Avoidable flaws in observational analyses: an application to statins and cancer