On the impossibility of inferring causation from association without background knowledge