International Choice Modelling Conference, International Choice Modelling Conference 2017

Font Size: 
Explore preference heterogeneity for the treatment preferences of people with type 2 diabetes: A comparison of random-parameters and latent-class estimation techniques
Mo Zhou, John F P Bridges

Last modified: 28 March 2017


Background: Diabetes affects nearly 30 million individuals in the US and results in significant morbidity and costs. As a prevalent chronic condition, diabetes requires meaningful patient engagement to improve outcomes, but poor adherence to known effective interventions persists. Since clinical guidelines often do not reflect the diverse needs among diabetic patients seen in practice, understanding patient preference and incorporating it in the guidelines may improve adherence, satisfaction, and quality of life. The aim of this study is to understand patients’ preference heterogeneity in diabetes medications and explore factors that potentially contributed to this variation.

Methods: A nationally representative survey was conducted among patients with type II diabetes in the US. The survey contained a discrete-choice experiment with 18 pairs of hypothetical diabetes medications. A Bayesian D-efficient design generated 48 choice tasks divided into three blocks of 16 choice tasks each. Two holdout tasks were added to each block. The medications in each task differed in six attributes, including effectiveness in reducing hemoglobin A1c, duration of blood glucose levels remaining stable, frequency of hypoglycemia, duration of nausea, treatment burden, and cost. Each attribute had three levels. Respondents were randomly assigned to a block and asked to choose the medication that they preferred in each choice task.

Data were first analyzed using mixed logit (MXL) and scale-adjusted latent class logit (LCL) models to detect preference heterogeneity. The appropriate number of classes was selected based on Bayesian Information Criterion (BIC), consistent Akaike Information Criterion (CAIC), segment size, and interpretability. Relative importance scores (RI) were calculated to measure the relative weights that respondents assigned to each attribute. A limitation of LCL model is that it assumes preference coefficients can only take a set of distinct values and are homogeneous within classes, which might not hold empirically. In the presence of overlap between classes or substantial within-class heterogeneity in the coefficients, LCL may lead to too many classes. As MXL, with an assumption of continuously distributed preference weights, can accommodate more extensive heterogeneity with fewer parameters, it may not be sufficient when there are sizeable subgroups.

As a remedy, we allowed random effects in the LCL model to capture within-class heterogeneity. The model reduces to LCL when the parameter variances equal to zero for all classes, and is the traditional MXL when there is only one class. The appropriate number of classes was selected based on BIC, CAIC, segment size, and interpretability. The model was estimated using the Expectation-Maximization and Newton-Raphson algorithms. It was compared to MXL and LCL based on BIC and the prediction errors. Multivariate analysis of variance (MANOVA) was used to explore factors (i.e. patient demographic, socioeconomic, and health status variables) associated with class membership.

Results: 552 respondents with type II diabetes filled out the survey with a 66% response rate. 543 completed all choice tasks and were included in the analysis. We oversampled patients who were African Americans and Hispanics because they have a high prevalence of diabetes and lower adherence rates to treatments. Respondents in the sample were comparable to the general population in most variables except that they were slightly more educated and had higher income level. Samples were rescaled using sampling weights.

All attributes had significant impacts on choices in MXL model (p<0.001). All effects were in the expected direction (e.g. higher effectiveness and lower cost were associated with higher utility weights) except that respondents gave similar utility weight to having one pill a day (0.334) and having two pills a day (0.374) (p>0.05). The standard deviations indicate that there were significant variations in the preference for all attributes (p<0.05) except the frequency of hypoglycemia (p=0.089).

The best-fitted scale-adjusted LCL model contained five preference classes and two scale classes. As in MXL, all attributes had significant impacts on choices (p<0.01). Preference weights for medication effectiveness, duration of stable blood glucose level, and treatment burden were significantly different across the five classes (p<0.05). Class 1 (24% of sample) considered cost the most important attribute (RI=0.47). Class 2 (24%) focused on treatment burden (RI=0.42) while gave little consideration to effectiveness (RI=0.04). Class 3 (23%) predominantly concerned about duration of nausea (RI=0.56) and also considered treatment burden (RI=0.16). Class 4 (18%) cared most about effectiveness (RI=0.48) and duration of nausea (RI=0.20). Class 5 (11%) considered duration of stable blood glucose level the most important attribute (RI=0.43) and cost not important (RI=0.04).

When we allowed random effects in the LCL model, three classes were identified. Compared to MXL and LCL, the model has lower BIC (8362.79 vs. MXL 8587.38 and LCL 8403.18) and lower predication error (12.69% vs. MXL 13.02% and LCL 15.69%). It also has higher degree of freedom (df=493) and higher Entropy R-squared (0.84) than LCL (df=477 and Entropy R-squared=0.77).

Based on the random effect LCL model, respondents in class 1 (60%) considered duration of nausea (RI=0.34) and effectiveness (RI=0.23) when making choices. Class 2 (20%) predominately focused on cost (RI=0.46). Class 3 (20%) considered treatment burden the most important attribute (RI=0.43) while also concerned about nausea (RI=0.20) and cost (RI=0.16). Respondents in class 2 were younger and had lower income (p<0.05). Respondents in class 3 were more likely to be single and have higher income (p<0.05). African Americans (p=0.001) and Hispanics (p=0.003) were more likely to be in class 1 and 3 after controlling for age, income, and marital status.

Conclusions: Preference heterogeneity exists among patients with type II diabetes. Younger and lower income people consider cost the most important factor when choosing medications, while single and high-income group focuses on treatment burden such as the number of pills they have to take per day and if there is any injection. Such heterogeneity in preference should be considered in clinical decision-making for diabetes treatment and self-management.

In the presence of significant within-class heterogeneity, LCL may lead to too many classes and random effect LCL can generate more parsimonious model. Random effect LCL provides the flexibility of LCL and the parsimony of MXL. Researchers can use it to test any over-fit of data in future studies.

Conference registration is required in order to view papers.