Author ORCID Identifier

Defense Date


Document Type


Degree Name

Doctor of Philosophy


Pharmaceutical Sciences

First Advisor

David A Holdford, R.Ph., M.S., Ph.D., FAPhA


Background: Medication adherence is a major obstacle to improving health care outcomes in long-term therapies for chronic diseases. According to the World Health Organization, interventions for improving medication adherence can have a higher impact on the health of the population than any other advance in medical treatments. Approximately 125,000 individuals die every year in the U.S. because of non-adherence to medication, representing societal costs of $100-289 billion. Previous research has successfully used group-based trajectories methods to identify similar longitudinal medication adherence trajectories. However, medication adherence is not an isolated behavior and is influenced by many factors that current interventions fail to confront. This study aims to (1) identify longitudinal trajectories of medication adherence of chronic diseases treated with oral medications, and (2) distinguish the predisposing, enabling, and need characteristics, which have been identified following an Andersen’s Behavior Model of health services use theoretical framework. Additionally, this study investigates the association between adherence trajectories membership and a posteriori consequences, that was examined by deploying two alternative predictive methods, one based on classic logistic regression and the other based on machine learning algorithms.

Methods: Participants of the Health and Retirement Study were linked to respective Medicare administrative health care claims between 2008-2016. Group-based trajectory models were used to elicit the number and shape of medication adherence trajectories, among a sample of 11,068 individuals taking hypertension medications, statins, or diabetes medications. Time-fixed and time-varying risk factors were examined using logistic regression and multi group-based trajectory modeling, respectively. The association between medication adherence trajectories and outcomes, including myocardial infarction, stroke, and diabetes-specific outcomes (ophthalmic complications, nephropathy, neuropathy, and peripheral angiopathy) was investigated by comparing logistic regression models with machine learning algorithms based on random forests. The outcomes were identified by respective ICD-9 and ICD-10 codes. Predictive ability of the logistic regression compared to machine learning algorithms was examined using the c-statistic.

Results: Group-based trajectory models were estimated for the sample population taking hypertension medications (n=7,272), statins (n=8,221), and diabetes medications (n=3,214). In the hypertension model, three trajectories were identified, following near-perfect adherence, slow, and rapid decline trajectories, accounting for 47.5%, 33%, and 19.5% of individuals in that group respectively. Five trajectories were identified in individuals taking statins, including near-perfect adherence (35.5%), slow decline (17.1%), low then increase adherence (23.6%), moderate decline (12.6%), and rapid decline (11.2%). The diabetes medications yielded the model with the greatest number of trajectories, including near-perfect adherence (24.2%), slow decline (16.9%), high then increase adherence (25.1%), low then increase (13.8%), moderate decline (10.7%), and rapid decline (9.3%).

Several socioeconomic factors were identified as predictors of non-adherence trajectories, which typically were indicative of lower socioeconomic status. While this study pioneered the use of multi group-based trajectories to identify time-varying predictors of medication adherence trajectories, no coherent trends were observed in the analysis. Nonetheless, loss of spouse was generally found to occur in parallel with decreases in adherence, or the opposite, in which regaining a spouse was met with increases or maintenance of high adherence.

Overall, based on the c-statistic, the logistic regression models exhibited better predictive ability than random forest machine learning algorithms in examining the relationship between medication adherence trajectories and outcomes. All non-adherence trajectories in all three models were found to be more likely to experience myocardial infarction compared to each respective near-perfect adherence trajectory. However, the same was not observed for stroke and diabetes-specific outcomes. All declining trajectories of patients taking hypertension medications were more likely to experience stroke. Additionally, only those in the rapid decline trajectory of statins model and those in the slow decline trajectory of the diabetes medications model were more likely to experience stroke compared to each respective near-perfect adherencetrajectory. In the diabetes medications model, only patients following declining adherence trajectories (slow, moderate, and rapid) were more likely to experience nephropathy and peripheral angiopathy than those following near-perfect adherence. No statistically significant differences were found for ophthalmic complications and neuropathy between the near-perfect adherence trajectory and all other non-adherent trajectories.

Conclusions: The GBTM models displayed a nuanced perspective of how participants in the Health and Retirement Study are adherent to their medication for hypertension, statins, and diabetes and how time-varying factors can be investigated to identify patients at risk of falling into non-adherent trajectories. However, non-adherent trajectories are not equally and statistically significantly found be more at risk of health outcomes than near-perfect adherent trajectories. Quality and health policy implications are discussed in light of the results of this research study.


© The Author

Is Part Of

VCU University Archives

Is Part Of

VCU Theses and Dissertations

Date of Submission


Available for download on Friday, August 11, 2023