We have recorded some kinematic data, and are looking at three measures derived from the movement data for each subject. In order to identify outliers, we are using the Mahanalobis distance. I got into a discussion with another researcher about whether it is OK to do the outlier detection for each kinematic measure alone (assuming they are more or less independent of each other), or whether outliers must be removed from all three measures. My feeling is that it is wrong to remove outliers for each measure separately, as then we are changing the sample of subjects. However, I could not find any good references to support this claim. Which approach is more reasonable / statistically sound? Can you suggest any references? Thanks.