Statistical Analysis

A time to event analysis was used to compare the risk of an end point between treatment groups, measuring risk time from the patient receiving their initial prescription and until the relevant event, emigration, death, or end of follow-up (whichever came first). An intent-to-treat approach was applied for the main analyses. This was supplemented by a continuous treatment analysis by censoring the follow-up if the patient was prescribed another treatment than what was initiated.

Crude incidence rates were calculated as the number of events divided by the person-time. Cox regressions with a robust variance estimator were used to compare event rates between the treatment groups, with warfarin as the primary reference. To address confounding by indication of treatment, an inverse probability of treatment weighted (IPTW) analysis was applied. We used weights that enabled estimates representing mean population treatment effects. The underlying propensity models included the following treatment predictors: age (continuous); binary indicators for sex, prior bleeding, vascular disease, hypertension, diabetes, renal disease, chronic pulmonary disease, heart failure, cancer, and a recent prescription for aspirin, β-blockers, nonsteroidal anti-inflammatory drugs, statins, or loop diuretics; and the HAS-BLED score.

Balances between treatment populations were evaluated by the standardized differences of all baseline covariates, using a threshold of 0.1 to indicate an imbalance.

To evaluate the potential for residual confounding, falsification analysis was performed by applying the propensity–weighted cohort in the analyses on (falsification) end points that a priori should be expected not to be associated with the effects of treatment. For this study, we considered pneumonia, hip fractures, cancer, and urinary tract infections.

We repeated the analyses on the subgroup with a hospital discharge diagnosis for AF either before or within 30 days of first receiving a prescription. In addition, as there may be confounding because of differences in the health status of patients, we also reported results of analyses in which patients with high-mortality conditions (heart failure, cancer, or chronic pulmonary disorder) at the time treatment was initiated were excluded. Eventually, the results of the IPTW analysis were compared with a trimmed analysis that removed 5% of the extreme weights as well as with an ordinary crude and adjusted analysis (data not shown).

The analyses on the entire population were supplemented by stratified analyses on the populations with 1 of the frequent risk factors who were older than 65 and patients with hypertension; additionally, the results of the analyses were reported for men and women separately. Stata/MP, version 14 (StataCorp) and R version 3.1.1 (The R Foundation) were used for the statistical analysis. A 2-tailed P value of less than .05 was considered statistically significant.