"which sorts...gain the most benefit from what sort of exercise" is really key.

An active area of research for me is analyzing (multivariate non linear) regression models to characterize sub populations in which an observed effect is strong or weak for large genomic studies. Correlations, p-values and univariate linear models are frustratingly insufficient in many cases and effects seen across the entire population may lead to recommendations that are bad for a large numer of people.

(I'm not saying everyone shouldn't go step away from the keyboard and move some iron...just talking stats.)