If the fundamental problem in causal inference is a missing data problem, then what conditions are necessary for the observed treatment effect (e.g., observed difference between treatment groups) to be an unbiased (and consistent) estimator of the average casual effect?