In the April 11 2015 issue of The Lancet, Richard Horton asks what we can do about the fact that “A lot of what is published is incorrect.” As a solution, Horton suggests a p-value criterion for "significance" of 10 to the -7. But in genetic epidemiology they already use something around 10 to the -7 and the results in that field are even less reproducible than the rest of epidemiology! No, clearly we need to go in the other direction and get rid of null hypothesis significance testing altogether. Rather, as eluded to earlier in the essay, we need to change the incentive structure so that one doesn't get rewarded for being published, but instead one gets rewarded for being right. Suppose that if your result were shown to be wrong, you had to give back the grant money. Then you'd see people checking for errors! In engineering, if your bridge falls down or your airplane crashes, you are gonna get fired. But do you see Walt Willett getting fired for the epidemiologic equivalent? Changing the p-value criterion before you can declare something to be "significant" is not going to help this problem.
|
|
RSS Feed