A SIMPLE OUTLIER DETECTION METHOD FOR INTRA-SUBJECT TIME-SERIES DATA

Abstract

Removal of outliers assists in improving the statistical representations of the general finding. Currently no simple method is advocated for detecting outliers in time-series data obtained in biomechanics. The aim was to demonstrate a 2-stage method for detecting outliers. The test data were the ankle and knee angles for the strides (n=41±2.8) from treadmill running (n=6). Stage 1 was an outlier detection of >±3.3SD from the mean at each time-point, and removing any stride with an outlier. Stage 2, with padding of k=3 points and mean-detrending, was a moving window SD for all strides across ±k data points, and removing strides with any point >±2.58SD. After removal of 5.2±3 (stage 1) and 2.0±1.4 (stage 2) strides, the mean was unchanged and the SD reduced (p<0.05). The method is simple and effective in removing outliers in intra-subject time-series data.