Mapping the FACT-G cancer-specific quality of life instrument to the EQ-5D and SF-6D

Cancer is now the leading cause of death in many developed countries including Canada [1]. It accounted for 30 per cent of all deaths in 2008, followed by heart disease (21%) and stroke (6%) [2]. Over the past 50 years, survival rates for many forms of cancer have improved markedly due to advances in surgery, radiation therapy, and chemotherapy [1]. However, emerging cancer therapies often come at high costs. This not only imposes an increasing financial burden on health care systems but also raises questions about how to assess best value for money [3, 4]. Economic evaluation is used by decision-makers in Canada and other countries to inform the allocation of scarce resources across health care interventions. Specifically, Cost-Utility Analysis (CUA) is increasingly becoming the main approach used to measure and value the impacts of interventions. CUA uses quality-adjusted life years (QALYs) to measure health outcomes by combining survival and health-related quality of life (HRQoL) into a single index. The use of CUA is recommended by the Canadian Agency for Drugs and Technologies in Health (CADTH), as well as peak regulatory bodies in Australia, the UK, and elsewhere in Europe [5‐10].

The Functional Assessment of Cancer Therapy-General (FACT-G) is one of the most widely used cancer-specific HRQoL instrument [11]. It has been validated across a wide range of different types of cancer patients, cultures, and languages, and can be used to assess the impacts of cancer and its treatment on the physical and psycho-social well-being of patients [11]. In spite of its widespread use in clinical trials, responses from the FACT-G cannot be readily used in economic evaluations because it does not provide a single preference-based index of HRQoL suitable for use in CUA. Health utilities, elicited using, for example, the EuroQol-5D (EQ-5D) and the Short Form-6D (SF-6D), provide preference weights from the general population that can be used to calculate QALYs for CUA; this can inform decisions of health-related resource allocation [12, 13]. However preference-based instruments typically cover broader health dimensions, and are not often administered in cancer clinical trials because many of the dimensions may neither be relevant nor sensitive to treatment effects [14].

When preference-based instruments are not administered in clinical trials, comparing effectiveness across different interventions becomes cumbersome. One option to overcome this limitation is to map, or ‘cross-walk’, the responses from a disease-specific instrument into a preference-based measure using regression modeling [15, 16]. In recent years, there has been growing interest in mapping in the literature, with a number of publications that have mapped responses on disease-specific instruments to generic preference-based measures [17‐26]; these studies are further described in two recent review articles [15‐27]. In the area of oncology, we have identified 13 studies that have mapped responses from cancer-specific instruments to yield utilities [28].

Our review revealed that only three studies mapped FACT-G responses to preference-based instruments. Two of these studies used responses from patients with a single tumor type (e.g., colorectal [29], prostate [30]; whereas, the remaining study performed the mapping exercise using a general cancer population with the EQ-5D as the target measure [22]. Given that the HRQoL of patients with different types of cancer (and different stages of disease) can vary considerably, it is possible that the results of mapping exercises may differ depending on the type of cancer patients included in the study. As such, further investigation is needed to explore whether a more appropriate mapping function can be developed using responses from patients with different tumor types compared to a specific population. The resulting algorithm may better facilitate future CUAs because it may be more applicable in general cancer populations.

Thus, the objective of this study is to develop a mapping function to impute both EQ-5D and SF-6D health utility values from responses on the FACT-G using a sample of patients with cancer from three different sites (breast, colorectal, and lung) with a range of disease severity. The influence of the tumor sites, as well as disease severity, was investigated.

HRQoL information was obtained from 367 patients with breast (n = 140, 38%), colorectal (n = 113, 31%), and lung (n = 114, 31%) cancer. Most of the questionnaires were completed face-to-face (n = 352, 96%). The average age of participants was 58.7 years (SD 11.5 years); 67% were women (Table 1). The patients reported their ethnic compositions as British/Irish (n = 164, 46%), Chinese (n = 53, 15%), or Other (n = 143, 40%). A good representation of disease severity, in terms of cancer stage and ECOG score, was observed in the study sample. There was no statistically significant differences in the demographic and clinical characteristics of the patients in the development (n = 184) and cross-validation samples (n = 183). The majority of patients answered all the items on the HRQoL instruments. The EQ-5D utility indices were negatively skewed, while the SF-6D utilities were mildly skewed; a ceiling effect was observed for the EQ-5D responses (23.4%) (Table 2). The FACT-G score was highly correlated with the EQ-5D (rho = 0.649) and the SF-6D (rho = 0.714) (Table 3).

Our results demonstrate that the estimation of both EQ-5D and SF-6D utility indices using the FACT-G responses from breast, lung and colorectal cancer patients can be achieved. We had a relatively large sample of patients with different disease severity levels. Our findings suggested that the CLAD and GLM models are best used to predict EQ-5D and SF-6D utilities, respectively.

Comparing predictions of the preference-based scores using FACT-G scales indicated that a better performance was observed for the SF-6D (adjusted R² = 0.651, MAE = 0.052, RMSE = 0.062) as compared to the EQ-5D (adjusted R² = 0.469, MAE = 0.062, RMSE = 0.104); these results are in agreement with previous studies [11, 25‐39]. Such studies reported that the physical and functional subscales were significant predictors of the utility scores. For both preference-based indices, including patients’ demographic and clinical characteristics improved explanatory power of the models.

Taking the RSME criteria as measures of prediction, the GLM model using FACT-G domain scores plus patient demographic (age, sex) and clinical (cancer stage) characteristics fitted the SF-6D data well. The GLM model was chosen despite non-Gaussian residuals as they produced the best predictions of utility scores on a natural scale. On the other hand, the CLAD model fitted the EQ-5D data well as compared to the OLS and GLM and it was investigated further.

The PWB and FWB sub-scales of the FACT-G were statistically significant predictors for both preference-based instruments. The EWB reached statistical significance only in the EQ-5D model. The SWB score of the FACT-G was not significantly associated with the EQ-5D and SF-6D and therefore, was not included in the regression model. Similar approaches have been used in previous studies and the similar results have been reported [22]. It is likely that the other dimensions can capture most of the information of other dimensions; as a result, additional information would not be needed. It is possible that only a negligible relationship between those sub-scales and HRQoL exists, and other aspects of illness are captured by them.

For both instruments, Model 3 produced more accurate predictions than Models 1 and 2 demonstrating that including patient characteristics improves predictions. There was a strong correlation between predicted and observed utilities for all three models. Utility scores derived from the FACT-G using the GLM approach were highly correlated with the observed scores (rho = 0.831 for the SF-6D).

The purpose of the current study is to examine the differential validity and prediction of the utility scores using cancer-specific instrument from a sample of breast, lung, and colorectal cancer patients. Our results showed that this is feasible. We also explored this further by assessing differences between observed and predicted utilities by disease severity. Our findings demonstrate that there are similar patterns of prediction by cancer stage, though there is a trend of over predicting values for patients with greater disease severity. Future research is needed to discuss utilization of mixed modeling for patients with different disease severities. While we anticipated that constructing a mapping algorithm using patients with three different tumor sites would be more applicable to facilitate future CUAs in general cancer populations, we found that adjusting for the type and stage of cancer increased the prediction of the model demonstrated by higher R².

The results from this current study are comparable to the one other study that mapped FACT-G responses to EQ-5D in a general cancer population [22]. Similar mean values for the EQ-5D and the FACT-G were observed in the current and previous studies (EQ-5D 0.82 versus 0.81; FACT-G 79.0 versus 81.1, respectively). For the current study, we observed a lower ceiling effect for the EQ-5D (23.4% versus 33.3%) and higher floor effect 0.28%. We observed similar goodness-of-fit values for our OLS Model 3 (R² = 0.47) compared to the previous study (R² = 0.45). In terms of accuracy, we reached same conclusion that predictions using CLAD was more accurate for the EQ-5D compared to the OLS; this was further confirmed when we ran the GLM model and found that the CLAD performs better.

We examined accuracy of predictions across the disease stage measures used in this study (cancer stage and the ECOG). Mean value predictions indicated that overall the GLM model performs better for the SF-6D and the CLAD model performs better for the EQ-5D compared to the alternate methods used in this study. All regression models tend to over predict lower observed utility scores and under predict higher scores.

The poor ability of the FACT-G to predict EQ-5D utilities may be influenced by the brevity of the preference-based instrument (i.e., three levels to define the five dimensions). Our previous work revealed that the EQ-5D was less discriminative between different levels of disease severities when compared to the FACT-G and SF-6D [46]. While the FACT-G, EQ-5D, and SF-6D assesses an individual’s HRQoL, the instruments are composed of different dimensions. For example, the FACT-G and SF-6D has a dimension describing vitality (i.e., an important health outcome for cancer patients), whereas, the EQ-5D does not. To adequately map disease-specific instruments to preference-based measures, a certain degree of overlap is required between the two descriptive systems [15]. If important dimensions are not covered in the preference-based measures, the mapping functions may be compromised. Performing the mapping exercise only using similar dimensions – for example, pain – would not permit the calculation of utilities to be used in CUAs.

The scoring functions of the preference-based instruments were derived from responses of the general population. The EQ-5D and SF-6D consist of dimensions related to individuals’ perceptions of their abilities within a social context (e.g., “I am able to perform my usual activities”). These dimensions capture the ability of the patients to adapt to their health state, which may not be a phenomenon that members of the general population are aware of when faced with an impaired health state [47]. The process of adaptation has been reported to enhance the difference between the utilities for health states provided by patients and members of the general population [48]. Lower utilities tend to be reported by the general population because they do not anticipate their ability to adapt when faced with life with an impaired health state, such as cancer [49]; this may be reflected by a lower magnitude of disutility in the population tariffs. While the use of general population tariffs raises concerns as to whether mapping functions can be estimated accurately for the population used in this study, generated results could be compared independently of the disease. This is one of the reasons why generic preference-based measurements of HRQoL are required for economic evaluations.

The ability of the cancer-specific instruments to predict utilities may be influenced by the psychometric properties of the preference-based measures. While responses on the FACT-G were obtained from Canadian cancer patients, the scoring functions of the EQ-5D and SF-6D are based on non-Canadian populations. These tariffs have been demonstrated that they are valid in a Canadian population [50‐53]. However, the constructed mapping algorithms from this study may not be transferrable to patients of other countries. While it is anticipated that the use of a different country’s tariff may only have a marginal effect on the resulting mapping function, the utilities generated may not be appropriate for information resource allocation decisions for the country of study. As such, there is a need to explore the ability of the FACT-G to predict individual dimensions of the EQ-5D and SF-6D; the predicted dimensions could be used to calculate an overall utility score. This has considerable advantages because the derived results are not country-specific and may be more applicable for guiding resource allocation decisions, especially for country that do not have country-specific tariffs for the EQ-5D or the SF-6D.

This study demonstrates that it is possible to facilitate economic evaluations for specific health conditions when a preference-based measure has not been administered and when it would be impractical to conduct a valuation survey using methods such as SG or TTO. This mapping approach offers a shortcut for policy-makers and researchers who need utility values for use in economic evaluation but do not have access to information on preference-based measures. The methodology presented in this study may be applied to other disease-specific instruments.