Causal Inferences in Natural Experiments: Some Complications in Matching Designs

In ex post facto designs, where randomization is impossible, one often selects individuals in such a way that dependent variables are inadvertently manipulated, with the result that inferences based on these designs can be misleading. Two specific cases are discussed: (1) where rare or deviant individuals are matched with a normal control group, and (2) where correlated independent variables are made unrelated in the sample in order to isolate their separate effects.