Evidence-Based Adherence Classes for Combination Antihypertensive, Lipid-Lowering, and Antidiabetic Therapy
Adherence classes were determined for patients taking multiple medications (ie, lipid-lowering, antihypertensive, antidiabetic) and were found to be related to serum profiles and risk of adverse events.
National databases and large population studies find that more than one-fourth to one-half of all Americans have multiple chronic conditions (MCCs).1,2 The prevalence of MCCs increases throughout adulthood3-6: among Medicare beneficiaries, 83% have at least 1 chronic condition and nearly a quarter have 5 or more.7 The numbers of chronic conditions strongly correlate to health risks, ranging from functional decline and hospitalizations to disability and mortality.3 Patients with 2 chronic conditions cost twice as much as those with only 1, and those with 3 or more chronic conditions cost twice as much as those with 2.2,8 Although identified nationally as a priority for healthcare research,9 studies of patients with MCCs are limited10—especially articles on patients with specific combinations of comorbidities.
As the number of chronic conditions increases, patients are treated with medication treatments of growing complexity and prescription cost.3,11 As a result, medication management becomes a challenge for patients with MCCs, and patients who poorly manage their medications also risk adverse events (AEs), including adverse drug reactions.12,13
A concern for patients with multiple medications is medication adherence, which can be defined as how well patients take their medications as prescribed.14,15 A recent review of medication adherence with multiple medications noted the limited literature on patients with 3 or more chronic conditions16—only a few studies have examined adherence among patients having the same chronic conditions. Further, most studies were on drug-induced morbidity; studies of chronic disease management were uncommon.16
A recent analysis of Medicare data identified diabetes as the most common condition among patients with MCCs.8 Older patients with diabetes commonly have hypertension and hyperlipidemia, with cardiovascular comorbidities becoming increasingly prevalent as well.17 The medication burden arising from these comorbidities can make managing diabetes especially difficult.18-20 Most patients with diabetes, hypertension, and hyperlipidemia do not consistently adhere to all of their medications, while those who do achieve good adherence can substantially reduce their risk of emergency department (ED) visits and hospitalizations.21 Poor adherence with hypertension, hyperlipidemia, and diabetes medications costs an estimated $105 billion per year.22
This study examines medication use among patients with hypertension, hyperlipidemia, and diabetes across a 3-year period. The objective is to understand patients’ patterns of medication adherence for all 3 conditions. Some patients, for example, might display similar adherence with medications for all 3 of the chronic conditions, while other patients might adhere better to some medications than to others.23 Although Rolnick et al reported substantial variation in adherence by condition among patients simultaneously prescribed medications for several conditions, they did not look for patterns of adherence to multiple medications.24 Among patients taking multiple medications for diabetes, Choudhry et al found that only about 1 in 3 patients achieved concurrently high adherence (ie, >80%) to 2 or more classes of oral hypoglycemics.25
We hypothesized that adherence patterns would vary across patients and applied a statistical approach to show the patterns. Unlike many studies that have defined medication adherence as taking medications 80% or more of the days prescribed,14,15 we chose instead to treat adherence as continuous, and posed 3 research questions: 1) “Can patients filling prescriptions to treat hypertension, hyperlipidemia, and diabetes be grouped into distinct patterns of medication adherence?” 2) “If so, do the adherence classes have clinically meaningful differences in health outcomes?” 3) “What patient characteristics are associated with class membership?”
The patients were members 25 years or older enrolled from 2007 to 2010 with the largest insurer in Hawaii. Both Hawaiian and national data show a very low prevalence of MCCs at younger ages.4,16 The patients were required to have at least 1 prescription per year for an antihypertensive medication, a lipid-lowering medication, and an oral antidiabetic medication. A total of 8452 patients met these criteria. Data analysis was restricted to the 2008 to 2010 interval. The prescription from 2007 was included as an eligibility criterion so that the population analyzed did not include new users who may have differed from established users in their medication adherence.26 The eligibility criteria were chosen to select patients who were managing 3 medications across multiple years. Continuous enrollment was required for the last 3 years; 6228 patients met this enrollment criterion. Further, only patients with annual measurements of both low-density lipoprotein cholesterol (LDL-C) and glycated hemoglobin (A1C) were included. The 4948 patients meeting these criteria were included in the analyses. The study was granted an exemption from institutional board review by the University of Hawaii.
Medication adherence was calculated as the proportion of days covered (PDC), a widely used metric27:
Number of days in period “covered” by medication /
Number of days of drug enrollment
Adherence was calculated by year for each of the 3 study medications. Other study variables included age, sex, residence on the most populous island of Oahu (ie, urban residence) or on a more rural neighboring island, and morbidity measured by the Johns Hopkins Adjusted Clinical Group Methodology.28 High morbidity was defined as a 4 or 5 on a 5-point scale. Previous studies have associated medication adherence with demographic characteristics and morbidity.29 Study outcomes were LDL-C levels of 100 mg/dl or higher, A1C levels of 9% and above, and the annual number of ED visits and hospitalizations. The LDL-C and A1C levels correspond to Healthcare Effectiveness Data and Information Set (HEDIS) measures for elevated LDL-C and poor glycemic control, respectively.30 Patients averaged 2.7 LDL-C measurements and 3.0 A1C measurements per year. Patient means per year were used in the analyses.
The analyses used a latent class model that both grouped patients into classes and estimated the rates of transition between classes over time.31 Patients were classified into their most likely class from the latent class model; transition rates were estimated as part of the regression results. Latent class models classify people based on multiple indicators of behavior.32,33 Ideally, indicators used to parse individuals into classes should vary substantially across people being classified. This study used medication adherence to antihypertensive, lipid-lowering, and antidiabetic medications as indicators to define the classes. Thus, a class was a group of patients with a similar pattern of adherence among the 3 medications groups.
In preliminary analyses, separate models were fit for each study year, resulting in similar class solutions. Therefore, to simplify interpretation, the associations of medication use with the classes were constrained to be the same for all 3 years. The optimum number of classes was decided by comparing models with increasing numbers of classes: the 3-class solution was better than the 1- or 2-class solution based on 3 measures of goodness of fit: a lower Bayesian Information Criterion, a significant bootstrap likelihood ratio test, and high entropy.33,34 Entropy is a measure of the strength of the classification; an entropy of 0.8 or better is considered high.35 The 4-class solution had a decrease in entropy and was less interpretable; it included a small class with only about 5% of the patients.
This article presents results from the 3-class solution. The primary model included the indicators of adherence and the 4 study outcomes (LDL-C and A1C levels and the annual number of hospitalizations and ED visits). The model included correlations of adherence within classes. Associations of classes with patient characteristics were analyzed in a second model using multinomial regression. This 2-step approach prevents the predictors of class membership from influencing the classes formed. From the primary model, individual patients received a probability of belonging to each of the 3 classes, and their most likely class assignment was used in the second step. All analyses were done using Mplus version 7.1 software (Muthén & Muthén, Los Angeles, California).
The average age of the 4948 patients was close to 65 years and the group included slightly more men than women. Over one-third of the patients had high morbidity. Medication adherence for patients with diabetes and hypertension averaged about 85% of the possible days whereas lipid-lowering medication adherence averaged slightly below 80% (
). Adherences with the 3 medications were moderately correlated (
The latent class analyses categorized the patients into 3 classes that varied substantially in adherence with lipid-lowering medications, antihypertensive medications, and antidiabetic medications (
). The entropy for the 3-class solution was 0.89, suggesting the data strongly supported the classification. The classifications are labeled "low," "intermediate," and "excellent" adherence for purposes of comparison. Patients in the excellent adherence class had filled prescriptions, on average, more than 90% of days for all 3 medications. The intermediate class remained adherent much of the time, averaging 71% to 85% of days with filled prescriptions. The low adherence class, by contrast, had filled prescriptions closer to half of the possible days, with adherence ranging from 45% for antihypertensive medications to 59% for antidiabetic medications. Fewer than 10% of the patients had low adherence; the intermediate and excellent adherence classes each contained approximately 40% to 50% of the patients (
From year to year, a proportion of the patients changed their adherence class (
). Most changes were to adjacent classes, such as going from low to intermediate, or from intermediate to excellent adherence. About 5% to 33% of patients had transitions from one year to the next; rates varied depending on the classes between which transitions occurred. The highest rate of transition was from excellent to intermediate adherence classes. Only about 5% of the patients moved between the low and excellent adherence classes in adjacent years.
Serum levels and rates of acute events varied significantly by adherence class (
). The percentages of patients with LDL-C levels >100 mg/dL decreased from 40.3% to 20.2% to 13.5%, going from low to intermediate to excellent adherence classes. The comparable percentages for A1C levels of 9% or higher were 18.8%, 8.6%, and 5.8%, respectively, among the 3 adherence classes. For both serum measures, the confidence intervals for adjacent classes did not overlap. Patients with excellent adherence had the lowest rates of ED visits and hospitalizations (0.33 and 0.18 per year, respectively). The rates of ED visits for the low and intermediate classes were 0.45 and 0.40 per year, respectively; both classes had hospitalization rates of 0.31 per year.
A secondary analysis examined the relationship between patient characteristics and class membership (
). Age, sex, morbidity, and residence on Oahu or a neighboring island were included as possible predictors. Age and sex were significantly associated with adherence class: older patients were less likely, and females more likely, to have low or intermediate medication adherence.
The study’s approach using latent class models distinguished 3 classes of patients: those with low, intermediate, and excellent overall adherence with antihypertensive, lipid-lowering, and antidiabetic medications. The models did not select classes with exceptionally good or poor adherence on some but not all 3 medications, suggesting that individuals adherent to 1 medication are likely to be adherent to other medications. The low adherence group was the most distinct with patients who had filled prescriptions only 44% to 57% of the days, with their best adherence being with oral antidiabetic medications. Adherence for the intermediate class ranged from 71% to 85%, which spanned the common threshold used for good adherence of 80% or higher. The excellent adherence class maintained more than 90% adherence with all 3 medications. The low adherence group was the smallest, averaging about one-tenth of the patients. The rest of the patients were divided almost equally between the moderate and excellent adherence groups. A subset of patients transitioned across the adherence categories from year to year, but changes occurred mainly between adjacent classes. The results suggest that most patients were persistent in how they managed their lipid-lowering, antihypertensive, and antidiabetic medications.
The entropy of 89% provided good support for the 3-class solution. Attempts to find more classes led to a decrease in entropy and classes with small percentages of patients. The 3 classes of medication adherence had clinically meaningful differences in serum levels of LDL-C and A1C, and rates of AEs. The percent of patients with LDL-C levels of 100 mg/dL or greater and A1C levels of 9% or greater decreased by half when comparing intermediate with excellent adherence. The percentages decreased 2-fold again when comparing patients with poor adherence to those with intermediate adherence. The excellent adherence group had the lowest rates of ED visits and hospitalizations and the intermediate class had lower rates than the low adherent class. However, the large overlap between the confidence intervals of the 2 classes suggests lack of a statistically significant difference.
The results indicate that increasing adherence, even among patients with relatively good adherence, might improve clinical outcomes. The low adherence class, with 40% to 60% adherence on the 3 medications, had the worst serum profiles and the highest rates of ED visits and hospitalizations. Identifying such patients may offer a first step toward improving their management. The intermediate class maintained adherence close to 80% of the days, suggesting that these individuals refilled their prescriptions with some regularity and maintained good engagement with their medication regimen. As a result, patients with intermediate adherence may not recognize nonadherence as a concern. The results, however, show that improving medication adherence to an excellent level provides added benefits. Furthermore, it questions the practice of relying on 80% adherence as a cutoff point for differentiating adherent from nonadherent individuals, as health outcomes among individuals with intermediate adherence (75%-85%) were significantly worse than among those with excellent adherence (exceeding 90%).
This study found male sex and older age to be significantly associated with improved adherence. Older patients may be more aware of their health status and potential adverse consequences of poor adherence or be more engaged with their healthcare providers. Poor adherence, however, can stem from a complex array of underlying factors. The World Health Organization proposed a multidimensional adherence model with 5 dimensions: socioeconomic, healthcare system—related, condition-related, treatment-related, and patient-related factors.36-38 Studies have reported variations in adherence by all 5 dimensions.10,18,24,39 Socioeconomic factors include minority and socioeconomic status and social support, system factors include convenience of pharmacy and communication with healthcare providers, condition-related factors include comorbidity and functional status, treatment-related factors include number of doses and timing of medications, and patient-related factors include language proficiency and health literacy. Small changes in daily routines can also affect medication adherence; 1 study reported breaks in adherence on weekends.40
The importance of improving medication adherence is well recognized. The Institute of Medicine, for example, lists medication adherence as 1 of its top 100 priorities for comparative effectiveness research (CER).41 Clinical trials, however, have achieved only limited improvements in adherence.42 This study show that excellent adherence is possible, even with multiple medications, but is not the norm for many patients.
Nationally, a major effort is ongoing to conduct CER and promote patient-centered care.43 Evaluating methods to conduct the research is important to reach the potential of the investment. This study assessed a method among patients taking medications for MCCs. The method successfully classified patients into 3 patterns of medication users. Expert panels have emphasized the importance of coordinating care to manage such complex patients.1,44
This study has some limitations to consider when interpreting its results. The patients belonged to a single insurer and the results may not generalize to other populations. In addition, by design, the results are specific to insured patients who are taking medications across multiple years. Adherence was defined as PDC; the limitations of using this method are well documented.15 For example, the study could not verify if the patients took the medications they filled. Moreover, the insurer’s database did not include samples provided in the doctor’s office. Also, using adherence over the entire calendar year may have misclassified patients with medications prescribed for shorter periods as nonadherent; this differential misclassification may have contributed to class transitions observed. However, the study chose 3 chronic conditions for which medications are generally prescribed long term and the actual rate of class transitions was low. Another limitation is that the categories (antihypertensive, lipid-lowering, and oral antidiabetic medications) are broad; there are many classes of medications within each of these groupings profiles. Blood pressure measurements, another important outcome, were not available in the database.
The study’s primary objective was to identify patterns of medication adherence. The analyses include predictors of adherence classes, but the data available precluded an in-depth analysis of likely underlying factors. Nevertheless, the statistical models used identified patients with differing patterns of adherence with multiple medications. The patterns observed showed associations with clinically important outcomes (LDL-C and A1C).
In summary, using latent class analysis, this study identified 3 patterns of adherence among patients taking lipid-lowering, antihypertensive, and antidiabetic medications. Based on the statistical criterion of entropy, the data provided good support for the patterns obtained. The approach is novel in letting the data decide the adherence classification as opposed to using predetermined cutoffs.
Patients in the highest adherence class, designated as excellent, had the best clinical profiles, with the lowest LDL-C levels, best glycemic control, and fewest ED visits and hospitalizations. Almost half of all patients achieved excellent adherence; they had filled prescriptions covering over 90% of the days for the 3 study medications. Published studies commonly classify patients with PDC >80% of the days as adherent.45 The convention of using an 80% cutoff to demarcate good adherence should be further explored, given that clinical outcomes in patients with intermediate adherence (75%-85%) in this study were significantly worse than those with excellent adherence. The results illustrate that excellent adherence is possible for many patients managing 3 comorbid chronic conditions.
With patient and physician agreement, excellent adherence might become the goal for others. The study adds to the literature by examining overall adherence with multiple medications considered simultaneously. Methods for adherence studies that reflect the underlying complexity of medication regimens may better reveal patient risks than those that treat each medication separately.