Use of multiple imputation on Linear Mixed Model and Generalized Estimating Equations for longitudinal data analysis: a simulation sStudy

Longitudinal studies are useful in medical and health sciences research to examine effects associated with time. However, longitudinal analysis may be complicated by the presence of missing values. The mixed effects model (MEM) and the generalized estimating equations (GEE) are common methods for analyzing incomplete longitudinal data. Both of them make use of all available data and thus are more appealing to other methods that cater subjects with complete data only. Alternatively, multiple imputation (MI) emerged as a method to facilitate the use of methods that do not accommodate missing values. Nevertheless, it was used together with MEM or GEE as a 3-step process: 1. created multiple datasets with missing values imputed; 2. perform MEM or GEE on each dataset; 3. combine results. There was recently an interest in comparing estimates with and without the use of different imputation methods on MEM and GEE. However, the literature has not examined the use of MI with MEM and GEE in a clinical trial setting when the objective is to determine the treatment effect at specific time epochs. Also, there has been no evaluation of using MI with MEM and GEE when missing values are not occurring at random. Therefore, we performed a simulation study to assess the performance of using MI on MEM and GEE under the three missing value mechanisms: missing completely at random (MCAR), missing at random (MAR) and missing not at random (MNAR). The additional use of MI generally led to overly estimated variances and may yield estimates more biased than the use of last observation carried forward (LOCF). Under MCAR and MAR, MEM or GEE alone yielded unbiased estimates and had good coverage. Under MNAR, there was serious bias in all methods but the use of MEM or GEE alone remained not the worst compared to other methods. Comparisons of the use of MI and LOCF on MEM and GEE were also made in a randomized controlled clinical trial