My favorite illustration of the problems with propensity-score matching comes from Facebook Their researchers had to use a massive set of 3,719 covariates to come close to matching the estimate produced by an experiment they ran 'Yeah, PSM works, just get 3,719 covariates, bro'