Results Summary
What was the research about?
If patients don’t take medicines as directed, the medicines don’t work as well for treating a health problem. It may also lead to more health problems. If doctors knew which patients were less likely to take medicines as directed, they could find ways to help these patients.
In this study, the research team wanted to learn if knowing who took medicines as directed in the past would predict if patients take a new medicine as directed. The team created two statistical models to predict if patients would take a medicine as directed. First, the research team created a model to predict if patients would take medicines to lower cholesterol. Then, they created a second model using data from these patients plus others who were taking medicines to lower blood pressure or strengthen bones.
What were the results?
The statistical model to predict if patients would take medicines to lower cholesterol worked well. But using the same model to predict if patients were taking other medicines didn’t work as well. The model that was based on data from more groups of patients worked better at predicting if patients were taking medicines to lower blood pressure or strengthen bones.
What did the research team do?
The research team looked at health record data from 89,490 patients. These patients started taking medicines to lower cholesterol during the study. Of these, the average age was 55, and 54 percent were women. The team also used data from patients who started taking medicines either to lower blood pressure or strengthen bones during the study.
The research team measured how often patients filled their prescriptions in the past. The team used these data to create a statistical model to predict if patients took medicines to lower cholesterol. Then, the team created another model to predict if patients would take more types of medicines. Both models took into account patients’ traits, such as age and health problems, that might also affect if patients take medicines as directed. The team compared how well the two models worked.
A group of 10 patients helped to design the study.
What were the limits of the study?
Other things, like family support, can affect if patients take medicines as directed. Health records often don’t include these data. Also, the team had data about how often patients filled prescriptions. But they didn’t know if patients took the medicines. Results may have been different if the team had these data.
To improve the models, future research could collect data on whether patients took their medicines.
How can people use the results?
Researchers can use these results when considering ways to predict which patients may not take their medicines as directed.
Professional Abstract
Objective
To develop and compare algorithms to predict medication adherence using measures of prior adherence
Study Design
| Design Element | Description |
|---|---|
| Design | Empirical analysis |
| Data Sources and Data Sets | Healthcare claims database for 176,666 insurance beneficiaries enrolled in commercial UnitedHealth Group health plans and patients with a Medicare supplement plan |
| Analytic Approach |
|
| Outcomes | Full adherence as a patient outcome and c-statistics to compare model performance |
Poor medication adherence leads to adverse clinical outcomes and substantial avoidable medical spending. Identifying patients who are likely to have low adherence may help develop effective interventions. This study developed algorithms to predict full adherence to a newly initiated medication based on patients’ prior medication adherence. The research team defined full adherence as at least 80% proportion of days covered (PDC), which is the proportion of days on which a patient has medication available during a year.
The study included data for 89,490 patients who had a statin dispensed during the study period and who had at least one different medication dispensed, used for measuring prior adherence. Of these patients, 54% were female, and the average age was 55. In addition, the study used data from 67,607 bisphosphonates initiators and 109,059 antihypertensives initiators.
The research team split the sample into training and testing cohorts. Using data from the training cohort, the team fit lasso logistic regressions to identify variables that could predict full adherence to statins other than prior adherence. With these variables and measures of prior adherence, the team used multiple logistic regressions to predict full adherence to statins. The team then applied each model from the training cohort to the testing cohort and selected the best performing algorithm.
To test whether the selected statin adherence prediction algorithm could apply to other drug classes, the research team applied the algorithm in cohorts of patients starting bisphosphonates or antihypertensives. The team compared the performance of the statin algorithm with algorithms they developed for bisphosphonates and antihypertensives. Finally, the team pooled all patients and developed a unified algorithm to predict adherence to any of the three drug classes and compared its performance to the class-specific algorithms.
A panel of 10 patients provided input into the design and conduct of the study.
Results
Among prior adherence measures for statins, combining mean prior PDC with baseline covariates like age, sex, and comorbidities yielded the best prediction. The likelihood of high adherence was greater for patients with high mean prior PDC (≥80%) compared with those with low (<25%) and moderate (between 25% and 79%) mean prior PDC.
The statin algorithm did not perform as well for the other two drug classes. Using drug-class-specific algorithms for the two drug classes improved prediction. The unified algorithm was better at predicting future medication adherence than applying the statin algorithm to other drug classes but not as effective as applying drug-class-specific algorithms.
Limitations
Claims data may not include information on factors that influence adherence decisions such as social support and beliefs about medications. Also, adherence measures are subject to measurement error because claims data record how often patients fill prescriptions rather than how often they take the medicine.
Conclusions and Relevance
Adding measures of prior adherence leads to modest improvements in the predictive ability of claims-based adherence prediction algorithms. Drug-class-specific algorithms performed better than a unified algorithm. However, a unified algorithm may be useful when developing drug-specific algorithms is not feasible.
Future Research Needs
Future research could assess the generalizability of these findings to other medications and populations. Researchers could also incorporate other data sources, such as electronic health records, to improve adherence prediction.
Final Research Report
View this project's final research report.
Peer-Review Summary
Peer review of PCORI-funded research helps make sure the report presents complete, balanced, and useful information about the research. It also assesses how the project addressed PCORI’s Methodology Standards. During peer review, experts read a draft report of the research and provide comments about the report. These experts may include a scientist focused on the research topic, a specialist in research methods, a patient or caregiver, and a healthcare professional. These reviewers cannot have conflicts of interest with the study.
The peer reviewers point out where the draft report may need revision. For example, they may suggest ways to improve descriptions of the conduct of the study or to clarify the connection between results and conclusions. Sometimes, awardees revise their draft reports twice or more to address all of the reviewers’ comments.
Overall, the reviewers found the report to be outstanding and clearly written. They found all of the conclusions to have support from the study’s findings. The reviewers provided a list of specific comments and suggestions primarily with the goal of enhancing clarity, and the researchers edited some of the language of the report in response and added suggested citations.