Longitudinal studies occupy an important role in scientific researches. The defining characteristic of the longitudinal studies is the sample unit. This sample unit is measured repeatedly over time. That is, the data are collected for the same set of units on two or more occasions. When analyzing longitudinal data, investigators are often confronted with missing values which produce potential bias, even in well controlled conditions. So, it is better to overcome missing data problem to analyze longitudinal data with complete dataset. That is, why the imputation methods are essential to learn. The effectiveness of these methods for different data structures has not well been studied. In this dissertation, eight commonly used imputation methods are compared. A real data set is applied to conduct a comparison between the imputation methods. Then, a simulation study is used in a bid to this purpose. The comparison evaluates the performance of each imputation method under a variety of circumstances. In the simulation, a complete dataset are generated, different missing data mechanisms (MCAR, MAR, MNAR) are created, imputation methods are used to predict the missing values, and analyzing the mean of empirical means for each datasets. The experiments are conducted by outlining the conditions for each imputation technique to produce reasonable and efficient statistical analysis. The main purpose of this dissertation is to emphasize the need for improving methodology to handle missing values especially, in the dropout missing for which the subject withdraws from the study.
|
