I would say sometimes you can control for this statistically. When the selection problem is exactly along the dimension you want to measure, (i.e., measure the impulse-control of children of parents with poor impulse-control, but only the part caused by angry yelling) statistical correction is less plausible.