When are probabilistic explanations possible?