Data dredging, bias, or confounding.

摘要: On 4 October 2002, women who were moderate drinkers received good news: their risk of breast cancer was not raised, according to a report in the Lancet that widely covered by British media.1 The bad news smoking at an early age now implicated as factor for cancer. However, after they had enjoyed guilt-free drinks (without cigarettes) only few days, on 13 November message reversed: alcohol did increase all, but declared innocent.2 press release proclaimed “Alcohol, tobacco and cancer: definitive answer.” A reader driven complain letters page Guardian (14 2002): “So let me get this right—alcohol's no anymore, if you smoked within five years getting your periods, that's too. Oh no, couple weeks ago; smoking's okay … Do things stop being us we just forget about them bit, do think?” This is familiar story—so much so Bristol set our medical students exercise examining “health scare week” appears each Friday, generally from study reported BMJ or .w1 The widespread perception epidemiological studies generate conflicting often meaningless findingsw2 has support recent randomised controlled trials, which have failed confirm even apparently robust findings observational studies. most topical these relates hormone replacement therapy. In 1991 meta-analysis results relating use therapy coronary heart disease concluded it halved risk, evidence statistically (relative 0.50; 95%