Validation of Propensity Score Calibration Method to Control for Unmeasured Confounding in Time-to-Event Analyses

• Activity Number: 360
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Rebecca Burne*+ and Michal Abrahamowicz
• Companies: McGill University and McGill University
• Keywords: Propensity Score Calibration ; Survival Analysis ; Pharmacoepidemiology ; Simulations ; Unmeasured Confounding

Abstract

Observational studies of the effects of treatments on clinical outcomes typically rely on large administrative databases, which usually lack information on important confounders, eg. clinical and lifestyle characteristics. However, such confounders are often recorded in smaller, clinical 'validation' datasets. Propensity Score Calibration (PSC) was proposed to control for confounding bias in such cases. The method uses the estimated relationship between 2 types of Propensity Scores (PS) based on, respectively, (i) all variables in the validation dataset (Gold Standard PS) and (ii) reduced set of variables available in the main database (Error-prone PS). This relationship is then used to predict the Gold Standard PS in the main database, and to use it as an adjustment variable for the effect of treatment. PSC relies, however, on an important 'surrogacy' assumption that the effects of all variables on the treatment are the same as their effects on the outcome. We perform simulations under various assumptions to validate and assess this method in Cox Proportional Hazards regression analyses of time-to-event data and assess the impact of the violation of surrogacy on the results.