Assessing Influence in Multiple Linear Regression with Incomplete Data

The problem of assessing influence and detecting influential cases in multiple linear regression with incomplete data is considered. A case is said to be influential if appreciable changes in fitted regression coefficients occur when it is removed from the data. A one-step influence measure is derived, based on the EM algorithm for detecting cases that are influential in the maximum likelihood estimation of the regression coefficients. Results are compared with the (complete data) Cook's distance measure. Techniques are demonstrated by examples.