The performance of the full information maximum likelihood estimator in multiple regression models with missing data

A Monte Carlo simulation examined the performance of a recently available f ull information maximum likelihood (FIML) estimator in a multiple regressio n model with missing data. The effects of four independent variables were e xamined (missing data technique, missing data rate, sample size, and correl ation magnitude) on three outcome measures regression coefficient bias, R-2 bias, and regression coefficient sampling variability. Three missing data patterns were examined based on Rubin's missing data theory: missing comple tely at random, missing at random, and a nonrandom. pattern. Results indica ted that FIML estimation was superior to the three ad hoc techniques (listw ise deletion, pairwise deletion, and mean imputatiom) across the conditions studied, FM parameter estimates generally had less bias and less sampling variability than the three ad hoc methods.