Methods
A retrospective review of the patients admitted to a single center from 2005 to 2012 was performed.  All patients with NVAF were selected who were on guideline directed anticoagulation therapy with warfarin (prior to approval of novel anticoagulants). Demographic, comorbidity and medical therapy data were collected and analyzed from a total sample size of 656 patients, Renal dysfunction was defined as GFR of <60 ml/min (339 patients), The MDRD method was used to calculate GFR. The patient group with GFR <60 ml/min was not further categorized into subgroups.
The selected cardiovascular outcomes of interest were myocardial infarction (MI), stroke (CVA), and death. MI was defined as NSTEMI type 1 or STEMI confirmed by in house cardiologist or by subsequent cardiac catheterization or a nuclear stress test. Patient with minimal troponin elevations in settings of demand ischemia were not labelled as having MI. Stroke was defined as radiographic evidence of non-hemorrhagic stroke or TIA confirmed by board certified neurologist at the time of discharge with the help of supporting studies done during the admission. Manual review of EMR was performed for all patients included in the study to avoid false positives and negatives.
All statistical analyses were performed with SAS version 9.4 (SAS Institute, Cary, NC), and statistical significance was set at 0.05. To check the distributions of all variables prior to analysis, categorical variables were summarized with frequencies and percentages while continuous variables were summarized with means, standard deviations, and medians and with histograms and normal probability plots. A Chi-square test, two-sample t-test, or Wilcoxon Rank Sum test was used to test for differences between the group of patients with renal dysfunction and the group of patients with GFR ≥60 ml/min (Table 1) in terms of demographic variables and other risk factors.
The association between GFR groups and myocardial infarction, stroke, death, or any of those events was tested using binomial logistic regression (Table 2).
To incorporate the element of time and adjust for covariates, a Cox proportional hazards model analysis was applied for each outcome variable (Table 3).
The time to the event was calculated as the time from the date of admission to the time of the event, or time of the first event in the case of the combination of all events, for patients who had an event and as the time from the date of admission to the end of the retrospective study period (12-31-2013) for patients who had no event over the course of the study period. Potential covariates were determined via a bivariate analysis of the variables listed in Table 1 with each outcome variable. Variables with a p-value <0.10 were included for model selection. Prior to model selection, the potential covariates were tested for multicollinearity using variance inflation factor (VIF) statistics, but no multicollinearity between variables was found. Various methods of model selection were then applied using the potential covariates with each outcome while always including GFR groups in the final model to determine if it continued to have a significant effect on the outcome when adjusted for other covariates. This included stepwise selection, backward selection, forward selection, and best subsets selection using 0.15 as the entry criteria and 0.05 as the stay criteria. The results from all methods were incorporated into making a decision about the composition of the final model. Hazard ratios were used to quantify the magnitude and direction of the effect of each significant variable on the outcome.