Materials and Methods
We used the General Practice Research Database (GPRD) in the UK to perform a population based cohort study. The GPRD contains prospectively collected medical information on about 3 million patients. This database has been used in multiple epidemiologic studies that have confirmed the completeness and validity of the recorded data [11-13]. We identified a total of 703,730 patients in the GPRD meeting the following conditions: 40–89 years old enrolled with a general practitioner for more than two years and a computerized prescription history of at least one year before January 1996. This was done to ensure that the patients have had at least one contact with the GP in the year prior to entering the study. Patients with a code for cancer (I CD 1400–2099) before January 1996 were omitted. From this source population we identified both study cohorts.

Study cohorts
AF cohort: We identified patients aged 40–89 years with a first ever recorded diagnosis of permanent/chronic atrial fibrillation in 1996. The diagnosis was confirmed by the general practitioners (GPs) through a questionnaire, this process is describe in detailed in a previous study [4]. We requested the GPs to confirm whether the episode of chronic AF was the first ever and to provide information on diagnostic tests, procedures and aetiology of the AF (patient confidentiality was always preserved). We excluded patients classified by the GP as having paroxysmal AF, 1,035 were finally confirmed. General Population cohort: Using the same source population in which the AF patients were identified and applying the same eligibility criteria as for the AF cohort, we sampled an age and sex matched cohort of 5,000 individuals free of AF.

Analysis
The two cohorts were followed up from date of first diagnosis of chronic AF (start date) and from a random date during 1996 in the general population cohort until the earliest of death, or end of follow-up (December 1999). Survival probability was computed in both cohorts and we estimated the relative risk of dying associated with AF using Cox proportional hazard regression to control for risk factors.
A nested case-control analysis was performed in the AF cohort to assess risk factors of death. We used all deceased AF patients as cases (n = 234) and used their death date as index date. The remaining alive AF patients (n = 801) were considered controls and we sampled a random date during their follow-up period that was used as their index date. Estimates of mortality risk and 95% confidence interval (CI) were computed using logistic regression. We collected recorded information on the following risk factors: smoking status, BMI, alcohol consumption, as well as prior history of heart failure (HF), ischaemic heart disease (IHD), cerebrovascular disease (CVD), hypertension and diabetes. We used both the information in the questionnaire and computerized files to ascertain the cause of death.

Drug exposure definition
We extracted from computerized records information on drugs used to treat AF between start date and index date: diuretics, beta-blockers, ACE inhibitors, calcium channel blockers, digoxin, aspirin, warfarin, NSAIDs, and other anti-arrhythmic drugs (disopyramide, procainamide, quinidine, flecainide, propafenone and amiodarone). We defined current use when drug supply lasted until index date or ended in the previous month. We considered short-term use when duration was less than 3 months, and long-term use when duration was more than 3 months. Past use was considered when the end of the most recent prescription was more than one month before index date and non use when there was no drug use between start and index date.