Data Sources
This study used five data sources from four countries and was a collaborative effort by partners from the University of Southern Denmark (Danish National Registers); the Centre for Pharmacoepidemiology, Karolinska Institutet (Swedish National Registers); RTI Health Solutions (Clinical Practice Research Datalink [CPRD]; UK); Optum (Optum Research Database [ORD]; USA) and Humana Healthcare Research (Humana Database; USA). The CPRD population was analysed as two groups, CPRD-linked (data from general practices that permitted hospital and mortality data linkage) and CPRD-unlinked (no linkage). The two US data sources were linked to the National Death Index.
Each research partner followed the same core protocol and statistical analysis plan, although methodological details varied at each site to account for differing data environments (see Electronic Supplementary Material [ESM]). Within each site, standard operating procedures and quality-control measures guided study conduct.
Study Design
This non-interventional, cohort study using secondary data included patients exposed to mirabegron or antimuscarinic medications from October 2012 through December 2018. Study periods within each source varied according to data availability. Exposure was based on prescription or dispensing data (hereafter prescriptions). The analysis incorporated a new user design and was conducted at the episode-of-use level as switching was expected to occur. Each mirabegron episode was matched to one antimuscarinic episode that started in the same year within each data source and according to year-specific propensity scores (PSs), with forced matching on age and sex. Incidence rates of major adverse cardiovascular events (MACE), acute myocardial infarction (AMI), stroke, CV mortality and all-cause mortality were estimated within person-time of current exposure of the matched mirabegron and antimuscarinic episodes. Hazard ratios (HRs) were obtained from Cox proportional hazard models that compared mirabegron with antimuscarinics.
Inclusion and Exclusion Criteria
Episode inclusion criteria were: a prescription for mirabegron or antimuscarinics (darifenacin, fesoterodine, oxybutynin, solifenacin, tolterodine, trospium) during the study period with no previous use of that specific medication in the prior 12 months (new use). Trospium was not available for use in Sweden during the study period. Additionally, patients were required to be ≥ 18 years old and have ≥ 12 months of prior enrolment in the data source. Prescriptions of two study medications on the same day and non-tablet antimuscarinics were excluded. Patients could contribute multiple episodes if each episode met the above criteria. Additional site-specific inclusion and exclusion criteria were implemented without impacting the overall study design (ESM).
Exposure Definition
Each person-day was classified as currently exposed to medication if the person-day was within the days of supply of the prescription plus a grace period of 50% [14]. The grace period accounted for variable adherence to medications where a patient still had some medication available beyond the days of supply due to forgetting some scheduled doses or changes in dosing schedule. Methods for calculating total days of supply varied according to the information available from each data source (ESM).
Current person-time for a treatment group was terminated upon a prescription of a medication from the other treatment group or following the end of the days of supply (after applying the grace period). When days of supply overlapped between prescriptions, the first prescription was truncated on the day before the subsequent prescription. Any patient could contribute current exposure person-time for mirabegron and antimuscarinics if they switched during the study period, but not simultaneously.
Study Outcomes
The main outcome was MACE, defined as the first occurrence of AMI, stroke or CV mortality. Additional outcomes included the individual events of AMI, stroke, CV mortality and all-cause mortality. AMI outcomes included both ST elevation myocardial infarction and non-ST elevation myocardial infarction [15] and stroke outcomes included haemorrhagic and ischaemic subtypes only [16, 17]. CV mortality was defined as death due to coronary heart disease or cerebrovascular disease, and all-cause mortality was defined as death due to any cause. Outcome identification and ascertainment methods varied across the data sources (ESM). It was assumed that any CV effects due to the OAB medications would present shortly after new use and continue while patients received medication, and that the effects would decline after discontinuation.
Follow-Up Period
Follow-up for outcome occurrence started on the day after the start of the episode and ended at the earliest of: occurrence of the outcome, end of the study period, last date with validated adverse CV or death outcomes, disenrolment from the data source (e.g. emigration, death), prescription of non-tablet antimuscarinics or a prescription for multiple OAB medications on the same day. Patients were eligible to contribute additional treatment episodes until the earliest of the end of the study period, last date of data with validated adverse CV or death outcomes or disenrolment from the data source.
Once an outcome occurred, patients were no longer followed for subsequent occurrences of the same outcome. Patients could experience multiple outcome types. However, the occurrence of CV or all-cause mortality censored the treatment episode and the study follow-up. For MACE, the treatment episode and the follow-up were censored at the date of the first targeted CV event.
Statistical Methods
Data were analysed separately for the five data sources and all coding was done independently by each partner. A meta-analysis was then conducted to combine the results.
Within each population, PSs were estimated for each eligible treatment episode by modelling the probability for treatment with mirabegron or antimuscarinics, conditional on the baseline covariates. A unique calendar year-specific PS was calculated for each eligible episode using relevant baseline variables from the 12 months prior to the start of each episode. In the USA, mirabegron was approved in June 2012, thus episodes from 2012 and 2013 were combined. The PS model was built using pre-specified variables from all databases (e.g. age group, prior OAB medication use), database-specific variables (e.g. length of health plan enrolment) and database-specific empirically defined variables (e.g. most frequent International Classification of Diseases diagnosis codes). Several checks were conducted to support the variable selection process, including correlations among pre-defined covariates and empirically defined variables and review of the PS distributions.
The mirabegron episodes were matched to the antimuscarinic episodes (1:1 ratio) based on the PSs using a greedy matching algorithm [18] with a calliper of 0.009. Matches were restricted to episodes from the same calendar year and patients of the same sex and age category (< 65 years, ≥ 65 years). If a patient contributed more than one antimuscarinic episode (even if the drugs were different), these episodes were joined together to form a single period of current exposure to antimuscarinics. This only occurred if the prescription date for the second treatment episode was within the grace period for the first episode.
Absolute standardised mean differences were calculated to assess the final PS model balance [19]. Variables with a difference ≤ 0.1 were considered balanced. If a variable had a difference of > 0.1 to ≤ 0.2, further assessment was required. A difference > 0.2 was defined as unbalanced and required further PS model modifications (e.g. interaction terms), or adjustment for the variable in the outcome model.
For the matched episodes, current exposure duration and the number of prescriptions per episode of current exposure were analysed, along with the number of matched episodes and episode type (mirabegron or antimuscarinics) contributed per patient. Baseline characteristics for the mirabegron and individual antimuscarinic episodes were evaluated before and after matching (data before matching not shown). Descriptive analyses were conducted using summary measures (frequencies, proportions, medians and interquartile ranges).
For each outcome, incidence rates per 1000 person-years and 95% confidence intervals (CIs) were calculated. Cox proportional hazard models were used to estimate HRs to compare outcome incidence rates between the mirabegron and the antimuscarinic episodes.
The incidence rate of each outcome within current exposure of the mirabegron episodes was compared with the antimuscarinic episodes as a group. This analysis was also conducted by age category, OAB medication use prior to the start of the treatment episode (naïve, non-naïve) and restricting episodes to those at high risk for CV events. Non-naïve users were defined as patients who received a prescription for another OAB medication during the prior 12 months. Episodes were defined as at high risk for CV events if the patient had at least one prior diagnosis for stroke, transient ischaemic attack, coronary artery disease, angina, AMI, heart failure, cardiac arrhythmias, chronic renal insufficiency or peripheral arterial or vascular disease or at least one prior percutaneous transluminal coronary angioplasty, coronary artery bypass graft or carotid endarterectomy. Episodes from patients with three or more of the following, hyperlipidaemia, hypertension, diabetes mellitus or age ≥ 80 years, were also considered to be at high risk. Secondary analyses were conducted that separated results according to sex and a prior history of AMI or stroke; investigated the potential effect of residual confounding by alcohol/substance abuse, obesity and tobacco smoking for each data source and censored all adverse CV outcomes upon occurrence of the first individual event. For the residual confounding analysis, assessments were performed prior to pooling estimates across data sources using a spreadsheet based on previously published formula and approaches [20].
Pooled HR estimates across data sources were generated using Comprehensive Meta-Analysis software (RRID:SCR_012779) Version 3.0 [21]. Given anticipated heterogeneity in episode characteristics, prescribing patterns and the availability of covariate information across data sources, both random-effects and fixed-effects models were implemented, with point estimates representing a weighted HR of the results from the individual data sources using an inverse-variance-weighting approach [22]. Heterogeneity across data sources was assessed using the I2 test, with I2 > 50% used to indicate substantial heterogeneity. Individual study population data were also generated, and the HRs were calculated using Cox proportional hazard models and the corresponding Wald-based CIs were derived from these models.