Abstract
We investigated the issue of selection bias in a case study of mammography utilization using a potentially endogenous predictor, the presence of gatekeeper requirements in an individual's health plan. Data from the 2000 Medical Expenditure Panel Survey and the linked 1998/1999 National Health Interview Survey were used. We employed and compared results from an extensive list of estimation methods: single equation estimations including standard probit and linear probability models, two-stage instrumental variables (IV), and bivariate probit models. Specification tests were used to inspect the relevance and validity of candidates for instrumental variables. Given the best instruments available to us, we did not find evidence that correcting for selection bias improved the estimates. Standard single equation estimations remained the preferred approach in our study. © Springer Science + Business Media, Inc. 2005.