Background:

Missing data is a common problem encountered in clinical trials. Researchers intending to analyze their data using Intention-to-treat principles commonly impute the missing data to achieve balanced data before analysis using analysis of variance (ANOVA) methods. A lesser known but advanced method of handling missing data is the Linear Mixed Effects model (LME) where balanced data is not required. A randomized controlled trial was recently conducted to determine the effect of oral mixed tocotrienols (T3) on the symptoms of school-going children with Attention Deficit/Hyperactive Disorder (ADHD). Between 30% to 53% of teacher-reported scores collected at three different time points during the trial were missing. This study showed improvement in symptoms for all children regardless of treatment groups.

Objectives:

To compare three methods of handling missing data in this longitudinal study on the changes of the teachers’ scores over time.

Methods:

Using a logistic regression model, we first determined that data was not missing completely at random. The three methods used to determine the change in teachers’ score over time were: 1) complete case method, 2) single imputation followed by ANOVA repeated measurement, and 3) LME. Imputation was based on multivariate normal distribution. The results from each method were then compared.

Results:

Methods 1 and 3 showed similar results in the effects of treatment but not time on the change in teachers’ score. Imputation in Method 2 did not consistently produce a similar data set each time it was done. Therefore, the results of the teachers’ score were different using Method 2.

Conclusions:

LME is superior, as it retains the power of the study and avoids the need for imputation. Therefore, if our findings are consistent in other settings, studies with high percentages of missing data handled using this model could be considered for inclusion in the meta-analysis of a Cochrane systematic review without increasing the risk of bias.