International Choice Modelling Conference, International Choice Modelling Conference 2017

Font Size: 
Estimating a preference-based single index measure for the impact of self-management on quality of life in diabetes for the UK general population
Alexander Michael Labeit

Last modified: 28 March 2017

Abstract


Objective: The Evaluating Self-Management in Diabetes (ESMiD) study involved the valuation of a classification system which includes both health related and self-management attributes for diabetes. A Discrete Choice Experiment incorporating duration (DCETTO) was used to value the descriptive system and produce a utility scale anchored on the full health (1) to dead (0) scale. The resulting utility scale can be used to estimate Quality Adjusted Life Years (QALYs) for use in cost utility analysis.

 

Methods: The Health and Self-Management in Diabetes (HASMID) classification system was developed as part of the ESMiD study to capture the impact of self-management on quality of life in diabetes. Four dimensions cover health-related quality of life (HRQoL) (mood, hypoglycaemic attacks, vitality and social limitations) and four other dimensions cover self-management (control, hassle, stress and support). The multi-attribute classification system has eight attributes, each with four levels (e.g. “never”, “sometimes”, “usually” and “always”) with the best level of an attribute as the reference level. The duration levels used in the survey were 1 year, 4 years, 7 years and 10 years. These levels have been successfully used in previous health state valuation studies of the SF-6Dv2 and EQ-5D.

An online survey collected data from the UK general population (n=1,493). Respondents were recruited via an existing online panel and were targeted to be representative of the UK population in terms of age and gender. Each DCETTO task presented a choice set including two different profiles, and each respondent completed 12 DCETTO tasks. Each profile consists of attribute levels selected from the classification system, and respondents are asked to indicate their preferred profile. Eight attributes from the classification system plus four duration levels resulted in 262,144 possible profiles and many more combinations of profiles into choice sets, meaning it is infeasible to value all. Profiles were selected for valuation using D-optimal methods in order to produce a design that enables estimation of the parameters in a pre-specified regression.

According to the approach of Bansback et al. (2012) coefficients are estimated for the interaction of the dimension level and duration attributes, and a separate coefficient for duration. To anchor the coefficients on the utility scale, the interacted coefficients are divided by the duration coefficient. Utility values for each state can then be estimated from the anchored coefficients for each level of each dimension.

The first econometric approach used a conditional fixed-effects logit model and the second econometric approach used panel latent class and mixed logit models as extensions of the conditional logit model which can incorporate preference heterogeneity. Latent class models have been estimated without and with class membership variables. Sex, age, having diabetes, belonging to a low or high household income group, EQ-5D have been used as class membership variables. Mixed logit models with two different specifications have been estimated: a first specification with all coefficients random except the duration variable and a second specification with all coefficients random including the duration variable; for the random coefficients a normal distribution has been selected. Information criteria (AIC, BIC and CAIC) was used to compare the statistical fit of the different models. The robustness of the results was also checked by excluding respondents based on the time taken to complete the survey: respondents with answer time less than 10 seconds, 5 seconds and with no exclusion of respondents.

 

Results: The valuation of the classification system produced consistent estimates for seven of the eight different attributes in the conditional logit fixed-effects model. There is one logically inconsistent coefficient (stress with level 2) and three insignificant coefficients for the milder levels of some other attributes (mood level 2, hypoglycaemic attacks level 2, social limitations level 2). The coefficients for the health and self-management attributes have the expected negative signs, showing that individuals have a preference to live in better health and self-management states. The absolute size of the coefficient increases as the severity levels of an attribute increase. The coefficients for the health and self-management attributes are of similar magnitude, suggesting that people value self-management as being as important as health improvement. The duration coefficient has also the expected positive coefficient showing that individuals have a preference to live longer. The anchored utility values of the conditional-fixed effects logit estimation ranged from 1 (for the best health state with no problems on each dimension) to -0.029 (for the worst health state with the highest level of problems on each dimension). Robustness checks show similar results if respondents with an answer time less than 5 or 10 seconds were excluded.

Preference heterogeneity is confirmed by the latent class models and mixed logit models, as both models shows an improved statistical fit in comparison to the conditional fixed-effects logit model (BIC: 22089). The latent class model with 3 latent classes as the preferred choice performs slightly better (BIC: 21431) than both mixed logit models specifications (BIC 22170 for the first specification and BIC 21649 for the second specification). Inclusion of the class membership variables brought no additional improvement of the statistical fit according to the information criteria.

 

Discussion: Our results show that the anchored coefficients of the four self-management attributes are similar in magnitude to more conventional health attributes like mood, social limitations and hypoglycaemia. The finding that the self-management attributes were important to the general population sample is important for policy reasons, because currently National Institute of Health and Care Excellence (NICE) uses the latter to value health in order to derive QALYs for assessing the cost-effectiveness of new interventions. Latent class and mixed logit models are preferred in comparison to the conditional fixed-effect logit model and utility values will be also derived.

 

Conclusion: Health and self-management can be valued in a single classification system using DCETTO to produce a single value by using QALYs. The results presented here can be used in economic models of the cost-effectiveness of new interventions to value the impact of interventions on HRQoL.


Conference registration is required in order to view papers.