The dashed line at 30 days indicates 30 days of use, which is the maximum drug use coverage for each unit.
eTable. Bayesian Information Criteria (BIC) and Proportion of Participants in Each Class With Posterior Membership Probabilities Above Selected Thresholds (0.7, 0.8, 0.9)
eFigure. Flowchart for Study Sample Selection
Customize your JAMA Network experience by selecting one or more topics from the list below.
Identify all potential conflicts of interest that might be relevant to your comment.
Conflicts of interest comprise financial interests, activities, and relationships within the past 3 years including but not limited to employment, affiliation, grants or funding, consultancies, honoraria or payment, speaker's bureaus, stock ownership or options, expert testimony, royalties, donation of medical equipment, or patents planned, pending, or issued.
Err on the side of full disclosure.
If you have no conflicts of interest, check "No potential conflicts of interest" in the box below. The information will be posted with your response.
Not all submitted comments are published. Please see our commenting policy for details.
Chen N, Brooks MM, Hernandez I. Latent Classes of Adherence to Oral Anticoagulation Therapy Among Patients With a New Diagnosis of Atrial Fibrillation. JAMA Netw Open. 2020;3(2):e1921357. doi:10.1001/jamanetworkopen.2019.21357
What are the longitudinal patterns of adherence to treatment with warfarin and direct oral anticoagulants among patients with atrial fibrillation who initiate anticoagulation therapy?
In this population-based cohort study of 16 969 Medicare beneficiaries with a new diagnosis of atrial fibrillation, only 4479 of 7491 patients using warfarin and 5043 of 9478 patients using direct oral anticoagulants belonged to the latent class characterized by continuous adherence. Adherence to a regimen of warfarin and direct oral anticoagulant use was associated with region of residence and HAS-BLED (hypertension, abnormal renal and liver function, stroke, bleeding, labile international normalized ratio, elderly, and drugs or alcohol) score, among other characteristics.
Among patients with atrial fibrillation who initiated anticoagulation therapy, more than 40% did not continuously adhere to therapy in the first year after diagnosis.
Less than half of US patients with a diagnosis of atrial fibrillation (AF) receive oral anticoagulation.
To identify patients with similar patterns of adherence to regimens of warfarin and direct oral anticoagulants (DOACs) in the first year after AF diagnosis and to evaluate associations between patient characteristics and membership in latent classes of adherence.
Design, Setting, and Participants
This retrospective cohort study used 2013 to 2016 Medicare claims data to identify 7491 patients with a new diagnosis of AF in 2014 to 2015 who initiated warfarin after AF diagnosis and 9478 patients with a new diagnosis of AF in 2014 to 2015 who initiated DOAC treatment after AF diagnosis, for a total of 16 969 Medicare beneficiaries. Participants were followed up for 12 months after AF diagnosis. Statistical analysis was performed from February 1 to November 30, 2018.
Treatment with warfarin or DOAC after AF diagnosis.
Main Outcomes and Measures
The main outcome was the proportion of days that patients received warfarin or DOAC, measured in 30-day intervals after AF diagnosis. Independent variables included patient demographic characteristics, socioeconomic status, region of residence, and clinical characteristics. Latent class mixed models were used to identify latent classes of warfarin and DOAC adherence, and polytomous logistic regression was used to assess the association between patient characteristics and membership in each latent class.
Among the 7491 patients receiving warfarin (4348 women), the mean (SD) age was 76.0 (10.0) years; among the 9478 patients receiving DOAC (5496 women), the mean (SD) age was 77.0 (8.5) years. Four latent classes of patients were identified based on warfarin adherence: late initiators (980 [13%]), early initiators who discontinued therapy at months 1 to 3 (1297 [17%]) or at months 5 to 10 (735 [10%]), and continuously adherent patients (4479 [60%]). Four latent classes of patients were also identified based on DOAC adherence: patients who initiated DOAC in months 1 to 5 (1368 [14%]) or months 6 to 11 (800 [8%]), patients with suboptimal and decreasing adherence (2267 [24%]), and continuously adherent patients (5043 [53%]). Membership in latent classes of warfarin adherence was significantly associated with sex, eligibility for Medicaid and income subsidy, region of residence, CHA2DS2-VASc (cardiac failure or dysfunction, hypertension, age 65-74 [1 point] or ≥75 years [2 points], diabetes, and stroke, transient ischemic attack or thromboembolism [2 points]–vascular disease, and sex category [female]) risk score, and HAS-BLED (hypertension, abnormal renal and liver function, stroke, bleeding, labile international normalized ratio, elderly, and drugs or alcohol) score. Membership in latent classes of DOAC adherence was significantly associated with race/ethnicity, region of residence, HAS-BLED score, and use of antiarrhythmic medications.
Conclusions and Relevance
This study found that, among patients who initiated anticoagulation therapy, 40% of those who initiated warfarin therapy and 47% of those who initiated DOAC treatment did not continuously adhere to therapy in the first year after AF diagnosis. Identifying longitudinal patterns of warfarin and DOAC adherence and the factors associated with them provides suggestions for the design of targeted strategies to mitigate suboptimal oral anticoagulation use.
Atrial fibrillation (AF) affects 2.3 million individuals in the United States and is the most common cardiac arrhythmia in older adults.1-4 Atrial fibrillation is associated with a 5-fold increase in the risk of stroke; this association between AF and stroke risk becomes more significant as age increases.5 Oral anticoagulant (OAC) therapy is associated with a reduced risk of stroke associated with AF by approximately 60%,6 but it is associated with an increased risk of bleeding.7 Clinical guidelines recommend the use of OAC therapy for patients with a moderate or high risk of stroke, which is generally defined as having a CHA2DS2-VASc (cardiac failure or dysfunction, hypertension, age 65-74 [1 point] or ≥75 years [2 points], diabetes, and stroke, transient ischemic attack or thromboembolism [2 points]–vascular disease, and sex category [female]) risk score of 2 or more.6 In spite of the important role of OAC therapy in stroke prevention, only 50% to 60% of US patients with AF recommended for OAC therapy actually receive these medications, and less than half of them adhere to them over time.8,9
Understanding the patterns of OAC use is important because prior research has shown that the underuse of OACs is not only caused by lack of initiation of OAC therapy but also by delayed initiation after diagnosis, therapy discontinuation, and suboptimal adherence while receiving therapy.9 Moreover, these patterns of OAC adherence have an association with stroke outcomes.10 However, it remains unknown whether patterns of adherence differ between patients who initiate OAC therapy with warfarin and patients who initiate OAC therapy with the new direct oral OACs (DOACs). Comparing patterns of adherence to a regimen of warfarin vs a regimen of DOACs is important because of the marked pharmacokinetic and pricing differences between them. It is possible, for instance, that patients who initiate use of DOACs are more likely to adhere to therapy because of the lack of monitoring requirement found in warfarin therapy. On the other hand, rates of discontinuation could be higher among patients taking DOACs because of higher costs and copayments.
To address this evidence gap, we applied latent class mixed models to a sample of patients with a new diagnosis of AF who initiated OAC therapy with warfarin or DOACs. We then assessed how patient characteristics were associated with membership in each latent class for each OAC type.
We used 2013 to 2016 claims data from a 5% random sample of Medicare beneficiaries obtained from the Centers for Medicare & Medicaid Services. The data sets included the Master Beneficiary Summary File; Part D event file; plan, prescriber, and pharmacy characteristics files; and Part A and B medical claims files. The data sets also contained information about demographic characteristics, zip code, Chronic Conditions Data Warehouse indicators and dates of first diagnosis of Centers for Medicare & Medicaid Services priority including AF, generic medication name, number of days of supply, and date of fill of pharmacy claims, and International Classification of Diseases, Ninth Revision and International Statistical Classification of Diseases and Related Health Problems, Tenth Revision diagnosis codes and dates of service for medical claims. This study was approved by the University of Pittsburgh Institutional Review Board as exempt. Patient consent waiver was approved by the University of Pittsburgh Institutional Review Board as data were deidentified. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.
We identified 72 306 patients with a new diagnosis of AF in 2014 to 2015, based on the Chronic Conditions Data Warehouse definition of AF as having 1 inpatient claim or 2 outpatient claims with an International Classification of Diseases, Ninth Revision diagnosis code of 427.31.11 We excluded patients with valvular disease (1453 patients) or with a CHA2DS2-VASc score less than 2 because they are not recommended for OAC treatment (3457 patients). Patients who died within 30 days of AF diagnosis (4394 patients) or who were not continuously enrolled in Medicare Part D Stand Alone Prescription Drug Plans during follow-up (28 848 patients) were also excluded. We excluded patients who did not fill any prescription for warfarin or DOAC in the year after first AF diagnosis (15 065 patients) or who had missing values for any covariate included in the study (603 patients). Finally, we excluded those who filled prescriptions for both warfarin and DOACs (1517 patients). The final study sample included 7491 participants who initiated a regimen of warfarin after AF diagnosis and 9478 participants who initiated DOAC therapy after AF diagnosis (eFigure in the Supplement).
We evaluated the association between time-invariant covariates and membership in latent classes of adherence. Covariates included patient demographic characteristics, variables that capture socioeconomic status, census division of residence, and clinical characteristics. Demographic characteristics included age, sex, and race/ethnicity. Socioeconomic variables included eligibility for Medicaid coverage and low-income subsidy (measured at the patient level), socioeconomic score (measured at the zip code level), and index of dissimilarity (measured at the Metropolitan Statistical Area level). We linked American Community Survey data obtained from the US Census Bureau and Medicare claims data using the zip code. We calculated the socioeconomic score using factor analysis. Specifically, factor analysis identified key census variables, which were then combined into a score representing socioeconomic status using z scores.12 The index of dissimilarity measured the fraction of black individuals (or white individuals) who would have to move from their neighborhoods to other neighborhoods to achieve perfect integration, and it was calculated using a previously defined formula.13,14 Clinical characteristics include CHA2DS2-VASc and HAS-BLED (hypertension, abnormal renal and liver function, stroke, bleeding, labile international normalized ratio, elderly, and drugs or alcohol) scores,15,16 Alzheimer disease and dementia, and use of nonsteroidal anti-inflammatory drugs and antiplatelet agents, as well as use of amiodarone, dronedarone, or verapamil. We used 12 months of claims data before AF diagnosis or Chronic Conditions Data Warehouse definitions to define clinical characteristics. The definitions of covariates have been previously published.9
The primary outcomes were adherence to warfarin use and adherence to therapy with DOACs, including dabigatran, rivaroxaban, apixaban, and edoxaban. For each participant, we extracted prescriptions for warfarin or DOACs after AF diagnosis. Using the dates the prescription was filled and the number of days of supply of medication, we calculated the proportion of days covered with warfarin or DOACs for each 30-day interval. We transformed the proportion of days covered by taking the arcsine of the square root to stabilize the variance of outcome measures.17
Statistical analysis was performed from February 1 to November 30, 2018. The latent class mixed model analysis was conducted in R software (R Project for Statistical Computing). Details of the model have been previously documented.18 We called the hlme estimation function using the lcmm R package to identify latent classes of warfarin and DOAC use over time. We included functions of time variable with up to fifth degree as the common fixed effects in the model, and an individual random effect on the intercept. We fitted a series of models from 2 through 6 latent classes. Models were estimated with the extended Marquardt algorithm to achieve maximum log likelihood.19 Initial values were not specified. We assessed the posterior membership probabilities of each model, which were computed using the Bayes theorem.20 In addition, we compared the Bayesian information criterion as well as the posterior classification rate across models to select the model with the best fit.
We reported patient characteristics at the time of first AF diagnosis across latent classes identified by the selected models. Specifically, class-specific characteristics of patients receiving warfarin and DOACs were reported as means and SDs for continuous variables and counts and percentages for categorical variables. We also compared the characteristics across classes using analysis of variance for continuous variables and χ2 tests for categorical variables and reported the P values. All P values were from 2-sided tests and results were deemed statistically significant at P < .05. Polytomous logistic regression was conducted to assess the association between baseline characteristics and membership in latent classes among those receiving warfarin and DOACs. The classes with the highest adherence to warfarin and DOAC therapy were treated as the reference group. Odds ratios (ORs) and 95% CIs comparing the other classes with the reference group were computed. We applied a Holm-modified Bonferroni correction to account for multiple comparisons.21
A total of 16 969 participants were included in our study. Among them, 7491 were warfarin users and 9478 were DOAC users, both with a mean follow-up of 11 months. The mean (SD) age was 76.0 (10.0) years for warfarin users and 77.0 (8.5) years for DOAC users. The proportion of male participants was 42% for both warfarin (3143) and DOAC users (3982); the proportion of female participants was 58% for both warfarin (4348) and DOAC users (5496). Most participants were white (87% of warfarin users  and 90% of DOAC users ).
The eTable in the Supplement shows the comparison of performance across the models with a different number of classes. For the modeling of both adherence to warfarin and adherence to DOACs, the Bayesian information criterion statistic decreased as the number of classes increased from 2 through 6, indicating a better fit of the model with a greater number of classes. Although the models with 5 and 6 classes had a lower Bayesian information criterion statistic, the high number of classes could represent overfitting and would decrease precision in the estimation of the association between patient characteristics and membership in trajectory classes. We thus selected the models with 4 classes for the modeling of adherence to both warfarin and DOACs.
Our final model identified 4 latent classes of patients using warfarin: late initiators (class 1; 980 [13%]), early initiators who discontinued therapy at months 1 to 3 (class 2; 1297 [17%]), early initiators who discontinued therapy at months 5 to 10 (class 3; 735 [10%]), and continuously adherent patients (class 4; 4479 [60%]) (Figure).
The mean (SD) age of warfarin users was approximately 76.0 (10.0) years across the 4 classes (Table 1). The proportion of male participants was slightly lower among the late initiators (357 of 980 [36%]) than among the other latent classes (class 2, 558 of 1297 [43%]; class 3, 314 of 735 [43%]; class 4, 1913 of 4479 [43%]). Patients eligible for Medicaid were underrepresented in the continuously adherent class (1229 of 4479 [27%]). Eligibility for low-income subsidy was lower among initiators who discontinued therapy at months 5 to 10 (265 of 735 [36%]) and among continuously adherent users (1763 of 4479 [39%]). The dissimilarity indexes were similar across classes. The continuously adherent users class had the lowest proportion of patients with a diagnosis of Alzheimer disease or other dementia (754 of 4479 [17%]) and of nonsteroidal anti-inflammatory drug or antiplatelet use (1048 of 4479 [23%]). The CHA2DS2-VASc and HAS-BLED scores were similar across classes.
Table 2 shows the OR for each covariate of belonging to a given latent class compared with belonging to the latent class characterized by continuous adherence (reference group). The reference class is often omitted in the text for the sake of simplicity.
Membership in latent classes of warfarin use was significantly associated with sex, eligibility for Medicaid coverage and low-income subsidy, region of residence, CHA2DS2-VASc score, and HAS-BLED score (Table 2). Male participants were less likely to be late initiators (OR, 0.75; 95% CI, 0.64-0.88). Patients who were eligible for Medicaid or for low-income subsidy were more likely to discontinue therapy in months 5 to 10 (Medicaid: OR, 1.94; 95% CI, 1.38-2.70). Compared with patients living in the Northeast, patients living in the Midwest were less likely to discontinue therapy in months 1 to 3 after diagnosis (OR, 0.76; 95% CI, 0.64-0.91). Higher CHA2DS2-VASc scores were associated with lower odds of discontinuing therapy in months 1 to 3 after diagnosis (OR, 0.92; 95% CI, 0.87-0.97). Higher HAS-BLED scores were associated with increased probability of late initiation (OR, 1.21; 95% CI, 1.10-1.33) and discontinuation in months 1 to 3 after diagnosis (OR, 1.10; 95% CI, 1.01-1.20).
Our final model also identified 4 latent classes of patients using DOACs: patients who initiated treatment with DOACs in months 6 to 11 (class 1; 800 [8%]), patients who initiated treatment with DOACs in months 1 to 5 (class 2; 1368 [14%]), patients with suboptimal and decreasing adherence (class 3; 2267 [24%]), and continuously adherent patients (class 4; 5043 [53%]) (Figure).
The mean (SD) age of DOAC users was approximately 77.0 (7.5) years across the 4 classes (Table 3). The proportion of male participants was slightly lower among patients who initiated treatment with DOACs in months 1 to 5 (528 of 1368 [39%]) and higher among patients with decreasing use (1036 of 2267 [46%]). Patients eligible for Medicaid (952 of 5043 [19%]) and low-income subsidy (1393 of 5043 [28%]) were less likely to belong to the continuously adherent latent class. The mean socioeconomic status score was the highest for continuously adherent users. The continuously adherent class had the lowest proportion of patients with a diagnosis of Alzheimer disease or other dementia (579 of 5043 [11%]) and of nonsteroidal anti-inflammatory drug or antiplatelet use (1283 of 5043 [25%]). The CHA2DS2-VASc score was similar across classes. The mean (SD) HAS-BLED score was slightly lower among continuously adherent users (2.8 [1.0]) than among other classes (3.0 [1.1]).
Table 4 shows the OR for each covariate of belonging to a given latent class compared with belonging to the continuously adherent latent class (reference group). The reference class is often omitted in the text for the sake of simplicity.
Membership in latent classes of DOAC use was significantly associated with race/ethnicity, region of residence, HAS-BLED score, and antiarrhythmic medication use. Male patients were more likely to belong to the latent class with suboptimal and decreasing adherence (OR, 1.14; 95% CI, 1.02-1.27) (Table 4). Compared with white patients, black patients were more likely to initiate DOAC therapy in months 1 to 5 after diagnosis (OR, 1.54; 95% CI, 1.23-1.92). Compared with patients living in the Northeast, patients living in the other regions (Midwest, Southeast, Southwest, and West) were more likely to belong to the class with suboptimal and decreasing adherence. Higher HAS-BLED scores and a diagnosis of Alzheimer disease or related dementia were associated with lower odds of belonging to the continuously adherent class.
We applied latent class mixed models to a nationally representative sample of patients with a new diagnosis of AF who initiated warfarin or DOAC use. We identified 4 latent classes of adherence to regimens of warfarin and DOACs, with 60% of warfarin users and 53% of DOAC users belonging to the class characterized by continuous adherence. Membership in latent classes of warfarin adherence was significantly associated with sex, eligibility for Medicaid and income subsidy, region of residence, CHA2DS2-VASc score, and HAS-BLED score. Membership in latent classes of DOAC therapy adherence was significantly associated with race/ethnicity, region of residence, HAS-BLED score, and use of antiarrhythmic medications.
Our findings are consistent with prior evidence suggesting that patterns of use and adherence to OACs are not only associated with patient clinical characteristics but also with demographic characteristics and region of residence.9 A previous study used group-based trajectory models to study OAC use among patients who received a diagnosis of AF, and it found that black race, HAS-BLED score, and residence in the South decreased the likelihood of initiation of OAC therapy.9 Factors that were associated with decreased odds of initiation of OAC therapy in the prior study were also associated with a higher probability of belonging to the classes with suboptimal adherence in the present study, which focused on patients who initiated OAC therapy.
Our study is an important contribution to the existing literature because we performed separate analyses for warfarin and DOACs. In doing so, we found that the proportion of continuously adherent patients was higher in the warfarin cohort than in the DOAC cohort (60% vs 53%). This difference was mostly owing to the higher proportion of patients initiating therapy late among the DOAC group. Although this finding could represent the use of medication samples provided in physician offices, which would not be captured in our study, lower adherence to DOACs could also represent financial barriers to the use of DOACs, which are protected by patents and are considerably more expensive than warfarin. Rates of therapy discontinuation were relatively similar between the 2 cohorts (27% for warfarin and 24% for DOACs), even when warfarin is considerably more inconvenient to use because it requires routine blood monitoring. However, patients discontinuing treatment with DOACs discontinued therapy sooner after AF diagnosis on average than those discontinuing warfarin therapy, which once again could represent barriers to DOAC access rather than the occurrence of clinical events such as bleeding associated with OAC use. Future analyses should leverage data sources with detailed clinical information to evaluate the reasons behind late initiation and early discontinuation of each type of OAC therapy.
Our study also has important implications from a methodological perspective. It demonstrates how latent class mixed models can be applied to describe patterns of medication adherence over time. We tracked adherence to treatment with warfarin and DOACs longitudinally, and through the measurement of the proportion of days covered with therapy every 30-day interval, we evaluated the use of warfarin and DOACs in 2 dimensions—adherence in each single unit and the longitudinal pattern composed of all the units. The extension of mixed models allowed for more variability across individuals through the inclusion of random effects. In addition, the model allowed for variations in the number of measurements across participants, so that participants with missing data for the outcome could be included and selection bias could be mitigated.
Our study is subject to several additional limitations. First, claims data contain information on the filling of prescriptions but not on the medications prescribed or on whether patients take the medications they fill. As a result, it is not possible to distinguish whether suboptimal adherence was a product of prescriber decision-making or of lack of patient compliance to the prescribed regimen. In addition, claims data do not contain prescriptions paid for with cash or information on the receipt of free samples, which could have led to an underestimation of adherence to taking warfarin and DOACs. Second, our methods did not allow for the estimation of how longitudinal patterns of adherence to warfarin and DOACs are altered by the occurrence of events during follow-up, such as stroke or bleeding events. Future analyses should assess longitudinal patterns of adherence to warfarin and DOACs by incorporating time-to-event measures for stroke and bleeding events. Third, our analyses did not evaluate patterns of switching across 2 types of OACs because we excluded patients who had a prescription for both warfarin and DOACs in the first year after AF diagnosis.
The application of the models also presented certain challenges. First, the choice of the number of classes is somewhat arbitrary. Models with a higher number of classes have a lower Bayesian information criterion statistic in this case; however, this may lead to overfitting the observed data. Second, because the analyses of clustering are based on the data provided, results might not be generalizable to other patient populations.22 Third, the global maximum likelihood was not guaranteed using the iteration algorithm from the model. In the future, multiple models with a different initial set of values could be created to evaluate the consistency of results. Fourth, our methods may not be sensitive enough to capture a single prescription or distinguish between patients with the same proportion of days covered but different use patterns in each time unit.
Among patients with a new diagnosis of AF who initiated anticoagulation therapy, 40% of those initiating treatment with warfarin and 47% of those initiating DOACs were not continuously adherent with therapy in the first year after AF diagnosis. Adherence to warfarin and DOACs was not only associated with clinical characteristics but also with socioeconomic status and region of residence. Identifying longitudinal patterns of warfarin and DOAC adherence and the factors associated with them could provide suggestions for the design of targeted strategies to mitigate suboptimal OAC use.
Accepted for Publication: December 17, 2019.
Published: February 19, 2020. doi:10.1001/jamanetworkopen.2019.21357
Open Access: This is an open access article distributed under the terms of the CC-BY License. © 2020 Chen N et al. JAMA Network Open.
Corresponding Author: Inmaculada Hernandez, PharmD, PhD, Department of Pharmacy and Therapeutics, School of Pharmacy, University of Pittsburgh, 3609 Forbes Ave, Rm 103, Pittsburgh, PA 15261 (email@example.com).
Author Contributions: Ms Chen and Dr Hernandez had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.
Concept and design: Chen, Hernandez.
Acquisition, analysis, or interpretation of data: All authors.
Drafting of the manuscript: Chen, Hernandez.
Critical revision of the manuscript for important intellectual content: All authors.
Statistical analysis: Chen, Hernandez.
Obtained funding: Chen, Hernandez.
Administrative, technical, or material support: Hernandez.
Conflict of Interest Disclosures: Dr Brooks reported that she is a member of the data and safety monitoring board for the Cerus Corporation. Dr Hernandez reported receiving consulting fees from Pfizer. No other disclosures were reported.
Funding/Support: The project described was supported by grant UL1TR001857 from the National Institutes of Health. Dr Hernandez is funded by grant K01HL142847 from the National Heart, Lung, and Blood Institute.
Role of the Funder/Sponsor: The funding sources had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.