The Paradox of Policy-Relevant RCTs and Natural Experiments