- Top of page
Objectives: The aim of this article is to map the European Organization for Research and Treatment of Cancer (EORTC) QLQ C-30 onto the EQ-5D measure to enable the estimation of health state values based on the EORTC QLQ C-30 data. The EORTC QLQ C-30 is of interest because it is the most commonly used instrument to measure the quality of life of cancer patients.
Methods: Regression analysis is used to establish the relationship between the two instruments. The performance of the model is assessed in terms of how well the responses to the EORTC QLQ C-30 predict the EQ-5D responses for a separate data set.
Results: The results showed that the model explaining EQ-5D values predicted well. All of the actual values were within the 95% confidence intervals of the predicted values. More importantly, predicted difference in quality-adjusted life-years (QALYs) between the arms of the trial was almost identical to the actual difference.
Conclusion: There is potential to estimate EQ-5D values using responses to the disease-specific EORTC QLQ C-30 measure of quality of life. Such potential implies that in studies that do not include disease-specific measures, it might still be possible to estimate QALYs.
- Top of page
The most popular form of economic evaluation, and the one recommended by the National Institute of Health and Clinical Excellence (NICE), is the cost-utility analysis . In this form of analysis, the outcomes are measured in terms of quality-adjusted life-years (QALYs), which combines life-years and quality of life into one index measure. Quality of life is measured on a scale ranging from 0 (death) to 1 (full health). Specific health states are located on this scale based on individuals' preferences elicited using methods such as the visual analog scale (VAS), time trade-off, and the standard gamble method. The majority view is that the preferences of the adult population (the taxpayers) are most appropriate given that they provide the funding for the National Health Service (NHS) . This means that population values need to be obtained for all of the relevant health states in the economic evaluation. Eliciting population's preferences for every single economic evaluation would clearly be prohibitively expensive and too time consuming. Fortunately, population values are readily available for several generic quality of life instruments including the EQ-5D  and the SF-6D. To estimate QALYs only the responses to say the EQ-5D instrument need to be elicited which can then be transformed into quality of life values by using the population “tariff.” Because these instruments use broad health domains, they can be applied to a wide range of conditions.
When the interest is in measuring the quality of life of patients, disease-specific instruments are often preferred to generic instruments . Disease-specific instruments focus on particular health problems and tend to be more sensitive to clinically important differences . Using the generic instruments preferred by economists, in addition to the disease-specific instruments, adds to the burden imposed on patients for completing questionnaires. Population values, however, are generally not available for disease-specific instruments. When using a disease-specific instrument only, QALYs can not be estimated and a cost-utility analysis cannot be performed.
One solution is to “map” disease-specific measures onto generic measures. The relationship between the two instruments can be established using regression analysis on existing data. Mapping would provide a model that would enable the estimation of QALYs in trials which did not include any generic instruments. Mapping between instruments has been attempted by previous studies with mixed results. Brazier  used regression analysis to attempt to map the disease-specific instrument, Impact of Weight on Quality of Life-Lite to the SF-6D. Although the models tended to perform well in terms of explanatory power, they were weak in terms of predictive ability. Tsuchiya et al.  explored possible ways of converting disease-specific measure (Asthma Quality of Life Questionnaire) to preference-based measure (EQ-5D) using a range of ordinary least squares (OLS) regression models. Although fairly stable functional relationships were achieved, their results showed limited ability to predict accurate EQ-5D values at an individual level. Brennan and Spencer  mapped the Oral Health Impact Profile onto the EQ-5D. The predicted values for the hold-out sample were substantially higher than the actual EQ-5D values.
The aim of this article is to map the European Organization for Research and Treatment of Cancer (EORTC) QLQ C-30 onto the EQ-5D measure to enable the estimation of health state values based on the EORTC QLQ C-30 data. The EORTC QLQ C-30 is of interest because it is the most commonly used instrument to measure the quality of life of cancer patients . It has been translated into more than 65 languages and is used widely internationally . Given its extensive use, successful mapping of this measure onto a generic measure could provide a substantial benefit. It is also hypothesized that mapping is likely to be more successful because the instrument includes more general domains in addition to cancer-specific symptoms. Some of these more general domains are similar to the dimensions of the EQ-5D. Regression analysis is used to establish the relationship between the two instruments, and the performance of the model is assessed in terms of how well the responses to the EORTC QLQ C-30 predict the EQ-5D responses for a separate data set.
- Top of page
Correlations between the individual responses for each of the five EQ-5D questions and for each of the 30 EORTC QLQ C-30 questions were first explored. Statistically significant correlations were found between scores for a number of dimensions which measure similar aspects of quality of life.
Table 2 shows the OLS regression results for the EQ-5D values. All statistical significance levels are at the 95% level. The coefficient on global health status is positive and statistically significant (P = 0.003). Three of the function scales are statistically significant, namely role functioning (P ≤ 0.000), emotional functioning (P ≤ 0.000), and cognitive functioning (P = 0.027). As hypothesized, pain is a statistically significant symptom scale (P ≤ 0.000). Fatigue is also statistically significant (P ≤ 0.000). Table 3 shows the ordered probit regression results for each of the five EQ-5D dimensions. In the case of mobility, role and cognitive functioning are statically significant (both P ≤ 0.000), as well as fatigue (P = 0.003) and dyspnea (P = 0.004). The same variables, apart from dyspnea, are statistically significant with respect to self-care [role functioning (P ≤ 0.000), cognitive functioning (P = 0.004), fatigue (P = 0.013)] with the addition of financial problems (P = 0.036). In the case of usual activities, global health status (P ≤ 0.000), role functioning (P ≤ 0.000), and fatigue (P = 0.004) are significant. In the case of pain, emotional functioning, fatigue, and pain are significant [(P = 0.002), (P = 0.03) and (P ≤ 0.000), respectively]. Finally, physical and emotional functioning are statistically significant with respect to anxiety and depression [(P = 0.006) and (P ≤ 0.000), respectively]. Models were also estimated using raw responses to the 31 EORTC QLQ C-30 questions. Similar results were found (full results can be obtained from the authors). In terms of the interaction terms between time and EORTC QLQ C-30 scales, only the interaction term with pain was statistically significant. This implies that, with the exception of one scale, the relationship between the scales and the EQ-5D values does not change as patients become more familiar with the instruments.
Table 2. Regression results for EQ-5D values
|Global health status*||0.0016||3.06|
Table 3. Regression results for EQ-5D dimensions
| ||Mobility||Self-care||Usual activities||Pain||Anxiety/depression|
|Global health status||−0.0048||−1.46||0.0027||0.90||−0.0110||−3.81||−0.0001||−0.04||−0.0009||−0.27|
|McKelvey and Zavoina's R2||0.473||0.370||0.621||0.668||0.565|
|Goodman and Kruskal λ||0.289||0.090||0.442||0.425||0.528|
The goodness of fit statistics for the different regression models are reported in Tables 2 and 3. The adjusted R2 for the OLS regression is relatively high, namely 0.611. The McKelvey and Zavoina's R2 ranges from 0.370 to 0.668 for the ordered probit regression. Most models have relatively high R2 indicating that a relatively large part of the variation is explained by the models. The models for mobility and self-care seem to perform less well in terms of goodness of fit. The Goodman and Kruskal λ for the ordered probit regression results, reported in Table 3, show that the models reduce prediction error by about 35% on average. The reduction in prediction error is highest for the anxiety and depression dimension (52.8%) and lowest for the self-care dimension (9%).
Table 4 shows the predicted EQ-5D values for the PRIME data set. The EQ-5D values predicted on the basis of the OLS regression are reported first. In terms of the overall sample, the predicted EQ-5D value is higher than the actual EQ-5D value. The mean difference is relatively small, namely 0.014. The actual EQ-5D value of 0.760 is just inside the 95% CI of the predicted EQ-5D value which ranges from 0.760 to 0.788. The difference of 0.014 is below an MCID of 0.03. Actual and predicted EQ-5D values are also reported for each time point. The difference between actual and predicted values seems to increase over time. Nevertheless, the actual EQ-5D values are all within the 95% CI of the predicted EQ-5D values. Moreover, all of the differences are below an MCID of 0.03. The mean QALYs are also reported in Table 4. Given that the predicted EQ-5D values are lower than the actual EQ-5D values, it follows that the mean QALYs estimated from predicted EQ-5D values are also higher than the mean QALYs estimated from actual EQ-5D values. The difference is equal to 0.018 QALYs. The actual mean QALYs is well within the 95% CI of the predicted QALYs.
Table 4. Predicted versus actual EQ-5D values and quality-adjusted life-years (QALYs)
| ||Actual||Predicted (95% CI)||Difference||Relative prediction error|
|Predictions based on predicted values from OLS regression|
|Mean EQ-5D||0.760||0.774 (0.760–0788)||0.014||1.84%|
|Mean EQ-5D by time point|
| Baseline||0.765||0.772 (0.746–0.798)||0.007||0.92%|
| 3.5 months||0.778||0.780 (0.752–0.807)||0.002||0.26%|
| 9 months||0.756||0.780 (0.751–0.808)||0.024||3.17%|
| 15 months||0.741||0.764 (0.735–0.794)||0.023||3.10%|
|Mean QALYs||0.945||0.963 (0.933–0.9937)||0.018||1.90%|
|Predictions based on predicted dimensions from ordered probit regressions|
|Mean EQ-5D||0.760||0.810 (0.797–0.822)||0.050||6.58%|
|Mean EQ-5D by time point|
| Baseline||0.765||0.801 (0.776–0.826)||0.035||4.71%|
| 3.5 months||0.778||0.814 (0.789–0.839)||0.036||4.63%|
| 9 months||0.756||0.824 (0.800–0.848)||0.068||8.99%|
| 15 months||0.741||0.800 (0.773–0.828)||0.059||7.96%|
|Mean QALYs||0.945||1.017 (0.980–1.054)||0.072||7.62%|
Table 4 also shows the EQ-5D values based on predicted EQ-5D dimensions from the ordered probit regression results. The models seem to predict less well as the difference between predicted and actual EQ-5D values is much larger, namely 0.050. Moreover, the actual value of 0.760 is outwith the 95% CI of the predicted EQ-5D value which ranges from 0.797 to 0.822. The difference of 0.050 is also greater than the MCID of 0.03. The difference between actual and predicted values again seems to increase over time. The models also predict less well in terms of QALYs as the difference between predicted and actual QALYs is equal to 0.072.
Results by Arm of the Trial
As highlighted before, the real question is whether the differences in predicted values between the arms of the trial are similar to the differences in actual values. That is, would the same conclusions have been reached regarding effectiveness and cost-effectiveness? In the PRIME study, no statistically significant differences in either the EQ-5D values or QALYs were found between the two arms of the trial. Table 5 shows the EQ-5D values and QALYs by arm of the trial based on the predictions from the OLS regression model. None of the differences in the predicted EQ-5D values between the two arms of trial are statistically significant. Moreover, the difference in predicted QALYs between the two arms of the trial is almost identical to the actual difference in QALYs, namely 0.017 versus 0.019.
Table 5. Actual and predicted EQ-5D values and quality-adjusted life-years (QALYs) by arm of the trial
| ||Radiotherapy||No radiotherapy||Difference between arms||t-value|
- Top of page
The aim of this article was to map the EORTC QLQ C-30 onto the EQ-5D measure to enable the estimation of health state values based on the EORTC QLQ C-30 data. Both the EQ-5D values as well as the EQ-5D dimensions were modeled as a function of the EORTC QLQ C-30 global health status, functioning, and symptom scales. The predictive ability of the model was tested on a separate data set. The results showed that the model explaining EQ-5D values predicted well. All of the actual values were within the 95% CI of the predicted values. More importantly, the predicted difference in QALYs between the arms of the trial was almost identical to the actual difference in QALYs. Using predicted rather than actual values would therefore not have changed the conclusions regarding the cost-effectiveness of the intervention. This indicates that there is potential to estimate EQ-5D values using responses to the disease-specific EORTC QLQ C-30 measure of quality of life. Such potential implies that in studies that do not include disease-specific measures, it might still be possible to estimate QALYs.
The prediction of the EQ-5D values based on predicted levels for each of the dimensions was less successful. This is most likely because of the fact that the analysis did not take any correlation between the dimensions into account because the dimensions were modeled separately. The use of multivariate analysis might improve the predictive ability. Unfortunately, multivariate ordered probit analysis is currently not available within standard software packages such as STATA and Limdep (Econometric Software, Plainview, NY).
The mapping was very successful in this study in that the model predicted the EQ-5D levels and QALYs well for a separate data set. The question arises whether these results are generalizable. The data used were collected from patients with esophageal cancer. An advantage of this data set was that there were sufficient numbers of patients in each of the levels of the five EQ-5D dimensions. This patient group, however, is unlikely to be representative of the “average” cancer patient group. As well as the type and stage of cancer factors such as age and sex may affect the predictive performance of the model. Although the results showed that the model did predict well for a group of patients with different type of cancer, namely breast cancer, the average age of the patients was similar in the two data sets. Further research exploring predictive performance for different patient groups is clearly required before the application of the model should become a recommended approach for converting the EORTC QLQ-C30 data into EQ-5D values.