The poverty–mental health gradient varies substantially across populations, far more than sampling variation alone would predict. We treat this heterogeneity as a primary object of empirical interest.