|
Abstract:
|
Large health care utilization databases are increasingly being used to study treatment effectiveness. But databases often contain only crude confounder measures, which may lead to residual confounding bias. Since collecting additional, more detailed confounder data is costly, it is useful to consider methods that involve collecting such data on a subset. We apply one such method, propensity score calibration (PSC) in a study of influenza vaccine efficacy (VE). The method 1) uses the full cohort to estimate a crude PS (based on crude confounders), 2) uses validation data to estimate a crude PS, a gold standard PS (based on accurate confounders), and a calibration model between them, and 3) performs regression calibration to correct the naïve VE estimate. We evaluate the ability of PSC to reduce bias using the 'control' period prior to influenza season when true VE is known to be zero.
|
