How should researchers exclude outliers: Across the data, or within conditions? In this paper, I show that when outlier exclusions are performed in a way that is not 'blind' to the researchers' hypothesis (e.g., within conditions), they increase Type I error to unacceptable levels.