Is in the abstract. They literally removed confounding variables. The only variable I can think of is that they are more acute to self check due past experiences.
> [..] all models were adjusted for sex, birth year, area-level urbanicity, area-level morbidity, area-level educational level, area-level employ- ment rate, school class size, school’s ninth grade size, parental educational level, parental income, and parental mental health.
So a contaminant or other environmental factor that didn't correlate perfectly with those factors wouldn't be corrected for. How could it be, without knowing what it is?
Another thing I'm not sure about -- I think some schools intentionally group troublemakers/slower-students into similar classes either officially or unofficially. I'm not gonna read the whole thing and see if they're accounting for that.
Another weird thing is that "behavioral disorder" isn't contagious if there's 1 other person who has it, but IS with 99.9% confidence if more than 1 other has it. The whole chart of P-values seems a bit odd, tbh.
I don't mean to entirely discredit this paper, but what I've learned is that a lot of academics LOVE datasets that they download and analyze without ever meeting a single person in that dataset (it's the quickest way to get papers out). This CAN be useful, but it can also be garbage-in garbage-out. Over my career I've learned that data, like code, should be presumed broken until proven otherwise.