Mapping The EORTC QLQ-C30 And QLQ-H&N35 To The EQ-5D For ...

Model development

Development of the best fitting mapping model was conducted in two phases. In the first phase, QLQ scales were preselected as potential predictors, resulting in three predictor sets based on theory (Set 1) and combined theory- and data-driven considerations (Set 2 and 3). This was done to retain parsimony of the model. In the second phase, statistical analyses were performed in three consecutive steps in order to select a model with the best fit, considering the different predictor sets. A schematic overview is given in Fig 1.

Download:

• PPTPowerPoint slide
• PNGlarger image
• TIFForiginal image

Fig 1. Flowchart of model development.

The grey rectangles display the three predictor sets with their criteria; the white rectangles the different models. The squares indicate the assessment of the models during model comparison; the rhombuses indicate the decision-making in the process. Assessment of the model performance is displayed with a dotted line. The EORTC QLQ-C30 and QLQ-H&N35 scales included in the predictor sets are highlighted in grey circles. The scales that were excluded are colored white. Abbreviations: AL, appetite loss; Cough, coughing; CF, cognitive functioning; CP, constipation; DH, diarrhea; DP, dyspnea; DM, dry mouth; EF, emotional functioning; EORTC, European Organization for Research and Treatment of Cancer; EQ-5D-3L, three-level EuroQol five-dimensional questionnaire; FD, financial difficulties; FG, fatigue; FI, felt ill; FT, feeding tube; GHS/QoL, global health status/quality of life; HRQoL, health-related quality of life; IS, insomnia; LS, less sexuality; NT, nutritional supplements; NV, nausea and vomiting; OLS, ordinary least-squares, OM, opening mouth; PF, physical functioning; PK, pain killers; Quality of Life Questionnaire-Core 30; QLQ-H&N35, Quality of Life Questionnaire-Head and Neck35; RF, role functioning; SC, trouble with social contact; SE, trouble with social eating; SpP, speech problems; SeP, senses problems; SL, swallowing; SF, social functioning; SS, sticky saliva; WG, weight gain; WL, weight loss.

https://doi.org/10.1371/journal.pone.0226077.g001

In this Methods section, both phases are described separately.

Preselection of QLQ scales.

Three sets of HRQoL outcomes were selected as potential predictors to map onto the EQ-5D. The first set of predictors was selected from EORTC QLQ-C30 scales based on the correspondence of these scales with the EQ-5D dimensions and its underlying construct. Correspondence was evaluated by matching EORTC QLQ-C30 scales to EQ-5D dimensions based on degree of overlap in content between items in both questionnaires. This predictor set functioned as a base for the model, and was retained in the model throughout the predictor selection from Set 2 and Set 3.

The second set of predictors included a number of the remaining EORTC QLQ-C30 scales, which were selected based on their ability to reflect on changes over time (a.k.a. responsiveness). A literature search was conducted to estimate the responsiveness of the scales. Studies were considered eligible when HNC patients had undergone a surgical and/or organ sparing intervention, the EORTC QLQ-C30 was completed at least twice by these patients at various time points within a timeframe of at least three months in which responsiveness of QoL was expected based on the treatment, and the sample size was ≥100. The search was restricted to studies published between January 2012 and July 2017. From the included studies, effect sizes (ES) were calculated for each EORTC QLQ-C30 scale, by dividing the mean difference of the score by the pooled standard deviation, and compared with the average ES of the EQ-5D calculated with the data used in this study [16].

A third set containing individual predictors consisting of EORTC QLQ-H&N35 scales was developed to explore whether use of HNC-specific HRQoL outcomes could improve the fit of the mapping model. Scales were assessed on intercorrelation, to limit overfitting as well as prevent multicollinearity. If a Pearson correlation coefficient of ≥ 0.7 between two individual EORTC QLQ-H&N35 scales was present, one of the scales was excluded based on theoretical considerations. Of the remaining EORTC QLQ-H&N35 scales, those that correlated with the dependent outcome (Pearson correlation coefficient ≥ 0.3) were included in the third set of predictors. The data showed no outliers, but were not entirely normally distributed. For completeness, we re-ran the analyses on the basis of Spearman's correlation results.

The predictors were tested one by one for their additional value to the model.

Statistical analysis.

The statistical analysis was conducted in three steps (Fig 1). The first step consisted of selecting the best fitting regression method using only the first set of predictors as input for the models. We considered four commonly used regression models:

1. Regression analysis using an OLS estimator (Model 1a);
2. Mixed-effects modeling approach (Model 1b) using a maximum likelihood solution, with a random intercept to take into account mutual correlation within repeated measurements present in our data;
3. Cox regression (Model 1c) with ‘censoring’ of all EQ-5D utility index scores <1.
4. Classical beta regression (Model 1d) modeling the dependent variable y in a unit interval 0 < y < 1. In order to include the full health (utility value of 1) in this interval, a transformation of y was applied [17, 18]:

in which y is the utility and n is the sample size. To create a generalized linear model, we applied the logit link function [17].

To select the overall best statistical approach, we compared the four models using the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) [19, 20]. The AIC and BIC can be used to compare non-nested models and reflect the relative quality of the models by assessing the goodness of fit while penalizing the number of model parameters. Models with lower BIC or AIC values are considered to be better fitting models, although there is much debate about how to interpret the numerical differences in outcomes between models. Published rules of thumb are: a between model difference in the AIC or BIC of 0 to 2 is considered to be weak, 2 to 6 to be positive, 6 to 10 to be strong and above 10 to be very strong [20, 21]. In this study, we considered the regression method with the lowest AIC and BIC to be the most appropriate base model to use for the subsequent statistical steps.

In the second step, we extended the base model selected in step 1 with the second set of predictors containing all responsive EORTC QLQ-C30 scales (Model 2). The added value of these predictors was assessed using the AIC, BIC and likelihood-ratio (LR) test with a cutoff p-value of 0.05. In case of significant outcome (p<0.05) of the LR test and lower AIC and BIC values, we used manual stepwise backward elimination of predictors of the second set, for parsimony of the model. Backward elimination was based on the p-value of the coefficients (using a cutoff of 0.1).

In step 3, we explored the added value of the selected EORTC QLQ-H&N35 predictors (third set) for each variable separately (Model 3). Each of these models was compared to the model with the best fit so far obtained after step 2. The same model fit statistics were used as described in step 2.