[gwern]

The return declines each year. See the section of the paper where they discuss how the two groups converge in outcomes as time passes.

(This is complicated by the high attrition in the control group: they lost track of a third to a half of the control group and so couldn't do an intent-to-treat analysis; they argue that the bias from this heavy attrition is probably in the direction of understating the benefits since it was mostly the ones who were worst-off at baseline who disappeared, but of course there's no way to be sure and one could easily argue the opposite.)