Perhaps the oldest recorded example

Another example: it's common to use the distribution of single-locus Fsts as a null, but Janis finds that this only works well for polygenic traits if one limits to single-locus Fst for common variants. Using rarer variants can send Type I error very high (7/n)

Janis tested a number of these in simulations and found that the details---some of which are subtle---can matter a lot. For example, using an "average of ratios" Fst with the Lewontin-Krakauer distribution can drive the type I error rate up close to 1 (6/n)

When using Pst, in particular, the fact that h2b is unbounded from above means that standard procedures can lead to conservative tests for selection. And it is easy to be misled by scenarios where selection and environmental changes drive a trait in opposite directions in one or more populations

Further, it has some odd properties---unlike what we normally expect from a variance decomposition, it is not bounded from above by 1 (it is not bounded from above at all, in fact), and it doesn't specify the direction of mean genetic or environmental differences among groups.

yes, in addition to its other virtues, this game teaches the value of a good remedy