Association between warfarin and COVID-19-related outcomes compared with direct oral anticoagulants: population-based cohort study

Study design

We conducted a population-based cohort study between 1 March 2020 and 28 September 2020.

Data source

Primary care records managed by the software provider TPP were linked to SARS-CoV-2 antigen testing data from the Second Generation Surveillance System, COVID-19-related hospital admissions from the secondary uses service, and Office for National Statistics death data through OpenSAFELY, a data analytics platform created by our team on behalf of NHS England [8]. The data set analysed within OpenSAFELY is based on 24 million people currently registered with primary care practices using TPP SystmOne software, representing 40% of the English population. It includes pseudonymised data such as coded diagnoses, prescribed medications and physiological parameters.

Study populations and exposure

We first identified all patients with a diagnosis of AF on or before study start (1 March 2020) (Fig. 1). People with missing data for sex, Index of Multiple Deprivation, < 1 year of primary care records, or aged < 18 or > 110 and prescribed injectable anticoagulants 4 months before study start date were excluded. In addition, people with a record of mitral stenosis or prosthetic mechanical valves, chronic kidney disease stage V (estimated glomerular filtration rate < 15 mL/min or on dialysis), or antiphospholipid antibody syndrome before study start were also excluded in this study because DOACs are not recommended for use in these patient groups.

We defined participants as DOAC users if they were prescribed a DOAC as their latest OAC prescription in the 4 months before study start. The comparison group was people who were prescribed warfarin as the latest OAC prescription in the 4 months before study start date. If both warfarin and DOACs were prescribed on the same day as the latest prescription (n = 32), we classified them as warfarin users as a conservative estimate because warfarin is hypothesised to have a harmful effect on severe COVID-19 compared with DOACs.

Outcomes and follow-up

The outcomes were (1) testing positive for SARS-CoV-2, (2) COVID-19-related hospital admission, and (3) COVID-19-related death (defined as the presence of ICD-10 codes U071 (confirmed COVID-19) and U072 (suspected COVID-19) anywhere on the death certificate). Testing outcomes were obtained from the UK’s Pillar 1 (NHS and Public Health England laboratories) and Pillar 2 (commercial partners) testing strategies and included results from polymerase chain reaction swab tests used to identify symptomatic individuals [9]. As pre-specified analyses, we also conducted negative control outcome analyses to examine the presence of residual confounding between warfarin and DOAC users. First, we anticipated that, within our population of people with non-valvular AF, there were unlikely to be marked differences in the likelihood of being tested for SARS-CoV-2 infection in relation to drug treatment with warfarin or DOAC. Therefore, we included being tested for SARS-CoV-2 as a negative control outcome to test our assumption. Second, we also included non-COVID-19 death, as differences in this outcome between DOAC and warfarin users could imply that differences in health characteristics had not been successfully controlled for. We conducted additional post hoc analyses to include cause-specific deaths as outcomes (i.e. death due to myocardial infarction, ischaemic stroke, venous thromboembolism, gastrointestinal bleeding and intracranial bleeding) to aid the interpretation of our results.

Follow-up for each cohort began on 1 March 2020 and ended at the latest of the outcome of interest in each analysis, deregistration from the TPP practice, death or study end date (28 September 2020) (Fig. 1).


Covariates were pre-specified, identified from a directed acyclic graph (DAG) approach (Additional file 1: Figure S1), including age, sex, obesity, smoking status, hypertension, heart failure, myocardial infarction, peripheral arterial disease, stroke/transient ischemic attack, venous thromboembolism, diabetes, flu vaccination, current antiplatelet use, current oestrogen and oestrogen-like therapy use, Index of Multiple Deprivation and care home residence. We identified covariates that are both associated with the exposure and the risk of severe COVID-19 outcomes either directly [8], or via venous thromboembolism [10, 11]. All codelists for identifying exposures, covariates and outcomes are openly shared at for inspection and reuse.

Statistical methods

Baseline characteristics in each study were summarised using descriptive statistics, stratified by exposure status. We present adjusted cumulative incidence/mortality curves using the Royston-Parmar model (Additional file 1: Figure S2). We estimated hazard ratios (HRs) with 95% confidence intervals (CIs) using Cox regression with time since cohort entry as the underlying timescale. We accounted for competing risk by modelling the cause-specific hazard (i.e. censoring other deaths for COVID-19 death analysis and censoring any death for other outcomes analysis). We used graphical methods and tests based on Schoenfeld residuals to explore violations of the proportional hazards assumption.

We performed unadjusted models, models adjusted for age (using restricted cubic splines) and sex, and DAG-adjusted models (stratified by general practice).

Quantitative bias analysis

We considered the possibility that if warfarin users had worse baseline health status they might act to lower their risk of SARS-CoV-2 infection through more risk-averse health behaviours (e.g. wearing face masks, avoiding close proximity to others) than DOAC users. Given that health behaviour is not captured in medical records, we conducted quantitative bias analyses to assess the sensitivity of our results to this potential unmeasured confounder.

We calculated the minimum strength of association required between an unmeasured confounder and one of exposure or outcome to move from the observed HR to a null bias-adjusted HR (aHR) (i.e. the E value) [12]. We also calculated the minimum strength of association required between unmeasured confounder and both of exposure and outcome to move from the observed HR to a null bias-aHR (i.e. the Cornfield condition) [12]. Furthermore, we calculated the minimum strength of association required to move from the observed protective associations to a bias-aHR of 1.2 because we hypothesised a harmful effect of warfarin in COVID-19-related outcomes.

Sensitivity analyses

Table 1 shows the list of other sensitivity analyses.

Data management was performed using Python 3.8 and SQL, with analysis carried out using Stata 16.1. All study analyses were pre-planned unless otherwise stated. All code for data management and analyses in addition to the pre-specified protocol ( are archived at:

Role of the funding source

The funder of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. AJW, CEM, SB, WH, CB, JC, LS and BG had access to the raw data. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Related Posts