‘Data snooping’ and the significance level in multiple testing