Data source

This study was based on data from the SNDS (Systeme National de Données de Santé) [14, 15], the national electronic health care database linked to vital statistics in France. The SNDS includes health care data for more than 67 million individuals living in France, from birth (or immigration) to death, and covered by the national health insurance system, which is mandatory in France. The SNDS contains individualized, anonymized data on demographics (sex, year of birth, date of death if relevant); on health status through long term disease (LTD) status resulting in full insurance coverage for the patient with at least one LTD; all reimbursed outpatient healthcare encounters (visits, medical procedures, lab tests, drugs, medical devices); all hospital procedures and discharge diagnoses both for public or private hospitals (with main diagnosis, related diagnosis, and as many associated diagnoses as necessary for one hospital discharge summary). Diagnoses identified in LTD and in hospital discharge summaries are coded according the 10th revision of the International Classification of Diseases (ICD-10). Information available for any prescribed and reimbursed drug is drug name, dosage, form, quantity and dates of prescription and dispensing in community pharmacies [14,15,16]. The SNDS has been extensively used in epidemiology and pharmacoepidemiology, in particular to assess cardiovascular disease management and outcomes [17,18,19,20].

For the purpose of this study, data were extracted from the SNDS in January 2017 after ethics and regulatory agreements, and covered the period from January 1st 2011 to December 31st 2014. Available information was demographics (including vital status), reimbursed drugs, LTD and hospitalisation data.

Patient selection

Subjects over 65 years old at the time of a first dispensing of any triptan between July 1st 2011 and June 30 2014 were selected and defined as new users of triptan (if nodispensing of any triptan had been given in the 6 months preceding the date of the firstdispensing identified in the study period). Subjects with adispensing of subcutaneous sumatriptan for cluster headache as well subjects concurrently prescribed ergot alcaloides and triptans were excluded. Each incident user of triptan was matched with four unexposed controls, on age, sex, and area of residence. The exclusion of the 6 first months in 2011 insured to include only incident users, as well the exclusion of the 6 last months in 2014 allowed to have at least 6 months of follow up for all subjects included in the cohort.

Study design

The data analysis was performed in 2 steps. First, from the exposed-unexposed cohort extracted from the SNDS, we implemented a propensity score-matched cohort study comparing the incidence of cardiovascular outcomes and death according to triptan exposure. Second, in order to address potential residual confounding, we conducted complementary analyses where subjects were their own controls which allows self-adjusting over a short period for individual time invariant characteristics that are not recorded in medico-administrative healthcare databases (such as diagnostic of migraine or cigarette smoking, alcohol consumption, overweight or obesity,…): a case-crossover (CCO) analysis and a self-controlled case series (SCCS) [21,22,23,24].

Primary analysis: propensity score matched cohort studySelection of triptan users and matched controls

The characteristics of the initial exposed-unexposed cohort identified in the SNDS were unbalanced, unexposed controls presenting more comorbidities and health care consumption than incident triptan users. A propensity score was computed trough a logistic regression model including baseline covariates available before cohort entry (index date being the date of the firstdispensing of triptan for triptan users and for the 4 matched-controls). These variables were sex, age, Charlson’s comorbidies score (CCS) estimated from ICD-10 codes identified in hospitalisation diagnoses, LTD, and/or drug dispensing [25,26,27,28] comorbidities known as risk factors of cardiovascular events (hypertension, cardiovascular disease, dyslipidemia and diabetes), number of medical visits and number of hospitalisations within the six months before index date. The list of codes used for identifying these covariates is provided in supplementary Table 1.

We performed a nearest neighbor matching with a caliper of width equal to 0.2 of the standard deviation. Distances before and after propensity score matching between triptan users and controls were investigated through Cohen’s d computation. Subjects in the PS matched cohort were followed for up to 90 days from the index date until predefined outcome.

Exposure

Triptan exposure was assessed through data available in the SNDS and included triptans available in France: sumatriptan (except subcutaneous sumatriptan for cluster headache), naratriptan, zolmitriptan, rizatriptan, almotriptan, eletriptan, frovatriptan. Quantities provided for each triptan at each dispensing were converted in defined daily dose (DDD) using the ATC/DDD index of the WHO Collaborating Centre for Drug Statistics Methodology [29]. According to this methodology, one DDD for a given drug is the assumed average maintenance dose per day for this drug used for its main indication in adults. Because data available in the SNDS do not allow to know the exact exposure in patients, we considered that they were exposed from the day of dispensing (index date) until the end of treatment period corresponding to the number of days of supply (number of provided DDDs, days of treatment at the assumed average of daily dose). For example, the DDD value is 2.5 mg for frovatriptan, naratriptan and zolmitriptan, or 40 mg for eletriptan (per os).

Outcomes

The main study outcome was acute ischemic vascular events, defined by an hospital admission within the 90 days following index date with a main diagnosis of a list ICD-10 codes, having a priori the best positive predictive values [17,18,19,20, 24, 30] for these events (supplementary Table 2). The secondary outcomes were death of any cause within the 90 days following index date and death occurring in the 30 days after hospital admission for an ischemic cardiovascular event.

Secondary analyses

Because the residual confounding in the exposed-unexposed cohort study, we added complementary analyses based on self-controlled designs. In that designs, only individuals with the event of interest are considered, which then act as their own control (i.e., they consist in within-patient comparison between different periods of time). Their main advantage is that time-invariant confounders that act multiplicatively on the event rates are inherently controlled for. These designs include the case-crossover (CCO) design and the self-controlled case-series (SCCS) and were developed in the early 90’s to study the short term effect of transient-exposures and abrupt onset events 22. Indeed, because of the self-matched design, the risk estimation includes only data for patients who switch their exposure status over time (i.e., from exposed to unexposed, or vice versa). Because exposure to triptans is occasional and study outcomes are acute by definition, these methods are appropriate to insure the robustness of results from the exposed-non exposed cohort.

The CCO designs were initially developed by Maclure in 1991 [21], with initial applications being the study of the triggering factors of myocardial infarction, or road injuries. They were then used in pharmacoepidemiology to study the association between drug exposure and the occurrence of an event.

All subjects with the event of interest (cases) are identified and included in the study. The exposure is collected over the so-called “risk period” immediately preceding the event, then over one or more previous periods so-called “control periods”. If an association exists between the exposure and the event, more frequent exposure should be observed during risk periods compared to control periods. This association is estimated through the calculation of a CCO odds-ratio (OR): the exposure rate in the case periods is reported to the exposure rate in the control periods in the same subject. Only the discordant pairs of subjects are considered, since patients whose exposure status does not vary over risk and control periods do not contribute to the calculation.

SCCS designs were initially introduced to investigate the potential relationships between Measles Mumps Rubella vaccination and aseptic meningitis, but have been widely used to study other vaccine risks, and extended to other adverse drug events [23–24].

As in the CCO design, only cases are included, but in contrast to the CCO, the entire exposure history inside a given time window is retrieved, not just exposure attributes of selected dates or periods. Other important features of SCCS are that the exposure history occurring after the event is included in the estimates. The event rate is compared between person-times at risk (risk periods with exposure) and person-times not at risk (baseline periods) in the same individual. Unlike CCO studies for which periods are defined a priori, SCCS expand the self-controlled method to periods with different sizes, and use all the information available during the observation period [23–24]. As with CCO studies, case selection must be done without knowledge of the potential exposure. The risk function for one individual is considered as depending on age and on his/her exposure at each time.

For the self-controlled analyses, the observation period to select cases began from 01/07/2011 (or at age 65) and ended on 31/12/2014 (or eventually the day of death). In this period, subjects in the triptan exposed cohort and hospitalized at least once for an ischemic event were identified. Due to the lack of a common formula for the population size (for CCO), we included as much as possible eligible subjects. For SCCS, with a level of significance at 5% and a power at 80%, and for an expected relative risk of 1.3 and a risk period/baseline periods ratio of 0.1, the required population was 1192 subjects.

Case-Cross Over (CCO)Definition of index date and risk periods

All subjects with an event of interest (hospitalization for an ischemic event) after inclusion were identified. Triptan exposure immediately before the event (risk period) was compared with triptan exposure in two previous and separated periods (control periods). The reference event was the first event observed. The index date was the day of hospitalisation (named T0) and the risk period was therefore the 10-day period preceding the index date. The control periods were two previous 10-day periods, separated by washout periods of 60 days, to ensure that the autocorrelation do not affect the results (Fig. 1).

Fig. 1

Study design of the case cross over analysis. All subjects with an event of interest (first hospitalisation for an ischemic event) and exposed to triptans before the event were selected (from the exposed cohort). They were observed from the beginning of the cohort (01/07/2011) or when aged 65, to the end (31/12/2014.). Exposure to triptans(beginning on day of dispensing and covering a number of days corresponding to the number of supplied DDD), was compared between risk period (yellow) and control periods (dark blue). Exposure during washout periods (light blue) was neutral. Duration of exposure was modified by adding 7, 14 and 28 days in sensitivity analyses. The risk period for an ischemic event was defined as 10 days before the day of hospital admission

Triptan exposure was defined by at least one day of triptan (one DDD) in risk or control periods. No exposure was defined by no overlap with the considered periods.

The analysis was performed by type of events (all ischemic vascular events, ischemic cardiac events, and ischemic cerebral events).

Self controlled case series (SCCS)

This analysis estimates the relative risk of the event of interest in risk periods compared with all other periods (baseline periods). Patients with at least one event of interest were selected in in the observation period, i.e. when they were aged 65 years old or from on 01/07/2011 to 31/12/2014 or on the date of death (Fig. 2).

Fig. 2

Study design of the self controlled case series. All subjects with an event of interest (first hospitalisation for an ischemic event) and exposed to triptans before the event were selected (from the exposed cohort). They were observed from the beginning of the cohort (01/07/2011) or when aged 65, to the end (31/12/2014). The rate of the event of interest was compared between risk periods (yellow) following triptan exposure (number of days in DDD plus 10 days) and baseline periods (dark blue)

For each subject, the “risk period” began the day of triptan dispensing and ended ten days after the last day of exposure. This period was added to ensure that triptans have been completely eliminated from the body, since the ½ life of all triptans and their active metabolites (except flovatriptan) ranges between 2 and 5 h, leading to a complete elimination between half a day and 36 h (flovatriptan with ½ life of 24 h may be eliminated in 7 days) To ensure an important assumption of SCCS that recurrent events in the same subject are independent, if an ischemic event (any) was followed by other events within the 6 months after the first event we restricted the analysis to the first event, both for the risk periods and the baseline periods.

Potential confounders

In addition to the covariates used in the cohort study, additional potential confounders were considered for CCO and assessed within the risk period and the control periods: exposure to monoamine oxidase inhibitors (MAOIs), selective serotonin reuptake inhibitors (SSRIs), non-steroidal anti-inflammatory drugs (NSAIDs), opioids, antiplatelet drugs, drugs for the cardiovascular system, caffeine-containing drugs, exposure to levothyroxine, and season of event occurrence (spring (as reference): March to May, summer: June to August, autumn: September to November and winter: December to February).

Statistical analysis For the primary analysis: propensity score matched cohort study

Descriptive statistics were used to describe the characteristics of the cohort, cases and matched controls. Continuous variables were described as mean and standard deviation (SD), qualitative variables were summarized using frequencies and percentages.

A Cox proportional hazard regression model, with stratification on matched pairs, was used to investigate the association between triptan exposure and outcomes. Univariate analyses were first performed to select the variables with a P value < 0.20, followed by a multivariate approach using backward selection and P value < 0.05 for statistical significance. Relevant interactions between covariates were checked. Proportional hazard assumption was tested for all covariates using interaction with time. The crude and adjusted hazard ratios (HRs) and their 95% confidence intervals (95% CI) were estimated.

Complementary analyses were performed on triptan exposure modelling, using cumulative triptan exposure since index date and considering triptan exposure as a time-varying variable. Sensitivity analyses were performed because the uncertainty related to definition of triptan exposure by adding 7, 14 and 28 days to the triptan duration (in DDD) and considering 7, 14 and 28 after the day of dispensing instead of exposure in number of DDD.

For the secondary analysis: CCO and SCCS

A conditional logistic regression model was used to estimate odds ratios (ORs) and their corresponding 95% CI in the CCO analysis, adjusted on the predefined covariates. A control-crossover analysis was performed to take into account the temporal trend of triptan use. This control sample included event-free triptan users matched with cases by age, sex and Charlson score (1:1). Index date (date of matching), case and control periods were the same as those in the CCO. This method estimated the odds ratio of the triptan-dispensing trend from 2011 to 2014.

In the SCCS analysis, the relative risk was estimated using a Poisson regression model, through the standard and pseudo-likelihood methods. The standard method took into account all exposure periods and events occurring during these periods. The pseudo-likelihood method took into account only the exposure periods before the event. Subsequent exposure periods were redefined as part of the baseline period and the number of events occurring during those exposure periods was adjusted to the number of events occurring in the baseline period.

The events of interest were analyzed simultaneously and separately (cardiac and cerebral events), to investigate whether the effect of triptans on receptors 5HT1B/D in different arteries may differ.

Sensitivity analyses were performed, by adding 7, 14 and 28 days to the period of exposure in the CCO and by excluding patients exposed to opioids in the SCCS All statistical analyses were performed using SAS 9.4® (SAS® Institute Inc, Cary NC, USA).

Data protection and ethics

This study protocol was approved and received all mandatory authorizations according to the French regulations (Institut des Données de Santé approval in June 2014 and Commission Nationale de l’Informatique et des Libertés authorization in December 2014). Given that data are anonymous, no informed consent was required for studies based on French health insurance databases. This study, called TRUE for « Triptan Use and serious vascular events elderly over 65 years », was registered in the post-authorization survey registry of the European Network Centers for pharmacoepidemioly and Pharmacovigilance (ENCePP) coordinated by the European Medicine Agency (EUPAS n°8976).