Outlier detection in a repeated measure design

Abstract

AbstractThis paper considered the problem of outlier detection in a repeated measure design (RMD) using the Estimate distance and the Liu and Weng's residual methods. The comparative performance of the two methods in identifying outlier in a RMD was evaluated through empirical and simulation studies. For the empirical study, two types of outliers (outliers I and II) were randomly introduced into the real‐life dataset from physiological research at a time. The results obtained revealed that the Liu and Weng's Residual and Estimate Distance tests correctly detected the outlier I randomly introduced into all the subjects at a time. However, the Liu and Weng's Residual test was only able to correctly detect the outlier II in subjects 1, 4 and 6 while the Estimate distance test could not detect the outlier II in any of the subject. In the simulation study, random samples of size 105, 120, 135, and 150 with a corresponding number of subjects (k = 7, 8, 9, 10), respectively, were generated from a multivariate normal distribution. Two types of outliers (outliers I and II) were randomly introduced into the simulated datasets. The results of the simulation study indicated that the Liu and Weng's residual and Estimate distance methods correctly detected the outlier I randomly introduced into the simulated data for all the sample sizes considered. However, the Liu and Weng's Residual test outperformed the Estimate distance test in detecting the outlier II. Thus, the Liu and Weng's Residual proved to be more powerful than the Estimate Distance test in identifying outlier in a RMD.