I like this article. To me, it even provides more evidence as to why Tukey's fences are not really the way to go for outlier detection (Probably not your intention). Especially when in the real world the data in itself is not a Gauss process I think in a case where you are analysing a sample of averages by all means do so (no pun intended!). In cases where you are testing for causality, be extremely careful. Just imagine a case where you have data that has high kurtosis, you could end up removing over 15% of your data. If you end up using Tukey's Fences as a form of outlier detection in such a scenario, how can one make a strong claim about causality after wiping out 15% of the data?? Again, good article.