Abstract
Background
The performance of the Beta Binomial (BB) model is compared with several existing models for mapping the EORTC QLQC30 (QLQC30) on to the EQ5D3L using data from lung cancer trials.
Methods
Data from 2 separate non small cell lung cancer clinical trials (TOPICAL and SOCCAR) are used to develop and validate the BB model. Comparisons with Linear, TOBIT, Quantile, Quadratic and CLAD models are carried out. The mean prediction error, R^{2}, proportion predicted outside the valid range, clinical interpretation of coefficients, model fit and estimation of Quality Adjusted Life Years (QALY) are reported and compared. MonteCarlo simulation is also used.
Results
The BetaBinomial regression model performed 'best' among all models. For TOPICAL and SOCCAR trials, respectively, residual mean square error (RMSE) was 0.09 and 0.11; R^{2} was 0.75 and 0.71; observed vs. predicted means were 0.612 vs. 0.608 and 0.750 vs. 0.749. Mean difference in QALY's (observed vs. predicted) were 0.051 vs. 0.053 and 0.164 vs. 0.162 for TOPICAL and SOCCAR respectively. Models tested on independent data show simulated 95% confidence from the BB model containing the observed mean more often (77% and 59% for TOPICAL and SOCCAR respectively) compared to the other models. All algorithms overpredict at poorer health states but the BB model was relatively better, particularly for the SOCCAR data.
Conclusion
The BB model may offer superior predictive properties amongst mapping algorithms considered and may be more useful when predicting EQ5D3L at poorer health states. We recommend the algorithm derived from the TOPICAL data due to better predictive properties and less uncertainty.
Similar content being viewed by others
Introduction
Mapping is a method where the interrelationship between a generic health related quality of life (HRQoL) measure such as the EuroQol EQ5D3L (EQ5D3L) and a condition specific HRQoL measure (e.g. EORTC QLQC30) is modelled so that utilities can be predicted (retrospectively) in studies where the generic measure was not used. Responses from the EORTC QLQC30 (QLQC30 thereafter) cannot be used directly in an economic evaluation because they are not measures of utility, although these can be obtained from external studies or algorithms. Therefore, a key objective of mapping is to estimate patient level utilities from which quality adjusted life years (QALY's) are determined which might otherwise not be available. The EQ5D3L is recommended by the national institute for health care excellence (NICE) in the UK for use in economic evaluation, in particular, cost utility analyses (CUA) [1].
Why use mapping
Mapping or "crosswalking" can be useful when patient level utilities are not available in a clinical trial. A statistical model, sometimes termed as 'mapping algorithm' is used to predict the EQ5D3L from a disease specific measure such as the QLQC30. If patient level EQ5D3L cannot be obtained then it becomes difficult to conduct a cost utility analysis with patient level data and reliance is made on published aggregate utilities. Mapping may therefore be the only way to estimate patient level utilities for a trial and can avoid the potential biases and uncertainties associated with using published aggregate utilities. Mapping can also offer an additional way of addressing sensitivity of estimated utilities (and QALYs) if there is concern about differences between the target population and the population from which utilities were estimated from valuation.
HRQoL outcomes from the QLQC30 are not based on preferences for given health states from a payer perspective, and therefore not used in CUA. However, a health state on the EQ5D3L of 12111 (for example, a score of 1 for Mobility, 2 for Self Care, 1 for Usual Activities, 1 for Pain/Discomfort and 1 for Anxiety/Depression) can be converted into a preference based utility value of 0.85 for generating a QALY.
The EQ5D3L currently used in many clinical trials, has 243 health states; for each state a corresponding utility value is available [2]. In this paper we use the UK tariffs based on the Time TradeOff (TTO) method [3]. The raw scores from the EQ5D3L are converted into an index ranging from 0.59 to 1, where 1 denotes 'perfect' quality of life, 0 for death and values below 0 as states 'worse than death'.
It is not uncommon to find a mapping algorithm used to predict patient level utilities for a clinical trial in a particular disease area (e.g. pain), although the mapping algorithm was developed using data from quite a different patient population [4]. Crott (2012) suggests that different algorithms or functional forms may exist for each cancer type [5]; similar issues have also been raised elsewhere [6]. Therefore, it is not often clear whether published mapping algorithms intend users to extrapolate to any patient population and how factors such as timing of measurements or presence of differential treatment effects influence predicted values. For example, where algorithms have been developed using only baseline data, it is often not immediately clear how useful they are for predicting post baseline EQ5D3L; users of algorithms are often interested in predicting differential (post baseline) treatment effects for costeffectiveness purposes [7].
Although the EQ5D3L is a relatively short instrument, it is nevertheless surprising that many studies do not collect patient level utility data. Recently, only 16% of studies in lung cancer collected preference based HRQoL data [8]. In some clinical trials, patient level utilities were not collected, although formal economic evaluations were conducted; 25% of HTA submissions to NICE used mapping in such situations [9]. In Australian HTA's, this was 24% [10]. There is no guarantee that future trials will continue to collect utilities from EQ5D3L as there is no obligation to do so. Moreover, the existence of the EQ5D5L (measured on a 5 point scale) suggests that the EQ5D3L may not be adequate to address concerns about sensitivity. These are some of the reasons why mapping may continue to play an important part in economic evaluation of new treatments.
In some early phase trials, preference based measures are not usually collected, but condition specific measures such as the QLQC30 are collected to provide early indications of symptom control for future phase II/III trial planning, particularly for reimbursement. A mapping function can be used to estimate EQ5D3L from (combined) early trial data for planning future costeffectiveness argument for a phase III trial. In situations where two identical trials are required for licensing purposes (e.g. multiple sclerosis), a useful mapping algorithm from one trial can be used to determine utilities in the other trial [11]. If the two trial designs and patient populations are identical, there might be a possibility for developing a mapping algorithm in the first trial and predict utilities for the second.
Although mapping can be useful (and sometimes necessary), it is preferable to collect EQ5D3L prospectively where possible [6]. Some particular problems identified with mapping include limited ability of models to predict utility at poorer health states (Rowen & Brazier [12]; the assumption that there is a conceptual relationship (overlap) between two measures (Round [13]) and that the predicted value is a true measure of HRQoL (Chuang [14]).
Previous work on mapping
Several models have been developed and published for mapping the QLQC30 to predict EQ5D3L [15][19]. These models have been compared using measures such as predictive power (R^{2}), predictive mean and residual mean squared error (RMSE). Models used for mapping include conditional mean or median regression models with varying degrees of success [6]. Of 119 models examined, ordinary least squares (OLS) approaches were the most common [6]. The authors concluded that more complex models only rarely had an impact on predicted values [6],[20].
Response mapping approaches Gray et al [21] where ordinal categorical responses are modelled have also been used with limited success. With such models, however, if there are a few responses at extremes, prediction at these extremes are likely to be imprecise, particularly with smaller sample sizes. Hernandez et al [22] compare linear and TOBIT models with adjusted censored models (which essentially modify the TOBIT) and similar models in a mixture modelling framework (but none were applied to cancer datasets and QLQC30 in particular). These models appear to work well, but seem to need larger sample sizes to work well. For smaller clinical trials (such as the SOCCAR trial), these models too may have limited ability to predict at the extremes. Basu & Manca (2012) use a bayesian form of a beta regression (using non informative priors) in a nonmapping context and compared with OLS regression [23]. The authors used a 2part version of the beta model to handle the over dispersion of values (e.g. over dispersion of 1's on the EQ5D3L scale) [23]. Some relatively recent approaches include more complex Bayesian networks (Quang et al [24]) where conditional and joint probabilities of responses using Bayes theorem may compute posterior probabilities of each EQ5D5L response and consequently the expected utilities [24]. However, in this model too, predictions at poor health states were reported as inadequate, although it performed better than other models. The predicted utilities would be dependent on the initial (prior) probability of response for each EQ5D3L domain.
Longworth (2013) suggests several alternative models including the Beta Binomial for investigation [9]. Crott (2012) also suggest research of more complex mapping algorithms as well as a need for greater validation [5].
One aspect of this 'complexity' might include adding several interaction and additional demographic variables, which might improve prediction, but on the other hand can result in an over complicated algorithm (e.g. rather than 15 QLQC30 variables, one may have a model with 30 factors, including interactions to predict EQ5D3L). Another way of thinking about 'complexity' might be a more complicated mathematical function, but with fewer independent variables.
One feature of almost all published algorithms, including those used in lung cancer (using QLQC30) has been the overprediction of utilities in patients with poorer health states. The commonly used models do not appear to have properly addressed overdispersion at the extremes of the distribution. The authors of several mapping algorithms suggest further research is needed to address the uncertainty of algorithms in a robust way. Some authors suggest that statistical evaluation may not be sufficient [25] and clinical indicators are needed to support the algorithm. For these reasons, and given that at the current time no mapping algorithm exists for this particular type of lung cancer group (elderly and unfit for chemotherapy), this research is needed.
Alternatives to mapping
Other approaches to predicting or estimating utilities from the QLQC30 include using valuation studies (such TTO) [12]. This raises a question on the usefulness of mapping the QLQC30 if utilities can be readily determined from published literature. However, as pointed out earlier, differences in disease (cancer) types can yield significant differences in predicted utilities. In our studies in non small cell lung cancer (NSCLC) patients, the average survival time was 3 months and patients were unfit for chemotherapy. In addition, patients from the trial were from the UK with relatively consistent practices for palliative care (often given to these patients).
Mapping is preferred over valuation when there is an absence of robust evidence from literature. Even when published utility measures are available, however, care should be taken that estimates of utility for an economic evaluation reflect those expected in the target population. In some cases valuation methods may focus on a reduced set of questions. For example, Rowen et al (2011) determine utilities associated with the QLQC30 in a cancer population with better prognosis and substantially longer median survival. The authors retain the question about a "long walk" for assessing physical function. In lung cancer patients, responses about "short walks" are likely to be just as if not more relevant [12]. The purpose of this research is not to compare utilities from valuation based approaches with mapping algorithms, but rather compare the more common mapping algorithms.
The value of the Beta Binomial (BB) model as a useful mapping algorithm is its flexibility and ability to model skewed and multimodal data measured on a zero to one interval [26]. The modelling context allows for clustering of data (to model correlations within and between subjects) and is shown to reported more precise and efficient parameter estimates [26]. In situations where responses are over inflated at extremes (ceiling effects), it is particularly useful because one can attempt to model extreme values rather than omitting them or considering them as outliers. Moreover, effect sizes in terms of odds ratios may be more meaningful to decision makers (particularly clinicians) than absolute mean differences.
Methods
Several mapping models applied to QLQC30 were identified. A useful recent review is provided by Longworth [9]. Of the five published algorithms which mapped QLQC30, four used linear models (OLS estimates) [20] and one used a Quadratic model [12]. Lung cancer data sets were used in two instances [18],[19].
Instruments
The EQ5D3L is a generic HRQoL instrument which is well documented [2]. It has 5 domains Anxiety/Depression, Mobility, SelfCare, Usual Activities and Pain/Discomfort measured on a 3 point scale from 1 to 3. The UK TTO tariff was applied to the raw EQ5D3L scores to convert into utilities [2]. The EORTC QLQC30 is an established instrument for measuring HRQoL in various cancers [27]. The QLQC30 has 15 domains, scored on a 0 to 100 scale. The scoring consists of 5 function scales: Physical Function (PF), Role Function (RF), Emotional Function (EF), Cognitive Function (CF) and social functioning (SF). There are also 9 symptom scales, Fatigue (FA), Nausea & Vomiting (NV), Pain (PA), Dyspnoea (DY), Insomnia (IN), Appetite Loss (AL), Constipation (CO), Diarrhoea (DI) and Financial Problems (FI); there is also a global health status score (QL). For the global health and function domains, high scores indicate better QoL. For the symptom domains, low scores indicate better symptoms.
Data
Data were from two national (UK) NSCLC clinical trials. Each trial received local ethics approvals and research was conducted in compliance with the Helsinki declaration (details in references provided). The first trial (TOPICAL) was a randomized phase III trial in 670 lung cancer patients which compare erlotinib (n= 350) with placebo (n= 320) [28]. Both the EQ5D3L and QLQC30 were collected monthly from randomization (baseline) until death. The analysis was based on using data over the first 12 months because nearly all patients had died by then.
The SOCCAR trial was a phase II randomized NSCLC trial comparing sequential chemotherapy followed by radical radiotherapy (experimental arm) versus concurrent chemoradiotherapy followed by chemotherapy in 130 patients (70 in Concurrent vs. 60 in Sequential) with inoperable stage III NSCLC and good performance status. EQ5D3L and QLQC30 were collected monthly from baseline for a period of at least 18 months in the SOCCAR trial [29].
Developing and testing alternative models
Separate mapping algorithms were developed using data from each of the TOPICAL and SOCCAR trials using BB regression and five other models (Linear, TOBIT, Quantile, Censored Least Absolute Deviation (CLAD), and Quadratic regression) for comparison. The five models selected are among the common mapping models reported in a review of the literature on mapping (Brazier et al [6]). Estimated utilities from each model were compared using several criteria including: RMSE, predicted distributions, MAE, confidence intervals, R^{2}, residual plots, proportion of predicted EQ5D3L outside the range 0.59 to 1.0, estimated QALYs and MonteCarlo simulation. The performance of each model was compared using independent data from the SOCCAR trial. In addition, each model was fitted using data from SOCCAR then tested with data from the TOPICAL trial.
Model specification and analysis methods
For each model, data were combined across time points and treatment groups following methods of previously reported mapping algorithms [15],[19]. One reason advocated for pooling across all time points is because more health states can be modelled. The models compared were:

(I)
Linear Mixed Effect Model

(II)
TOBIT Mixed Effect Model

(III)
Quadratic Mixed Effects Model following Crott [15]

(IV)
Quantile Fixed Effects Model

(V)
Censored Least Absolute Deviation (CLAD): Fixed Effects Model

(VI)
Mixed Effects Beta Binomial Regression Model
The linear mixed model is a regression model with subject as a random term. The linear mixed and Quadratic models, model the mean of the EQ5D3L utility distribution. The quadratic model (III) of Crott (2010) was included because it is considered as a nonlinear model with squared terms for some QLQC30 domain scores. The Quadratic model has PF, EF, SF SL and DI as squared terms, following Crott [15]. The TOBIT models the mean 'plus' the remainder of the distribution in a mixed effects context, however the slope parameter is adjusted by the probability of censoring. The censoring in the TOBIT can be from either below (censoring EQ5D3L to 0) or above (censoring EQ5D3L to 1). The BB models the distributions of EQ5D3L responses. Models (I)  (III) are not described in detail because a review of the common features of these and other models have been discussed elsewhere [6].
(IV) quantile regression
In linear regression, estimation problems can exist when a response variable such as the EQ5D3L is skewed, truncated or discrete [20]. The regression line (in the case of one independent variable) passes through the mean values of the response (EQ5D utilities), for given values of the independent variable. It assumes that the relationship between EQ5D and QLQC30 will be the same (changes by a constant slope), even if patients have very low or very high QLQC30 scores. QLQC30 may be predictive of EQ5D, but the predictions may be different for poor quality of life (low EQ5D scores) compared to better quality of life (high EQ5D). Quantile regression might therefore be useful for predicting utilities in specific groups of patients (e.g. with worse health states); that is it examines how the relationship between EQ5D and QLQC30 changes, depending on the values of EQ5D, in this application.
In quantile regression, a line would pass through the median or a specific quantile of the response for specified values of the independent variable (QLQC30 scores). Predictions from the Quantile regression model for individuals subject are estimates of the (conditional) median when τ = 0.5 in the regression equation below:
Y_{i} are observed EQ5D utilities for each subject, X is a matrix of 15 QLQC30 values including an intercept column, β_{ τ } is a vector of parameters associated with the 15 scores of the QLQC30 and ε_{i} is the absolute deviation which is to be minimized, known as the least absolute deviation (LAD) [30],[31] for estimating values of β_{ τ }.The expected median EQ5D, when τ= 0.5 is assumed to increase for increasing QLQC30 scores. No distributional assumptions are made about ε_{i}, whereas in linear regression, ε_{i} are assumed normally distributed with constant variance. In practice, one might get several regression lines, one for each quantile. The relationship might be stronger at any one of these quantiles.
In order to predict a patient level median EQ5D3L utility (Y_{i}^{*}) from 3 given scores (such as PF, RF and EF), for example, a quantile (linear) regression model when τ= 0.5 would be:
When the values of PF, RF and EF are all zero, Y_{i}^{*} is an estimate of the sample median.
(V) Censored Least Absolute Deviation (CLAD)
CLAD extends quantile regression with an emphasis on the 50^{th} percentile (when τ= 0.5). In addition, predicted estimates of EQ5D utilities which are either >1 or between 0 and 0.59 are restricted (censored) to 1 and 0 respectively (Powell; Khan & Powell) [32],[33]. The model is described below:
The term Xβ refers to the predicted median EQ5D3L plus some deviations, ε_{i,} which are assumed to follow any distribution with a median of zero. The estimators of β are unbiased and consistent though not efficient [34]. Conditional medians are estimated at the patient level in a similar way as in (IV) except that estimates are restricted to lie between 0 and 1. The mean of all the individual (conditional) medians can be used as before for deriving QALYs. The mean is the statistic of choice for decisions relating to health technology assessment [35],[36]. The population mean and median are approximately equal for normally distributed data.
(VI) Beta binomial regression
The BB distribution is often used in probabilistic sensitivity analyses (PSA) in health economic modelling for utility measures such as the EQ5D3L [37]. One reason for use in PSA appears to be convenience of assuming a scale from 0 to 1 for utility (although there is no EQ5D3L tariff which is exactly equal to zero). In particular, the BB regression can model responses which are unimodal or bimodal with varying levels of skewness [38][40] ; utilities are often reported as having skewed or truncated distributions [20]; in addition, the BB estimates the mean of the distribution whereas some other models estimate the median. Therefore the BB approach may be a suitable model to test for developing a mapping algorithm.
An important feature of the BB approach is that mean predicted estimates of EQ5D3L can be estimated while restricting the range between 0 and 1. Although responses are required to be in the (0,1) interval, the BB can still be used in any interval (a, b) for a < b using the transformation Ya/ba. For example, if the observed EQ5D3L value is 0.1, then 0.1 – (0.53)/1 (0.53) would give a transformed value of 0.28. However, it may be difficult to correct for both asymmetry and heteroscedasticity resulting in difficult interpretations of parameter estimates in terms of the original response [41]. Moreover, there may also be potential difficulties in interpreting mean QALYs. For example, a utility of 0.34 transformed to a value of 0.124 generates a different interpretation of the QALY. One reasonable assumption in the use of the BB model is that observed values <0 are set equal to 0 because there were <0.5% of values with EQ5D3L responses <0 in each data set, hence, the potential for bias is likely to be small. A transformation was therefore not used.
Using the BB can be more complicated than linear models, depending on the need to model the variance. Without modelling the variance (overdispersion), modelling mean response (EQ5D3L) as a function of the 15 QLQC30 variables requires using a simple logit function (because values are assumed to lie between 0 and 1), as in a logistic regression model.
Notation for beta binomial
A response variable (EQ5D3L) is assumed to follow a Beta (α,β) defined by:
The mean and variance of a y of (1) are α/(α + β) and αβ/(α + β)^{2}(α + β + 1), where α and β are the shape and scale parameters, respectively. The parameters α and β can be estimated from the observed mean and variance using the method of moments. For example, the values of β from the TOPICAL and SOCCAR data are <1 for EQ5D responses, using the relationship:
When μ and σ^{2} are population mean and variance; sample estimates can be used as estimates for these.
The above description of the BB distribution in (1) is not useful for regression modelling and first requires reparameterization (Ferrari & CribariNeto 2004), so that a response can be defined along with a set of predictors to form a regression model [42]. This is similar to simple linear regression where the responses are normally distributed with mean μ and variance σ^{2}. In simple linear regression we express μ, the predicted mean in terms of a set of predictors x_{1}….x_{15} (i.e. μ = a + bx_{1} + bx_{2} + …..b_{15}x_{15} for the QLQC30).
If we set μ= α/α + β and ϕ = α + β in (1) then this becomes:
The expression in (2) is a beta distribution: y ~ Beta (μ, ϕ) and the mean, μ is expressed as a link function (to model the mean) in terms of some predictor variables. Typically, the link function g(μ)= Xβ is such that g(μ)= log(μ/1 μ). With this logit link function, the mean response is: μ= e^{ω}/1+ e^{ω}, where ω= α + Xβ. The value of μ is restricted to a 0 to 1 scale. In simple linear regression, g(μ)= μ.
The BB regression form for modelling EQ5D3L is:
where the part log(μ_{i}/1 μ_{i}) are the transformed values of the EQ5D responses to a logit scale. One main difference with the logistic regression model is that there is no need for responses to be dichotomous (the values μ_{i} are continuous).
In order to compute the patient level predicted EQ5D3L, the logistic function, with X as the set of independent QLQC30 variables with β as the parameter vector yields:
A second model (such as a log function) could also be used to model the dispersion in terms of a set of QLQC30 variables (not necessarily all 15 variables, because the variance might depend on some important ones).
The additional precision parameter ϕ, like the mean parameter, μ, involves an equation which relates the variance to a set of predictors. This equation often assumes a log function ln[h(ϕ)] (because the variance is >0). Therefore, h(ϕ) describes the relationship between the variance and W
Since the relationship between variance may not depend on all 15 variables (in the case of QLQC30), these are labelled W with the corresponding parameters δ.
The addition of the second (precision) parameter, ϕ, allows greater flexibility to model any over dispersion of EQ5D values [42].
Hence, two sets of equations are associated with the QLQC30 variables; one through the mean EQ5D3L and one through the variance. These (two) equations provide the basis to determine estimates of the parameters β.
The responses y (i.e. the EQ5D3L) are therefore Beta ([(g(μ), h(ϕ)], with likelihood function:
where,
If we compare the above function (3) with (1):
α1 = s1 and β1 = t1 in (3) relates the observed EQ5D3L (y values) with the QLQC30 predictors through both the mean and variance.
This form of parameterization allows a powerful way of modelling utilities not only for mapping, but in a generalized mixed modelling context for estimates of utilities and HRQoL which can be scaled to a 0,1 interval. This approach uses a nonlinear mixed modelling approach where the general likelihood has been programmed directly using the SAS software (version 9.3) [40]. An example SAS code is provided in Appendix I for implementing the BB model.
One reasonable assumption we impose is that observed values <0 are set equal to 0. There were <0.5% of values with EQ5D3L responses <0 in each data set, hence, the potential for bias is likely to be small. A transformation was therefore discarded.
Although the BB regression can be very flexible, a limitation of the model is that observations existing at either 0 or 1 must be scaled away from these values. That is, when there are many 0 or 1 responses, estimation with a standard BB regression can become problematic. Therefore, a zeroone inflated BB model was used to account for this overdispersion [40].
Testing algorithms with lung cancer data
For each of the models, the predicted and observed values were compared. Models were developed using the larger TOPICAL data set and validated with SOCCAR. The observed health states were ordered (from 11111 to 33333) and the mean predicted EQ5D3L was computed for each algorithm and plotted. In addition, we estimated the proportion of individual predicted EQ5D3L responses from each model within +5% to +30% of the observed EQ5D3L. Estimated utilities from each model were compared using model statistics described earlier.
Model checking and adequacy
In all models, adequacy of fit was considered using residuals, tests for homoscedasticity and Aikakes Information Criterion (AIC). The AIC was used to compare models for the same dataset.
Simulations
MonteCarlo simulations (10,000) from a multivariate distribution for the EQ5D3L and QLQC30 scores using the method of Fleishman [43],[44] with the observed correlation structure were carried out to assess uncertainty of predicted means from each model. The method of Fleishman uses higher order moments (Skewness and kurtosis) as a way of simulating data that approximates the sampling distribution. Each data set simulated contained 670 and 130 patients with 2038 and 1002 observations for TOPICAL and SOCCAR respectively.
Addressing over prediction of EQ5D3L at the 'poorer' health states
We investigated the over prediction at 'poorer' health states using a health state of 11321 (EQ5D3L utility 0.433) as a cutoff for 'Poor' and 'Good' health states in TOPICAL and 22222 (EQ5D3L utility 0.516) in SOCCAR. The selected health states cutpoints were chosen because this is where the observed and predicted EQ5D3L values start to diverge.
Impact on QALY estimates
For each model, patient level QALYs were generated using model estimates of patient level utilities and multiplying them by observed survival times. For PSA, simulation was used to estimate the mean overall survival (OS), progression free survival (PFS) and postprogression survival. OS and PFS are important outcomes in cancer trials often calculated from the time from treatment allocation or randomization until death (OS) or disease progression (PFS). The exponential model was chosen to fit the empirical KaplanMeier curve for OS and PFS. Using the relationship: OS= 1*Log(1x_{i}))/λ,where x_{i} are randomly generated from a uniform distribution and λ is the observed hazard rate. In economic evaluation, the objective is to simulate survival times which approximate the mean OS and PFS (not the median).
For each realization the mean (area under the survival curve) OS and PFS were determined for each treatment group. For TOPICAL, there were no censored data and for SOCCAR, the censoring distribution was taken into account (because patients were still alive) such that simulations resulted in PFS < OS. The pre and post progression utilities were determined from each of the simulations described earlier. Hence a total of 10,000 mean OS, PFS, preprogression and postprogression utilities were generated to determine QALYs. QALYs were estimated as weighted sums of preprogression and postprogression mean EQ5D3L for each treatment group.
Results
A total of 2038 and 1002 data points with 84 and 54 health states were observed for each of TOPICAL and SOCCAR respectively. The average (median) number of observations per health state were 3 for TOPICAL and 2 for SOCCAR. The most frequent health state in TOPICAL was 21222 (12%); For SOCCAR, the most frequent health state was 11111 (25%) followed by 21222 (8%). Patients in the SOCCAR trial had better performance status (Table 1) compared to TOPICAL patients. Less than 0.5% (3/2038 observations in TOPICAL and 1/1002 in SOCCAR) of EQ5D3L observations had values <0 (corresponding to 3 health states in TOPICAL and 2 health states in SOCCAR). The correlation between EQ5D utilities and each of the 15 domains ranged from 0.32 (FI) to 0.69 (PF), suggestion that mapping was possible (some overlap in terms of responses was present).
The observed mean (SD) EQ5D3L for TOPICAL and SOCCAR were 0.61 (0.29) and 0.75 (0.23) respectively over all postbaseline time points (Additional file 1: Table S1). Figure 1 shows the distributions of EQ5D3L confirming presence of nonnormality, skewness and multimodality. If α and β (both are shape parameters which are used to model the distribution – for example if α and β are the same, the distribution tends towards symmetry) are <1, data are considered nonnormal, multimodal or skewed. The KolgomorovSmirnoff goodness of fit rejects normality (P= 0.0093) and P < 0.001 for TOPICAL and SOCCAR respectively).
Comparison of models
The models were first tested on the same dataset used to develop the mapping algorithm (Table 2 and Additional file 1: Table S2, Figure 2a and 2b). All terms in the model were retained (regardless of statistical significance); even if some terms are not statistically significant, since they can still be relevant [5],[12],[16]. The standard errors for the BB were smallest (Table 2). The AIC values were smallest for the BB model and ranged from 2215 (BB) to 864 (Quantile) for TOPICAL and 1529 (BB) to 587 (Quantile). Smaller AIC values suggest better fit (Table 2). For the six models, the estimated R^{2} ranged from 0.53 (CLAD) to 0.75 (BB); R^{2} was highest (TOPICAL R^{2}= 0.75) for the BB model (Table 2). Estimated RMSE ranged from 0.09 (BB) to 0.18 (Quadratic, CLAD). The proportion of predicted values of EQ5D3L >1 were highest for the quantile model (3.5%), whereas for the Quadratic model predicted values <0 were more common (5%), despite only 0.3% of observed values <0 were set to zero for modelling purposes (Table 2). Only the TOBIT, CLAD and BB did not predict outside the 'observed' range.
Mean predicted EQ5D3L distributions are also shown in Figure 3a and 3b. The predicted means were 0.608 for BB, 0.584 (Linear), 0.631 (TOBIT), 0.635 (Quadratic), 0.633 (Quantile) and 0.593 (CLAD) in TOPICAL; For SOCCAR these were 0.749 (BB), 0.771 (Linear), 0.771 (TOBIT), 0.774 (Quadratic), 0.766 (Quantile) and 0.782 (CLAD). Predicted mean EQ5D3L was closest to the observed with the BB model (Table 2).
Testing models using independent data
The model developed from TOPICAL was tested on SOCCAR data (Figure 3a, 3b and Table 3). Figure 1 compares 'out of sample' predicted and observed EQ5D3L distributions. In particular, Figure 3a shows the predicted vs. observed distributions for the model developed from TOPICAL and tested using the SOCCAR dataset; the BB model predicts the overdispersion at values of Zero and One better than other models: in SOCCAR, 25% of EQ5D3L responses were one: the BB has predicted these very well. The CLAD and Quantile also predict these with some success whereas TOBIT and Linear were less accurate. Figure 3b shows the predicted mean EQ5D3L by health state compared to observed values. The BB also overpredicts EQ5D3L at poorer health states, but the extent of the overprediction is less severe.
The R^{2} values were highest with the BB (R^{2}= 0.75) when the model developed from TOPICAL data was tested on SOCCAR data and R^{2}= 0.61 for the model developed from SOCCAR data and tested on TOPICAL; the RMSEs were higher compared with other models (Table 3). Mean predicted EQ5D3L for SOCCAR was 0.747 (95% CI: 0.733, 0.760) for the BB. Hence, the BB developed from TOPICAL data predicted mean EQ5D3L to within 0.4%. From 10,000 simulations, 77% of the 95% confidence intervals for the predicted mean EQ5D3L contained the observed SOCCAR mean EQ5D3L value of 0.75.
When the model was developed using SOCCAR data and then tested on TOPICAL, the mean predicted mean EQ5D3L was 0.622 (compared with the observed 0.61) and 59% of the 95% confidence intervals contained the mean EQ5D3L value of 0.61 observed in TOPICAL. On possible reason for the lower proportion of coverage is that SOCCAR patients had less severe NSCLC patients and consequently 'better' health states compared to patients in the TOPICAL trial. Normal probability plots of the model tested on SOCCAR data show residuals closest to the line with the BB model (Additional file 2: Figure S1).
Patient level predictions using independent data
Comparing mean predicted EQ5D3L with observed mean may not always be the best way to judge model performance because the distributions of the prediction values tend to cluster around the observed mean [6]. For example, an observed mean utility of 0.61 compares with a predicted mean of 0.593 using CLAD (Table 2), a difference of 0.017 (3%). However, about 40% of individual predicted values differed from the observed mean by about 10%. Therefore, for each patient, percent differences within ±5% to ±30% of observed values were calculated. About 28% (BB) of predicted EQ5D3L were within ±5% (Additional file 3: Figure S2) of the observed EQ5D3L compared to 20% (Linear), 23% (TOBIT), 24% (Quadratic), 22% (Quantile) and 22% (CLAD) with SOCCAR data. Predictions were in general better with the BB model (the curve is above all others). The median prediction error for the BB model is about 10% for both TOPICAL and SOCCAR. Highest prediction errors are observed with the Linear model (median of 15% error for both TOPICAL and SOCCAR). The QLQC30 responses ranged from 0 to 100 for 14 out of the 15 domains (scores for the financial domain ranged from 30 to 80).
Overprediction in worse health states
Mean predicted EQ5D3L at observed health states for each algorithm are shown in Figures 4a and 4b. The BB model had mean predicted EQ5D3L estimates closest to the observed values at a given observed health state for TOPICAL and SOCCAR respectively. Differences between observed and predicted mean EQ5D3L for most models occur at about health states of11321 (value on xaxis of 32 in Figure 4a) for TOPICAL and about 22222 (xaxis value of 26) for SOCCAR. The Quadratic model underpredicted mean EQ5D3L at less severe health states compared to the BB model (Figure 4a). With SOCCAR data, the BB model approximates mean EQ5D3L at each health state better than all other models (Figure 4b). Figure 1b shows a similar plot using independent data from the TOPICAL developed algorithm.
Relationship between health states and adverse events
The relationship between adverse event frequency for different definitions of 'Good' and 'Poor' health states were briefly investigated. The results suggest that patients with 'Poor' health (defined roughly here as >11321 for TOPICAL and >22222 for SOCCAR) are also those with a higher frequency of adverse events. In the TOPICAL trial, 24% of patients in 'Poor' health states (i.e. worse than 11321) experienced more than 2 grade 34 adverse events compared with 15% for patients in 'Good' health states (health states 11321 or better); for SOCCAR this was 66% vs. 44%.
One reason why mapping algorithms overpredict EQ5D3L at 'Poor' health states might be because treatment related toxicity is not directly captured into the mapping algorithm, resulting in estimating a QoL higher than it actually is. A similar pattern was also seen for other cutoffs which define 'Poor' and 'Good'. For example, when the cutoff for 'good' and 'poor' was defined as health states 21321 (EQ5D3L of 0.364) and >22111 respectively, fewer patients in 'good' health states had AEs compared with patients with 'poor' health states: 17% vs. 26% of patients in 'good' vs. 'poor' health states had at least 2 adverse events in the TOPICAL trial. A similar pattern was observed for SOCCAR data. This suggests that there may be a more complex underlying mapping algorithm between EQ5D3L, QLQC30 and toxicity which might better explain the variability and prediction of EQ5D3L, particularly in patients with 'Poor' health states.
Impact on QALY estimates
Table 4 compares observed and expected QALYs from each of the models for SOCCAR and TOPICAL separately. The Observed QALY difference was 0.051 for TOPICAL (Erlotinib vs. Placebo) and 0.164 for SOCCAR (Concurrent vs. Sequential). Predictions from the BB model generated closest QALY estimates in both trials with a mean QALY difference of 0.053 for TOPICAL and 0.162 for SOCCAR (Table 4). QALY predictions from other models ranged from 0.041 (Linear) to 0.072 (Quadratic) for TOPICAL and 0.153 (Linear) to 0.208 (Quantile).
Adjustment for demographic variables
Several additional factors were added to the model. Although R^{2} changed slightly from 0.75 to 0.78 in TOPICAL with the inclusion of ECOG (P < 0.001) and Gender (P < 0.001), the underlying pattern of prediction shown in Table 2, Table 3 and Figure 4 did not change. Adding demographic variables does improve model fit slightly but does not have a major impact on predicted means and their standard errors (available on request from authors).
Discussion
Superior predictive properties have been demonstrated with a nonlinear BB mapping algorithm developed and tested using data from two independent lung cancer patient populations. Two mapping algorithms for different types of lung cancer patients (poor and better prognosis) have been shown to perform better than commonly used models. Either algorithm could be used, but our preference is for the one derived from the larger TOPICAL trial due to smaller uncertainty (Table 3) and better model fit (Table 2). Simulations assessed the uncertainty of mean estimates of EQ5D3L utilities. The degree of overprediction of mean utilities at poorer health states was less with the BB model compared to other models. QALY estimates from models were also closer to the observed values with the BB model. Our findings confirm previous untested assertions that the relationship between the EQ5D3L and QLQC30 may be better understood with a nonlinear model structure [5],[9],[15].
Previous mapping models have mostly used OLS forms, considered inadequate or over simplistic [6]. Most reported mapping models suffer from overprediction at poorer health states. Previously reported models have reported R^{2} values (using QLQC30) ranging from 0.23 to 0.83 [6]; In our study, we report similar values for these models. Rarely do models yield values of R^{2} above 70%. Other models (e.g. multinomial, ordinal) have also been used, but shown to be inadequate [5]. Estimates based on absolute deviation (CLAD, Quantile and adjusted censored models) predict patient level medians, whereas the statistic of interest is the mean. The Quadratic model takes into account non linearity (by having squared terms in the model) but is essentially a linear model (linear in parameters because the coefficients are interpreted in the same way as linear models).
Improving model fit by "discarding" or "weighting" outliers with extreme values [5] is not an optimal solution if the extreme outliers with can be modelled (rather than excluding observations). Moreover, the choice as to which variables to square and then combine with nonsquared variables in quadratic models has many possibilities. For example, squaring all 15 domain scores of the QLQC30 is one possible choice, as is squaring 14 and have only one nonsquared term remaining. Without doing a very large number of tests and increasing the type I error, it is difficult to understand the relative merits of one set of variables over another. Therefore, this can lead to some arbitrary selection of combinations of terms to improve model fit.
There are several advantages of the BB approach. The BB model applied to this data confirms some superior statistical properties in terms of accuracy and efficiency [38],[41]. The clinical interpretation of the model is still reasonably clear. Clinical relevance is still important since if as some argue [45] that a generic measure is sufficient to provide both estimates of utilities and a clinical effect size of HRQoL, estimates based on odds ratios are likely to be more easily interpretable: for example the clinical relevance of a mean treatment difference of 0.012 on the EQ5D3L is difficult to judge. However, if this was equivalent to an odds ratio of 1.2 (a 20% improvement in HRQoL), a way of relating both clinical effects and utilities becomes possible. This can also be extended to the response domains; treatment effects can be interpreted using a ratio scale and mean differences also derived. This makes the BB a powerful and flexible mapping algorithm relevant to health economic evaluation, policy and clinical decision making.
The strength of our research lies in our approach to validating the model using independent data and extensive multivariate simulation from correlated EQ5D3L and QLQC30 data. We also explore some plausible reasons why overprediction at the poorer health states occur by using adverse event data (collected in all trials) often ignored in mapping algorithms. It is possible that joint relationships between adverse events and EQ5D3L might offer a plausible explanation for overprediction, because higher toxicity was observed in the poorer health states. Some researchers suggest that EQ5D3L responses have a bimodal distribution and therefore 2 separate mapping algorithms might be needed Veerstegh [25] for patients in 'Poor' and 'Good' health states. The nature of the bimodality could be explored using baseline clinical data (e.g. using baseline ECOG).
Our research has several limitations. We have not exploited the impact on results if the true values of α and β are different to the sample estimates. In our application, we set α and β so that the mean EQ5D3L could be modelled; other possibilities might include searching for α and β which might optimize R^{2}, minimise MSE and improve predictions. Secondly, we assumed a scale of 0 to 1, which might be suitable for some disease but may not be suitable for others where states worse than death are likely to be more important. Surprisingly, even in this NSCLC population, the proportion of such cases were low. Thirdly, model validation has also been limited to lung cancer data and further testing would be useful in both lung cancer (to see whether algorithms are tumour specific) and nonlung cancer data sets (to check for generalizability). Finally, we did not compare the BB with other models such as Bayesian network models which report superior predictive properties compared to the more common models. However with such Bayesian models, the choice of the initial (prior) estimates of probability of EQ5D3L responses can influence the predicted utility.
There are several concerns when using mapping functions, a point repeated in previous research [5],[6],[15],[19]. One key concern is that it is unknown whether the predicted utilities are close to the observed values unless we know both. Secondly, there are questions as to what exactly is being measured [13] or estimated because some key information in one instrument is not included in the other, particularly when predicting EQ5D3L from clinical measures alone. One approach might be to look at the psychometric properties of the two instruments and also check correlations. Weak correlation (Spearman's or linear) might explain a poor mapping algorithm.
The first concern regarding mapping can be partially answered with the use of simulation by quantifying uncertainty in how well the predicted approximates the observed can be quantified as in Table 3. This does not tell us what the predicted EQ5D3L is actually measuring, but it is assumed that the closer the predicted values are to the observed numerically, the preferences become 'essentially similar'. If in 90% of simulations, the observed and predicted values are close, it may be reasonable to assume that the mapping algorithm provides estimates that are measuring aspects of "essentially similar" preferences, which for practical purposes might be acceptable. Moreover, the statistical significance of several predictors might also indicate the nature of preferences; for example pain was a highly significant predictor in the model (p < 0.0001) for the BB model, hence the 'nature' of predicted preferences includes pain (also measured with EQ5D3L). If a model predicts every EQ5D3L perfectly, then we may wish to conclude that the model has correctly predicted the 'essential nature' of the preferences (ultimately contained in a single index), or remain sceptical and seek additional evidence to confirm that the 'essential nature' of preferences are captured by the model.
Other concerns with mapping involve time points used when developing and applying an algorithm. For example, including baseline data in a model which aims to predict postbaseline treatment differences may lead to misleading estimates. Assumptions that the rates of change in EQ5D3L (the coefficients of the QLQC30) are constant from one cancer type to another are also unlikely to hold. If baseline or demographic variables are used, the relevance of these for the target population using the algorithm will also be important. Finally, the mapping algorithm should offer reasonable clinical interpretation. In lung cancer, for example we might expect dyspnoea to be an important predictor of HRQoL. In some models, dyspnoea (Additional file 1: Table S2) was not statistically relevant for predicting EQ5D3L, although it is an important symptom in lung cancer patients. For increasing symptoms scores we might expect EQ5D3L utilities to decrease which is not always the case.
Conclusions
The BetaBinomial regression approach shows superior performance compared with published models in terms of predicting the observed EQ5D3L from QLQC30 in these lung cancer trials. This nonlinear approach may offer advantages over existing models for mapping and as a general modelling procedure for utilities. We also confirm recent results that mapping algorithms have shown to overestimate the HRQoL at the poorer heath states. The reasons why current algorithms persistently overpredict at poorer health states requires further interrogation, perhaps incorporating adverse event information into the models. Guidelines on using algorithms may also be useful. Mapping may be useful however there are still concerns as to whether the predicted utilities are essentially the same as the observed values.
Consent
Written informed consent was obtained from patients for the publication of this report and accompanying images.
Authors' contributions
IK proposed the research hypothesis, was involved in the collected of the data and its analysis. SM and IK contributed to the writing, review and approval of the final manuscript.
Appendix I: SAS Code for ZeroOne Inflated BetaBinomial
Example SAS code
Additional files
References
Guide to the methods of technology appraisals. [website]: NICE; 2004 [cited 2012 1st October]; Available from: , [http://www.nice.org.uk/aboutnice/howwework/devnicetech/technologyappraisalprocessguides/GuideToMethodsTA201112.jsp]
Dolan P: Modeling valuations for EuroQol health states. Med Care 1997, 35(11):1095–1108. [Research Support, NonU.S. Gov't] 10.1097/0000565019971100000002
Drummond MA: Economic Evaluation in Health Care. 2001.
Dunlop W, Uhl R, Khan I, Taylor A, Barton G: Quality of life benefits and cost impact of prolonged release oxycodone/naloxone versus prolonged release oxycodone in patients with moderatetosevere nonmalignant pain and opioidinduced constipation: a UK costutility analysis. J Med Econ 2012, 15(3):564–575. [Research Support, NonU.S. Gov't] 10.3111/13696998.2012.665279
Crott R, Versteegh M, UyldeGroot C: An assessment of the external validity of mapping QLQC30 to EQ5D preferences. Qual Life Res 2013, 22(5):1045–1054. Epub 2012 Jun 29 10.1007/s1113601202209
Brazier JE, Yang Y, Tsuchiya A, Rowen DL: A review of studies mapping (or cross walking) nonpreference based measures of health to generic preferencebased measures. Eur J Health Econ 2010, 11(2):215–225. [Research Support, NonU.S. Gov't Review] 10.1007/s101980090168z
Barton GR, Sach TH, Jenkinson C, Avery AJ, Doherty M, Muir KR: Do estimates of costutility based on the EQ5D differ from those based on the mapping of utility scores? Health Qual Life Out 2008, 14: 6.
Jäkel A, Plested M, Dharamshi K, Modha R, Bridge S, Johns A: A systematic review of economic evaluations in second and later lines of therapy for the treatment of nonsmall cell lung cancer. Appl Health Econ Health Policy 2013, 11(1):27–43. doi:10.1007/s40258–012–0001–1
Longworth L, Rowen D: Value Health. 2013, 16(1):202–210. 10.1016/j.jval.2012.10.010
Scuffham PA, Whitty JA, Mitchell A, Viney R: The use of QALY weights for QALY calculations: a review of industry submissions requesting listing on the Australian Pharmaceutical Benefits Scheme 2002–4. Pharmacoeconomics 2008, 26(4):297–310. 10.2165/0001905320082604000003
CHMP: Guideline on Clinical Investigation of Medicinal Products for the Treatment of Multiple Sclerosis. 2005.
Rowen D, Brazier J, Young T, Gaugris S, Craig BM, King MT, Velikova G: Deriving a preferencebased measure for cancer using the EORTC QLQC30. Value Health 2011, 14(5):721–731. doi:10.1016/j.jval.2011.01.004
Round J: Capturing Information Loss in Estimates of Uncertainty That Arise from Mapping Algorithms. ?: Health Economists' Study Group; 2008.
Chuang LH, Whitehead SJ: Mapping for economic evaluation. Br Med Bull 2012, 101: 1–15. 10.1093/bmb/ldr049
Crott R, Briggs A: Mapping the QLQC30 quality of life cancer questionnaire to EQ5D patient preferences. Eur J Health Econ 2010, 11(4):427–434. 10.1007/s1019801002337
Jang RW, Isogai PK, Mittmann N, Bradbury PA, Shepherd FA, Feld R, Leighl NB: Derivation of utility values from EORTC QLQC30 values in lung cancer. J Thorac Oncol 2009, 4(9):S420S421.
Kim EJ, Ko SK, Kang HY: Mapping the cancerspecific EORTC QLQC30 and EORTC QLQBR23 to the generic EQ5D in metastatic breast cancer patients. Qual Life Res 2012, 21(7):1193–1203. [Research Support, NonU.S. Gov't] 10.1007/s111360110037y
Kontodimopoulos N, Aletras VH, Paliouras D, Niakas D: Mapping the cancerspecific EORTC QLQC30 to the preferencebased EQ5D, SF6D, and 15D instruments. Value Health 2009, 12(8):1151–1157. 10.1111/j.15244733.2009.00569.x
McKenzie L, van der Pol M: Mapping the EORTC QLQ C30 onto the EQ5D instrument: the potential to estimate QALYs without generic preference data. Value Health 2009, 12(1):167–171. 10.1111/j.15244733.2008.00405.x
Kharroubi SA, Brazier JE, Roberts J, O'Hagan A: Modelling SF6D health state preference data using a nonparametric Bayesian method. J Health Econ 2007, 26(3):597–612. 10.1016/j.jhealeco.2006.09.002
Gray AM, RiveroArias O, Clarke PM: Estimating the association between SF12 responses and EQ5D utility values by response mapping. Med Decis Making 2006, 26(1):18–29. 10.1177/0272989X05284108
Hernández Alava M, Wailoo AJ, Ara R: Tails from the peak district: adjusted limited dependent variable mixture models of EQ5D questionnaire health state utility values. Value Health 2012, 15(3):550–561. 10.1016/j.jval.2011.12.014
Basu A, Manca A: Regression estimators for generic healthrelated quality of life and qualityadjusted life years. Med Decis Making 2012, 32(1):56–69. 10.1177/0272989X11416988
Le QA, Doctor JN: Probabilistic mapping of descriptive health status responses onto health state utilities using Bayesian networks: an empirical analysis converting SF12 into EQ5D utility index in a national US sample. Med Care 2011, 49(5):451–460. 10.1097/MLR.0b013e318207e9a8
Versteegh MM, Rowen D, Brazier JE, Stolk EA: Mapping onto EQ5D for patients in poor health. Health Qual Life Out 2010, 8: 141. 10.1186/147775258141
Paolino P: Maximum likelihood Estimation of models with betadistributed dependent variables. Polit Anal 2001, 9(4):325–346. 10.1093/oxfordjournals.pan.a004873
EORTC Quality of life. EORTC:[cited 2012 10th October]; Available from: ., [http://groups.eortc.be/qol/eortcqlqc30]
Lee SM, Khan I, Upadhyay S, Lewanski C, Falk S, Skailes G, Marshall E, Woll PJ, Hatton M, Lal R, Jones R, Toy E, Chao D, Middleton G, Bulley S, Ngai Y, Rudd R, Hackshaw A, Boshoff C: Firstline erlotinib in patients with advanced nonsmallcell lung cancer unsuitable for chemotherapy (TOPICAL): a doubleblind, placebocontrolled, phase 3 trial. Lancet Oncol 2012, 13(11):1161–1170. Epub 2012 Oct 16 10.1016/S14702045(12)704126
Maguire J, Khan I, McMenemin R, O'Rourke N, McNee S, Kelly V, Peedell C, Snee M. SOCCAR: A randomised phase II trial comparing sequential versus concurrent chemotherapy and radical hypofractionated radiotherapy in patients with inoperable stage III NonSmall Cell Lung Cancer and good performance status. Eur J Cancer . 2014 Oct 7. doi:10.1016/j.ejca.2014.07.009
Honore B, Khan S, Powell JL: Quantile regression under random censoring. J Econometrics 2002, 109(1):67–105. 10.1016/S03044076(01)001427
Koenker R: Quantile Regression (Econometric Society Monographs). 2005.
Khan S, Powell JL: Twostep estimation of semiparametric censored regression models. J Econometrics 2001, 103(1–2):73–110. 10.1016/S03044076(01)000409
Powell JL: Least absolute deviations estimation for the censored regressionmodel. J Econometrics 1984, 25(3):303–325. 10.1016/03044076(84)900046
Kaambwa B, Billingham L, Bryan S: Mapping utility scores from the Barthel index. Eur J Health Econ 2013, 14(2):231–241. Epub 2011 Nov 2 10.1007/s1019801103645
Willan AR: Statistical analysis of costeffectiveness data from randomized clinical trials. Expert Rev Pharmacoecon Outcomes Res 2006, 6(3):337–346. 10.1586/14737167.6.3.337
Willan AR: Sample size determination for costeffectiveness trials. Pharmacoeconomics 2011, 29(11):933–949. 10.2165/1158713000000000000000
Briggs AH, Claxton K, Sculpher MJ: Decision Modelling for Health Economic Evaluation. Oxford University Press, Oxford; 2006.
Ospina R, Ferrari SLP, CribariNeto F: A general class of zeroorone inflated beta regression models. Comp Stat Data Analysis 2004, 31(7):799–815.
Kieschnick R, McCullough BD: Regression analysis of variates observed on (0,1): percentages, proportions and fractions. Stat Model 2003, 3(3):193–213. 10.1191/1471082X03st053oa
Swearingen CJ, Castro MSM, Bursac Z: Inflated Beta Regression: Zero, one and Everything in Between. 2012.
Ospina R, Ferrari SLP: A general class of zeroorone inflated beta regression models. Comput Stat Data Anal 2012, 56: 1609–1623. 10.1016/j.csda.2011.10.005
Ferrari SLP, CribariNeto F: Beta regression for modelling rates and proportions. J Appl Stat 2004, 31(7):799–815. 10.1080/0266476042000214501
Fleishman AI: A method of simulating nonnormal distributions. Psychometrika 1978, 43: 521–531. 10.1007/BF02293811
Pourahmadi M, Daniels MJ, Park T: Simultaneous modelling of the Cholesky decomposition of several covariance matrices. J Multivariate Anal 2007, 98(3):568–587. 10.1016/j.jmva.2005.11.002
Kind P: Measuring the value of quality of life in cancer: an index based on EORTC QLQC30 presentation (abstract), ASCO 2005 Annual Meeting. J Clin Oncol 2005, 29(16S, Part I and II).
Acknowledgements
We would like to thank Cancer Research UK as funders of the original trials: grant numbers C1438/A8972 and C1438/A4147 (TOPICAL) and 11922/A4558 (SOCCAR). We also thank the chief investigator for TOPICAL, Dr SM Lee, Dr J Maguire and the Trial Management Groups for agreeing the use of data for publication. We are grateful to SJ Sarkar, Allan Hackshaw and Jeff Round for making valuable comments to earlier versions of this work.
Author information
Authors and Affiliations
Corresponding author
Additional information
Competing interests
The authors declare that they have no competing interests.
Electronic supplementary material
12955_2014_163_MOESM1_ESM.doc
Additional file 1: Table S1.: Summary Statistics of EQ5D and QLQC30. Table S2: Summary of Models: Coefficients. (Pvalues). (DOC 80 KB)
Authors’ original submitted files for images
Below are the links to the authors’ original submitted files for images.
Rights and permissions
This article is published under an open access license. Please check the 'Copyright Information' section either on this page or in the PDF for details of this license and what reuse is permitted. If your intended use exceeds what is permitted by the license or if you are unable to locate the licence and reuse information, please contact the Rights and Permissions team.
About this article
Cite this article
Khan, I., Morris, S. A nonlinear betabinomial regression model for mapping EORTC QLQ C30 to the EQ5D3L in lung cancer patients: a comparison with existing approaches. Health Qual Life Outcomes 12, 163 (2014). https://doi.org/10.1186/s1295501401637
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s1295501401637