Dispensing pharmacy chains and direct anticoagulants: Potential associations with patient outcomes

Pharmacy chains can differ with respect to the characteristics of their patient populations as well as their nonprescription products, services, and practices, and thus may serve as a surrogate for potential unmeasured confounding in observational studies of prescription drugs. This study evaluates whether a single‐source drug can have different patient outcomes based on the dispensing pharmacy chain.

• Dispensing pharmacy chains may have the potential to confound some studies, due to either the differences in their patient populations or in their services, practices, and available products.
• We evaluated whether dispensing pharmacy chains were associated with differences in clinical outcomes for Medicare beneficiaries prescribed the anticoagulants apixaban or rivaroxaban.
• There were no differences in clinical outcomes between pharmacy chains for apixaban; however, for rivaroxaban, there were differences in outcomes, including ischemic stroke.
• These results suggest that dispensing pharmacy chains may have the potential to act as a confounder of associations between drug exposure and outcome in some observational studies.

Plain language summary
There have been a variety of studies showing that pharmacist interventions can impact healthcare delivery.Studies have assessed the benefits of pharmacy-associated programs in areas that include immunizations, medication management, adherence and counseling.Although a study has directly compared chain and independent pharmacies for medication adherence, we were unable to find studies that directly compared pharmacy chains for drug-related patient outcomes.To evaluate this, we compared drug-related patient outcomes for two anticoagulant drugs across three large pharmacy chains.For the anticoagulant apixaban, the dispensing pharmacy chain did not make a difference; however, for the anticoagulant rivaroxaban, there were some differences in outcomes.
Of note, one pharmacy chain had a lower rate of stroke than another pharmacy chain.The methods used in this study are typically used to compare drugs in epidemiology studies.The results suggest there may be certain situations where dispensing pharmacy chains can impact the results of a welldesigned epidemiology study.This is the first study that directly explores the effects of pharmacy chains on drug-related outcomes.Although this is an exploratory study, it raises important questions about direct and indirect effects of pharmacy chains that should be addressed.

| PURPOSE
The pharmacy chain where a patient receives their prescription drug may act as a confounder of association between treatment and outcome, especially when the product is tightly correlated with pharmacy chain.This coupling of pharmacy chain and manufacturer/distributor, as well as differences in patient demographics by pharmacy chain, was demonstrated in a prior study on generic anticoagulants and central nervous system stimulants. 1 Pharmacy chain associations may confound studies in two distinct ways: first, there may be unmeasured differences in the patient populations in factors not captured in observational studies which influence outcome risk; second, pharmacy chains themselves may have differences in dispensing practices, services offered, and over-the-counter products available.In the first case, pharmacy chain may act as a surrogate for hidden population differences without actively influencing outcomes.In the second case, pharmacy chain may directly influence patient outcomes.Clinical pharmacy services have been shown to be beneficial in diabetes management, 2 warfarin anticoagulation 3 and direct-acting oral anticoagulation. 4Pharmacy type (chain vs. independent) has been associated with drug adherence. 5Dispensing practices can vary between pharmacy chains and have been criticized in the lay press. 6,7 evaluate whether pharmacy chains are associated with differences in patient outcomes, we selected two brand (single-source) direct oral anticoagulants, rivaroxaban and apixaban.For each product, we compared the risk of certain clinical outcomes between patient cohorts that received their prescriptions at different pharmacy chains, after controlling for typical study factors using propensity score weighting.These products were only available from one manufacturer during the study period, ensuring that no differences between manufacturers could influence study outcomes.S1 and S2).Each product was available from only one manufacturer during the study period; a generic apixaban was approved the week before closing of the study period, and so was not included.The National Plan and Provider Enumeration System (NPPES) was used to identify pharmacy chains from national provider IDs on Part D claims.

| Study populations
Separate populations by product were created for Medicare patients aged ≥65 years, who filled at least one prescription for rivaroxaban 20 mg daily or apixaban 10 mg daily (5 mg twice a day) in the study period.The evaluation period began on each drug's approval date (November 4, 2011 for rivaroxaban; December 28, 2012 for apixaban), and extended through December 31, 2019 for both drugs.Index date was defined as the date of the first qualifying study drug prescription for each patient.In the 183 days prior to their index date, all patients were required to have continuous FFS enrollment in Medicare Parts A, B, and D; a diagnosis of atrial fibrillation; no anticoagulant use; and no deep vein thrombosis, pulmonary embolism, valve repair or replacement, joint replacement, or mitral stenosis.Additional study inclusion criteria are listed in Appendix S1 (Expanded Methods).
Cohorts were assigned based on the dispensing pharmacy chain of the patient's index prescription.The three pharmacy chains with the largest number of claims for rivaroxaban and apixaban were selected for primary comparison (Walgreens, CVS, and Walmart).As this was an exploratory study, the pharmacy chains were assigned a letter (A, B or C) in the results.

| Episode construction and censoring criteria
For all primary analyses, a 3-day prescription coverage gap allowance was used to define continuous treatment periods for evaluation; this was chosen due to the short therapeutic half-life of apixaban and rivaroxaban and our desire to increase the likelihood that patients included in analyses were therapeutically anticoagulated.In addition to timing of refills, pharmacy chains may impact other patient behaviors and have different populations.To better evaluate these potential differences, we wanted to compare cohorts with a potential for continuous anticoagulation.The maximum of days of supply was used if multiple prescriptions for the same drug had the same dispensing date.
Follow-up for all patients started on the day after the index date and continued until the occurrence of a primary or secondary study endpoint, a gap in medication coverage (considered as treatment stopped), a switch in anticoagulant/dosage/dispensing pharmacy chain, admission to a Skilled Nursing Facility/Hospice/Nursing Home, disenrollment from Medicare Parts A/B/D or enrollment into Medicare Advantage (instead of FFS), a kidney transplant or dialysis, or the end of the study.

| Outcomes and study power
The primary study endpoints were ischemic stroke, intracranial hemorrhage, acute myocardial infarction, gastrointestinal (GI) bleeding, and all-cause mortality.Aside from all-cause mortality, only the hospital setting was considered for outcome events.Major GI bleeding was evaluated as a secondary study endpoint.Outcomes were defined using previously validated algorithms 8 based on ICD-9 codes, which were then mapped to ICD-10 equivalents and examined for consistency of usage across ICD-9 to ICD-10 eras (Appendix S1: Expanded Methods).

| Covariates and weighting
Patient demographics and medical comorbidities were captured at baseline and evaluated for balance across pharmacy chain cohorts.
Covariates with standardized absolute mean differences (SAMDs) less than 0.1 were considered balanced. 9In order to estimate the average treatment effect, within each population a model was run across all three pharmacy chains using generalized propensity scores to calculate inverse probability of treatment weights (IPTWs) for each patient; this allowed for the construction of pseudo-populations that were balanced across pharmacy chain cohorts for the included covariates. 10 In the first step of the primary analysis, the set of covariates included in the propensity score model was adapted from a prior study on anticoagulants; these included index year, medications, medical conditions, demographic and socioeconomic characteristics, and healthcare visits (all covariates are displayed in Tables S3   and S4). 11Variables that could impact drug metabolism were also considered.Both rivaroxaban and apixaban are dependent on P-glycoprotein for absorption and CYP3A4 for metabolism; however, concomitant use of drugs that alter P-glycoprotein and CYP3A4 activity was very low in patients at any pharmacy chain in this study.
The SAMDs between pharmacy chains for these drugs were less than 0.03 (Table S5), so these variables were not included in the model.
In addition to the patient demographics and medical comorbidities, a second step of the primary analysis included pharmacy location-related socioeconomic covariates, including the census tract median income range, census tract percent of population below poverty line, and area deprivation index (ADI) percentile categories, which is a measure of socioeconomic deprivation comprised of socioeconomic indicators originally sourced from Census data. 12

| Statistical analyses
In the first step of the primary analysis, a weighted Cox proportional hazards model was used to estimate time-to-outcome with pharmacy chain as the predictor.In the second step of the primary analysis, the covariates capturing pharmacy location characteristics were added to the weighted model.Robust standard errors were calculated to account for the use of weighting. 13The models simultaneously estimated the HRs for each pairwise comparison between pharmacy chains, and statistical significance was determined when 95% confidence intervals (CI) did not overlap with one.
Sensitivity analyses included a Cox regression model adjusting for covariates in the weighted cohorts, unadjusted and adjusted Cox regressions in the unweighted cohorts, and an increased 14-day gap allowance.Pharmacy chains were also compared using Kaplan-Meier curves weighted as per the primary analysis.Post-hoc analyses included multiplicity evaluations, pharmacy retention rate summaries, a competing risk analysis for death, a 30-day gap allowance, and an intention-to-treat analysis.1 and 2 display abbreviated summaries of pre-weighting covariates for users of each anticoagulant across these three pharmacy chains, and the full set of preweighting covariates are displayed in Tables S3 and S4.SAMDs for each pairwise comparison are shown for each covariate.

| Descriptive analyses
The most notable differences between pharmacy chains in the unweighted cohorts were in socioeconomic and geographic variables for both apixaban and rivaroxaban populations.Specifically, pharmacies with a metropolitan statistical area classification of rural were more common among individuals in pharmacy chain C compared to chains A and B, and the census tract median income range tended to be lower in pharmacy chain C relative to A and B. Other demographics between pharmacy chains were similar except for small imbalances in race and average age in the apixaban population.Medical covariates were generally well-balanced across pharmacy chains, with only a few small differences in physician office visits in the past 0-30 days (in both anticoagulant populations) and malignancy (in the rivaroxaban population).Additionally, outcome counts and corresponding incidence rates in the unweighted populations are detailed in Tables S6 and S7.After using IPTW, the pharmacy chain cohorts for both rivaroxaban and apixaban were balanced for all covariates (all SAMDs <0.033; Tables S8 and S9).Propensity score model coefficient differences are in Table S10.

| Primary analysis
The HRs for the primary analyses are shown in Tables 3 and 4 for pharmacy chains B and C versus A, and for pharmacy chain B versus C. For beneficiaries who received apixaban (Table 3), there were no significant pharmacy chain differences for any primary outcome evaluated.For beneficiaries who received rivaroxaban (Table 4), there was a statistically significant difference when comparing pharmacy chain C versus A cohorts for the primary outcome of ischemic stroke (HR 0.57, 95% CI 0.38-0.87)and a marginally significant difference for the primary outcome of GI bleeding (HR 0.83, 95% CI 0.69-1.00).The model for GI bleeding failed the proportional hazard assumption test.No statistically significant differences were observed for the other three outcomes.
The addition of pharmacy location-related socioeconomic covariates in step two of the primary analysis yielded similar results to step one (Table S11), although the rivaroxaban GI bleeding model no longer failed the proportional hazard assumption test.
For the secondary outcome of major GI bleeding, no significant pharmacy chain differences were seen with apixaban.For rivaroxaban, both pharmacy chains C (HR 0.76, 95% CI 0.59-0.98)and B (HR 0.76, 95% CI 0.62-0.94)were significantly different from pharmacy chain A. This finding was robust to the inclusion of pharmacy locationrelated socioeconomic variables (Table S11B).All the primary and following additional analyses are displayed in Figures S1 and S2.
The occurrence of outcomes over time are shown in weighted Kaplan-Meier curves in Figure S3.

| Sensitivity analyses
The significant difference in ischemic stroke risk between pharmacy chain C and A in the rivaroxaban population was robust across all sensitivity analyses using a 3-day gap allowance (Figure S2B and Table S12B).However, the differences seen in the primary and secondary bleeding outcomes were not significant in at least one sensitivity analysis with a 3-day gap allowance (Figure S2D,F and Table S12B).
The sensitivity analyses that allowed for an increased gap of 14 days between anticoagulant prescriptions impacted several comparisons in the rivaroxaban population (Table S12B).The hazard ratio for ischemic stroke for pharmacy chain C versus A was no longer statistically significant (HR 0.86, 95% CI 0.62-1.18).A time-stratified analysis (Table S13) only had an initial marginal difference.This pattern was also seen with pharmacy chain B versus A for the major GI bleeding outcome (HR 0.90, 95% CI 0.75-1.07).There were no statistically significant pharmacy chain differences for beneficiaries receiving apixaban (Table S12A).The distributions of censoring reasons for both the primary 3-day gap allowance and 14-day gap allowance analyses are displayed in Tables S14 and S15.

| Post-hoc analyses
We also evaluated a 30-day gap allowance and conducted an intention-to-treat analysis; in both approaches, point estimates generally either approached null compared to the primary analysis, or remained similarly null (Table S16).For major GI bleeding, comparing chains C and A, these analyses remained significant.
We evaluated pharmacy chain differences in retention over time (Table S17A).Pharmacy Chain C had a somewhat lower retention rate than pharmacy chain A or B. Population demographic and renewals patterns for the planned analyses were evaluated over time (Table S17B and Figure S4).An analysis accounting for the competing risk of death was also performed; results remained very similar to the primary analysis (Table S18).
Based on the large number of tests, adjusting for multiplicity using the study-wide Bonferroni correction led to no significant results.In a post-hoc evaluation treating each anticoagulant as a separate study with 18 tests (3 pair-wise comparisons of 6 outcomes), the Benjamini-Hochberg Method 14 with a false discovery rate of 25% led to 4 positive signals (Table S19).Thus, only 1 of 4 signals is likely to be a false discovery.All 4 tests that were positive signals in this T A B L E 1 Baseline characteristics (abbreviated) of beneficiaries on apixaban by the pharmacy chain (preweighting).particular post-hoc analysis were also positive in the primary analysis ( p < 0.05 without multiplicity adjustments).

| DISCUSSION
For beneficiaries receiving apixaban, the dispensing pharmacy chain had no statistically significant association with any of the studied outcomes in the primary analysis or sensitivity analyses.For beneficiaries receiving rivaroxaban, we found an association in the primary analysis between pharmacy chain and two of the primary study outcomes, with the pharmacy C cohort appearing to have reduced risk of ischemic stroke and GI bleeding (and the related secondary outcome of major GI bleeding) as compared to the pharmacy A cohort.Only the secondary outcome of major GI bleeding showed a statistically significant reduced risk for the B versus A comparison.However, these results were not robust across sensitivity analyses (although for ischemic stroke, the only discordant sensitivity analyses were when the gap allowance between prescriptions was increased).The other three primary study outcomes did not have statistically significant results for any of the pairwise pharmacy chain comparisons.While these few statistically significant results are not conclusive as to the impact of pharmacy chain on these types of analyses, additional research is needed to determine if there are specific characteristics of particular drugs, outcomes, or pharmacy chains that make them susceptible to pharmacy chain differences.
One surprising observation from this study is that risk of both GI bleeding and ischemic stroke appeared to be reduced for pharmacy chain C-dispensed rivaroxaban as compared to A, suggesting an apparently paradoxical lower risk in both over-and under-coagulation.
This finding has been observed in prior work on unstable warfarin levels. 15Although this observation was with warfarin, not direct anticoagulants, factors that increase the variability of anticoagulant levels could potentially lead to similar outcomes with other anticoagulants.
T A B L E 3 Weighted outcome counts and incidence rates by the pharmacy chain, and hazard ratios between pharmacy chains, for apixaban in the primary propensity score analysis.Improved patient education and ongoing management may decrease such variability.
An additional hypothesis for this observation is that there could be differences in control of drug-drug interactions that might impact the stability of anticoagulant levels.However, the very low concomitant usage of P-glycoprotein-or CYP3A4-acting drugs across all pharmacy chains make this unlikely.
Differences in patient adherence to recommended dosing intervals could also impact the stability of anticoagulant levels.Differences in patient adherence would have the greatest impact when the dosing regimen has little tolerance for deviation.The once-daily dosing of rivaroxaban has been questioned based on its relatively short half-life. 16Thus, the impact of errors in the timing of doses would be much greater for rivaroxaban than for a more frequently dosed anticoagulant, with a similar or longer half-life, such as apixaban. 17Although a recent meta-analysis of controlled clinical trials suggests there is no impact of once versus twice daily dosing of direct acting anticoagulants, 18 a less controlled real-world scenario may better reveal the impact of dosing differences.
In our sensitivity analyses with increased gap allowance, we observed no statistically significant association between the ischemic stroke outcome and pharmacy chain, while the association was significant in the primary analysis.This could be due to the larger allowed gaps in anticoagulation prescriptions permitting more untreated time in the study, which would tend to bias the results towards the null.
Alternatively, a longer gap allowance may increase the duration of follow-up and/or change the retention of beneficiaries in ways that could impact the differences between pharmacy chains.A timestratified analysis for the ischemic stroke outcome using the longer gap allowance did not provide strong evidence for a time-dependent effect.Although we noted some trends toward pharmacy chain differences in retention, renewal patterns, and demographics over time, additional study is needed to further evaluate this.
This was an exploratory study without a pre-specified plan for controlling the overall type I error rate.Other limitations included insufficient sample size to perform stratified sensitivity analyses with adequate power and a lack of access to complete information about pharmacy characteristics and practices.Additionally, we examined only two drugs, making it difficult to generalize to additional anticoagulants or other types of drugs.However, the few statistically significant findings that we did observe raise the possibility that pharmacy chain associations may occur only under certain circumstances, making them challenging to identify.In the case of future rivaroxaban generics, if one manufacturer was contracted exclusively to pharmacy chain A and another to C (as has been observed with other generics 1 ), future research comparing those two manufacturers could falsely suggest a manufacturing-based difference in product performance for bleeding or stroke outcomes.Further study is needed to understand when an imbalance in dispensing pharmacy chains could impact a study outcome, so that researchers can appropriately mitigate in their study designs.Additionally, the potential for direct effects of pharmacy chains in patient outcomes warrants further investigation, ideally with additional information on pharmacy chain practices and services.If there are best practices that could improve safe use and patient outcomes even with a small subset of drugs, these should be understood and encouraged.
Medicare claims for Fee-for-Service (FFS) enrollees with Part A (inpatient hospital), Part B (outpatient medical care), and Part D (prescription drug) coverage were evaluated.Products were identified using National Drug Codes (NDCs) in Part D claims (Tables There were 184 506 individuals who met the criteria for the apixaban population, of which 64 723 were in pharmacy chain A, 70 471 in pharmacy chain B, and 49 312 in pharmacy chain C.There were 111 058 individuals who met the criteria for the rivaroxaban population, of which 37 679 were in pharmacy chain A, 46 177 in pharmacy chain B, and 27 202 in pharmacy chain C. Tables Baseline characteristics (abbreviated) of beneficiaries on rivaroxaban by the pharmacy chain (preweighting).
Weighted outcome counts and incidence rates by the pharmacy chain, and hazard ratios between the pharmacy chains, for Rivaroxaban in the primary propensity score analysis.
a Hazard ratio.b Confidence interval.T A B L E 4 a Hazard ratio.b Confidence interval.