Maximum Likelihood Inference for Multiple Regression with Missing Values: A Simulation Study

Maximum likelihood inference for the coefficients of a multiple regression with missing values is studied by simulation, using artificially generated multivariate normal observations with randomly deleted missing values. The method is compared with least squares on the complete observations by mean squared error. Two methods of estimating the covariance matrix of the m.1. estimates are compared, and the applicability of large sample theory to obtain tests and confidence intervals for the regression coefficients is considered for moderate size samples.