One thing slowing down work here is that some discriminatory behaviors are wildly different from one another (violence vs. avoidance). A model that is predictive for one may not be for another. But work involving behavior is costly, so a challenge is knowing to what your results generalize