differentiation of local dependency arising due to multidimensionality (trait dependency) from local dependency due to method effects (response dependency) [43]. If Chi-square fit for item-trait interaction is non-significant (p > .05, Bonferroni adjusted), and if Smith’s test confirms the uni-dimensionality of this domain-level subtest solution [44], then this higher-order WHOQOL structure is confirmed. Misfit was investigated with domain level analyses [32] before repeating subtests. Lastly, Differential Item Functioning (DIF) examined how much subtests perform differently by gender, age and living alone.

Discriminative validity was estimated by comparing ‘extreme’ groups for socio-demographic characteristics. Means, standard deviations (SD) and a significance test (F) were calculated for domains. We were unable to make health status comparisons between sick and healthy subgroups, as all participants were diagnosed with chronic conditions. Construct validity was established by examining concurrent validity and through Rasch modelling. Concurrent validity involved correlating WHOQOL-Combi domains with EQ-5D-5L domains [23].

The sample contained 2833 participants, mean age 74.91 (SD 6.37), of which 46% were men and 52% women (2% missing from gender analysis). Thirty-four % lived alone, and 40% without a spouse/partner. Nineteen % had professional qualifications; 7% a university degree. Although 83% were retired from paid work, 5% received pay; the remainder did voluntary work, could not work, or were family carers. The majority identified as white British (94%), 3% as ‘other’ white, and 3% mixed ethnicity, Asian or Caribbean. Missing data was low and largely from social (4%) and spiritual (0.5%) domains.

Table 2 shows the 43 facet items tested for inclusion in the WHOQOL-Combi. A general WHOQOL-BREF facet (containing two items on overall QoL, and general health) was excluded from specific item analyses. The 41 specific facets were comprised of 24 WHOQOL-BREF items, and 17 items extracted from facets in three short-form modules: the WHOQOL-HIV-BREF [28], WHOQOL-Old [15] and WHOQOL-SRPB-BREF [18]. The preliminary WHOQOL-Combi was scored as: physical, psychological, social, environment and spiritual domains. As the WHOQOL-Old scored its modular items in an ‘Old’ domain [15], these were reallocated to a domain nominated during the pilot study, for reassessment. The position of 17 module items (italics) are shown in Table 2: sensory functioning (physical); achievement, fear of the future, autonomy and freedom (psychological); social inclusion, use of time, intimacy, and blame (social); none for environment, and eight new facets on spiritual, religious and personal beliefs (SRPB) with the single original WHOQOL-BREF spiritual item, totalling nine.

Table 3 displays findings that guided preliminary item reduction for 41 items. Item means > 3.0 indicated good QoL. Most showed acceptable skew and kurtosis (< 1.00). Of the items with mild deviations from normality, a large SD was found for spiritual connection; and blame had elevated skewness. Ceiling and floor effects were identified where > 10% of the sample endorsed one extreme response option on the 5-point scale. The items spiritual connection, spiritual strength, blame and faith showed unacceptable floor effects; social inclusion had a ceiling effect.

Table 3 shows Cronbach’s α when each item was removed. Where values matched or exceeded alpha for the full scale (.950), then the item risked being redundant and did not contribute to reliability [45]. Unacceptable items were: dependence on medication or treatment, blame, spiritual connection, inner strength and faith. For 24 WHOQOL-BREF specific items, ICR = .931, so both instruments showed excellent ICR.

Corrected item-total correlations were largely acceptable (r > .40) for facet items except spiritual connection, blame, faith and inner strength. Intimacy (.49) and sex life (.47) had similar item-total correlations. A strong association between achievement and the psychological domain (.82) confirmed its position. Items correlated more highly with an alternative domain were: freedom (.57), social inclusion (.57) and awe (.53) with the environment domain, and sensory functioning (.47) and use of time (.70) with the psychological domain. Fear of the future correlated − .38 with social QoL and with its designated domain. No WHOQOL-BREF items were more strongly associated with an alternate domain.

Item correlations with general QoL ranged from .12 (faith) to .67 (autonomy; achievement). Blame, spiritual connection, faith and inner strength revealed spiritual domain weakness. Strong associations with the core were found for autonomy, achievement, peace, hope and time use.

In summary, 17 facet items were selected from three international short-form modules. Five facets showed two or more measurement problems, so were candidates for exclusion: blame, fear of the future, spiritual connection, faith and spiritual strength. Both sex life and intimacy were retained. Although four WHOQOL-BREF core items were weak, they had been previously endorsed by cross-cultural consensus [5, 6], so were retained. Departures from the global structure are expected for individual cultures, so these UK results do not threaten the integrity of the international measure.

The CFA goodness-of-fit indices (Table 4) resulted from testing the baseline five-factor model (Model 1) based on known factor structures from previous instruments, with additional modular items assigned through conceptual similarity. Sample B results were slightly better than Sample A; the CFI > 0.950 criterion for Model 1 was met for Sample A, but RMSEA and SRMR still indicated substantial misfit in both. As modification indices and factor loading patterns did not reveal misfit sources, auxiliary Rasch results with Sample C were considered before continuing with CFA.

Rasch analysis commenced with an initial 41-item model (excluding general items), and no proposed domain structure. As Model A was a significant misfit (χ²(205) = 1794.47, p < .001), the subsequent step tested the five-factor structure using a subtest approach. When Model B combined all relevant items into domain subtests adequate fit was still not achieved (χ²(25) = 71.50, p < .001). In domain 5, three items on spiritual connection (Q9), faith (Q11), and spiritual strength (Q26) stood out with significantly elevated fit residuals (> 10.00, when the acceptable criterion is − 2.50 to 2.50). In the next iteration (Model C), these three spiritual items were discarded from the SRPB subtest. While overall fit was no longer significant (χ²(25) = 26.42, p > .05), there were substantial residual correlations. The final Model D, resolved residual correlations between domain 1 and 2 subtests by merging them into a further subtest. This strategy is consistent with previous WHOQOL-BREF research [42]. It achieved a non-significant fit (χ²(20) = 11.25, p > .05), evidencing an overall higher-order solution.

Model 2 (CFA) confirmed the tenability of the final model proposed after Rasch analysis where stand-out items (Q9, Q11, and Q26) had been removed from domain 5 (Table 4). Compared to Model 1, all fit indices substantially improved and now SRMR also met criteria of very good fit in both samples [41]. Although RMSEA still exceeded 0.060, such cut-off criteria may vary for asymptotically distribution-free methods, so it remains adequate in the context of other fit indices [41]. A marginal factor loading of 0.37 for blame (Q7), supported its deletion. Lastly, as fear of the future (Q8) showed a factor loading of 0.49, it was identified for removal to strengthen the measure. After deleting items 7 and 8, Model 3 continued to display excellent fit (Table 4; Fig. 1) and Rasch evidenced adequate unidimensional fit for this solution (χ²(20) = 11.36, p > .05), providing reassurance of robustness.

WHOQOL-Combi domains correlated moderately with EQ-5D-5L domains [23] in Table 5. This set of results was stronger than for the WHOQOL-BREF. Physical and psychological domains showed strongest concurrent validity (r > .50). Compared with the WHOQOL-BREF, WHOQOL-Combi domains were strengthened by the inclusion of new facets. This improvement was especially important for the social domain, which had been weakest. As expected, the WHOQOL-Combi spiritual domain associated most weakly with the EQ-5D-5L domains, highlighting its unusual contribution to generic assessment.

Final WHOQOL-Combi inter-domain correlations (Table 6) were moderate to good. The spiritual domain was strengthened by removing three items. The social domain was now moderately correlated with other domains. WHOQOL-Combi domains correlated moderately strongly with general QoL. Correlation range is weaker for the WHOQOL-BREF in Table 6, possibly due to fewer items.

Table 7 presents means and SDs for key socio-demographic characteristics. When comparing QoL across age bands (60–69, 70–79, ≥ 80), all domains except social, significantly decreased over time. Environmental QoL declined only in over 80′s. Subsequent analyses therefore controlled age as a covariate (ANCOVA). Men reported higher physical and psychological QoL; women higher social QoL. Poorer QoL in all domains except physical, was associated with living alone. Holding a professional qualification was consistently linked to higher QoL. Retirees without paid work had lower physical QoL than paid workers.

The WHOQOL-Combi SPSS syntax algorithm for scoring recodes five negatively worded items, calculates domain means from items, then multiplies by 4.00. We recommend calculating means only where individuals complete a majority of items in each domain (maximum two items missing per domain). The profile of domain scores is transformed (0–100). Permission to use the WHOQOL-Combi and its scoring syntax, is obtainable from Dr Christine Rowland (christine.rowland@manchester.ac.uk).