Predicted outcomes as virtual controls: five conditions for valid treatment effect estimates
Researchers have formally defined the five conditions under which comparing observed outcomes with model-predicted counterfactuals yields valid treatment effect estimates.
