While it seems obvious that patients need to take their medications to benefit from them as shown in clinical trials, several studies and systematic reviews have analyzed the association between adherence to evidence-based pharmacotherapy and outcomes in chronic diseases such as coronary ischemic disease (McDermott et al., 1997; Rasmussen et al., 2007; Bitton et al., 2013; Choudhry et al., 2014), hyperlipidemia (De Vera et al., 2014; Karlsson et al., 2018), diabetes (Hood et al., 2009; Khunti et al., 2017), hypertension (Dragomir et al., 2010; Xu et al., 2017), osteoporosis (Nikitovic et al., 2010) and other conditions (DiMatteo et al., 2002; Simpson et al., 2006). Not surprisingly, virtually all the published evidence has shown that high adherence to appropriate treatment reduces mortality, several adverse clinical events, readmissions, healthcare utilization and costs. Oral anticoagulant (OAC) treatment in atrial fibrillation (AF) patients is not an exception and several studies (Deitelzweig et al., 2013; Shore et al., 2014; Thorne et al., 2014; Spivey et al., 2015; Alberts et al., 2016; Yao et al., 2016; Borne et al., 2017; Deshpande et al., 2018) have shown the valuable impact of high levels of adherence to OAC medication on various clinical outcomes, particularly the prevention of cardioembolic stroke.

Although there are several methods for assessing adherence (metabolite or biologic marker detection, self-reporting, pill count, electronic monitoring devices and other), retrospective studies based on data from refill databases have been favored in recent years because health organizations have extensively introduced health information systems with a unique patient identifier and, also, because this methodology has some advantages regarding the alternatives: it is unobtrusive (thus patient sensitization is avoided), objective (it produces quantifiable and reproducible data for each individual), and pragmatic (it is easy to use and analyze, inexpensive and applicable to different settings and drugs) (Rudd, 1979; Steiner and Prochazka, 1997; Schneeweiss and Avorn, 2005; Suissa and Garbe, 2007; Wettermark et al., 2013).

Despite the presence of an important methodological heterogeneity, studies linking adherence and health outcomes using refill databases usually adopt a common analytical approach: patients are categorized as adherent or non-adherent according to a threshold on medication days’ supply over a definite period, and the relationship between these two groups (patients under and over the threshold) and outcomes is evaluated. Generally, in the literature on adherence to OAC (Shore et al., 2014; Alberts et al., 2016; Yao et al., 2016; Borne et al., 2017; Deshpande et al., 2018) and to other chronic treatments, patients are defined as adherent when they exceed a cut-off point of 80% in the Proportion of Days Covered (PDC) or in other analogous adherence measures based on days’ supply, as the Medication Possession Ratio (MPR).

The 80% threshold (PDC80) originated in blood pressure studies carried out in the 70s of the last century (Sackett et al., 1975), but does not have an empirical basis for dichotomizing adherence to OAC treatment in AF patients, nor in many other conditions (Watanabe et al., 2013; Lo-Ciganic et al., 2015; Gellad et al., 2017; Morrison et al., 2017; Stauffer et al., 2017) [although some studies give some support to the 80% threshold in some conditions (Hansen et al., 2009; Karve et al., 2009)], introducing an element of arbitrariness with regard to the consideration of “sufficient” adherence and also to the interpretation of the relationship between adherence and outcomes.

Moreover, adherence measures based on days’ supply such as PDC or MPR can be calculated in several ways (Steiner and Prochazka, 1997; Andrade et al., 2006; Nikitovic et al., 2010; Vrijens et al., 2012; Arnet et al., 2016; Krueger et al., 2018). This in turn generates a huge variability in the real exposure to treatment of patients categorized as “adherent” or “non-adherent” using an apparently equal cut-off point value and, therefore, in the meaning of a given threshold or in the relationships between adherence (defined as PDC ≥ 80% independently of the method used for its calculation) and outcomes.

Differences in the design of adherence studies and in the operational definitions employed in these studies can importantly skew the adherence estimates based on days’ supply. In this way, the relationship between adherence thresholds and outcomes may be more sensitive to the way in which PDC (or MPR) has been calculated than to the actual patient exposure to the medication. These differences commonly arise from: (1) the exclusion (total or partial) from adherence studies of non-adherent patients (e.g., designs where two or more prescriptions filled or a minimum of covered days are required as an inclusion criteria) (Crivera et al., 2015), (2) the inclusion of patients more likely to be adherent, e.g., studies including experienced users (in the field of OAC, those can include long term users of an index drug, or switchers to a different OAC drug) (Ray, 2003; Danaei et al., 2012; Maciejewski et al., 2013; Lund et al., 2015; Li et al., 2018), (3) the definition of the index date (the date of first prescription, filled or not, or alternatively the date of the first prescription filled), (4) the censoring of cases at the date of the last refill, avoiding the use of a fixed time-window (e.g., 12 months) for follow-up and calculation of days’ supply based estimates (Kozma et al., 2013), (5) the censoring of cases (or the classification as non-adherent) when patients switch to a different treatment of the same class (Zhu et al., 2017), (6) the consideration (or not) of stockpiling (Greevy et al., 2011), (7) the consideration (or not) of immeasurable time (Dong et al., 2017; Palmaro et al., 2017), and (8) some other variants such as the estimation of the daily dose, the duration of the look-back period (Riis et al., 2015; Roberts et al., 2015; Nakasian et al., 2017), or the censoring (or not) of deceased patients or those who lose coverage.

Finally, in most adherence studies the information about doctors’ prescription is absent from refill databases. This is especially relevant because the concept of adherence refers to the “extent to which patients take medications as prescribed by their health care providers” (Sabaté, 2003; Osterberg and Blaschke, 2005; Cramer et al., 2008). In absence of a doctors’ prescription information, studies based on refill databases do not capture patients who do not fill any prescription (therefore, they may overestimate adherence as fully non-adherent patients are not accounted for) or may misclassify some patients as non-adherent when in fact the gap in days’ supply is caused by treatment discontinuation decided by their physicians (in this case, these studies may underestimate adherence due to the erroneous classification as non-adherent of patients not on treatment anymore).

The aims of this study, using a retrospective cohort of all patients with Non Valvular Atrial Fibrillation (NVAF) newly prescribed an OAC drug in the Valencia Region during nearly 5 years, are: (1) to estimate the variability in days’ supply adherence measures (PDC and PDC80) using a doctors’ prescription-based design and two common dispensation-based designs regarding real drug exposure, and (2) to determine the impact of adherence (defined as PDC ≥ 80%) on health outcomes according to each method of calculation of PDC.

Baseline patient characteristics of the three cohorts were similar (Table 2), not surprisingly as the three cohorts share most of their patients. Mean age was 74 years, slightly less than a half of the patients were women, CHA2DS2-VASC score was 3.7 and HAS-BLED score 2.8. Comorbidities were those expected in this population, with a high prevalence of hypertension (79%), diabetes (34%), coronary disease (17%), heart failure (16%) and previous ischemic stroke or transient ischemic attack (14%). Almost 1% of the patients had one hospital admission for stroke and another 1% for a bleeding episode during the year employed to calculate adherence.

Mean PDC ranged from 96.2% for the PD cohort to 88.5 and 84.7% for D2 and D1 cohorts, respectively (Table 3). The percentage of patients with PDC ≥ 80 ranged from 94.2% (PD cohort), to 79.6% (D2 cohort) and 74.7% (D1 cohort). Patients in the PD cohort had a mean of 306 days covered by medication (74% of the follow-up days; 316 for adherent patients and 149 for non-adherents), compared to 309 days covered in the D1 cohort (85% of the follow-up days; adherent patients: 352; non-adherent patients: 181) and to 323 in the D2 cohort (88% of the follow-up days; adherents: 352; non-adherents: 209). The 295 patients excluded from cohort D1 with respect to the PD cohort (patients who did not refill any prescription during the year after the first doctors’ prescription) had a mean of 41 covered days (most part of patients were fully incompliant but some patients restarted after a year from the first prescription), and the 4,706 patients excluded from cohort D2 (patients with less than two refills or less than 180 days between refills) had a mean of 89 covered days.

When observing specific drugs (Table 3), differences in mean PDC—although significant in some cases—were negligible in the PD cohort (from 96% for acenocoumarol to 94% for rivaroxaban), but widened in the dispensation-based cohorts (from 89% for apixaban to 83% for dabigatran in the D1 cohort and from 94% for apixaban to 88% for acenocoumarol in the D2 cohort). Regarding patients with PDC ≥ 80%, all drugs showed high adherence figures (93–94%) in the PD cohort. In the D1 cohort, these figures declined notably for acenocoumarol and dabigatran (73–74% of patients with PDC ≥ 80%) and to a lesser extent in the case of rivaroxaban (81%) and apixaban (84%). The D2 cohort showed even greater differences between drugs: from 90% of patients with PDC ≥ 80% when apixaban was the index drug (or 88% for rivaroxaban) to 82% for dabigatran or 78% for acenocoumarol. Therefore, while in the PD cohort acenocoumarol showed the highest number of patients with PDC equal to or greater than 80% (95%), in both D1 and D2 cohorts, rivaroxaban and especially apixaban showed the best adherence results, with acenocoumarol falling to 78% (D2 design) or 73% (D1 design).

A 20% of patients classified as adherent (PDC ≥ 80%) in the PD cohort had less than 80% days covered by medication, by virtually none in cohorts D1 and D2 whose classification of “adherence” is very close to the actual exposure to OACs (Table 4). D1 and D2 cohorts offer a very similar adherence classification, even at the expense of excluding a significant proportion of non-adherent patients in the case of the D2 cohort.

Crude events rate by 100 person-years during follow-up (1.9 years for mortality and 1.4 for stroke and bleeding) were about 6.7 for death, 0.9 for stroke and 1.4 for bleeding (Table 5). Unadjusted proportional hazard regression showed a significant protective effect of adherence on mortality (Hazard Ratio [HR] between 0.79 and 0.93 depending on the cohort) and also on stroke in the D1 and D2 cohorts (HR of 0.62 and 0.69, respectively) but not in the PD cohort, with no effect on admission for bleeding episodes. Multivariate proportional hazard regression models (Table 5 and Supplementary Tables S2–S4) showed similar results, with adherence associated with a significant reduced risk of death for the three cohorts (adjusted HR between 0.80 and 0.86). Stroke risk reduction was significant for D1 and D2 cohorts (adjusted HR between 0.61 and 0.64), and very close to statistical significance for the PD cohort (adjusted HR 0.66, 95CI: 0.43–1.02). Adherence had no effect on bleeding.

Our study highlights the importance of patient adherence to OAC medications. Irrespective of the approach used for measuring PDC, adherence to OAC in NVAF patients was associated with a substantial reduction in mortality and (except in the PD cohort) in the incidence of thromboembolic stroke, without a significant increase in admissions for hemorrhagic episodes. Significance of stroke risk reduction in the PD cohort was probably not achieved because a substantial number of patients with low drug exposure due that physician-initiated treatment discontinuation (20% of patients classified as adherent, with PDC ≥ 80%, had less than 80% of days of exposure) were classified as adherent.

The association between high levels of medication adherence and better outcomes is, in general, consistent with previous research (all with D1 or D2 designs) (Shore et al., 2014; Alberts et al., 2016; Yao et al., 2016; Borne et al., 2017; Deshpande et al., 2018) evaluating the association between adherence to OACs and health outcomes using the PDC80 threshold. This consistency, despite methodological differences (some of the referenced publications include experienced users, censor switchers or censor follow-up at the time of the last refill, use variable time-windows and other process differences), also strengthens the importance of adherence to approximate the real-life OAC benefits to those obtained in clinical trials.

To the best of our knowledge our study is the first using a prescription-based design to assess OAC adherence (and of the very few who have used this methodology in any other condition). This specific design evidences that patients (at least in OAC treatment in the Valencia region public healthcare setting) are highly adherent to medical recommendations, with 94% of patients above PDC80 when adherence is assessed against doctors’ prescription. Nevertheless, and due to doctor’s discontinuation, this high “adherence” translates into a mean of only 74% of patients with at least 80% of the follow-up days covered by medication. This finding suggests that physician-initiated treatment discontinuation is a major contributor to reduced OAC exposure. When using the dispensation-based D1 design “adherence” figures represent medication exposure more accurately (PDC80: 75%; days’ supply: 74%), but at the expense of misestimating “real adherence” to the medication prescribed, attributing to patients a supposed loss of adherence when it is actually treatment discontinuation by doctors. This finding is of especial relevance because most interventions aimed to improve medication adherence are focused on patient’s behavior. In the light of our results, and especially for essential treatments and in some contexts the appropriateness of medical discontinuation should be also considered when tackling the issue of real-world drug adherence. Moreover, dispensation-only designs to assess adherence classifies some adherent patients as non-adherent, and thus, targeting improvement interventions to these patients would be inefficient.

Adherence to medical recommendations was high and similar for the most used VKA in Spain (acenocoumarol, PDC80 = 94%) and any of the marketed NOAC (PDC80 between 92 and 94%), contrasting the worst adherence to NOAC (vs. VKA) speculated by some authors based on the lack of close monitoring (INR control) (Rodriguez et al., 2013; Hendriks et al., 2017; Burn and Pirmohamed, 2018). In the same way, adherence to VKA—always taking the medical prescription as reference—does not seem inferior to NOAC adherence, contrary to what is suggested by studies assessing the comparative adherence to VKA and NOAC (McHorney et al., 2017). However, doctors discontinue treatment to a greater extent in patients initiating with acenocoumarol (PDC80 in D1 Cohort: 73.4%) or dabigatran (PDC80 in D1 cohort: 74.1%) than rivaroxaban or apixaban (PDC80: 80.6 and 84.9% in the D1 cohort). This may partially explain why in the dispensation-only cohorts used by several adherence comparative studies the latter medications appear to have better adherence results than VKA or dabigatran (Crivera et al., 2015; Hanemaaijer et al., 2015; McHorney et al., 2015, 2017; Alberts et al., 2016; Beyer-Westendorf et al., 2016; Brown et al., 2016, 2017; Coleman et al., 2016; Yao et al., 2016; Borne et al., 2017; Manzoor et al., 2017; Maura et al., 2017; Mueller et al., 2017). Note that these differences are not due to switching to another OAC (an aspect not assessed in our study, in which, by design, the days’ supply of the new OAC after switching would continue to be considered in the index drug) but to discontinuation of treatment without restart during the follow-up period.

Finally, our study showed that adherent patients (defined as patients above PDC80) and non-adherent patients have different medication exposure depending on the method employed to calculate PDC (with mean days’ supply ranging from 10 to 12 months for adherent patients and from 5 to 7 months for non-adherent patients, depending on the cohort). Also, we showed that, in D1 and D2 cohorts, the design fails to include relevant groups of low-adherent patients as non-adherent patients or patients with only one refill, and also periods of non-adherence (as primary or early non-adherence) are ignored (Rodriguez-Bernal et al., 2018). Classification of patients as adherent or non-adherent based in the PDC80 threshold is sensitive to the method used for calculating PDC, and the assessment of the relationship between adherence and health outcomes compare different groups regarding drug exposure (e.g., in some designs outcomes are compared between patients with 10 vs. 4.6 months of average treatment, while in others designs outcomes are compared between patients with 12 vs. 7 months of average treatment). These results emphasize the need to standardize study designs and methods of calculation of days’ supply-based adherence measures with refill databases.

In this regard, the PQA method to estimate adherence requiring two prescriptions filled by at least 180 separate days excludes a relevant part of non-adherent patients artificially improving adherence with respect to the D1 design and without any additional advantage over this method. In fact, some studies linking doctor’s prescription to pharmacy dispensation show that up to one in three patients either never fill their initial prescription or discontinue the drug after a single fill (Fischer et al., 2010; Shrank et al., 2010; Raebel et al., 2012).

Studies assessing adherence with data from refill databases and using a 80% threshold on days’ supply measures tend to exclude low-adherent patients and periods of non-adherence and, also, misclassify periods of treatment discontinuation by doctors as patient non-adherence. Designs based on doctors’ prescription solve these problems, but require additional information about the actual exposure to treatment since these designs classify as adherent (to the medical prescription) patients with very low drug exposure whose doctors have discontinued treatment, a major reason for low OAC exposure.

The 80% threshold to classify patients as adherent or non-adherent can include very different groups of patients according to the methods used for the construction of days’ supply measures, highlighting the need to standardize (and transparently detail) the procedures used for the construction of these measures. In any case, high adherence and high exposure to evidence-based treatments, such as OAC in NVAF patients, seem to be associated with better outcomes, strengthening the need to develop strategies (focused on the patient, doctors and/or health organizations) to bring real-world outcomes closer to those of the clinical trials.

GS-G, AG-S, and SP conceived the study and wrote the first draft of the manuscript. GS-G and SP are guarantors and take full responsibility for the integrity of the data and the accuracy of the data analysis. IH-N, AG-S, CR-B, YS-A, SP, and GS-G participated in the study design and methodology. IH-N and SP performed the analysis. All authors participated in interpretation of data, and contributed to the critical revision of the manuscript for important intellectual content. All authors agree to be accountable for all aspects of the work and have read and approved the final manuscript.

The FISABIO foundation (a non-for-profit research institution depending on the Valencia Ministry of Health) had a Collaboration Agreement with Boehringer Ingelheim to conduct non-conditioned independent research in chronic health care, pharmacoepidemiology and medical practice variation (2013–2014), and received an unrestricted research grant from Daiichi-Sankyo (2017), both not related with the current study. GS-G participated in 2014 in an advisory meeting of Boehringer Ingelheim. None of the mentioned institutions/firms were involved in the design and conduct of the study; the collection, management, analysis, and interpretation of the data; or the preparation, review or approval of the manuscript, or in the decision to submit it for publication. The views presented here are those of the authors and not necessarily those of the FISABIO Foundation, the Valencia Ministry of Health or any other institution.