Psychometric properties of the Spanish version of the European Organization for Research and Treatment of Cancer Quality of Life Questionnaire (EORTC QLQ-C30)

The EORTC QLQ-C30 version 3.0 [1] includes 30 items comprising five multi-item functional scales: physical (PF), role (RF), cognitive (CF), emotional (EF), and social (SF); 3 multi-item symptom scales: fatigue (FA), nausea and vomiting (NV), and pain (PA); 6, single-item additional symptoms commonly reported by cancer patients: dyspnea (DY), insomnia (SL), appetite loss (AP), constipation (CO), and diarrhea (DI), as well as the perceived financial impact (FI) of the disease and treatment, and a two-item global health status/QoL scale. The questionnaire uses a 1-week time frame and 4-point Likert-type response scales (“not at all,” “a little,” “quite a bit,” and “very much”), with the exception of the two items of the overall QL scale, which are rated 1–7. The Spanish version of EORTC QLQ-C30 presented by the EORTC was used for data collection. A sample can be downloaded at the following URL: http://​groups.​eortc.​be/​quol/​erotc-qlq-c30. Following standard European Organization for Research on the Treatment of Cancer procedures, all scores were linearly converted into a 0 to 100 scale, with higher scores indicating a higher level of functioning, better HRQOL for global health status, and more severe symptoms [1]. It was administered to the patients prior to initiating chemotherapy and, again, 6 months later.

The Brief symptom Inventory (BSI-18) consists of 18 items categorized into three dimensions (depression, anxiety, and psychological distress) and rated on a five-point scale [20] and has been adapted to Spanish cancer patients [21]. Raw scores are converted to T-scores based on sex-specific normative data. Alpha coefficients were between 0.75 and 0.88 for the Spanish version among cancer patients [21].

Patient and tumor characteristics were obtained from the interview and clinical history. The following variables were compiled: sex, age, marital status, education level, employment status, tumor site, stage, and treatment.

The data are derived from a prospective cohort of patients with non-metastatic colon cancer from the multicohort NEOCOPING study promoted by the Continuous Care Group of the Spanish Society of Medical Oncology (SEOM) and conducted in 17 Spanish medical oncology departments. Individuals > 18 years of age with resected, non-metastatic colon cancer and eligible for adjuvant chemotherapy were consecutively enrolled. Those who had received preoperative chemotherapy or radiotherapy, and those with any condition that impeded comprehension of or participation in the study were excluded. Patients completed the questionnaires at home and returned them at the following appointment for adjuvant cancer treatment to be initiated. Subjects were followed until adjuvant treatment was ended and questionnaires were filled out again after approximately 6 moths. The flowchart comprising inclusion and exclusion criteria of participants is given in Fig. 1.

Ethical approval was obtained by the Institutional Research Ethics Committee of each hospital and the Spanish Agency for Medicines and Health Products (AEMPS) (number L34LM-MM2GH-Y925U-RJDHQ); informed consent for voluntary participation was obtained in writing from all subjects prior to performing any study procedure. STROBE guidelines were used to ensure the reporting of this study [22].

A five-stage approach was used in the analyses. First, the dimensionality and structure of the EORTC were assessed in the entire sample through exploratory factor analysis. To evaluate dimensionality, we computed parallel analysis (PA); [23] and essential unidimensional indices [24]: unidimensional congruence (Unico), explained common variance (ECV), and mean of item residual absolute loadings (MIREAL). PA compares the dimensions obtained in the sample dataset to the ones obtained in random datasets in which the null model holds: the number of dimensions in the sample dataset that account for more common variance than the ones based on random datasets is the number of factor advised to be extracted in the sample dataset. The indices to assess essential unidimensionality help to decide whether a strong, dominant factor exists: when the value of the three indices computed crosses their particular threshold, the unidimensional factor solution is advised. The threshold values that determine that the factor solution is essentially unidimensional were: UniCo > 0.95, ECV > 0.85, and MIREAL < 0.30. As the outcome of these indices were inconclusive, we computed an exploratory bifactor model to decide between a single-factor model and a two-factor model [25]. The bifactor model allows the hypothesis of a general factor, while addition common variance is modeled using group factors. When a formal hypothesis does not exist for the overall bifactor model (i.e., the general factor and a particular number of group factors), Pure Exploratory Bifactor models (PEBI) can be inspected [25]. Second, inasmuch as a clear, cross-validation-resistant, factorial structure was finally identified in the first stage, multiple-group confirmatory factor analyses (CFAs) were performed to gauge measurement invariance in groups defined by gender and tumor localization. Third, the reliability and appropriateness for individual assessment of the scores derived from the chosen FA solution were examined using a multi-faceted approach [26]. Fourth, the sensitivity of treatment-induced change was assessed by fitting two-wave (pretest, posttest) structural equation modeling. Finally, convergent validity was appraised by means of a full structural model in which the CFA was extended to include external variables.

Table 1 shows the means and standard deviations of the Spanish version of QLQ-C30 (version 3.0) administered pre-chemotherapy and 6 months later. Mean item scores ranged from 77.1 (social) to 86.0 (physical) on the functional scales. The functional scale score distribution displayed negative asymmetry on all the scales: scores often exceeded 70.0 (that is, many patients exhibited acceptable functioning overall). The symptom scales and the single-item measures reveals a positive asymmetrical distribution, with scores close to 0 (no symptoms), except for insomnia and fatigue with mean scores of 30.1 and 20.0, respectively.

The EORTC items are ordered-categorical with only four response points and, in this study, were administered to a large sample. Furthermore, the results of the previous section revealed that the distribution of some items is greatly skewed in both directions. In this scenario, the non-linear item factor analysis model, based on an underlying variables approach, (see e.g., [27]) was deemed the most suitable to fit the EORTC data. To fit this model, we followed the procedure described in [28]: the inter-item correlation to be analyzed was the polychoric correlation matrix, and was fitted using the robust unweighted least squares method. This general setting was applied in all structural analyses conducted in this study (EFAs, CFAs, two-wave, and extended validity model) and proved feasible in all cases. EFAs were performed using the FACTOR¹ software [29]. The remaining analyses were conducted using Mplus version 5.1 [30].

The data were found to be appropriate to fit the EFAs (and subsequent analyses). The KMO index (0.810) and Barlett’s test (χ2 = 10.582.3, df = 435, p < 0.001) suggested that there was enough inter-item consistency to fit the FA model.

PA results suggested a two-factor solution as the most suitable for the sample data. However, when a single factor was extracted, goodness-of-fit indices indicated that the model was almost marginally acceptable (RMSEA = 0.072; CFI = 0.970; GFI = 0.956). Furthermore, the values of indices assessing essential unidimensionality (Unico = 0.956; ECV = 0.840, and MINREAL = 0.217) suggested that the single-factor solution is close to essential unidimensionally, but not perfectly, and that fitting more than one dimension is advisable. The situation in which a set of items is essentially unidimensional, but in which a (generally small) subgroup of items share specific variance, is relatively common in practice. This alternative model can be examined using an exploratory bifactor model (PEBI). The outcome of PEBI in our sample data is presented in Table 2.

As depicted in the table, all the items in the scale have a substantial loading value in the general factor (Quality of Life). Moreover, the 5 items related to physical function, the 2 items concerning role function and 2 items addressing symptoms of fatigue define a group factor (Physical Health, loading values > 0.40). This factor combines aspects connected to physical status and fatigue-related symptoms, and could be sensitive to treatment toxicity. As for model data fit, goodness-of-fit indices suggested that the model could now be deemed acceptable (RMSEA = 0.053; CFI = 0.985; GFI = 0.977).

The stability and replicability of the proposed solution, as well as its appropriateness for a confirmatory solution, were examined using a double cross-validation schema. Once the sample had been randomly split into two halves, the factorial congruences between the obtained solution in each half sample were always above 0.98 for both the general and group factors. Furthermore, when a bifactor CFA solution was specified in the second half sample, based on the exploratory bifactor solution obtained in the first half, the specified CFA solution was always the same and fitted well. Given these results, a confirmatory bifactor solution in which the general factor was defined by all the EORTC items and the group factor (physical health) was defined by the 9 items described above was considered the most appropriate for the entire group, and served as the basis for all the subsequent CFA analyses.

Based on the common CFA solution described above, strong invariance multiple-group solutions were fitted using sex and tumor site as grouping variables. As discussed in detail elsewhere [21], if strong measurement invariance is achieved, it can be assumed that the same two factors are measured in the different groups and that the EORTC items function in the same way in these groups. Thus, differences in mean group scores can be validly interpreted as reflecting ‘true’ group differences in the general and group dimensions being measured. The strong invariance solution was identified by using the delta parameterization for the residuals (see [30]) and fixing the factor means and variances to zero and 1, respectively, in the first group. Preliminary, weaker-invariance solutions (non-invariant loadings, thresholds, or both) were tried and we found that the strong solution fitted better in relative terms (as indicated by the RMSEA) and also according to the parsimony indices (the Bayesian information criterion). Because the strong solution also fitted acceptably well in overall terms (see Table 3), we only report the group estimates derived from it, which are displayed in Table 3.

So as to interpret the mean differences in the table and for identification purposes, the means are always fixed to zero in the first group and are freely estimated in the remaining. The results can be summarized as follows. First, no mean differences in any of the factors (Quality of Life and Physical Health) were detected by sex. Second, mean differences in both factors were found for tumor location. This might be due to the fact that there are patients with chemotherapy regimens that entail more side effects than other and with the possible effect of surgery prior to treatment, given that the sample consisted of individuals with different types of tumors.

As the results uphold the assumption that the EORTC items function in the same way in both groups, accuracy and appropriateness measures of the EORTC scores can be obtained for the whole group. First, marginal reliability measures were attained for two types of scores: (a) factor score estimates obtained by using all the available information in the CFA solution and (b) the simplest raw scores achieved by adding all the EORTC items (Quality of Life scale) and the salient items in the group factor (Physical Health scale). Reliability estimates for the factor scores were 0.96 (QoL) and 0.88 (Physical Health). The corresponding reliability estimates (omega coefficients) for the raw scores were 0.94 and 0.86. Hence, for both factors, all the scoring schemas are highly reliable, particularly for the scores intended to quantify the general factor. As expected, the most informative factor score estimates are more reliable than raw scores, albeit the latter can already be considered reliable enough to be used in clinical assessment. To assess this issue further, the ordinal coefficients of fidelity were also computed for the raw scores. These coefficients estimate the correlation between each set of raw scores and the factor they intend to measure [31]. Fidelity estimates were 0.97 (QoL) and 0.91 (Physical Health). To sum up, QoL and Physical Health scale scores can be obtained for the EORTC items and these scores are both accurate and representative of the factors they seek to measure.

The bifactor CFA solution used in the previous section was extended to a two-wave panel model (Pretest, Posttest) using similar constraints as those in the previous multi-group analyses. More specifically, strong invariance was imposed here but, instead of across groups, it was over time. If this restriction is found to be acceptable, inasmuch as the items function with the same properties at both time points, mean group changes in both factors can be validly and univocally assessed. Thus, strong invariance over time was assumed and the mean of each dimension at Time 1 was set to zero. Results are illustrated in Table 4 and can be summarized as follows. First, the models exhibited an acceptable fit. Second, significant pretest–posttest changes were observed for both factors, especially the QoL factor, that consisted of decreased group means on posttest measures. While these changes are highly significant (given the high power of the test), they would be qualified as small in terms of effect sizes (Cohen’s d).

The basic bifactor CFA solution used in the previous analyses was extended here to include three BSI scale scores (depression, anxiety, and total) that were used as external variables. Regarding the substantive results, psychological distress and depression are clearly the two factors that are most strongly and negatively correlated with the QoL factor. Anxiety was the scale that most closely correlated with the physical health factor. Since total BSI score is the sum of the other two subscale scores, a separate model was fitted for the former. The fit of the extended models was very similar to that of the base model; therefore, fit results are not reported here. The validity estimates are summarized in Table 5 and include disattenuated standardized regression coefficients (i.e., ß weights) and squared multiple correlations for each external variable. The first coefficient quantifies the predictive power of each factor (QoL and Physical Health) relative to the external variable. The squared multiple R quantifies the variance of the external variable that can be explained by both factors together.

To facilitate clinical interpretation by psychologists and oncologists, we have introduced a normative table to convert raw scores to T-scores and centiles. The overall quality of life (QOL) and physical health scores are interpreted as the higher the score, the better the quality of life and physical status, see Table 6.