COMPARISON OF IMPUTATION METHODS FOR MISSING VALUES IN LONGITUDINAL DATA

ABSTRACT
Longitudinal data are common in various sectors where repeated measurements on a dependent variable are collected for all subjects. Missing data pattern are caused when most planned measurements are unavailable for some subjects. The dropout process may cause three missing values mechanism, namely: Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR). The missing values have influence on quantitative study that can be serious, leading to biased estimates of parameters, information loss, reduced statistical power, increased standard errors, and weakened generalization of findings. This thesis compared the performance of seven (7) techniques of imputing missing values under the assumptions of MCAR and MAR mechanisms. The study adopted the little’s test to check whether a dataset with missing values is MCAR or MAR. The techniques for solving missing values problems were compared using the Generalized Estimating Equation (GEE) model for the complete dataset, the coefficient of determination and root mean squared error (RMSE). The study discovered that when large (above 10%) or small (below 10%) values are missing at random (MAR), it is important to use multiple imputation or expectation maximization to replace missing values in the dataset. The pairwise deletion is the best under MCAR mechanism. Listwise deletion and thehotdeckimputationmethodsperformedpoorlyundertheMCARmechanism. It is recommended that researchers should understand the patterns of missing values in dataset and clearly recognize missing data problems and the situations under which they occurred. However, further research is needed to find a better method for imputing missing not at random (MNAR) with multiple imputation. This thesis focused on missing values in a longitudinal dataset. However, future research using categorical data is a step in right the direction.