Statistical Analyses
Descriptive statistics were used to summarize the demographic and
clinical characteristics of patients, according to the type of OAC used.
The follow-up periods and the level of adherence were reported as the
mean with 95% confidence interval (CI) or the median with interquartile
range (IQR). The adherence to treatment in the year of follow-up was
calculated by dividing the total number of days of treatment by 365.
When the dispensing periods overlapped, the full length of each filled
claim was accounted for, and the start date of the second claim was
shifted to the end of the previous claim.
For the main analyses of the primary effectiveness and safety composites
in an on-treatment, we used an
inverse probability of treatment weighting (IPTW) approach to account
for differences in patient characteristics between treatment
groups.38,39 Four IPTW cohorts were created: (i)
rivaroxaban 15 mg vs. warfarin; ii) rivaroxaban 20 mg vs. warfarin;
(iii) apixaban 2.5 mg vs. warfarin; (iv) apixaban 5.0 mg vs. warfarin.
We then used a multivariable logistic regression model to estimate the
observed probability (according to propensity score matching) of being
in the treatment group (rivaroxaban 15 mg, rivaroxaban 20 mg, apixaban
2.5 mg, and apixaban 5.0 mg), based on all the baseline covariates, and
the impact of temporal trends accounted in the analysis by including the
date of cohort entry in the IPTW matching. By approximating the
randomization used in RCTs, the IPTW approach establishes a
pseudo-population, balances the treatment groups according to the
covariates included in the model, and thus minimizes the impact of
confounding biases in observational studies. All weights were stabilized
by multiplying the IPTW weight by the marginal probability of being in
the treatment group. Descriptive statistics were also used to summarize
the baseline characteristics of each IPTW cohort. For baseline
characteristics, only absolute standardized differences of 10% or more
between the unadjusted cohort and the IPTW-adjusted cohort were
considered
meaningful.38We reported the outcomes per 100 person-years for each treatment in each
IPTW population. Hazard ratios (HRs) with 95% CIs associated were
estimated using Cox proportional hazards models for each of the four
IPTW cohorts described above.
Patients were censored at the time of enrolment in a non-governmental
drug coverage plan, admission to a long-term care facility, hospital
admission (for more than two weeks), the occurrence of a safety or
effectiveness endpoint or death (whichever occurred first). The
patient’s OAC exposure and censored status were updated every 30 days.
For the sensitivity analyses of the primary effectiveness and safety
composites, we first estimated Cox proportional HRs for outcomes
in an intent-to-treat analyses in
which we removed the censoring criteria of drug discontinuation or
switching, so that all patients were followed up for 365 days unless
they were censored for another reason. We used an IPTW approach to
account for differences in patient characteristics between treatment
groups. We reported the outcomes per 100 person-years for each treatment
in each IPTW population. HRs and 95% CIs associated were estimated
using Cox proportional hazards models for each of the four IPTW cohorts
described above.
Secondly, we provided a negative control outcomes analyses using the
risk of diabetes complications (primary code of hospitalization (ICD-9:
250.1-250.9, 357.2, 366.41; ICD-10: E10-E14 excluding E10.9, E11.9,
E12.9, E13.0, E14.9). Lastly, we calculated an E-value to assess the
impact of unmeasured confounding.40 The E-value
indicates how strongly an unmeasured confounder would have to be
associated with use of apixaban 2.5 mg, or apixaban 5.0 mg vs. warfarin
and the outcomes to reduce the observed effect to the null, depending on
the measured covariates. All analyses were performed using SAS software
(version 9.4, SAS Institute, Cary, NC).