Comparison of regression coefficient estimates of linear mixed model on missing longitudinal data

Abstract

The aim of this research is to compare the regression coefficient estimates of linear mixed models on different mechanisms of missing longitudinal data namely missing completely at random (MCAR), missing at random (MAR), and missing not at random (MNAR). In this study, longitudinal data were initially generated where the random error were normally distributed. Then, the comparison was carried out through the simulation study where the missing rates of each mechanism were set to be 10% and 20%, and the sample sizes were 5, 10, 20, and 50 with 5 repeated measures. The simulation process was replicated 1,000 times using R program in which the maximum likelihood estimation was employed. In overall, by considering both pseudo-bias and pseudo-root mean square error (RMSE) we found that this estimation method is suitable only when the data were missing completely at random (MCAR), but not when the data were missing not at random (MNAR), especially in the large sample size.