Healthcare Costs and Impact of Medication Adherence on Outcomes in Patients on Novel Anticoagulant Therapy

Introduction: Atrial fibrillation (AF) is a very common condition that causes cardiac rhythm disturbance and affects 2.7 to 6.1 million individuals in the United States (US). Warfarin, which is considered as a gold standard anticoagulant for the last 50 years to treat AF has limitations pertaining to the risk of bleeding, interaction with drugs and requires frequent monitoring. Novel Oral Anti-Coagulants (NOAC including dabigatran and rivaroxaban) are new promising drugs which have shown better or similar efficacy to lower stroke risk and fewer side effects compared to warfarin in the clinical trials. To compete with warfarin, NOACs may need to demonstrate substantial real-world evidence in regards to improving clinical outcomes and cost savings. Objective: The study was designed to evaluate the extent of undertreatment (adherence), and its predictors along with the impact of adherence on clinical outcomes, including ischemic stroke, bleeding, and Deep Vein Thrombosis and Pulmonary Embolism (DVTPE). Furthermore, the analysis helped to estimate the economic burden of NOACs vs. warfarin and identify specific subgroups with high-cost drivers and Healthcare Resource Utilization (HCRU) to achieve optimal benefits and devise strategies for cost-savings. The objective was achieved by conducting the following studies: Study 1 To examine patterns of medication adherence (measured by Proportion of Days Covered [PDC]) in patients with atrial fibrillation taking NOACs vs. warfarin for 6 or 12 months (post index). Furthermore, the study examined the short and long-term factors predicting adherence to the NOAC therapy after controlling potential confounders. Study 2 To examine the impact of adherence on the short and long-term risk of ischemic stroke, bleeding DVTPE and recurrent DVTPE in patients with AF taking NOACs during a one-year period (post index). The impact of adherence on outcomes was investigated by comparison of risk among propensity-matched adherence (adherent vs. non-adherent) cohorts. Study 3 – To describe and compare the economic burden (cost and HCRU) in patients using NOACs vs. warfarin therapy. Furthermore, the study aimed to identify specific subgroups and key drivers of high-costs and HCRU. The final aim of the study was to explore if there are any differences in cost/HCRU between adherent and non-adherent NOAC patients. Methods: The research utilized a retrospective cohort study design. Atrial fibrillation patients (ICD-9-CM codes 427.31/32), with ≥2 prescription fills for NOAC or warfarin, CHA2DS2VASC score ≥1, and 6-months pre-index continuous enrollment from the Optum® ClinformaticsTM Data Mart (Optum Insight, Eden Prairie, MN) (Jan 1, 2010 and Dec 31, 2012) were included. The index date was the first prescription claim for NOAC or warfarin. Adherence was calculated using Proportion of Days Covered (PDC) over a 1-year period. Predictors of adherence (PDC ≥ 80%) were examined using a logistic regression model controlling for the covariates including age, gender, stroke risk, comorbidities, insurance type, region, pre-index cardiac drug use (beta-blocker, Angiotensin II receptor blockers [ARB] or Angiotensin-converting enzyme [ACE] inhibitor, statin), etc. For the second study, adherent (PDC ≥ 80%) and non-adherent patients were matched on the above covariates using propensity score (Inverse Probability Treatment Weighting). The adjusted risk estimates were obtained at 6 and 12 months using a Cox proportional hazards model or generalized linear models (Poisson, negative Binomial) and compared across adherence based matched cohorts. In the final study, the economic value in terms of adjusted healthcare costs (inpatient, outpatient, and drug costs) and HCRU was estimated using a GLM model with gamma distribution and compared between patients taking NOACs vs. warfarin. Unadjusted costs were presented using descriptive analysis by subgroups based on demographic and clinical characteristics (age, gender, Charlson’s comorbidity index (CCI), insurance type, CHA2DS2VASC score, region, pre-index cardiac drug use, including beta-blocker, ARB-ACE inhibitor, statin use). Cost specific to bleeding events were calculated as an exploratory analysis. Similarly, the costs and HCRU were descriptively compared between the adherent and non-adherent patients taking NOACs. Results: A total of 5057 (N=1770 NOAC vs. N=3287 warfarin) patients with mean age of 66 years were included in the cohort based on the inclusion and exclusion criteria. For a 12-month follow-up, the proportion of adherent (PDC ≥80%) patients were higher among NOACs users (78.42%) compared to warfarin users (61.88%). Using multivariate logistic model controlling for the confounders; Age, CCI and statin use were major predictors of both short (6-month) and long-term (12-month) adherence to NOACs. The CHA2DS2VASC score was significantly associated with the short-term adherence while but not associated with the long-term adherence. For 12-month of adherence assessment, the three cohorts for bleeding, ischemic stroke, and DVTPE included 1617, 1651, 1739 patients (N=3440 for recurrent DVTPE at 6month assessment). For 12-month drug use, the incidence of bleeding, ischemic stroke, and DVTPE was 4.21%, 3.11%, and 1.11% respectively. Based on the multivariate analysis at 6 and 12 months of adherence assessment, the non-adherence was significantly associated with 1.72 and 1.94 times increase in the stroke risk respectively. Similarly, non-adherence was found to be significantly associated with elevated risk of recurrent DVTPE 3 and 6 months and DVTPE risk at 3, 6, 9 months. The risk of bleeding in non-adherent patients was slightly lower (HR= 0.84 – 6 months, HR= 0.94 – 12 months) but not significant compared to the risk of bleeding in adherent patients. High annual drug cost for NOAC users ($4988 vs. $331) was offset by higher medical (inpatient and outpatient) costs for warfarin users (Total annual cost for warfarin $31,400 vs. $22,134). The mean of annual ER visits (14 vs. 13) and office visits (76 vs. 49) was also higher for warfarin users compared to the patients taking NOACs. Overall, among warfarin users, female patients had higher HCRU, patients from the South had higher medical costs and office visits. Highest cost drivers for drug cost for warfarin users was patients from Northeast. Conversely, highest cost drivers for medical cost were patients less than <65 years and patients with CCI +3. For NOACs, the highest cost driver for the drugs was user who were 65 and above, from Northeast, CHA2DS2VASC >2 (mod-high risk), and independent insurance. Additionally, medical cost was driven by EPO insurance and CCI+3. Although medical costs and HCRU were lower for adherent vs. non-adherent patients taking NOACs, the differences were non-significant. Conclusion: Use of NOACs due to its better adherence compared to warfarin may help prevent inadequate anticoagulation and complications. Determining the factors influencing the adherence such as age, CCI, and stroke risk can help plan targeted approaches and interventions to improve adherence. Our results can help healthcare providers and managed care organizations to recognize the importance of adherence to NOAC medications among patients to prevent clinical risks including stroke, DVTPE and bleeding events. The study provides a valuable estimate of the economic burden in AF patients using NOACs and warfarin. These cost estimates can be further used as inputs in the studies involving cost-effectiveness analysis and indirect treatment comparisons. We found the higher drug costs for NOACs were offset by lower inpatient costs, outpatient costs, and HCRU; which can lead to overall monetary savings to the patient taking NOACs and to the healthcare system. Overall, the conducted research provides comprehensive evidence to help support NOACs as an optimal treatment choice for the AF patients.

Results: A total of 5057 (N=1770 NOAC vs. N=3287 warfarin) patients with a mean age of 66 years were included in the cohort based on the inclusion and exclusion criteria. For a 12-month follow-up, the proportion of adherent (PDC ≥80%) patients were higher among NOACs users (78.42%) compared to warfarin users (61.88%). Similarly, the patients using NOACs were consistently more adherent than warfarin users for 3, 6, 9 and 12 months of adherence assessment. The proportion of adherence among NOAC users was high at 3-months assessment (84.30%) and declined over time [6 months (82 and statin use were positively associated (p≤ 0.05) with an increase in medication adherence for 12 months among NOAC users. For short term NOAC use (6 months), patients with low-risk (based on the CHA 2 DS 2 VASC score of 1,2) were 27% less likely to adhere to the NOAC treatment (OR = 0.725 95% CI 0.580-0.907) compared to the high-risk patients (CHA 2 DS 2 VASC score of ≥3).

Conclusion:
Overall, patients taking NOACs have better (short and long-term) adherence to the therapy compared to warfarin users. Age, CCI and statin use were major predictors of both short and long-term adherence to NOACs. The CHA 2 DS 2 VASC score significantly associated with short-term adherence while statin use was specifically associated with long-term adherence to NOACs. The results obtained from the study will help clinicians and healthcare providers to understand the use of these (especially NOACs) drugs in a real-world setting and help implement therapy in practice to provide optimal benefits to the patients. Further research on subgroups and their treatment patterns (including switching) is warranted.

Atrial Fibrillation and Prevalence
Atrial Fibrillation (AF) is a common condition causing cardiac rhythm disturbance due to structural or electro-physical abnormality resulting in abnormal impulse formation. 1 In 2010, the prevalence of AF in the United States (US) was 2.7 to 6.1 million and is expected to grow between 5.6 and 12 million in 2050. 2,3 Approximately 70% of patients with AF are of the age between 65-85 years. 4 AF can be caused by ischemic heart disease, heart failure, hypertension while other causes of AF may include hyperthyroidism, acute infection, alcohol withdrawal, or post-surgery. The symptoms manifested in AF may include palpitation, dizziness, sweating, and shortness of breath. 5 AF is one of the key risk factors for ischemic stroke, increasing the risk up to 5-fold and accounts for one-third of all hospitalizations in the US for cardiac rhythm disturbances. 6

Treatment of Atrial Fibrillation
Treatment options for AF primarily include antiplatelet, anticoagulant, beta blockers, calcium channel blockers, sodium and potassium channel blockers. 7 Anticoagulants significantly decrease symptoms and health outcomes in AF leading to greater benefits. 8 Anticoagulants prevent blood clots and are used to treat existing blood clots.

Anticoagulants and their use in AF
Warfarin is an oral vitamin K-antagonist approved in 1954 and has been a gold standard of care for more than 50 years. It acts on multiple sites in the clotting cascade by preventing the synthesis of main coagulation factors, including II, IX, VII, and X by inhibiting vitamin K-dependent γ-carboxylation to work as an anticoagulant. 9 Warfarin as an anticoagulant has shown to reduce the risk of stroke, myocardial infarction, and death. In a randomized clinical trial (RCT) of 973 patients aged 75 years or over in the United Kingdom (UK), the risk of hemorrhage was significantly (p≤0.05) lower for warfarin (1.4%) compared to aspirin (1.6%) 10,11 .
The variable dosing, frequent dose adjustments and narrow window for therapeutic use in warfarin have prevented it's the widespread use in patients. Moreover, drug-interactions with concomitant medications, change in the diet, and the need for periodic monitoring has made warfarin use challenging for clinicians and patients. Due to these restrictions, many AF patients cannot use warfarin. 12 In some patients, warfarin cannot be administered due to other factors such as non-response, poor adherence, unwanted sideeffects, etc. The unmet need can be fulfilled using Novel Oral Anticoagulants (NOACs) which have recently proven to exhibit better efficacy, safety, and convenience compared to the existing warfarin treatment.

Novel Oral Anti-Coagulants (NOACs)
NOACs include two newly approved oral drugs, dabigatran "Pradaxa" (2010) and rivaroxaban "Xarelto" (2011). Dabigatran was the first oral anticoagulant approved in the US by the Food and Drug Administration (FDA) in 50 years. It acts as a thrombin inhibitor and is indicated for reducing the risk of stroke, systemic embolism, treatment of deep venous thrombosis (DVT) and pulmonary embolism (PE) in AF patients. In an RCT involving 18,113 patients with a primary outcome being the stroke (of any type), the risk of hemorrhagic strokes with dabigatran was also significantly (74%) lower than that of warfarin. 13 Dabigatran is given as a fixed dose of 110 or 150 mg twice daily, and requires negligible monitoring and has a peak effect in 1-2 hours as opposed to 4-5 days in warfarin. 14 Rivaroxaban was the first oral factor Xa inhibitor approved by the FDA. In ROCKET-AF trial (Rivaroxaban Once-Daily, Oral, Direct Factor Xa Inhibition Compared with Vitamin K Antagonism for Prevention of Stroke and Embolism Trial in Atrial Fibrillation) with 14,264 patients, rivaroxaban was non-inferior to warfarin in annual rates of stroke and systemic embolism (1.7% versus 2.2%) 15 . Similar to dabigatran, rivaroxaban has a rapid onset of action, requires twice daily dosing and has lower side effects compared to warfarin. Apaxiban and edoxaban, both factor Xa inhibitors were approved in December 2012, and January 2015 respectively. Based on the data availability for our study (2010 to 2012 only), apaxiban and edoxaban will not be included as a part of the hypothesized research plan.
Overall, NOACs have shown better or similar efficacy compared to warfarin in the clinical trials. Few benefits of NOACs include quick time-to-peak effects, fewer drug to drug and dietary interactions, fixed dosing regimens, and requires little monitoring. Now the reversal antidotes for NOACs (idarucizumab, andexanet alfa) are available.
Disadvantages of NOACs include an inability to administer to patients with a prosthetic heart valve or stage V chronic kidney disease. 12

Adherence and NOACs
Adherence has been defined as "An active, voluntary, and collaborative involvement of the patient in a mutually acceptable course of behavior to produce a therapeutic result". 16 and direct method (directly observed therapy, metabolite in blood, and biological marker in the blood) 17 .
Since their (NOACs) launch in 2010-2011, there has been limited literature published on "real-world adherence to the NOACs" using large observational studies. For oral anticoagulants based on the "Randomized Evaluation of Long-term Anticoagulant Therapy" (RELY) trial, the discontinuation rates at 1 year (using pill count) for dabigatran, and warfarin were 15.5%, and 10.2%, respectively. In the ROCKET AF trial, discontinuation rates were similar between rivaroxaban and warfarin groups (15.7% vs. 15.2%). Although it is essential to acknowledge that trial data has some limitations as the adherence reported in the trials is calculated in a controlled environment.

Study Rationale and Justification
A small number of studies have reported adherence using different methods (MPR, PDC, Persistence gap of 60 to 180 days) for NOACs. [18][19][20][21][22] Although the clinical guidelines recommend the use of NOACs for anticoagulation, the utilization of dabigatran and rivaroxaban remains sub-optimal in the real-world. Adherence to the medication therapy is crucial in improving the efficacy, reducing the costs and hospitalizations. Since the NOAC therapy is relatively new, it is not yet widely accepted and prescribed as warfarin by clinicians and healthcare providers.
Not many studies have examined the medication adherence of NOACs and its patterns over time. This observational study captured medication adherence (NOACs vs. warfarin) and its trend using Proportion of Days Covered (PDC) over a period of 12 months in a real-world setting. Furthermore, short and long term predictors of adherence to NOAC therapy were evaluated. The results obtained from the study will help clinicians and healthcare providers to understand the use of these drugs in a real-world setting and help make a suitable therapeutic choice in practice to provide optimal benefits to the patients.
Hypothesis: Ho = There is no statistical difference in estimate of medication adherence between AF patients taking NOACs and warfarin.

Study Design
This was a retrospective cohort study to compare adherence between NOAC and warfarin users and examine its patterns over a one-year period.
The study was conducted using commercial insurance de-identified claims data from January 1, 2010, to December 31, 2012, using a large-scale US managed care health plan affiliated to Optum® Clinformatics™ Data Mart (Optum Insight, Eden Prairie, MN) database. The primary outcome of the study was adherence calculated using PDC at 3, 6, 9 and 12 months and assessment of predictors of adherence at 6 and 12 months.
Data Source: The database mainly includes medical claims and pharmacy claims data. It contains details on dates of service, place of service, International Classification of Diseases, Ninth Revision, (ICD-9) diagnosis codes, provider type, National Drug Code (NDC), drug quantity dispensed, days supplied, charges, deductibles, and copayments.
The large US health plan database includes 14 million patients and 500,000 Medicare enrollees. The member file constitutes the demographic data and eligibility information.
The database also includes a drug file, a medical file (outpatient), standard pricing file (cost) and a confinement file for inpatient data. The database comprehensively covers diverse geographical areas of the US. All study data were accessed using Health Insurance Portability and Accountability Act (1996) compliant protocols. To ensure the patient confidentiality, no identifiable protected health information was used or analyzed during the study. 23 The data was accessed using the server at the University of Rhode Island (URI) and analyzed using SAS EG 7.1. The study was also approved by an Institutional review board at URI.
Sampling design/procedures: All patients between January 1, 2010, to December 31, 2012, were identified. Patients with warfarin, dabigatran or rivaroxaban (NOACs) were identified using the NDC codes and brand name using REDBOOK. The index date was defined as the date of first prescription fill of the NOAC or warfarin in their respective drug cohort. Patients with at least two claims for the study drugs in the post-index period were included. Patients were included based on at least one AF or atrial flutter diagnosis claim identified using the medical file (inpatient or outpatient) with an ICD-9 code of 427.31/427.32 during the pre-index period or 30 days within the index date. In addition to the AF patients, subjects with atrial flutter were also included since a large proportion of patients with atrial flutter also suffer from atrial fibrillation (overlap) and the recommended treatment is similar for AF and flutter for the prevention of stroke. Patients with age ≥ 18 years were included. Patients with concomitant use of warfarin and NOAC during the post-index assessment period were excluded. Few patients previously used warfarin in the pre-index period prior to starting the NOACs. Since, NOACs are also prescribed to fulfill the unmet need in few patients with prior warfarin use, to avoid exclusion of any NOAC user (and sample size considerations), the inclusion of these patients was based on the definition of "warfarin naïve." Based on the definition of 'warfarin naïve' in RELY trials, a patient was defined 'warfarin naïve' if there was no use of warfarin 2 months before the index date (first fill) of the NOAC or if the NOAC was used for a duration of at least 5 or more months. 24 This criterion was to ensure that we capture all NOAC users and avoid any potential bias in regards to prior warfarin therapy for assessment of outcomes. For sensitivity analysis, we also examined no use of warfarin 100 days before the index date as a threshold. Thus, the index date of these "warfarin naïve" users was based on the first prescription fill of NOACs. Data based on RELY trials showed no heterogeneity between patients who had prior warfarin use (based on the above definition) and those with no prior warfarin therapy. Age was used as a continuous variable. Furthermore, patients with CHA 2 DS 2 VASC score ≥1 (1-9) were identified using ICD-9 codes. CHA 2 DS 2 VASC characterizes the risk of stroke based on a score composed of (congestive heart failure, hypertension, age >75 years, diabetes, prior stroke, pulmonary or vascular disease, age [65-74 years], sex [as female]) (Please refer to Appendix Table 1 for ICD-9 codes and variable categorization). The patients with at least 6 months of pre and post index continuous eligibility with a permissible gap of 45 days were included in the cohort. Patients with hyperthyroidism (ICD-9 242.9) were excluded from the patient cohort since it may be the probable cause of AF but is not related to cardiac pathways.

Measurement of Adherence
The medication possession ratio (MPR) and PDC are the most frequently used measures to estimate adherence. The denominator in MPR is defined based on the difference between first and the last fill and doesn't account for discontinuation of the drug.
Furthermore, overestimation of MPR might occur due to early refill and if patients take concurrent medications of the same class. PDC is the most favored method to measure adherence in the recent years since it accounts for non-persistence where the denominator is days between first fill and end of the study. 25

Definition of Outcome
Adherence was calculated using PDC as follows =

Data Analysis:
Adherence was presented for NOACs users and warfarin users as PDC ranging from 0-100%. PDC above 100 was truncated to 100% (e.g., If a patient has a PDC of 110%, the patient was still considered as 100%. No patient was excluded based on the PDC truncation). 27 Categorical analysis was performed to present frequency and percentage of patients with adherence (PDC) ≥80% across the 2 groups (NOACs users vs. warfarin users). Adherence was assessed at 3, 6, 9, 12 months of use.
Secondly, the descriptive characteristics were analyzed and compared across adherent and non-adherent cohorts using NOAC therapy.
Pre-index cardiac medications were selected based on American Heart Association (AHA) medication classes recommended for AF therapy. Furthermore, these drug classes (beta-blockers, Angiotensin-II Receptor Blocker [ARB] or Angiotensin-converting enzyme [ACE] inhibitors, and statins) were as covariates in the pivotal dabigatran trial.
The type of insurance was also evaluated as possible confounders. First, the adherence (DV) was tested separately against each covariate using bivariate analysis. Univariate logistic regression was performed for the continuous variables and association with categorical variables was tested using a chi-square test. Possible confounders were identified, and their association with adherence was examined (with p-value ≤0.25). Due to the clinical significance, CHA 2 DS 2 VASC score was considered as a major confounder and was retained in the final model.
Next using the basic model, the multicollinearity between the variables of interest was tested to check whether the regressor variables are similar (in direct linearity) to other regressor variables. The collinearity was examined using the condition number (if >30 then collinearity exists) and the proportion of variance statistics for the eigenvalues. The multicollinearity was tested by examining the Variance Inflation Factor (VIF) for the variables. A VIF > 5 indicates collinearity among variables. Only one of the two or more collinear variables was selected for inclusion in the final multivariate model.
Moreover, the predictive model was primarily built using 2 methods 1. Automatic backward elimination process and 2. Manual Elimination process for sensitivity analysis.
Using the automatic backward elimination process, the model with the preliminary set of variables was refined by sequentially removing variables which do not contribute to the model. It was performed as an iterative process to predict adherence by examining Wald p-values, with confirmation through likelihood ratio testing (p-value > 0.05 confirming exclusion). This backward elimination process was repeated until a basic model was obtained. Similar to backward elimination, variables were manually evaluated by elimination of the non-significant variable based on Wald p-values. Once eliminated, the model was iteratively re-ran with the remaining variables and the process was continued until a parsimonious model with only significant covariates was obtained.
Furthermore, all two-way interactions between CHA 2 DS 2 VASC and other independent variables were also investigated for possible synergistic relationships. Significant interaction terms were retained. As a part of sensitivity testing, the model with independent variables + interaction term of CHA 2 DS 2 VASC score with each independent variable (one at a time) was examined. Comparison of models was also performed using Akaike Inclusion Criteria (AIC -lower is better). Moreover, the goodness of fit of the models was examined by Hosmer-Lemeshow test. 29 If p-value ≥ 0.05, the model fit is good. Model fit statistics were evaluated examining -2 log-likelihood estimate. Larger -2log-likelihood estimate indicated a poorly fitted model.
Results from the final parsimonious model were presented with the help of adjusted odds ratios along with their 95% CI, and p-values obtained from two-sided tests with a significance level of p-value ≤0.05.
For sensitivity analysis, predictors of short-term adherence (at 6-month assessment) to NOACs were also investigated.

RESULTS
We found a total of 25,120 users of NOACs and 149,359 users of warfarin within the study period. A total of 14,618 NOAC and 120,607 warfarin users had 2 or more prescription fills. Out of 14618 NOAC patients, 4332 patients had a prior warfarin use in the pre-index period, and the definition of 'warfarin naïve' was used to screen the patients. Based on the definition of 'warfarin naïve', 1107 out of 4332 patients were included as NOACs. The sensitivity analysis using a threshold of 100 days instead of 60 days led to a very minor change in the sample size and hence threshold of 60 days was retained and used. A total of 1032 patients were excluded due to an overlap (concomitant use of warfarin and NOACs) in the post-index period. Based on the other inclusion criteria (diagnosis of atrial fibrillation in pre-index period or 30 days within the index date, ≥18 years, continuous enrollment for 6 months pre and post-index, and CHA2DS2VASc ≥1), a total number of warfarin and NOAC patients were 8130 and 4758 respectively. Based on 6 months of drug use a total number of 3,453 NOAC users and 5596 warfarin users were included in the analysis. At 12 months of the assessment period (drug use), the study sample consisted of 5057 patients. A total of 1770 NOAC patients and 3287 warfarin patients met the inclusion and exclusion criteria and were included in the final analysis cohort.

Baseline Characteristics among NOAC and Warfarin patients
Based on the study sample of 5057 patients, the mean age of the sample was 66 years with more men (66.7%) than females. Most of the patients were either from the South or the Midwest (65%). Most of the patients (65%) were categorized as moderate to high-risk of stroke based on CHA 2 DS 2 VASC score >2. Over 80% of patients in the cohort had CCI above 0. For medication use, statins were the most frequently used drugs (50% of patients) followed by beta-blockers used by more than 25% of the sample.
Based on the chi-square test, most of the patient characteristics were different across the NOACs and warfarin. More than 50% of NOAC users were from the South (vs. 37% for warfarin users). For the stroke risk based on the CHA 2 DS 2 VASC score, a higher proportion of warfarin users had a moderate-high risk (more severe) compared to NOAC users (67% vs. 60%). Most of the patients had CCI score > 1(85%) where patients on warfarin therapy were slightly severe (with a higher proportion of 3+ comorbidities) compared to the NOAC users (48% vs. 34%). The use of statins was high in both cohorts (>50% of the patients). Please refer to

Baseline Characteristics among Adherent and Non-Adherent NOAC patients
For the cohort based on the drug usage for 12 months, the patient demographic and clinical characteristics were summarized for adherent vs. non-adherent NOAC users at baseline (index date). Age, CHA 2 DS 2 VASC score, type of insurance, and use of statins were significantly (p≤0.05) different across the adherent and non-adherent patients. The mean age of patients was 65 years with adherent (66 years) patients being older than the non-adherent (62 years) patients. The cohorts consisted of more men (69.3%) than women (30.7%). The majority of patients were from the South (51.3%) or the Midwest (21.9%), and more than 60% of the final cohort received point-of-service (POS) insurance.
There were 39.4% patients with a CHA 2 DS 2 VASC score of 1-2 (low risk) and 38.8% with a CHA 2 DS 2 VASC score > 2 (termed as a moderate-high risk of stroke). Patients with moderate-high risk of stroke (based on the CHA 2 DS 2 VASC score) were more adherent to medications compared to the low-risk patients. The CCI scores were well distributed across adherent and non-adherent patients. In regards to the cardiac medication use, statins and beta-blocker use were higher among adherent patients compared to the non-adherent patients. Table 1.3 in Tables and Figures I describes the patient demographic and clinical characteristics in detail among adherent and nonadherent NOAC patients.

Bivariate Analysis
The bivariate analysis between adherence and independent variables was initially performed to select the variables needed in the multivariable model. Variables with an association and a cut-off threshold of p ≤ 0.25 were included for further analysis. Age, insurance type, region, CHA 2 DS 2 VASC score, statin and beta-blocker use was associated with the adherence to the therapy. Table 1.4 in Tables and Figures I describes the significance estimates for the bivariate analysis.

Multivariate analysis
A multivariate logistic regression was modeled to examine the predictors of adherence for 12-month assessment in NOAC patients. The dependent variable was adherence (1,0), and age, CHA 2 DS 2 VASC score, Charlson's comorbidity index (CCI), region, insurance type, statin use, beta blocker use was selected as covariates based on the bivariate analysis. CHA 2 DS 2 VASC Score, Charlson's comorbidity index were included irrespective of association (in the bivariate analysis) due to their clinical significance. The parametric assessment was performed on the variables to assess the distributional characteristics. The CHA 2 DS 2 VASC score was categorized as low risk (1-2) and moderate-high risk (>2). The CCI was also categorized as (0, 1-2 and ≥3). Based on the preliminary backward elimination model, age, CCI, insurance type, statin use and monthly drug cost were significantly associated.  Although not statistically significant at 12 months, low-risk patients based on the CHA 2 DS 2 VASC score were less likely to be adherent to NOAC therapy compared to high-risk patients.

Sensitivity analysis at 6 months
Using the same model building procedure, predictors of adherence were examined for short-term NOAC use (6 months). At 6 months, age, CCI and statin use remained consistently significant as seen in multivariate analysis for 12 months. Additionally, it was found that CHA 2 DS 2 VASC score and region were significant predictors (p ≤ 0.05) of adherence to NOACs for short term use. The patients with low risk (based on the CHA 2 DS 2 VASC score of 1,2) were 27% less likely to adhere to the treatment (OR-0.725 95% CI 0.580-0.907). Please refer to Table 1.5 Tables and Figures and Appendix I Table   3 for the complete model.

DISCUSSION
Our analysis found higher adherence to NOAC therapy as compared to warfarin over a 1year period. This result was consistent over the short and long-term when examined at 3, 6, 9 and 12-month interval. The adherence decreased over time in both the cohorts (NOAC vs. Warfarin).
Unadjusted estimates suggested age, insurance type, region, CHA 2 DS 2 VASC score, statin, and beta-blocker use was associated with the adherence to the therapy. For multivariate analysis controlling for the covariates, an increase in age, fewer comorbidities, and statin use led to better adherence whereas low-risk CHA 2 DS 2 VASC led to lower adherence. It was interesting to know that CHA 2 DS 2 VASC score and region was significantly associated with short-term adherence to NOACs, but not long-term, whereas statin use was not influential predicting adherence based on both short and long term use.
Our study was the first to examine the patterns of short and long-term NOAC use and assess predictors of adherence among a large nationwide database in a real-world setting.
Many observational studies have shown an estimate of adherence to NOAC therapy is variable ranging from 40-88%. 30 In an observational study of Veteran Affairs (VA) cohort of 5,376 patients with AF initiated on dabigatran, 72% of patients had the proportion of days covered (PDC) ≥ 80%.
Moreover, the mean PDC reported was 84% ± 22%. 34 In another retrospective study evaluating electronic medical and pharmacy records within the University of California (UC) Davis Medical Center with 400 patients, the mean MPR for patients using dabigatran was 0.63. Overall, 43% of the patients taking dabigatran had an MPR of < 0.80. The study found gender (female), and PRN (prescription as necessary) medication use as predictors of low adherence among patients. 18 In another nationwide observational study in Denmark with 2960 patients, 1-year PDC for dabigatran was 84%. 19 The PDC for rivaroxaban and dabigatran using Healthcare claims from the Humana database between July 2013 and December 2014 was >65%. 33 Another study using IMS Health's LifeLink Health Plan Claims Database from 2010 to 2012 found mean MPR as 0.73 and 40% patients with adherence above ≥ 0.80. 30 In a study using US Department of Defense administrative claims data, the persistence rate (with a gap defined by ≥ 60-day discontinuation) for dabigatran versus warfarin was 72% versus 53% for 6 months and 63% versus 39% for 1 year respectively. 20 In a German study of 1204 patients, discontinuation rates of rivaroxaban in daily care for stroke prevention in atrial fibrillation (SPAF) patients were approximately 15% in the first year and very low thereafter. 22 Furthermore, persistence (no gap of => 60 days) was compared between rivaroxaban and warfarin in a US based study using MarketScan data (2010)(2011)(2012)(2013). Patients were more persistent to rivaroxaban (77%) as compared to warfarin (58%). 35 Our study found higher age, the risk of stroke (CHA 2 DS 2 VASC score >2), statin use, and lower CCI scores as major predictors of adherence consistent with the previous literature. 20,22,36,37 Younger age, male as a gender, low stroke-risk, poverty, higher education and poor cognitive function, have also been found to be associated with lower adherence. Another recent study based on the Danish patient registry reported an overall 1-year PDC equal to 83.9 % and found that females (OR-1.06), patients using cardiovascular drugs and CHA 2 DS 2 VASC ≥2 (OR-1.12) were major predictors of adherence among dabigatran users. 19 Numerous studies have reported discontinuation of NOACs is primarily due to bleedingrelated side effects. In a study on 467 Chinese patients in a clinic, dyspepsia was the most common cause of discontinuation, followed by adverse events and bleeding events including gastrointestinal bleeding and intracranial hemorrhage. Furthermore, dosing frequency, lack of laboratory monitoring, fear of side effects, and cost were other minor causes of discontinuation of therapy. 38 In addition to understanding the relative impact of anticoagulants-associated complications (such as bleeding), future research should emphasize on creating greater awareness and partnership between patient and physician for better decision making around anticoagulation. 39 The OPTUM database is a large nationwide database and provided sufficient sample size to interpret results robustly. However, claims data can lead to selection bias due to imprecise billing codes. Moreover, it should be acknowledged that OPTUM is mostly a commercial database under-represented by the elderly population (above 65 years). Over 65% the sample was represented by males, predominantly from the South or the Midwest and the database lacked information in regards to the race, ethnicity, and reason for discontinuation of therapy.
Although the clinical variables (e.g. INR values, ventricular ejection fraction, body mass index) were not included in the dataset, clinical determinants such as CHA 2 DS 2 VASC and CCI helped to control for disease severity by considering hypertension, prior cardiovascular disease, diabetes and other co-morbidities. The claims data lacks information regarding reasons for discontinuation or side-effects due to the drug which might help explain non-adherence estimates. Furthermore, adherence assessment based on 3, 6, 9, and 12-month windows might lead to truncation of the data. Therefore, the adherence assessment windows were kept close at every 3 months. It is also important to understand that the dosing of warfarin is variable and frequently adjusted. We also looked at the distribution of days of supply to explore a potential bias. The distribution of the days of supply for warfarin and NOACs was primarily around 30 and 60-day dosing which substantiated that the therapies might be comparable. Prior use of cardiac drugs was also accounted, and selection of the drugs was based on AF therapy recommended by American Heart Association (AHA). 40 These drugs were also used as covariates to understand the individual effects in the dabigatran pivotal trials. However, aspirin use was not comprehensively captured in the claims database due to its availability as over the counter (OTC) drug. The differences in the descriptive characteristics might be explained by the fact that NOACs might be prescribed to patients who have unmet need after warfarin therapy, this might lead to potential channeling or selection bias, in our study we did not control the selection bias using propensity scores.
Sensitivity analysis helped to confirm the results over a short and long-term period. To eliminate the bias in regards to variable follow-up time, adherence was assessed for patients who had medication use at regular intervals up to 3, 6, 9, and 12 months.

CONCLUSION
Overall adherence to NOACs is suboptimal and decreases over time. Patients taking NOACs have higher (short and long-term) adherence to the therapy compared to warfarin users. Age, CCI, and statin use were major predictors of both short and long-term adherence while CHA 2 DS 2 VASC score was associated with short-term adherence but not with long-term adherence. The short and long-term estimates of adherence to NOACs and associations observed in our study may help the healthcare providers and managed care organizations to strategize and provide optimal care to the patients by improving adherence to reduce clinical complications, healthcare resource use, and costs. Further   Based on the multivariate analysis at 6 and 12 months of adherence assessment, the nonadherence was significantly associated with 1.7 and 1.9 times increase in stroke risk respectively. Similarly, non-adherence was found to be significantly associated with elevated risk of recurrent DVTPE 3 and 6 months and DVTPE risk at 3, 6, 9 months.
The risk of bleeding in non-adherent patients was slightly lower (HR 0.84 -6 months, 0.94 -12 months) than risk in patients who are adherent to the NOACs.
Conclusion: Impact of adherence on the reduction of stroke and DVTPE risk is noteworthy. The risk of bleeding is not significantly different between adherent and nonadherent patients. Further studies on longer follow-up are warranted.

BACKGROUND
Atrial Fibrillation (AF) is a common condition causing cardiac rhythm disturbance due to electro-physical or structural abnormality resulting in abnormal impulse formation. 1 In 2010, the prevalence of AF in the United States (US) was 2.7 to 6.1 million and is expected to grow between 5.6 and 12 million in 2050. 2,3 Approximately 70% of patients with AF are between 65-85 years of age. 4 AF is one of the key risk factors for ischemic stroke, increasing the risk up to 5-fold. 5

Stroke Risk and AF
Stroke is the number four cause of mortality and a leading cause of long-term disability in the US. In the US, the stroke led to 1 in every 19 deaths in 2009, and a total treatment cost is estimated to be $38.6 billion per year. 6 A stroke is caused due to an abrupt disruption of blood supply to the brain, which may be caused due to the bursting of the blood vessels (hemorrhagic stroke) or blocking by a clot (ischemic stroke). 7 The risk factors for developing a stroke include, but not limited to, older age, cigarette smoking, obesity, cardiovascular disease and diabetes. One such heart-related condition, AF, an irregular heart rhythm, is a significant risk factor. 8 Recent estimates in the US has AF affecting an estimated 2.6 to 3 million Americans with more than 795,000 people having a recurrent or a new stroke each year. 9,10 The risk of stroke due to AF also increases with older age, rising from 1. Patients with AF and stroke tend to develop DVTPE. Hence, the use of anticoagulants as a preventative therapy for DVTPE and recurrent DVTPE is recommended. 13 Low molecular weight heparin, fondaparinux and VKA anticoagulants (warfarin) are also used as pharmacological treatments. 14 Lately, NOACs have shown promising results in a trial setting and are now approved for prevention of recurrent DVTPE.

Bleeding Risk
The use of anticoagulants is always associated with risk of bleeding-related complications. The difference in risk of bleeding was found to be non-significant between rivaroxaban and warfarin in the ROCKET-AF clinical trials (3.32 vs. 3.57). 15 Similar results with higher (but not significantly different) rates of bleeding with dabigatran were found compared to the warfarin. 16 Major bleeding (including intracranial hemorrhages, GI bleeding, etc.) and its related costs lead to an enormous burden on the healthcare system. Also, based on a recent assessment by FDA adverse event systems, the NOACs have been associated with a high number of bleeding-related adverse events but the information on risk factors and age is unavailable. A limited number of real world studies have presented the evidence of a non-significant difference between warfarin and NOACs in regards to the bleeding risk. 17 It is important to quantify the bleeding risk in NOACs using real world data and understand the impact of adherence on the risk of bleeding.

Risk Stratification
The CHADS2 score has been used to quantify the risk of stroke or VTE in AF patients.
It has been updated to CHA 2 DS 2 VASc (scored 0-9) which adds three more factors: vascular events or disease (V), age 65-74 years (A), and female gender (Sc). The risk stratification is necessary for the AF patients since underuse of anticoagulants is prevalent in the high-risk population. 20 Mostly clinician and patients tend to prefer to prescribe aspirin which is a safer and cheaper alternative to OAC. Therefore, it is essential to assess the risk, for selection of appropriate treatment to achieve maximum benefits.

Use of NOACs in stroke and DVTPE
Medication options in AF include antiplatelets, anticoagulants, beta blockers, calcium channel blockers, sodium and potassium channel blockers. 21 Anticoagulants significantly decrease symptoms and health outcomes and are responsible for improved overall health. 22 Warfarin is an oral vitamin K-antagonist approved in 1954 and has been a gold standard of care for more than 50 years. The variable dosing, frequent dose adjustments and narrow window for therapeutic use in warfarin therapy have prevented its the widespread use in patients. Moreover, interactions with concomitant medications, dietary restrictions, and the need for periodic monitoring has made warfarin use challenging for prescribers and patients. 14 NOACs include newly approved oral drugs: dabigatran "Pradaxa" (2010) and rivaroxaban "Xarelto" (2011), apaxiban and edoxaban. Overall, NOACs have proven to have better or similar efficacy in terms of reduction of stroke risks and non-inferior to warfarin for risk of bleeding complication. NOACs are also prescribed to prevent recurrent DVTPE. 23 The American Academy of Family Physicians and the American College of Physicians guidelines for VTE recommends 3 to 6 months of anticoagulant therapy following the first occurrence of a DVTPE.

Relationship of Adherence with Stroke and Bleeding Risk
Retrospective observational studies have shown evidence that adherence to cardiovascular medications can reduce the hazard of stroke. [24][25][26]

Study Rationale and Justification
Adherence is pivotal to the success of the therapy and is a crucial to ascertain the riskbenefit of a recommended treatment. Although several studies have reported adherence to NOACs ranging from 40-94%, there is inadequate data on the impact of adherence on the stroke, bleeding, DVTPE and recurrent DVTPE risks. This study examined an association between adherence to NOACs and risk of ischemic stroke, bleeding and DVTPE and recurrent DVTPE over a short and long-term period. This study will be the first to match the adherent and non-adherent NOAC users based on propensity score to compare the stroke, bleeding, DVTPE and recurrent DVTPE risks between the two cohorts.
Hypothesis: Ho = There is no statistical difference in risk of stroke/bleeding/recurrent DVTPE/DVTPE between equivalent cohorts based on adherence, matched using propensity score in AF patients taking NOACs.

Study Design
A retrospective cohort study design was utilized to calculate the adherence to NOACs using PDC and examine its impact on stroke, bleeding and DVTPE risk in NOAC patients.
The study was conducted using medical and pharmacy claims data from January 1, 2010, to December 31, 2012, using a large-scale US managed care health plan affiliated to Optum® Clinformatics™ Data Mart (Optum Insight, Eden Prairie, MN) Inc. database.
The primary outcome of the data was a stroke, bleeding, and recurrent DVTPE risk compared across propensity score-matched adherent and non-adherent cohorts.
Data source: The OPTUM database mainly includes medical claims, including inpatients and outpatient files and pharmacy claim data. It contains details on dates of service, place of service, International Classification of Diseases, Ninth Revision, Clinical Modification ICD-9-CM/ICD-10 diagnosis codes, provider type, National Drug Code-NDCs, drug quantity dispensed, days supplied, charges, deductibles, and copayments.
The large US health plan database includes 14 million patients and 500,000 Medicare enrollees. The member file constitutes the demographic data and eligibility information.
The database comprehensively covers diverse geographical areas of US with most of its enrollees from the South and the Midwest. All study data access was compliant with the Health Insurance Portability and Accountability Act.
To ensure the patient's confidentiality, no identifiable protected health information was used or analyzed during the study. 32 The data was accessed using the server at the University of Rhode Island (URI) and analyzed using SAS EG 7.1. The study was also approved by the Institutional review board at URI.
Sampling design/procedures: All patients between January 1, 2010, to  Since, NOACs are also prescribed to fulfill the unmet need in few patients with prior warfarin use, to avoid exclusion of any NOAC user (and sample size considerations), the inclusion of these patients was based on the definition of "warfarin-naïve." Based on the definition of 'warfarin naïve' in RELY trials, a patient was defined 'warfarin-naïve' if there was no use of warfarin 2 months before the index date (first fill) of NOACs or if the NOAC was used for a duration of at least 5 or more months. 33 This criterion was to ensure we capture all NOAC users and avoid any potential bias in regards to prior warfarin therapy for assessment of outcomes. Thus, the index date of these "warfarinnaïve" users was based on the first prescription fill of NOACs. For sensitivity analysis, we also examined no use of warfarin 100 days before the index date as a threshold. Data based on RELY trials has shown no heterogeneity between patients who have prior warfarin use (based on the above definition) and those with no prior warfarin therapy.
Age was used as a continuous variable. Furthermore, patients with CHA 2 DS 2 VASC score ≥1 (1-9) were identified using ICD-9 codes and included. CHA 2 DS 2 VASC characterizes the risk of stroke based on a score composed of (congestive heart failure,

Definition of Outcome
The index date was defined as the date of the first prescription of NOACs. For calculation of PDC (exposure assessment), the patients were followed from one month after the index date to the end of assessment period (6,9,12 months) and were censored at the occurrence of an outcome (stroke, recurrent DVTPE, DVTPE and bleeding), disenrollment from the healthcare plan, death, end of the study, whichever came first.
Subjects with an outcome, such as ischemic stroke, DVTPE, bleeding between one month after the index date to the end of adherence assessment period (See figure 1) were excluded. These patients were excluded to avoid assessment of exposure and outcomes in the same time frame and to set a sequence where exposure (adherence) precedes the outcome to establish a causal relationship. Adherence was counted after first 30 days of drug use since it overlapped with the window for diagnosis of atrial fibrillation.
Outcomes were calculated from the end date of adherence assessment to outcome date, end of the study period or end of enrollment/death whichever occurred first. The patients were eligible to enter the cohort only once and were counted only for the first incident stroke, bleeding or post-index DVTPE episode. For assessment of recurrent DVTPE, only patients who had a previous DVTPE episode in the pre-index period (prior to the start of NOAC) were included in the cohort. Cumulative incidence of the study outcomes was calculated, and relative risk (RR= Risk of outcome among exposed/Risk of outcome among unexposed) and its 95% CI was computed as a risk estimate.
The outcomes (ischemic stroke, major bleeding, DVTPE and recurrent DVTPE) were diagnosed based on ICD-9 codes (Please check the Appendix II Table 2.1 for a list of ICD-9) codes.

Measurement of Adherence
The medication possession ratio (MPR) and PDC are the most frequently used measures to estimate adherence. The denominator in MPR is defined based on the difference between first and the last fill and doesn't account for discontinuation of the drug.
Furthermore, overestimation of MPR might occur due to early refill and if patients take concurrent medications of the same class. PDC is a preferred method to measure adherence in the recent year since it accounts for non-persistence where the denominator is days between first fill and end of the study. 34 Adherence was calculated using PDC.
Adherent patients were defined based on a cutoff of PDC ≥ 80%.

Propensity Score (PS) Matching
Patient cohorts (adherent vs. non-adherent) were matched based on the propensity score to reduce any confounding due to study covariates and control selection bias. There are few ways to use the propensity score to compare outcomes: matching or stratifying patients on the PS, inverse probability of treatment, weighting using the PS, and covariate adjustment in subsequent multivariate regression models. 37 Matching is the most suitable method when the cases and control group are expected similar in size. 38 The patients were initially matched with 1:1 ratio using a specified caliper distance (e.g. equal to 0.02 of the standardized deviation of the logit of the propensity score). 39-42 Since the adherence was high (>70%) among the cohorts, to prevent loss of sample size; the patients were matched "with replacement" of controls. Matching was evaluated by plots and by comparing descriptive characteristics. In our data, due to large number of pseudocontrols, optimum matching was not achievable using the caliper method.
Inverse probability Treatment Weighting was then used where the patients in the treatment (adherent) group were assigned a weight of the direct inverse of their propensity score, and the control group (non-adherent) group patients were assigned the inverse of the propensity score subtracted from 1. 43 The propensity weights were also trimmed for extreme values greater than 10. The truncation of extreme values helped achieve better matching of the weights. 44 The matching cohorts were compared using the plots and tabular results of the matched variables. The IPTW weights thus obtained were used in the multivariate models (Cox and GLM Models) to predict the effect of the exposure (adherence) on the outcomes (stroke, bleeding, DVTPE).

Multivariate models
Using the IPTW matched data for each cohort, the incidence of ischemic stroke, bleeding, DVTPE and recurrent DVTPE and mean follow-up time were calculated. To quantify the association of adherence to the ischemic stroke, bleeding, DVTPE and recurrent DVTPE risk multivariate model were used. The occurrence of an outcome (stroke, bleeding, DVTPE) was the dependent variable in the model and adherence was the primary independent variable. The Cox proportional hazard model or GLM models are preferred to calculate the relative risk. 45 The relative risk was compared and graphically plotted using Kaplan Meier graphs. For GLM models, binomial models restrict the probabilities of an outcome to be greater than or equal to zero and thus leads to convergence issues. 46 Since the occurrence of stroke, DVTPE and bleeding are rare (with <10% of the sample); to avoid the convergence issues, negative binomial, and Poisson distributions were preferred for GLM models. 47 The models were checked for over-dispersion by checking the deviance and Pearson's chi-square and were scaled in the model as required.
To control for different follow-up time for outcomes for each patient, a semi-parametric Cox proportional hazard model was used by matching the cohorts using IPTW weights.
The data was right-censored. The Cox models were prior tested for proportionality of hazards by checking the Schoenfeld residuals, plots for log negative log of time and were evaluated for interaction with the adherence and time-varying covariate to check for the non-significant interaction term.

RESULTS
We To ascertain individual discontinuation dates (non-competing) for assessment based on each distinct occurrence of an outcome (stroke, bleeding, recurrent DVTPE and DVTPE) for each subject separately, 4 cohorts based on 4 outcomes were created. For cohort with the outcomes such as a major bleeding, the total number of patients at 6, 9 ,12

Overall Baseline Characteristics
The mean age of the overall cohort was 65 years with more men (>70%) than females.
The majority of the patients were either from the South or the Midwest (54%

Matched Cohorts
First, the propensity score matching was performed using a caliper matching of 0.2 and "without replacement" which led to very low sample size due to fewer controls (high adherence). Propensity score matching was tried using caliper matching of 0.2 and "with replacement" of controls. A higher proportion of adherence (>70%) led to multiple pseudo-matching controls with poorly matched cohorts. To remove the bias with minimal loss of patients, IPTW method was used to match the patients based on the propensity score. The covariates used in the propensity score model (Age, gender, insurance, region, CHAD2VASC score, CCI, cardiac drug use -Appendix Table 2) were compared between the matched adherent and non-adherent cohorts. Based on the non-

Bleeding Risk and Adherence
The overall incidence of bleeding in the cohort was 5.91% and 4.21% based on 6 and 12month adherence assessment. The proportion of bleeding after 6 months of adherence assessment was slightly higher in adherent patients compared to non-adherent patients (6.34 vs. 5.29 p=0.165), but the association was not statistically significant. Similarly, after long-term use (12 months), adherent patients were more likely (4.41 vs. 3.99 p=0.676) to have an occurrence of bleeding as compared to the adherent patients, but the association was non-significant. On an average, the total follow-up time for outcome evaluation was 8.8 months after 6 months of adherence assessment and 6.7 months after 12 months of adherence assessment. Please see Table 2.3 in Tables and Figures II for frequencies and relative risks based on IPTW matched data.
The association was further examined using multivariate Cox proportion hazards models (to control for the variable time) and GLM models adjusted by the IPTW weights.  Tables and Figures II). This trend was also consistent for short-term (6-month) use all three models (Cox and GLM-Please see Table 2.4/2.5 in Tables and Figures II). Overall, the adherence was modestly associated with elevated bleeding but was not significantly different between adherent and non-adherent patients. Hence, higher adherence in NOACs might not significantly increase the risk of bleeding.

Risk of Ischemic Stroke and Adherence
The incidence of ischemic stroke was 3.11% based on the adherence assessment period of 12 months. The proportion of ischemic stroke were lower among adherent patients at 6, 9, and 12 months ( Tables and Figures II).
Overall, adherence to NOACs is protective for an incidence of ischemic stroke over the short-term while providing less significant effect over a long-term drug use.

Risk of DVTPE and Adherence
At 12 months, the incidence of DVTPE was 1.11%. The follow-up time for outcomes post 6 and 9-month adherence assessment was 9.08 months and 6.82 months.

Risk of recurrent DVTPE and Adherence
Although the incidence of recurrent DVTPE was very low (0.87% at 3 months' drug use and 0.55% for 6 months' drug usage), adherence was significantly associated with the reduction of recurrent DVTPE in patients who already had a pre-index occurrence. This association was observed for both 3-month and 6-month adherence to NOACs ( preventive effect for recurrent DVTPE over a period of 3-6 months after the therapy, although the low incidence of outcomes in the study might be responsible for the inflated estimates.

Model Testing
Overall, the model fit examined with the help of deviance for GLM models using Poisson distribution was good with Pearson's chi-square <1 (no re-scaling using adjusting of Pearson's or deviance residual was required). Schoenfeld residuals to assess the proportionality of hazard were mostly parallel for hazards over time for adherence as a covariate. The models were also tested by adding an interaction term for adherence and follow-up time to evaluate the effect; non-significant estimates on the interaction term helped confirm the assumption of proportional hazards. The GLM models using negative binomial distribution mostly converged except models for recurrent DVTPE at 3 and 6 months of therapy.

DISCUSSION
Based on large real-world cohort, our results indicate that there is a significant association of adherence to NOAC therapy and a reduction in ischemic strokes and DVTPE at 6 and 12 months and recurrent DVTPE after 3 and 6 months of NOAC treatment. Furthermore, based on our results, higher adherence might not lead to significant increase in bleeding events.
The adherence to NOACs in our study was greater than 70%, which is consistent with the literature 48,49 . The overall follow-up to assess outcomes was more than 8.5 months and 6.5 months for a 6-month and 12-month adherence assessment (exposure) period respectively, for the NOAC drugs. This will be the first study to examine the impact of adherence on outcomes such as major bleeding, ischemic stroke, DVTPE and recurrent DVTPE using propensity score-matched cohorts. Since the NOACs have already shown better efficacy results against warfarin, evidence that adherence is more beneficial to reduce the outcomes may lead to even greater cost savings. Our study not only concurs with the previous findings in hypertensive drug classes that better adherence to the drug leads to reduced stroke risk and other cardiac-related events but also translates the evidence to NOACs (anticoagulants). 25,50 Overall, we found a higher incidence of ischemic stroke and bleeding in our real -world study as compared to the estimates found in the clinical trials. It was noteworthy that the number of bleeding events was not significantly different between adherence cohorts for the short and long-term use of NOACs. Medicare claims database (N= 64 935) was 1.73 per 100 person-years. 57 The slightly higher incidence of stroke in our study may be due to a higher proportion of patients with moderate-high risk of stroke (60%) in the overall population.
A study based on combined data from MarketScan and OPTUM reported the incidence rate of DVTPE as 0.58. 58 The incidence of recurrent DVTPE in our study ranged from 0.55-0.87 over 3 and 6 months of NOAC use.
We used large claims data since our hypothesis and study designed required to have a large sample. To address the selection bias and residual confounding, the adherence based cohorts were matched based on stroke risk (CHAD 2 SVAS 2 C) and other covariates including CCI, age, gender, insurance, region, prior cardiac drug use. The issue in regards to the multiple pseudo-controls due to less control as compared to cases was addressed using IPTW, which helped to achieve better matching. The study design helped to achieve a sequential order of exposure preceding the outcome, but this led to an exclusion of outcomes occurring between index and end of adherence assessment. Furthermore, adherence assessment based on 3, 6, 9, and 12-month windows might lead to truncation of the data. Therefore, the windows were kept close at every 3 months. ICD-9 codes were selected based on an in-depth literature survey where only those reported to have a value of ≥ 90 Positive Predictive Value (specificity and sensitivity) were included. This criterion helped us to have a better confidence in the selection of patients and avoid misclassification. Sensitivity analysis using the Poisson model helped to validate the results over a short and long term (6, 9, 12-months

CONCLUSION
Adherence to NOACs is suboptimal for anticoagulation control and decreases over time.
Overall, adherent NOAC patients had lower rates of ischemic stroke, DVTPE, and recurrent DVTPE compared to non-adherent patients. Short and long term risk of bleeding was not significantly different between adherent and non-adherent patients.
The short and long-term estimates of adherence to NOACs and its effect on the risk of bleeding, stroke, and DVTPE as observed in our study may help the healthcare providers and managed care organizations to compare risk-benefits of prescribing NOACs. The findings will further help to provide optimal care to patients by improving adherence and reduce HCRU and healthcare costs. Future research on adherence to NOACs including the newer drugs and longer follow-up times is warranted.

TABLES AND FIGURES II Figure 2.2: Cohort Selection Based on Inclusion and Exclusion Criteria
Patients with ≥2 fills NOAC = (N=14618) and Warfarin (N=120607)

Methods:
As an exploratory analysis, the HCRU and costs were compared across adherent (>80% PDC) and non-adherent NOAC patients.
Result: Annual drug cost was higher for NOAC users ($4988 vs. $331) was offset by higher medical costs for warfarin users (Total annual cost for warfarin $31,400 vs. $22,134). The mean annual ER visits (14 vs. 13) and office visits (76 vs. 49) for warfarin users was slightly higher compared to NOAC users.
Highest cost drivers for drug cost for warfarin users was patients from Northeast.
Conversely, highest cost drivers for medical cost were patients less than <65 years and patients with CCI +3.
For NOACs, the highest cost driver for the drugs was user who were 65 and above, from Northeast, CHA 2 DS 2 VAS C >2 (mod-high risk), and independent insurance. Additionally, medical cost was driven by EPO insurance and CCI+3.
Although ER visits (13.78 vs. 14.47), and inpatient costs ($20,756 vs. $23,208) were lower among the adherent patients, there was no significant difference in estimates between adherence and non-adherent patients.
Conclusion: Although drug cost was higher among NOAC users, the total cost was offset by higher cost for warfarin users in inpatient and outpatient setting. The costs presented by subgroups can help to target specific patient groups (higher CCI index, high stroke risk, patients from the Northeast, POS insurance) for greater cost savings. Adherence to NOACs is slightly helpful to reduce costs and HCRU but might not lead to substantial monetary benefits to the patient and the provider.

BACKGROUND
Atrial Fibrillation (AF) is a common condition causing cardiac rhythm disturbance due to electrophysical or structural abnormality resulting in abnormal impulse formation. 1 AF is one of the key risk factors for ischemic stroke, increasing the risk up to 5-fold 2 . In 2010, the prevalence of AF in the United States (US) was 2.7 to 6.1 million and is expected to grow between 5.6 and 12 million in 2050. 3 In the US, AF accounts for a total of >467,000 hospitalizations annually and leads to >99,000 deaths per year. AF is also responsible for adding an amount of $26 billion to the US healthcare spending annually, which is mostly driven by the inpatient and outpatient costs. 1

Use of NOACs in AF
Treatment options for AF primarily include antiplatelet, anticoagulant, beta blockers, calcium channel blockers, sodium and potassium channel blockers. 4 Anticoagulants significantly decrease symptoms and health outcomes leading to significant patient benefits. 5 Warfarin is an oral vitamin K-antagonist approved in 1954 and has been a gold standard of care for more than 50 years. The variable dosing, frequent dose adjustments and narrow window for therapeutic use in warfarin have prevented its the widespread utilization in patients. Moreover, interactions with concomitant medications, change in the diet, and the need for periodic monitoring has made warfarin use challenging for the clinicians and patients. 6 NOACs include newly approved oral drugs: dabigatran "Pradaxa" (2010), rivaroxaban "Xarelto" (2011), apaxiban and edoxaban. Overall, NOACs have shown better or similar efficacy compared to warfarin in the clinical trials. Few benefits of NOACs include: quick time-to-peak effects, fixed dosing regimens, require little monitoring, and have a fewer drug to drug interactions. Although an antidote is now available for dabigatran and its copays have been lowered since the launch of the drug, the total cost of the branded NOACs compared to the generic warfarin is very high. The challenge for the policy decision makers and drug reimbursement is to weigh if the high drug costs for NOAC's offsets the other related medical costs, outcomes and quality of life.

NOACs and Healthcare costs
NOACs have been widely prescribed and covered by the insurance providers and Medicare Part D although the copays may vary from $30-$120. 7 According to a claims database study by Desai et.al on 6893 patients, NOACs accounted for 62% of new prescriptions and 98% of anticoagulant-related drug costs in 2014. 8 In some cases, on the formulary, NOACs may require prior authorization (if less expensive drugs might work better), and be used as " Step therapy" to start with the drug after generic alternative.
Most of the providers have placed NOACs as Tier 2 (drugs are designated preferred brand because they have been proven to be effective, be safe, and favorably priced compared to other brand drugs) or 3 (have the highest copay or coinsurance, generally not found cost-effective). But it is important to consider that a significant share of the healthcare cost comes from inpatient and outpatient setting and is crucial to guide decision makers in regards to the right treatment selection.
There are some data published on the Healthcare Resource Utilization (HCRU) and economic outcomes of NOACs as a therapy. In the RELY trial, the hospitalizations for dabigatran were lower compared to warfarin (2311 vs. 2458, p <0.003). 9 A study comparing the HCRU between rivaroxaban and warfarin using a Humana claims database reported fewer hospitalizations in rivaroxaban users compared to warfarin (AFrelated, 2.11 vs. 3.02 days; all-cause, 2.71 vs. 3.87 days). 10 Furthermore, using the same database, the healthcare costs were comparable to warfarin with mean hospitalization costs for rivaroxaban slightly lower than warfarin (all-cause: $5411 vs. $7427), although pharmacy costs were slightly higher for rivaroxaban $5316 vs. $2620 but were not significantly different. 11 For a study based on the HealthCore data, pharmacy costs per month for dabigatran were higher than the warfarin cohort $455(SD,429) vs. $328(SD, 517) but medical costs were comparable $2,696 (SD, 6,699) vs. $2,893 (SD, 6,819). There was no difference in the adjusted total healthcare costs between the two cohorts (dabigatran vs. warfarin: $2949 vs. $2959) 12 . In an economic analysis by Deitelzweig et.al. 2012 using 10,000 Monte-Carlo iterations, it was demonstrated that 92.6% of the time, the one-year medical cost for dabigatran was less than warfarin. Similarly, 79.8% of the time, the one-year medical cost for rivaroxaban was less than warfarin. The study also examined one-year medical costs of major bleedings (excluding hemorrhagic stroke) with dabigatran and rivaroxaban (+$31 and +$108, respectively) compared with warfarin. 13 Based on the literature review of cost in AF patients, the total annual cost in 2013 ranged from $18,454 to $38,270 while inpatient cost was $7,841 to $22,582 per patient. 14 Another database study by Fonseca et.al, the total cost of patients taking dabigatran and warfarin after propensity score matching was 14,794 vs. $16,826. 15 Overall based on the recently published literature, the NOACs tend to demonstrate better or comparable economic outcomes than warfarin.
The real-world data regarding cost differences and events rate among NOAC treatment is limited. No published study has examined the differences in cost and HCRU outcomes in detail across the clinical subgroups based a real-world data. Also, comparison of the various component costs across adherence cohorts will help explore and generate a hypothesis in regards to the impact of adherence on clinical outcomes and its related HCRU.

Study Rationale and Justification
Although the cost estimates have been predicted by numerous simulated economic modeling studies and meta-analysis, there is limited real-world evidence to further explain the distribution of the costs and healthcare resource utilization for outcomes in NOACs. This study examined the cost along with its sub-components (inpatient, outpatient, and drug cost) and HCRU across (sub-grouped by) clinical factors like age, gender, CHA 2 DS 2 VASc, CCI. Furthermore, possible relationship of adherence to inpatient costs and HCRU was explored. This evidence is critical as the NOACs move towards competing against warfarin (generic) as an anticoagulant. Comparison between subgroups will help identify the cost drivers and recognize the difference in regards to costs in the real world to help generate hypothesis for in-depth analysis which can target specific subgroup and lead to higher cost savings. The results regarding cost will help understand the landscape and the economic burden of NOACs on the healthcare system.
Several inputs from the results can be used for detailed cost-effectiveness analysis and other economic modeling studies.
The proposed study aims to test the following hypothesis Hypothesis: Ho = There is no statistical difference in cost and HCRU between AF patients taking NOAC and warfarin.
Other analyses focusing in the subgroups and comparison of cost and HCRU between adherent vs. non-adherent patients are aimed to generate a hypothesis.

Study Design
This is a retrospective cohort study design to compare costs and HCRU between NOAC vs. warfarin patients across different subgroups.
The study was conducted using medical and pharmacy claims data from January 1, 2010, to December 31, 2012, using a large-scale US managed care health plan affiliated to Optum® Clinformatics™ Data Mart (Optum Insight, Eden Prairie, MN) database. The primary outcome of the data was adjusted inpatient, outpatient, drug cost and HCRU (inpatient, ER and outpatient visits).
Data source: The OPTUM database mainly includes medical claims, including inpatients and outpatient files and pharmacy claim data. It contains details on dates of service, place of service, International Classification of Diseases, Ninth Revision, Clinical Modification ICD-9-CM/ICD-10 diagnosis codes, provider type, National Drug Code-NDCs, drug quantity dispensed, days supplied, charges, deductibles, and copayments.
The large US health plan database includes 14 million patients and 500,000 Medicare enrollees. The member file constitutes the demographic data and eligibility information.
The database comprehensively covers diverse geographical areas of US with most of its enrollees from the South and the Midwest. The data access was compliant with the Health Insurance Portability and Accountability Act.
To ensure the patient's confidentiality, no identifiable protected health information was used or analyzed during the study 16 The data was accessed using the server at the University of Rhode Island (URI) and analyzed using SAS EG 7.1. The study was also approved by the Institutional Review Board at URI.
Sampling design/procedures: All patients between January 1, 2010, to  Since, NOACs are also prescribed to fulfill the unmet need in few patients with prior warfarin use, to avoid exclusion of any NOAC user (and sample size considerations), the inclusion of these patients was based on the definition of "warfarin-naïve." Based on the definition of 'warfarin naïve' in RELY trials, a patient was defined 'warfarin-naïve' if there was no use of warfarin 2 months before the index date (first fill) of NOACs or if the NOAC was used for a duration of at least 5 or more. 17 This criterion was to ensure we capture all NOAC users and avoid any potential bias in regards to prior warfarin therapy for assessment of outcomes. Thus, the index date of these "warfarin-naïve" users was based on the first prescription fill of NOACs. Data based on RELY trials has shown no heterogeneity between patients who have prior warfarin use (based on the above definition) and those with no prior warfarin therapy. Age was used as a continuous variable. Furthermore, patients with CHA 2 DS 2 VASC score ≥1 (1)(2)(3)(4)(5)(6)(7)(8)(9) were identified using ICD-9 codes and included. The CHA 2 DS 2 VASC characterizes the risk of stroke based on a score composed of (congestive heart failure, hypertension, age >75 years, diabetes, prior stroke, pulmonary or vascular disease, age [65-74 years], sex [as female]).
Please refer to Appendix Table 1 for ICD-9 codes. The patients with at least 6 months of pre and post index continuous eligibility with a permissible gap of 45 days were included in the cohort. Patients with hyperthyroidism (ICD-9 242.9) were excluded from the patient cohort since it may be the probable cause AF but is not related to cardiac pathways.

Follow-up per iod
Data Analysis: The data were analyzed using SAS EG version 7.1. The generalized linear models with gamma distribution were used to obtain and compare the adjusted annual and monthly per-patient costs between warfarin and NOAC users. Age, CCI, CHA 2 D 2 VASC, region, insurance type, gender, and cardiac medication use were used as covariates. The total cost along its sub-components and HCRU was compared between the NOAC vs. warfarin patients at a significance level of p ≤0.05.
The univariate statistics were presented for the subgroup analysis. The mean AF-related and bleeding related costs (all-cause, inpatient, medical, drug) and HCRU for NOAC vs.
warfarin users across subgroups were measured. As an exploratory analysis, unadjusted HCRU and costs were compared between patients who were adherent vs. non-adherent. Based on 6 months of drug use, a total number of 3,453 NOAC users and 5596 warfarin users were included in the analysis. At 12 months of the assessment period (drug use), the study sample consisted of 5057 patients. A total of 1770 NOAC patients and 3287 warfarin patients met the inclusion and exclusion criteria. Figure 1 in Appendix III describes the cohort sample selection in detail.

Baseline Characteristics among NOAC and Warfarin patients
Based on the study sample of 5057 patients, the mean age of the sample was 66 years with more men (66.7%) than females. The majority of the patients were either from the South or the Midwest (65%). Most of the patients (65%) were categorized as moderate to high-risk of stroke based on CHA 2 DS 2 VASC score >2. Over 80% of patients in the cohort had a CCI above 0. For medication use, statins were the most frequently used drugs (50% of patients) followed by ACE-ARB inhibitors used by more than 25% of the sample.
Based on the chi-square test, most of the patient characteristics were different across the NOACs and warfarin. The females preferred to use NOACs compared to patients using warfarin (69% vs. 64%). More than 50% of NOAC users were from the South (vs. 37% for warfarin users). For the stroke risk based on the CHA 2 DS 2 VASC score, a higher proportion of warfarin users had a moderate-high risk (more severe) compared to NOAC users (67% vs. 60%). Most of the patients had CCI score > 1(85%) where patients on warfarin therapy were slightly severe (with a higher proportion of 3+ comorbidities) compared to the NOAC users (48% vs. 34%). The statin use was high in both cohorts (>50% of the patients). Please refer to proportion of adherence among warfarin users at 3 months was 77.43% (N=5224/6747) followed by 72.61% and 61.88% for at 6 and 9 months respectively.

Baseline Characteristics among Adherent and Non-Adherent NOAC patients
Based on the 12-months drug usage, the patient demographic and clinical characteristics were summarized for 1388 adherent and 382 non-adherent patients at baseline (index date). Age, CHA 2 DS 2 VASC score, type of insurance, monthly drug costs and use of statins were different across the adherent and non-adherent patients. The mean age of patients was 65 years with adherent (66 years) patients being older than the nonadherent (62 years) patients. The cohorts consisted of more men (69.3%) than women (30.7%). The majority of patients were from the South (51.3%) or the Midwest (21.9%), and more than 60% of the final cohort received point-of-service (POS) insurance.
There were 39.4% patients with a CHA 2 DS 2 VASC score of 1-2 and 38.8% with a CHA 2 DS 2 VASC score > 2 (termed as moderate risk). Patients with moderate-high risk of stroke (based on the CHA 2 DS 2 VASC score) were more adherent compared to the low-risk patients. The CCI scores were well distributed across adherent and non-adherent patients. In regards to the cardiac medication use, statin and beta-blocker use was higher among adherent patients compared to the non-adherent patients. Table 3.3 in Appendix III describes the patient demographic and clinical characteristics in detail among adherent and non-adherent NOAC patients.

Inpatient, Outpatient and total) Costs [NOAC vs. Warfarin cohort]
The annual AF related all-cause healthcare costs (inpatient, medical, total) were calculated based on the GLM model using the gamma distribution since we expected a skewed distribution of the cost and the gamma model being acceptable approach handling the cost data. 18 As expected, the annual drug cost for warfarin users was significantly lower compared to the NOAC users (331 vs. 4988) since the NOACs are branded drugs and warfarin is generic on the market. However, the medical cost which included all inpatient and outpatient costs were significantly higher ($31,400 vs. $22,134) for patients on warfarin therapy as compared to the NOAC users. The individual inpatient cost for warfarin users was $25,405 compared $15,362 for patients taking NOACs.
The total healthcare cost consists of the sum of all drug cost (including copays and deductibles), inpatient cost, outpatient cost (including the professional fees, and cost of INR monitoring specifically for the warfarin users). The total annual adjusted AF-related cost for warfarin users was significantly higher than NOAC users ($32,157 vs. $26,803).
Overall, the high drug cost for NOAC users was offset by higher inpatient and outpatient costs (medical costs) for warfarin users.
A similar trend was observed for the monthly all-cause total cost (+ $388) and medical costs (+ $773) significantly greater for warfarin users compared to the NOAC users.
Please refer to Table 3.4 and 3.5 in Appendix III for detailed results.

Comparison of Healthcare Resource Utilization (NOAC vs. Warfarin cohort)
The adjusted estimates for annual ER visits for warfarin were similar compared to NOAC users (14 vs. 13, p=0.4607). Although, there was a significant difference in the number of annual office visits between warfarin and NOAC (76 vs. 49) users. Similar results with a higher ER (1.14 vs. 1.04) and office visits (6.37 vs. 4.11) was observed for the monthly estimates for warfarin vs. NOAC users. Please refer to Table 3.4 and 3.5 in Appendix III for detailed results.

Subgroup Analysis
The different components of the all-cause annual cost (including the drug, inpatients, medical costs) and HCRU were described across the subgroups. Age, gender, region, insurance type, CHA 2 D 2 VASC, and CCI were the major subgroups.

Drug Cost
The drug cost was highest in the Northeast region for both warfarin and NOACs. Annual drug costs for the older patients was slightly higher compared to younger patients on NOACs (5284 vs. 5064). Patients with an independent coverage for NOACs and POS for warfarin had highest drug costs compared to other types of insurance (including HMO, PPO, EPO, and Others). As expected, the drug cost for NOACs increased with the stroke risk. The patients on NOACs with moderate to high-risk CHA 2 D 2 VASC had higher drug cost compared to low-risk patients (5205 vs. 4941). A similar trend of increase in drug cost based on higher CCI was observed for the NOAC users. Please refer to Table 3

.6 in
Appendix III for the detailed results.

Medical Cost (Inpatient and Outpatient costs)
Annual overall medical costs for NOAC or warfarin users was higher in males, patients with EPO insurance, 3+ comorbidities based on CCI and patients from the South. The numbers might be due higher proportion of males and patients from the South in the entire population. The medical cost for NOAC patients increased with severity by CCI ($19820 for CCI=0, $26396 for CCI=1-2, and $41144 for CCI ≥3).
On the contrary, the patients in the low-risk CHA 2 D 2 VASC group had slightly higher medical cost compared to the patients in the higher risk groups (36024 vs. 29247). Please refer to Table 3.7 in Appendix III for the detailed results.

Healthcare Resource Utilization
Annual ER visits for patients using warfarin were highest in the Midwest (16.6. For patients on the NOAC therapy, highest ER visits were in patients from the Northeast (18).
The mean annual ER visits were higher in patients with age < 65 years' in both NOAC (13 vs. 7) and warfarin (16 vs. 10) Table   3.9 and 3.10 in Appendix III for detail on HCRU by subgroups.

Overall Results
Highest cost drivers for drug cost for warfarin users was patients from Northeast.
Conversely, highest cost drivers for medical cost were patients less than <65 years and patients with CCI +3.
For NOACs, the highest cost driver for the drugs was user who were 65 and above, from Northeast, CHA 2 DS 2 VAS C >2 (mod-high risk), and independent insurance. Additionally, medical cost was driven by EPO insurance and CCI+3.
Although ER visits (13.78 vs. 14.47), and inpatient costs ($20,756 vs. $23,208) were lower among the adherent patients, there was no significant difference in estimates between adherence and non-adherent patients.

Bleeding Related Costs
A total of 558 and 224 patients had bleeding related costs. Annual bleeding related drug cost was significantly higher for NOACs compared to warfarin ($6057 vs $3737). This trend was also consistent for monthly drug costs for NOACs vs Warfarin ($505 vs $311).
Please refer to Table 3.14 in Appendix III for detail on HCRU by subgroups. Similar to all cause AF related costs, the bleeding related medical cost for NOACs was lower ($1741 vs $6021) compared to warfarin.
The main cost drivers for bleeding related medical costs for warfarin were patients with Age <65 and patients with EPO insurance while the main drivers for NOACs users were higher CCI and EPO insurance. Please refer to Table 3.12 and 3.13 in Appendix III for detail on HCRU by subgroups.

Comparison of Annual HCRU and Costs by Adherence
As an exploratory analysis to generate a possible hypothesis, cost and HCRU were compared across adherent and non-adherent cohorts for NOAC users. The HCRU was compared between the adherent and non-adherent NOAC patients based on the 12months of drug use. Although ER visits (10.59 vs. 12.68), inpatient costs ($24760 vs. $30549) and all-cause total cost ($34854 vs. $37821) was lower among the adherent patients, there was no significant difference between the estimates for adherence and nonadherent patients. Conversely, adherent patients had non-significantly higher office visits compared to the non-adherent patients. Please refer to Table 3.11 in Appendix III.

DISCUSSION
Our study found the economic burden of atrial fibrillation on anticoagulant users based on total annual healthcare cost was substantial (>$25,000). Moreover, the total annual healthcare costs for warfarin users was higher compared to the patients taking NOACs.
The inpatient and outpatient costs for NOAC users were significantly lower compared to patients with warfarin therapy, which offset the higher drug costs for NOACs. Moreover, warfarin patients had higher ER and office visits (HCRU) compared to NOAC patients.
Approximately, at least 60% of the total cost was attributed to the inpatient setting.
Our study was the first to investigate each of the cost (Medical, inpatient) and HCRU (ER and office visits) components among NOAC vs. warfarin users across clinically important subgroups. The results provided valuable insights to identify specific patient groups with high cost and HCRU and can help to plan targeted approaches and interventions. Also, the real world cost estimates can be used as cost input in further budget impact and cost-effectiveness studies.
Our study was also the first to explore an impact of medication adherence on cost and HCRU in AF patients. Although our study did not find any association of adherence with a decrease in cost, it indicates that increase in adherence does not lead to any additional use of economic resources. Further in-depth analysis using matched cohorts is warranted.
Our estimates of the healthcare costs (drug, medical, total) were consistent with the existing literature. Based on United States Department of Defense (DOD) Military Health System data, drug costs were higher ($4369, p < 0.001) for dabigatran compared to warfarin which is similar to our estimated drug costs. 19 24 Furthermore, in a study based on the Medicare data, the total outpatient visits for a 12-month period were 53. 25 Our study found higher estimates which might be attributed to the inclusion of a higher proportion of patients with moderate to high-risk of stroke and CCI index of +3. Higher ER and office visits in warfarin might be due to the suboptimal utilization of INR testing lead to complications in regards to inadequate anticoagulation.
A study on warfarin users found no significant differences in costs and HCRU between adherent and non-adherent patients. 26 Based on our results, the above inference relates to the entire class of anticoagulants including the NOACs.
Our study was the first to compare the HCRU and cost between the NOAC and warfarin taking patients and to quantify the estimates across the demographic and clinical subgroups. The Northeast region had highest drug costs, which could be due to better healthcare and insurance availability (coverage) in the region. Annual overall medical costs for NOAC or warfarin patients was higher for patients with 3+ comorbidities based on CCI, and for patients from the South. Our study found higher cost was associated with lower age (<65) which is consistent with previous cost studies in AF. 22,27 It is important to note that younger patients might be aggressively treated than older patients for AF.
Our study identified higher burden of cost and resource utilization is incurred on specific subgroup of the patients taking NOACs e.g. total cost and ER visits were higher in males.
The drug cost, inpatient costs, and HCRU were highest for patients in the Northeast region. The medical cost and HCRU were highest for the patients with POS insurance.
Higher stroke risk and CCI score in NOAC users translated into higher costs and HCRU.
These results provide valuable insight on the utilization of healthcare resources in regards to targeting specific population groups.
One of the advantages of the study was the use of large commercial database represented by population across the United States. All the cost reported were adjusted to 2016 according to the latest consumer price index (CPI) data released by US government. It should be considered that even though generic warfarin is cheaper than warfarin, the cost of regular INR monitoring is an additional cost for the warfarin users. Our study found a total of 2464 patients (74%) who checked their INR and had an average cost of $120 per person per year. We also included INR monitoring costs for calculating the total healthcare costs. The healthcare costs, including the inpatient, medical, and total cost were adjusted using the gamma model; the crude confidence intervals were obtained by exponentiating the estimates.
The aim of the subgroup analysis and comparison of adherence based cohorts was targeted at hypothesis generation, and thus, any disease specific (GI related, stroke related) costs were not examined. All-cause costs were estimated to understand the landscape and economic burden in the AF-related population. The estimates for drug cost, HCRU, and medical costs can be applied to the economic models, while the estimates based on the comparison between adherent and non-adherent samples can be helpful for meta-analyses and indirect comparisons. Further studies can be planned in regards to event-specific estimates (related to stroke, bleeding, DVTPE) and costeffectiveness analysis.
It should be noted that the costs estimated in our subgroup analysis were not adjusted.
Although cost estimates (inpatient, medical, total cost) were adjusted using GLM models, there was no matching performed on NOAC vs. warfarin cohort, so there might be a possibility of selection bias due to unmeasured factors in regards to the sample populations. Use of a large administrative database may not be generalizable as it is mostly represented by commercially insured population. Furthermore, use of claims data may lead to reliance on diagnosis, coding and lack of clinical details. In our cohort with 12 months of follow-up assessment, most of the drug prescriptions were accounted for the years 2010(2010-Warfarin -1438, NOAC-178 and 2011warfarin -1829, with very few patients in 2012 due to criteria for a follow-up of 12 months. Hence, further analysis with longer follow-up time may help to capture details and specific patterns.
Accepting these limitations, our results can be used to help support reimbursement decisions and help generate hypothesis for more comprehensive studies. Inputs can be helpful in the further economic analysis and meta-analyses.

CONCLUSION
The study helps to estimate the substantial economic burden of AF in warfarin and NOAC users. The higher drug cost of NOACs was offset by the lower inpatient and outpatient costs and HCRU for the NOAC users as compared to warfarin users. There is no significant difference in costs and HCRU (except office visits) between adherent and non-adherent NOAC patients. The study provides valuable insight, identifying specific subgroups (e.g. Patients from the Northeast/South, less than 65, HMO/POS and a higher severity based on CHA 2 D 2 VASC and CCI) with a higher burden of cost and resource utilization in warfarin and NOAC users. These results could help the decision makers to balance the risk over benefit, and consider the cost associated with an optimal therapeutic choice of anticoagulants. Further research on the use of NOACs in regards to the specific outcomes and its cost-effectiveness is warranted.         398.91,402.01,402.11,402.91,404.03,404.11,404.13,404.91,404.93,and 428.x,518.4 Diabetes 250.x,357.2,362.0,and 366.41 Hypertension 401.x,402.x,403.x,404.x,and 405.x Stroke TIA 433,434,435,436 Vascular disease 410,411,412,413,414,443.8,443 Mercaldi, Catherine J., et al. "Long-term costs of ischemic stroke and major bleeding events among Medicare patients with nonvalvular atrial fibrillation."Cardiology research and practice 2012 (2012).