Confounder Adjustments by Propensity Score and Disease Risk Score in the Early Stage of Post-Marketing Drug Safety Surveillance

• Activity Number: 263
• Type: Contributed
• Date/Time: Monday, August 1, 2016 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Tae Hyun Jung* and Jessica Kim
• Companies: Yale University and FDA
• Keywords: FDA Mini-Sentinel project ; Propensity score ; Disease risk score ; Iteratively reweighted least square ; Type I error rate ; Empirical power

Abstract

Propensity score (PS) and disease risk score (DRS) are used to reduce multiple confounders into a single score and balance the difference between groups. A recent FDA Mini-Sentinel pilot project investigated the performance of PS and DRS using the full-cohort (full-cohort DRS) under a sequential design. However, the DRS using the unexposed group (unexposed-only DRS) has not been evaluated yet despite its clinical implications. Also, automated variable selections employed to estimate PS and DRS models may drop important variables and possibly compromise the validity of the results. In this study, we assumed the outcome is rare (0.5%) and improved the PS and DRS model estimation by using the iteratively re-weighted least square while the time-varying exposure is ranged from 5% to 30%. Also, we accommodated the unexposed-only DRS in sequential monitoring framework. Matching (1:1 and 1:4) and stratification (5 strata) of PS and two types of DRS were implemented in the conditional logistic model. Three scenarios related to different confounding strength were considered to compare the performance evaluated by type I error rate, empirical power, and time to signal detection.