Baseline characteristics were cross tabulated by each pair of canagliflozin or its comparator. To control for imbalances in patient characteristics between cohorts, we calculated exposure propensity scores as the predicted probability of receiving the treatment of interest (ie, canagliflozin v each comparator) conditional upon the subjects’ baseline covariates using three separate multivariable logistic regression models.18 (link) All variables were included and no further selection was conducted. We 1:1 matched cohorts on their propensity score using a caliper width equal to 0.2 of the standard deviation of the logit of the propensity score.19 (link) Covariate balance between the cohorts before and after propensity score matching was assessed using standardized differences; meaningful imbalances were defined as a standardized difference greater than 0.1.20 (link) For each comparison and for all outcomes, we calculated unadjusted and propensity score matched number of events, incidence rates, and hazard ratios with 95% confidence intervals. We assessed the proportional hazards assumption by testing the significance of the interaction term between exposure and time, and confirmed that it was not violated.21 (link)
We conducted several sensitivity analyses to test the robustness of our primary findings. First, among the patients with baseline HbA1c levels available (approximately one third of the total population depending on the cohort), we re-estimated the propensity score adding HbA1c level in addition to the other baseline covariates to further account for underlying glucose control. Second, to address the potential for unmeasured confounding associated with the high risk for recurrence, we restricted to patients who had not been admitted to hospital for heart failure, acute coronary, or cerebrovascular events during the 60 day period before entry to the cohort. Third, to address potential informative censoring, we carried forward the exposure to the initiated drug for 365 days without considering drug discontinuation or switching, to mimic an intention to treat approach.22 (link)
In addition, we conducted subgroup analyses stratified by presence of heart failure or cardiovascular disease at baseline for the primary outcomes of heart failure admission to hospital or the composite cardiovascular endpoint respectively (see web appendix 3 for the definition of subgroups).
All analyses were performed using SAS 9.3 Statistical Software (SAS Institute Inc, Cary, NC).
We conducted several sensitivity analyses to test the robustness of our primary findings. First, among the patients with baseline HbA1c levels available (approximately one third of the total population depending on the cohort), we re-estimated the propensity score adding HbA1c level in addition to the other baseline covariates to further account for underlying glucose control. Second, to address the potential for unmeasured confounding associated with the high risk for recurrence, we restricted to patients who had not been admitted to hospital for heart failure, acute coronary, or cerebrovascular events during the 60 day period before entry to the cohort. Third, to address potential informative censoring, we carried forward the exposure to the initiated drug for 365 days without considering drug discontinuation or switching, to mimic an intention to treat approach.22 (link)
In addition, we conducted subgroup analyses stratified by presence of heart failure or cardiovascular disease at baseline for the primary outcomes of heart failure admission to hospital or the composite cardiovascular endpoint respectively (see web appendix 3 for the definition of subgroups).
All analyses were performed using SAS 9.3 Statistical Software (SAS Institute Inc, Cary, NC).
