I illustrate that the simple regression-adjustment estimator is inconsistent for the average treatment effect when the random effects affecting treatment assignment are correlated with the random effects that affect the potential outcomes. I present a simple parametric estimator that is consistent in this case.