Identification and Estimation for Nonignorable Missing Data: A Data Fusion Approach