Data source

The source population was identified using the UK's Clinical Practice Research Datalink (CPRD; previously known as General Practice Research Database (GPRD)),14 ,15 the world's largest computerised database of anonymised primary care medical records on a population of over 12 million patients enrolled with more than 650 general practices. The geographic distribution of the practices, as well as age and sex distributions of patients, broadly reflects those of the UK population.14 Information collected includes demographic characteristics, lifestyle factors, medical symptoms and diagnoses, laboratory tests, prescriptions and referrals to specialists and hospitals. Prescriptions dispensed by the general practitioner are automatically recorded on the computerised file. Read codes are used to enter medical diagnoses,16 and a coded drug dictionary based on the UK Prescription Pricing Authority Dictionary is used for recording prescriptions. General practitioners are specifically trained for the recording of data, and practices fulfilling predefined quality criteria are labelled ‘up-to-standard’ (UTS).17 Completeness and quality of data as well as consistency with medical files are checked regularly, and numerous studies have shown the high quality of recorded data,18 ,19 including diagnostic codes for stroke and MI.20

Study design and cohort definition

We conducted a retrospective cohort study that was analysed using a nested case–control approach,21 ,22 separately for the two outcomes. This nested case–control approach is computationally more efficient than a time-dependent survival analysis while producing equivalent estimates.23 ,24 The cohort was formed of all patients in the CPRD aged 18 years or older who were members of a UTS practice, and who received at least one prescription for one of the following AEDs between 1 January 1990 and 30 April 2013: beclamide, carbamazepine, ethosuximide, gabapentin, lacosamide, lamotrigine, levetiracetam, oxcarbazepine, phenobarbital, phenytoin, piracetam, pregabalin, primidone, rufinamide, sodium valproate, tiagabine, topiramate, vigabatrin and zonisamide. Cohort entry (time zero) was taken as the date of the first AED prescription within the study period. Patients with less than 1 year of information in the database prior to the date of cohort entry were excluded, as were patients who had been dispensed an AED before cohort entry in order to include only new users of AEDs.25 Patients with a history of stroke at any time before cohort entry were excluded when studying stroke, whereas patients with a history of MI were excluded when studying MI. Patients were followed until the date of their first stroke or MI, depending on the outcome under study, the date they transferred out of the UTS practice, death or end of the study period (30 April 2013), whichever occurred first.

Case definition

For both outcomes, we identified all participants within our cohort with a first-time diagnostic code recorded for stroke or MI after cohort entry. The index date was defined as the calendar date of the first recorded ischaemic stroke, or MI, depending on the outcome under study. Potential cases of stroke were identified using codes referring to ischaemic, haemorrhagic and unspecified types of strokes (eg, cerebrovascular accident). An algorithm integrating examinations, procedures and treatments was developed to classify strokes not further specified as ischaemic or haemorrhagic. The computerised medical records of a random sample (10%) of the potential cases of stroke were reviewed by a neurologist (CR), blinded to exposure status, and the corresponding algorithm was modified until complete agreement between the computer and the reviewer's diagnostic classification was achieved. Following this process, strokes whose type could not be determined (30.6%) were classified as ischaemic strokes. Patients with a first code for haemorrhagic stroke were censored at the date of diagnosis and only ischaemic strokes and strokes not otherwise specified were considered as cases.

Control selection

For each case, up to 10 controls were randomly selected among the cohort members in the risk sets defined by the case, after matching on sex, age (±1 year), indication for AED prescription (epilepsy, pain, psychiatric disease, other/unknown), date of cohort entry (±1 year) and duration of follow-up. When no controls were found for a case, matching criteria were relaxed for cohort entry (±2 years) and for age (±5 years). The matching on calendar time permits controlling for trends over time in the use of AEDs and changes in incidence of the outcome, whereas matching on indication of AED is essential to control for the baseline vascular risk that is likely to differ between indications. Controls being selected from the risk set defined by each case, they were necessarily alive, active in the practice and event-free when matched to their corresponding case. The date resulting in the same duration of follow-up for the case and controls defined the index date for the controls.

Definition of AED exposure

For each matched set of cases and controls, all AED prescriptions issued between cohort entry and the index date were identified. The duration of each prescription was calculated from the number of prescribed tablets combined with dosage instructions. In order to assess if the risk of ischaemic vascular event varies according to enzyme-inducing properties, AEDs were classified into the three following classes: inducing AED (carbamazepine, phenytoin, phenobarbital and primidone), inhibiting AED (sodium valproate) and non-inducing AED (all other AEDs). Patients were considered current users if the duration of their last prescription included the index date or ended within 90 days prior to the index date. We further distinguished between those who had no AED prescription in the year before the index date except on the 90 days before index date (initiators) and those whose current AED exposure started before the past 90 days (current users). This distinction ensures a relevant exposure time window as it is unlikely that an AED initiated shortly before the event would be responsible for the event based on the biological hypothesis. Moreover, it avoids a potential protopathic bias, especially when studying stroke, where starting an AED could be linked to early manifestations of cerebrovascular disease. Past users were patients whose last AED prescription ended between 365 and 90 days prior to the index date. Non-users were patients with no AED prescription in the year preceding the index date. Patients currently exposed to more than one AED class, for instance with prescriptions for both inhibiting and inducing AEDs, were classified into a separate category (multiple users). Consequently, current users, initiators, past users, multiple and non-users represented mutually exclusive exposure categories. In a subsequent analysis, patients currently exposed (based on the exposure definition above) were categorised according to duration of use during the entire follow-up from cohort entry to index date. The duration of continuous use of an AED was defined by successive prescriptions for the AED class of interest with a grace period of 90 days added to the end of a prescription to allow for refill time. If within the calculated duration a new prescription of the same AED class was issued, exposure was considered continuous, lasting until the last day of the last prescription. Duration of use was stratified into three groups according to the tertiles of the distribution of use in the controls.

Covariates

In addition to the inherent adjustment by the matching variables, the following potential confounders were included in the multivariable analysis: obesity (body mass index≥30), smoking status, alcohol abuse, transient ischaemic attack, diabetes, hypertension, hyperlipidaemia, atrial fibrillation, coronary artery disease, congestive heart failure, peripheral vascular disease, chronic obstructive pulmonary disease, chronic kidney disease, cancer, blood disorders and number of medical consultations. In addition, history of stroke was included in the analysis when studying MI and history of MI when studying stroke. All variables were measured 366–730 days prior to the index date and identified with relevant diagnostic codes except hypertension, diabetes and hyperlipidaemia which were defined using diagnostic codes or related medications. Medications prescribed within 1 year before the index date, such as oral anticoagulants, aspirin and other antiplatelet drugs, non-steroidal anti-inflammatory agents, antidepressants, oral contraceptives and hormone replacement therapy, were also included as covariates. Missing data were expected for body mass index and smoking, so a separate category was created to classify this missing information; in a sensitivity analysis, we used multiple imputations for these missing values. We also performed a supplementary analysis excluding hyperlipidaemia as it is likely in the causal pathway between AED exposure and outcome. We also removed transient ischaemic attack or coronary artery disease from the potential confounders adjusted for in the multivariate model, depending on the outcome under study, as they may represent minor forms of the outcome of interest.

Data analysis

We calculated crude incidence rates for ischaemic stroke and MI, along with 95% CIs based on the Poisson distribution. Incidence rate ratios (RRs) of ischaemic stroke and MI associated with current use of enzyme-inducing AEDs and enzyme-inhibiting AEDs relative to non-enzyme-inducing AEDs and their corresponding 95% CIs were estimated from the ORs computed using conditional logistic regression. This approach provides unbiased estimates of the RRs that would be obtained from a Cox regression analysis, with little or no loss in precision, and is computationally more efficient, particularly when analysing time-varying exposures in large cohorts.24 In secondary analyses, we stratified cases and their matched controls according to the indication for AED to assess effect modification by the underlying reason for AED prescription. We also determined separately the risk of stroke and MI associated with current exposure to the different inducing AEDs as compared with non-inducing AEDs. Finally, we quantified the risk of stroke and MI according to the duration of AED use in an analysis restricted to current users ever exposed to one AED class only to exclude potential persistent effects of other AED classes.

We also calculated an approximation of absolute adjusted rate differences (RDs) and corresponding 95% CIs as measures of the excess risk associated with the inducing AED. Adjusted absolute RDs were calculated as R₀×(RR−1), where R₀ was the incidence rate of outcome among the non-inducing AED person time in the cohort (which was estimated to be 4.68/1000 persons per year for ischaemic stroke and 3.61/1000 persons per year for MI), and RR was the adjusted RR for the inducing AED versus non-inducing AED.

Sensitivity analyses were performed to assess the robustness of our results. To explore the effect of potential misclassification of current exposure on the estimated RRs, the duration of the current exposure time window was shortened to 30 days prior to the index date for all cases and controls. Thus, some participants classified as currently exposed to AEDs at the index date based on the previous definition of current use in our main analysis became past or non-users. To assess the potential adjustment for covariates in the causal pathway, we repeated the primary analysis while adjusting for covariates measured at cohort entry. Finally, to assess the potential for residual confounding, we repeated the primary analyses (enzyme-inducing vs non-enzyme-inducing AED and enzyme-inhibiting vs non-enzyme-inducing AED) using propensity scores26 (PS) and high-dimensional propensity scores27 (hdPS). The PS were estimated in controls and applied to cases using covariates measured between 1 and 2 years before the index date. To ensure overlap in the distribution of PS for each of the exposure categories, 5% of cases and controls were excluded at either end of the distribution of the PS. HdPS were estimated in controls using an algorithm described in detail elsewhere.27 Briefly, after prioritising variables according to their potential to bias the exposure–outcome association, the 300 variables most likely to cause bias were included in the hdPS model. Estimated PS (or hdPS) were then included as quintiles in the conditional logistic regression model to assess the risk of stroke and MI associated with the different AED classes. CIs were calculated using a significance level of 5%. All computations were performed using the SAS software V.9.3 (SAS Institute Inc, Cary, North Carolina, USA).