Hypothesis-generating procedures and unmasking novel associations in large observational studies: are we doing harm while doing good?