Embolism

A Markov cohort model was developed, in accordance with good modelling practices,^{10 (link)} conceptualizing the course of the disease in terms of mutually exclusive health ‘states’ and the possible transitions among them, accruing direct healthcare costs, life years, and quality-adjusted life years (QALYs) (Figure 1). The health states considered were NVAF (i.e. the starting health state for all patients), ischaemic or unspecified strokes (referred hereafter as ischaemic stroke), systemic embolism, myocardial infarction (MI), intracranial haemorrhage, other major bleed, clinically relevant non-major bleed, cardiovascular hospitalization unrelated to the events modelled or death. A detailed description of the model can be found in Supplementary material online, Appendix SC.
Figure 1

Schematic representation model. ASA, aspirin CRNM, clinically relevant non-major; ICH, intracranial haemorrhages; NVAF, non-valvular atrial fibrillation; AC, anticoagulant; IS, ischaemic stroke; HS, haemorrhagic stroke. ‘M’ represents a Markov process with 11 health states that are identical for each of the treatment options. All patients remain in the ‘NVAF’ state until one of stroke, bleed, SE, MI, treatment discontinuation, or death occurs. The transition probabilities of these events depend on the treatment. For patients on second-line aspirin ‘NVAF subsequent ASA’ the events are identical however patients cannot experience any further discontinuation. Triangles indicate which health state the patient enters after an event. Health states coloured in blue are permanent health states, with the remainder being transient health states occurring for a maximum period of 6 weeks before returning to the prior or subsequent health state.

The inverse probability of treatment weighting (IPTW) approach was used and standardized to the rivaroxaban cohort to estimate an average treatment effect on the treated (ATT) [41 (link)]. Weights were calculated based on a propensity score (PS), defined as the conditional probability of being treated with rivaroxaban based on observable baseline characteristics. The PS was then used to create a pseudo-population such that the distribution of covariates in the control group (i.e., warfarin cohort) mimicked the distribution of covariates in the treatment group (i.e., rivaroxaban cohort) [42 (link)]. The probability weight was 1 for patients in the rivaroxaban cohort, and calculated as PS/1–PS for those in the warfarin cohort, and then normalized (i.e., dividing each weight by the mean of the weights per cohort) to preserve cohort size.
Baseline characteristics used in the PS calculation included age, sex, year of index date, region, insurance plan type, Quan-Charlson comorbidity index (CCI) score, CHA₂DS₂-VASc (congestive heart failure, hypertension, age ≥ 75 years, diabetes mellitus, stroke or transient ischemic attack [TIA], vascular disease, age 65–74 years, sex category) score, HAS-BLED (hypertension, abnormal renal and liver function, stroke, bleeding) score, diabetes complications severity index (DCSI) score, DCSI-related complications (i.e., cardiovascular complications, nephropathy, neuropathy, peripheral vascular disease, cerebrovascular complications, and retinopathy), stroke/systemic embolism (SE), major bleeding, baseline medication use (i.e., non-oral anticoagulants, antihypertensives, antihyperlipidemics, other cardiovascular agents, antiplatelets, and antidiabetics), cardiovascular procedures (i.e., percutaneous coronary intervention, catheter ablation, and coronary bypass graft), other comorbidities of interest with prevalence ≥ 5%, history of cancer diagnosis and treatment, HRU (IP, ER, and OP), and costs (IP, ER, OP, and pharmacy). To prevent outliers from skewing the results of our analyses, observations assigned extremely high weights were truncated at the 99^th percentile of the distribution, whereby all weights higher than the 99^th percentile value were replaced by that threshold value.
Patients’ baseline characteristics (unweighted and weighted) were reported by treatment cohort using descriptive statistics. Differences in baseline characteristics between cohorts were assessed using standardized differences. A standardized difference < 10% was considered a negligible difference between cohorts [43 (link)].
HRU rates and healthcare costs were reported per patient-year (PPY), calculated as the number of events and the costs divided by the patient-years of observation, respectively. This approach is commonly used in non-experimental study settings to account for different lengths of observation periods between patients. HRU rates were compared between cohorts using weighted rate ratios (RR) obtained from Poisson regression models and weighted odds ratios (OR) from logistic regression models. Healthcare costs from a payers’ perspective were reported as the weighted mean (standard deviation [SD]), and weighted cost differences between cohorts were calculated. All costs were inflated to 2021 US dollars based on the medical care component of the Consumer Price Index [44 ]. Because HRU and cost data have positive values that follow a non-normal distribution and also often have zero values, non-parametric bootstrap procedures were used to estimate 95% confidence intervals (CI) and p values [45 (link)]. All analyses were conducted using SAS Enterprise Guide v.7.15.