Discussion of AnneTheAgile's comment: (Very minor) To further concretize the point, one could include an extreme example. Suppose an experiment's outcome is 'age' and the control leg has a very large range of measurement, eg sigma=40. Compared to the test leg, how much improvement was there when the difference in age between the two was 2? Clearly, little differred. When I was doing yield analysis, I naturally hit on the effect size metric, without giving it a name. It was so clear that differences between my samples could be dwarfed by the noise.