Julian Parris suggests that in cases where the outlier is clearly not a result of a collection mistake, it can be useful to label them and run the analyses two ways - with and without outliers - to see if the outliers change the relationships among the data or help you glean information from the data.