A Comparison of Missing Data Imputation Methods in Within-Subject Repeated Measure Design

Abstract

Within-subject repeated measure design is an experimental design conducted by collecting data from the same sample unit at different times or with other conditions. It is popular in medical or public health research. This article presents a comparison of missing data imputation methods in within-subject repeated measure design when missing values are missing completely at random. The imputation methods were applied by the Mean Substitution method, CopyMean Trajectory method, CopyMean LOCF method and Artificial Neural Network method by using 3 assessment criteria such as MAD, RMSD, and Bias. All these methods were tested on both real dataset and artificial datasets when mean and variance in each variable were equally defined. The results revealed that, in the most cases, the artificial neural network method performed the best in real dataset and in artificial datasets with no correlation or low correlation (0, 0.3, and 0.5). However, in artificial datasets with high correlation (0.7 and 0.9), the CopyMean Trajectory method was the best method in the most cases.