Missing covariates in multiple linear regression when the data is missing at random

Abstract

The occurrence of missing values is frequent in data collected for different uses such as in surveys, censuses, balanced experiments. On the other end most statistical analysis methods have been developed for complete rectangular data. This paper uses a simulated data set to examine the performance of recently available methods for treating data with missing values. Multiple imputations (Ml) and maximum likelihood (ML) methods through expectation maximization (EM) were compared with the complete case (CC) analysis which is the default method in statistical computer packages. The effects of the data treatment methods were examined on the regression coefficients. The results indicate that ML through EM and Ml methods arc both superior than the commonly available complete case analysis (CC).