Observable effectiveness can be demonstrated statistically. The only option I see is if it's unobservably ineffective, which can't be ruled out experimentally.