Predictive models to estimate utility from clinical questionnaires in schizophrenia: findings from EuroSC

Economic evaluation allows comparing several alternative therapies in terms of benefit brought by a new treatment and associated costs. The comparison is often conducted by the calculation of incremental cost-effectiveness ratio (ICER). Cost-utility analysis is a common type of economic evaluation in which product effectiveness is expressed in quality-adjusted life years (QALYs). QALYs are determined by multiplying life years gained using the therapy by utilities, corresponding to patient’s quality of life (QoL) during this period. Utility of a certain health state accounts for different aspects of QoL and corresponds with the desire or preference that individuals exhibit for this state. Unlike psychometric measures that consider only the degree of abnormality of an impairment of health, utility preference-based measures allow estimating the significance of the impairments and so can be used in economic evaluation to assess the value of interventions from the perspective of the patient. A number of instruments have been developed to measure preference-based utilities directly, or through generic or disease-specific preference-based questionnaires. However, in the context of developing cost-utility models, utility measurements are rarely available and are sometimes predicted using mapping extrapolation from a clinical questionnaire and other parameters [10].

Schizophrenia is chronic mental disorder which significantly impacts patient’s QoL [11, 12, 34]. It affects approximately 1 % of the general population, and onset usually occurs before the age of 25 years [28]. Schizophrenia symptoms are generally devised in positive symptoms (psychotic behaviours not seen in healthy people, i.e. hallucination, delusions, etc.), negative symptoms (decrease or loss of normal functions, i.e. apathy, lack of emotions), and cognitive symptoms. The key component of schizophrenia treatment is antipsychotic medications which are mainly used to manage positive symptoms. There exist two generations of antipsychotic drugs called typical (first generation) and atypical (second generation) antipsychotics.

There is a large range of symptoms and treatment-related adverse events, and these are associated with serious functioning, social, and QoL alterations in patients with schizophrenia. This makes the utility assessment challenging. In 2010, Mavranezouli reported that seven utility or cost-utility analyses were performed for schizophrenia treatment [29]. Out of these analyses, none of the studies used utility values for schizophrenia which had been generated by using EQ-5D, otherwise widely used in cost-utility analyses and a preferred instrument of the National Institute for Health and Clinical Excellence NICE [32]. Therefore, developing a predictive equation to estimate utility measures based on clinical, functioning, and QoL variables would support and simplify health economics assessment of therapies designated to treat patients suffering from schizophrenia.

Several studies have investigated the independent predictors of QoL in people with schizophrenia. These studies report that clinical factors, such as positive and negative symptoms, depression, and extra-pyramidal symptoms, are associated with low QoL [2, 14, 22, 27, 33, 36, 38, 39, 43].

The objective of this article is to generate a predictive equation for EQ-5D utility in schizophrenia which would allow its extrapolation from other instruments and clinical parameters, using data from the European Schizophrenia Cohort (EuroSC). Two predictive models were developed by using two sets of variables in order to take into account differences in the availability of clinical data to practitioners. Additionally, a predictive model for SF-6D utilities was developed to study mapping mechanisms and to determine whether mapping the same set of predictors to the utilities collected, using the two different instruments, conforms to a different algorithm.

Descriptive results are presented in Table 1. The mean patient age was 41.8 years. Males were more prevalent in the cohort (63 vs. 37 %). The mean PANSS was about 55, corresponding to “mildly ill” state.

The correlation structure of the explanatory variables is summarized in Table SS1 (see Online resource). A strong correlation was observed between the PANSS_POS and PANSS_PSY scores (0.71), the PANSS_NEG and PANSS_PSY scores (0.69), and the PANSS_NEG and GAF scores (−0.5). The Hausman’s test showed that the correlation between the variables led to the co-linearity and would cause estimation biases, justifying the choice of the procedure for variable selection.

No significant difference was observed in the variables means between the sample without missing values and the sample with at least one missing value for all patients (see Table SS2 in Online resource). Both sub-samples were very close, thus confirming the absence of selection bias.

This study aimed to build a model to map the demographic and clinical measures of patients with schizophrenia to EQ-5D index. Although the EQ-5D is currently recommended by NICE [32] for use in economic evaluation, it remains largely under-utilized in clinical trials for schizophrenia. As suggested by the NICE [32], mapping can be used when EQ-5D is not included in clinical trials. The proposed mapping functions can constitute the first step in promoting the assessment of utility values in schizophrenia as required in cost-utility analyses. Our findings suggest that the mapping relationship between the socio-demographic, clinical characteristics, and EQ-5D is reliable and robust.

From a clinical perspective, age, gender, PANSS psychopathology score, and CDSS score are the most important predictors associated with EQ-5D. These findings are consistent with those of the studies focusing on non-preference-based health status measures. Age negatively affects the utility measure. This is in line with Kemmler et al. [25] results, showing that social problems, isolation, and stigmatization of patients with schizophrenia tend to increase with age. Male gender positively affects utility measures. This finding appears to be consistent with the general literature, in which the quality of life of female patients is often reported to be lower than that of men, especially with regard to psychological and mental health domains [21, 35]. Regarding the influence of the PANSS scores, PANSS psychopathology factor was found to negatively affect the utility (p < 0.001), whereas the PANSS positive and negative factors do not affect utility measures. Similarly to the PANSS psychopathology factor, CDSS score negatively affects the utility (p < 0.001). These findings concur with those of several studies and meta-analyses that reveal that symptoms have only a modest relationship with quality of life and that general psychopathology symptoms (e.g. anxiety and depression) were the most important predictors [8, 16, 20]. Finally, extra-pyramidal symptoms (BAS) are associated with lower utility measures. In point of fact, the burden of the side effects has been extensively explored as a predictor of poor medication adherence, relapse, and poor QoL [3, 18, 21].

From a methodological aspect, the high co-linearity among the explanatory variables led to the inappropriateness of variable selection based on the statistical significance. A specific procedure employing the principle of PCA was then developed and applied. Inconsistent parameter estimates were obtained for the EQ-5D random effect model based on PANSS score and additional covariates because of the correlation between some explanatory variables and the errors terms. The estimates were then recalculated using the fixed effects panel model, which provided consistent results. Additionally, a number of other specification tests were performed in order to provide the evidence that the proposed mappings model is well specified and reliable. Finally, the proposed approach was also tested with SF-6D. The results were coherent with those obtained for EQ-5D confirming the robustness of the method. In their paper, McCrone et al. [30] showed that from an analytical perspective, the SF-6D has advantages over the EQ-5D due to its normal distribution and the lack of ceiling effect. However, both measures produced similar mean utility scores, and further comparisons of the EQ-5D and SF-6D were required. Additionally, the prediction errors were calculated using RMSE and MAE. Their values remained moderate. For the EQ-5D model, RMSE was around 0.25 and MAE was approximately 0.18. SF-6D model prediction errors were quite lower (RMSE and MAE were about 0.12 and 0.10, respectively). The cross-validation results confirmed the stability of the models. It may be concluded that the models are applicable to predict utility measures.

It should be noted that the EQ-5D value theoretically range from minus infinite to one. In the original data, the range is [−0.429; 1.000]; however, the range of predicted values is [0.114; 0.930]. A similar pattern is observed for SF-6D. Predictive models generate predictions of the conditional mean, but not of the conditional variance and thus may compress the range of individual predicted values. The variance can be accounted for by using the predicting interval instead of the predicting value. Siani et al. [37] show how to account for uncertainty in the context of mapping prediction.

Regarding the analysis limitations, the representativeness of the sample should first be discussed. Although the sampling procedure for the EuroSC aimed to provide a representative patients sample, this cohort included mostly paranoid schizophrenia and is characterized by long-term illness. Moreover, the difference in severity between excluded and included patients was observed (with higher clinical severity in excluded patients). Further analysis is therefore required, using larger and more diverse groups of patients. However, the large sample size of the presented study and the longer follow-up allowed overcoming the limitations of past studies [26].

Because treatments for schizophrenia have significant effects on the quality of life of patients, reliable methods for economic evaluations are needed to account for effects of treatment, to assess utility values, and to calculate QALYs for further cost-utility analyses. This paper proposes reliable, accurate, and easy-to-apply mapping models for EQ-5D index based on demographic and clinical measures in schizophrenia. An advantage of these mapping functions is the use of generally available in clinical trials data, such as PANSS score, which expands its applicability.