Many health plans wish to monitor racial/ethnic disparities among their enrollees, but few acquire self-reported racial/ethnic data from their entire membership. When these data are unavailable, surnames and neighborhood contextual measures can provide useful surrogate information regarding race/ethnicity. We describe and evaluate a hybrid Bayesian method for combining surname and geocoded residential address information to infer race/ethnicity using data from 1,973,362 enrollees of a national health plan. This Bayesian approach substantially outperforms geocoding alone and an alternative combination of both information sources in terms of predictions of race/ethnicity, prevalence estimates and racial/ethnic disparity estimates.