We conducted a population-based, propensity score-matched, retrospective cohort study of residents of Ontario, Canada aged 66 years or older who initiated combination antihypertensive therapy between April 1, 2004 and December 31, 2014. Individuals were followed for 5 years or until March 31, 2015. Provincial health insurance in Ontario covers physician and hospital services for all residents and prescription drugs for those over 65 years of age. This study was conducted using a prespecified analysis plan approved by the Sunnybrook Health Sciences Centre Research Ethics Board. This study is reported as per the RECORD guidelines (S1 Checklist).

We used the Ontario Drug Benefit claims database, which records prescription medications dispensed to all Ontarians over the age of 65, to determine exposure to combination antihypertensive therapy. Data pertaining to hospitalizations and emergency department use were obtained from the Canadian Institute for Health Information Discharge Abstract Database and National Ambulatory Care Reporting System. The Registered Persons Database was used to obtain basic demographic information and date of death for all Ontario residents. Demographic and specialty data for all physicians practicing in Ontario were obtained from the Institute for Clinical Evaluative Sciences (ICES) Physicians Database, and the Ontario Health Insurance Plan claims database was used to identify claims for all insured physician services. These databases have excellent data completeness and quality [17], and they were anonymously linked using encrypted person-level identifiers, as in previous studies [18–20].

We identified a cohort of new users of combination antihypertensive medication who were prescribed one angiotensin-converting enzyme inhibitor (ACEI) or angiotensin II-receptor blocker (ARB) plus one thiazide diuretic, either as FDC or as a multipill combination. These medication combinations were selected because they are common guideline-recommended options for initial antihypertensive therapy [4,5] and to avoid potential confounding when comparing combinations of various medication classes. We included only new users of antihypertensive medications to avoid selection bias based on prior medication adherence. New users were defined as receiving no prescription for any antihypertensive medication in the year prior to study enrollment.

The index date for study enrollment was defined as the date of first prescription of antihypertensive medications. To match FDC therapy, in which both medications are taken together, the multipill combination group included only individuals who were dispensed both medications on the same index date. Initiating multipill combination therapy on separate days might reflect treatment intensification in response to failure of monotherapy and select for a higher-risk population, therefore these individuals were not included.

We excluded individuals with any hospitalization for stroke, transient ischemic attack (TIA), heart failure, or myocardial infarction in the year prior to study enrollment to reduce selection bias that might arise from differential prescribing of FDC versus multipill combinations after hospitalization. We also excluded individuals with any emergency department visit for stroke or TIA in the year prior to study enrollment because the combination of ACEI and thiazide may be used for secondary stroke prevention even among nonhypertensive adults [21]. Finally, we excluded individuals who were prescribed any antihypertensive medications in addition to the initial combination therapy on the day of study enrollment. See cohort flow diagram for details (Fig 1).

To minimize selection bias, we used high-dimensional propensity score matching to identify comparable groups [22]. Each multipill combination user was matched with one FDC user based on the dose of index antihypertensive medications and the propensity score (allowing no more than a difference of 20% of the SD of the logit of the propensity score between matched pairs). The following datasets were used to develop the high-dimensional propensity score based on data from the year prior to study enrollment: prescription drug claims, diagnosis codes, and procedure codes from all hospitalizations and emergency department visits as well as insurance claims and diagnosis codes for physician services. In total, this represented 7 dimensions of data, and the 200 most prevalent codes from each dimension were retained as candidate covariates. All potential covariates were sorted in descending order by the magnitude of the log of the multiplicative bias term as previously described [22], and the top 500 covariates were included in the propensity score model. We also included the following prespecified covariates: age, sex, income quintile, year of the index date to account for possible changes in clinical practice over the duration of the study period, Charlson comorbidity index, the number of outpatient physician visits in the year prior to study enrollment, and any cardiology visit in the 3 months prior to study enrollment. The high-dimensional propensity score was calculated using version 2.4.4 of the freely available macro developed at Harvard University [23].

The dose of index antihypertensive medications was used as a proxy for severity of hypertension at baseline. As described in the supporting information, each medication was categorized into high- and low-dose categories based on the usual daily dose range described in clinical practice guidelines (S1 Table) [24]. Individuals were then categorized into 3 groups based on the dose of their index antihypertensives: “low” if both medications were low-dose, “high” if both medications were high-dose, and “intermediate” if only one medication was high-dose.

We conducted 2 complementary analyses to disentangle whether differences in outcomes between FDC and multipill regimens might be related to improved adherence or rather, differences in the actual effects of medications between groups. First, we performed an on-treatment analysis, censoring patients when they first discontinued treatment. Therefore, outcomes were assessed only during active treatment, which removed the effect of adherence. Second, we conducted an intention-to-treat analysis that followed individuals irrespective of disruptions in treatment to quantify the difference in drug adherence between groups and assess its impact on clinical outcomes.

We present 2 measures to describe adherence to antihypertensive medications based on medication dispensing: the time to the first instance of discontinuation and the proportion of total days covered. Discontinuation of antihypertensive medication was defined as no repeat prescription within 150% of the previous days’ supply. For example, a medication would be considered discontinued if the index prescription was dispensed for 30 days and a second prescription was not dispensed within 45 days. For the multipill combination group, repeat prescription of both medications within the 150% grace period was necessary to be considered continuous use. Medication class switching between ACEI or ARB or between different thiazides was not considered discontinuation. The on-treatment analysis followed individuals until the first instance of discontinuation. Because individuals may subsequently receive the index antihypertensive medications after an initial disruption in therapy, we also calculated the total duration of use of the index antihypertensive medications during follow-up. This was reported as the proportion of days of follow-up covered by the index antihypertensive medications in all periods of continuous use.

We hypothesized that, if the benefit of FDC therapy was related to improved medication adherence, the intention-to-treat analysis would demonstrate significant between-group differences in clinical outcomes whereas the on-treatment analysis would not.

The primary outcome was a composite of all-cause death and hospitalization for acute myocardial infarction (AMI), heart failure, or stroke. The administrative diagnostic codes (S6 Table) used to define the primary outcome have good sensitivity and positive predictive value, respectively: AMI, 89% and 87%; heart failure, 79% and 85%; and stroke, 76% and 97% [25,26]. Secondary outcomes included each individual component of the primary outcome, hospitalization for hypokalemia or hyponatremia to assess for safety, and the first instance of discontinuation of antihypertensive medication. Because we expected no association between antihypertensive treatment and cataract surgery, this was used as a “tracer outcome” to assess for residual confounding between the groups.

Propensity score matching resulted in balanced groups with no baseline covariates differing by more than 0.1 standardized difference except for index medication, which differed because ARBs were more common in FDC formulations than ACEI. Adjusting for covariates with less than 0.1 standardized difference imbalance provides negligible benefits in addressing residual confounding [27]. Thus, no additional covariates were included in the regression models. In the on-treatment analysis, individuals were censored when they first discontinued antihypertensive medications. The median time to the first instance of medication discontinuation was compared between the FDC and multipill combination groups using Kaplan-Meier estimators and the log-rank test. The proportion of days covered by antihypertensive medications in both groups was compared using the Wilcoxon signed rank sum test. Our primary analysis was a time-to-event analysis in the matched cohort using Cox proportional hazards regression including a robust variance estimator that accounted for clustering within matched sets [28]. When analyzing each component of the primary composite outcome separately, individuals were also censored for death. Proportional hazards assumptions were tested for the primary outcome using a time-dependent covariate by including an interaction term between the antihypertensive group and time in the Cox models. All of the above analyses were prespecified, with the exception of the calculation of the proportion of days covered by index antihypertensive medications, which was a post hoc analysis conducted to better describe medication use after we recognized that many participants continued to receive index medications after an instance of medication discontinuation.

The following additional analyses were provided in response to peer review comments. First, we conducted a sensitivity analysis by defining medication discontinuation as no repeat prescription within 300% of the previous days’ supply. Second, to better understand the role that concomitant cardiovascular risk-modifying treatments may have played in influencing our findings, we described the receipt of other cardiovascular medications in the last 90 days of follow-up in the FDC and multipill groups. A standardized difference less than 0.1 between groups was considered to be well-balanced. Finally, to better understand patterns of medication use, we reported antihypertensive and other cardiovascular medications dispensed in the last 90 days of follow-up among individuals who did and did not discontinue their index antihypertensive medications, defined as having any break in therapy of greater than 150% of the previous days’ supply.

All analyses were performed using SAS software (version 9.4).

After propensity score matching, we identified a cohort of 13,350 individuals (6,675 in each group) who were new users of combination antihypertensive therapy with an ACEI or ARB plus a thiazide diuretic. In the intention-to-treat analysis, median follow-up time was 1,826 days in both the FDC group (interquartile range 1,163–1,826) and the multipill group (IQR 1,142–1,826).

The 2 groups were well-balanced on all baseline characteristics except medication class at index (Table 1) and were relatively similar before matching (S3 Table). The median age at index was 71 years (IQR 68–77). In both groups, 42.7% of individuals received low-dose medication, 43.0% received intermediate-dose, and 14.3% received high-dose. FDC users were more likely to receive an ARB (65.1% versus 23.3% in multipill group) than an ACEI and more likely to receive hydrochlorothiazide (88.2% versus 82.9% in multipill group) than other thiazides.

The median time to the first instance of medication discontinuation was 191 days (IQR 45–741) in the FDC group and 150 days (IQR 45–446) in the multipill group (P < 0.01, Table 2). Medication discontinuation, defined as any break in therapy of greater than 150% of the previous days’ supply, occurred in 88.7% of individuals receiving multipill treatment and 83.1% in the FDC group (HR 0.80, 95% CI 0.77–0.83, P < 0.01). Individuals often resumed use of their index medications after a period of discontinuation (S4 Table). When examining use of the index antihypertensives over the entire study period, the proportion of days covered was 70% in the FDC group (IQR 19%–98%), which was significantly greater than 42% in the multipill group (IQR 11%–91%, P < 0.01). The proportion of days covered was similar in the sensitivity analysis using the less stringent definition of medication discontinuation (S5 Table).

Use of other cardiovascular risk-lowering medications was well-balanced between the FDC and multipill groups at baseline (Table 1) and in the last 90 days of follow-up (Table 3). Individuals in the FDC group were more likely to receive their index antihypertensive in the last 90 days of follow-up, but use of other antihypertensive medications was well-balanced (Table 3).

In the on-treatment analysis following individuals until the first instance of medication discontinuation, there were no significant differences in primary or secondary outcomes between groups (Table 4). The composite primary outcome occurred at a rate of 2.4 events per 100 person-years in both the FDC and multipill groups (HR 1.06, 95% CI 0.86–1.31, P = 0.60).

There were also no significant differences between the groups with respect to hospitalizations for hypokalemia or hyponatremia, which occurred in fewer than 0.5% of cases (Table 2). There was no significant difference in the tracer outcome of cataract surgery between the groups (HR 0.99, 95% CI 0.85–1.14, P = 0.83).

In a sensitivity analysis using the less stringent definition of medication discontinuation, the results were similar, with the primary outcome occurring at a rate of 2.2 events and 2.1 events per 100 person-years in the FDC group and the multipill group, respectively (HR 1.09, 95% CI 0.92–1.30, P = 0.34).

The composite primary outcome in the intention-to-treat analysis occurred at a significantly lower rate in the FDC group than the multipill group (3.4 versus 3.9 events per 100 person-years; HR 0.89, 95% CI 0.81–0.97, P < 0.01; Table 5, Fig 2). In our analysis of secondary endpoints, the hazard of death was significantly lower among individuals in the FDC compared to the multipill group—2.4 versus 2.8 deaths per 100 person-years, respectively (HR 0.85, 95% CI 0.77–0.94, P < 0.01). No significant differences were observed in the hazards for other components of the composite endpoint (Table 5).

There were no significant differences between the groups with respect to hospitalizations for hypokalemia or hyponatremia, which occurred rarely (Table 5). There was no significant difference in the tracer outcome of cataract surgery between the groups (HR 1.01, 95% CI 0.93–1.10, P = 0.78).