Assistant Professor Department of Pharmaceutical Outcomes and Policy, University of Florida College of Pharmacy, United States
Background: Propensity score (PS) matching is commonly used in observational studies to balance the distribution of patient characteristics between exposed and unexposed groups. However, studies with a non-active comparator raise challenges. Data‐driven approaches such as high‐dimensional (HD) PS have been developed to guide covariate selection, and adjustment in pharmacoepidemiologic studies. We propose to use the method to assist in non-active comparator selection. Treatment of left atrial appendage occlusion (LAAO) with the WATCHMAN device has been demonstrated as an alternative of warfarin to reduce the stroke risk in patients with in nonvalvular atrial fibrillation (AF). Yet there is a lack of evidence on its performance in real-world patients.
Objectives: To compare 1-year mortality between nonvalvular AF patients who received the WATCHMAN device and those who did not in two cohorts using: 1) traditional PS; 2) integrating traditional PS with information from HD adjustment algorithm.
Methods: A retrospective cohort study using Medicare Fee-For-Service claims data from 2011-2017. First, patients entered the base cohort once diagnosed with AF. Then, exposed and unexposed patients entered the study cohort upon receiving WATCHMAN or at any outpatient visit with a diagnosis for AF, respectively. The date of the procedure or outpatient visit was defined as index date. We included patients 1) with 1-year continuous enrollment of Medicare A, B and D, and 2) were >=65. Patients who received palliative care prior to the index date were excluded. To improve comparator selection in cohort 2, we conducted HD algorithm, and included selected variables in eligibility criteria and PS model. We used PS-matching with 1:3 ratio between patients in the exposed and unexposed groups. Patients were followed until censoring due to the outcome (all-cause mortality), disenrollment from the Medicare, or 1-year follow-up, whichever occurs first. Cox proportional hazards model was used to estimate adjusted hazards ratios with 95% confidence intervals (CI).
Results: In the first cohort, we identified 1,159 and 3,477 patients with a mortality rate of 0.0802 and 0.0897 in exposed and unexposed groups (hazard ratio 0.87, 95% CI: 0.69 to 1.09). With the support of HD algorithm, we identified malignant cancer and frailty as two important factors for mortality. Thus, we excluded patients diagnosed with malignant cancer, and added frailty score to the PS model. Accordingly, we identified 953 and 2,859 patients with a mortality rate of 0.0745 and 0.0769 in exposed and unexposed groups (hazard ratio 0.95, 95% CI: 0.73 to 1.24).
Conclusions: HD adjustment algorithm provides an opportunity to help improve covariate adjustment and non-active comparator selection in traditional PS analysis.
