Well, sure, but generally when your hypothesis demands a highly non-linear distribution function to make sense, it's just wrong. That might be true; the math could be made to work. But if it were, that is the result the study would be pushing and not the bland thing about smoking.

It's a smell test thing, basically.

Why shouldn't it be non-linear? Non-linear is what I would expect--the educated know enough to avoid a range of stupid behaviors that lower life expectancy. You see few of the life-destructive behaviors in the degreed group. You also would have seen the same effect in many of the non-degreed individuals but the data didn't separate them. More degrees, more of the careful people move from the non-degreed pool to the degreed pool and the gap between the pools rises.

There's also the factor that simply getting a degree screens out many of the people that engaged in such behaviors.

I think that's basically correct, but you don't even have to put it in quite those terms. Life expectancy measures are dominated by early deaths. The group of people without a college degree probably includes most of those who are likely to die early (teen moms, structurally unemployed, lumberjacks or electrical linemen, etc.) It also includes a bunch of people without those risk factors (administrative assistants, etc). So as more of the latter group get a degree, the high-risk population comprises a greater and greater share of the group without a degree.

I don't think it would require a weird distribution. Average life expectancy is dominated not by people living a really long time, but by a minority of people dying really early (due to child/maternal mortality and work-related or accidental deaths). If those people are concentrated among those without a college degree, then you'd see the life expectancy of people without a college degree declining as a greater share of the population goes to college.