Assistant Professor University of South Carolina Columbia, United States
Background: Classifying patients into medication adherence subgroups provides important information for targeting suboptimal medication adherence. Medication adherence (psychological process) is a complex concept that cannot be measured directly but can be inferred from the observed data. Thus, it's necessary to model the heterogeneity of medication adherence through latent variable framework.
Objectives: To identify medication adherent subgroups and evaluate patient characteristics associated with specific medication adherence subgroup using latent mixed model.
Methods: Latent variable analysis is patient-centered statistical method that models heterogeneity by classifying patients into unobserved (latent) subgroups that share similar characteristics. In latent variable framework, the subgroup membership is not observed but can be inferred from the observed indicator variables. Each patient will have an estimated conditional probability assigned to each latent class. These probabilities reflect the degree of certainty and precision of classification. Patient's class membership can be determined by highest conditional probability. The indicator variables can be continuous or categorical. New users of statins from South Carolina State Health Plan and Medicaid Database will be used to demonstrate the use of latent class mixed model. A series of latent mixed models were built with different number of potential adherence classes. The goodness-of-fit statistics were assessed to find the optimal number of latent classes. After number of adherent classes was determined, the association between patient characteristics and class membership was evaluated. Multinomial logistic regression was used to assess the significance of predictors on membership of adherent class. The longitudinal model was used to evaluate the predictors in each class. The class-specific trajectories of adherence were displayed by latent mixed model via 'hlme' function in 'lcmm' R package.
Results: A total of 5,342 new users of statins were included in the study cohort. Four latent adherence classes were: 2,231 (42%) patients in class 1 (best adherent); 511 (9%) in class 2 (consistent but not persistent); 1,768 (33%) in class 3 (not consistent but persistent); 832 (16%) in class 4 (worst adherent). From class-membership model with class 4 as reference, patients with age 18-24 is more likely to be in classes 1 and 2 compared to patients with age ≥35 (p=.0001; p=.007).
Conclusions: Identifying longitudinal patterns of adherence and associated factors can suggest target groups for future customized adherence interventions. Latent mixed model approach can be used in medication adherence research for other therapeutic areas.
