This nationwide study retrieved data from Taiwan's National Health Insurance Research Database (NHIRD) for the period from 2012 to 2020. The NHIRD includes demographic information and medical claims for outpatient, inpatient and emergency care services of about 23 million people (nearly 99.9% of the entire population of Taiwan) [13, 14]. Several diagnosis codes for major diseases in the claims database have been validated by previous studies, such as ischemic stroke [15], atrial fibrillation [16], and depression [17]. We linked the NHIRD to the Cause of Death Registry which covers the entire population in Taiwan, to capture information about completed suicide. The data for the Cause of Death Register is sourced from government agencies, including the household registration authority and the Ministry of Health and Welfare. This data is based on the cause of death listed on official death certificates, which are legally recognized by the authorized department, ensuring high accuracy of the data. National Cheng Kung University Human Research Ethics Committee approved this study (NCKU HREC-E-110–453-2).

We hypothesized an open labeled, randomized, pragmatic trial to evaluate the risk of suicide-related outcomes among patients with atrial fibrillation, receiving different oral anticoagulants [18, 19]. The key components of the trial emulation are presented in Additional file 1: Table. S1. Specifically, we included patients with atrial fibrillation aged 20 years and older who newly received an oral anticoagulant, and who had no previous suicide-related events and no contraindications for NOACs. We randomly allocated patients to either NOACs or warfarin. The primary outcomes were suicide-related outcomes, including attempted suicide and completed suicide. The causal contrasts of interest were the intention-to-treat effect and the per-protocol effect [20].

We emulated the trial using Taiwan's NHIRD. The study design diagram is provided in Additional file 1: Fig. S1. All eligibility criteria were analog to the target trial. Specifically, the index date was defined as the first date of oral anticoagulant prescription. We retrieved data covering one year before the index date to ensure patients were new users of oral anticoagulants, had no previous suicide-related events, and had no contraindications for NOACs. We confirmed patients’ diagnosis of atrial fibrillation by International Classification of Diseases, Ninth Revision, Clinical Modification codes (ICD-9-CM) codes 427.31 or ICD-10-CM codes I48.0, I48.1, I48.2 and I48.9.

We classified patients into the NOACs or warfarin groups based on their first prescription for oral anticoagulants. To emulate the randomization of the target trial with the two comparison groups sharing the same probability of treatment assignment, we implemented a propensity score with fine stratification weighting approach. The details of this approach are described elsewhere [21, 22]. Briefly, we calculated the propensity score using logistic regression based on the measured covariates listed in Additional file 1: Table S3. These covariates were defined based on literature review and experts’ opinions, including patient age at index date, sex, insurance premium level, index year, specialty of prescriber and hospital level where the oral anticoagulants were prescribed. We also considered patients’ baseline CHA2DS2-VASc score, history of major bleeding events, comorbidities and medication use, which were based on the medical records from one year before the index date (Additional file 1: Fig. S1). After trimming observations from non-overlapping regions of the propensity score distribution, we created 50 equally sized strata based on the PS distribution of the entire population. We then determined the individual patient weights and the average treatment effect among patients receiving NOACs (ATT), according to the numbers of patients in each stratum [21].

Just as in the target trial, the primary outcome of the study was suicide-related outcomes, including attempted suicide and completed suicide. Attempted suicide is identified by ICD-9-CM codes E850-E854, E858, E862, E868, E950-E959, 965, 967, 969, 9779 and 986 or ICD-10-CM codes X40, X41, X42, X44, X46, X47, X60-X84, Y87, Y10-Y12, Y16-34, T39, T40, T423, T424, T427, T43, T509 and T58 from both outpatient and inpatient services. These definitions have been used and validated by previous studies [23]. Completed suicide is identified by cause of death Case Type 03 in the Cause of Death Registry. We employed as-started analysis, an analog of intention-to-treat effect, and followed patients from the index date until the occurrence of study outcomes, loss to follow up, death other than suicide or end of the study period (December 31st, 2020), whichever occurred first. Also, we employed on-treatment analysis to emulate the per-protocol effect, which additionally censored patients whose treatments changed, including switching to a different anticoagulant and discontinuation (i.e., no prescription refill for more than 90 days from the last day of the previous prescription).

We calculated the standardized mean difference (SMD) to assess differences in the covariates between the two groups. An SMD value of less than 0.1 was considered to signify no difference between the groups [24]. We estimated the cumulative incidence curves for each study outcome using the Kaplan–Meier function. To address possible competing risks from mortality other than suicide, we applied the cause-specific hazard model and the Fine and Gray subdistribution hazard model to compare the risk of suicide-related outcomes between the two groups and we obtained the hazard ratio (HR) and 95% confidence intervals (CIs) for each outcome [25]. All statistical analyses were conducted with SAS software, version 9.4 (SAS Institute).

We conducted a series of sensitivity analyses to test the robustness of the analysis. We determined the individual patient weights and the average treatment effect among the entire population (ATE), differing from the targeted population of the main analysis. We classified the NOACs treatment group into four subgroups, i.e., the dabigatran, rivaroxaban, apixaban and edoxaban groups, and evaluated the relative risk for each individual NOAC, compared to warfarin. We repeated the analyses among patients with a history of depressive disorders and among different subgroups by patient age (< 65 or ≥ 65 years), sex (male or female), and index year (2012 to 2014, 2015 to 2017, or 2018 to 2020), to examine if the risk profiles changed [26,27,28]. Additionally, because clinicians may consider prescribing warfarin for patients with renal impairment, we excluded patients with chronic kidney disease to minimize the potential confounding effect.

We repeated the analyses with positive and negative control outcomes, for which their associations with the drugs were well-known, to examine potential residual confounders. The positive control outcomes included ischemic stroke or systemic embolism, cardiovascular mortality and all-cause mortality [29, 30]. We hypothesized that patients taking NOACs were associated with a reduced risk of these positive control outcomes, compared to warfarin. The negative control outcomes included oral health conditions, which were assumed not to be related to the use of oral anticoagulants. We also performed an analysis of outcomes occurring within 90 days after the index date. We assumed that oral anticoagulants require an induction period of more than 90 days to have an influence on suicide-related outcomes [31]. Any observed difference in the risk of suicide-related outcomes between the groups during this period would likely have resulted from the patients’ baseline risk rather than from the use of oral anticoagulants.