Thursday July 7, 2005
Â· There I was two weeks ago, making sarcastic jokes about how Bad Science was just a cover for the [coughs] popular statistics lecture series I secretly yearned to give, and now I’m about to try to explain funnel plots to you, in a national newspaper, and without any diagrams. Here we go. Publication bias makes the published scientific evidence on something look skewed. Let’s say lots of people do studies on something. Let’s say … we have a theory that rubbing semen into your face prevents skin cancer. Just go with it.
Â· So lots of groups run studies where people cheerfully rub semen into their faces, and then they measure the rates of skin cancer years later. The studies where they find that semen protects you get written up and published. Plenty more studies find that semen has no effect: but the researchers can’t be bothered to write up these unintriguing, negative findings, or journals can’t be bothered to publish them, so nobody hears about them. If all the studies were published, the evidence would show that semen has no effect, but because of publication bias, it looks as if it works. Everyone starts rubbing semen on their face, and for none of the usual reasons.
Â· You can get a hint of whether there is a publication bias in the evidence on something by doing a funnel plot. I’m not really going to explain funnel plots to you here, although they are jolly interesting and you can read about them at tinyurl.com/75fa7. The basic principle is this: all other things being equal, trials that show a smaller error in their result, perhaps the larger ones, will tend to cluster more around the true answer of how useful something is. The less accurate studies, perhaps the smaller ones, should scatter about this point randomly, some overestimating efficacy, some underestimating it. On a funnel plot, you can see whether there seem to be more positive trials than you might expect: if there are, then you have a non-random distribution, and publication bias might be the explanation.
Â· There are circumstances in which you don’t need to do a funnel plot. A study in 1998 found that there had never been a single trial published in China that found a treatment to be ineffective. Perhaps that’s another “tradition” in traditional Chinese medicine. And the BMJ this month published a paper called A Systematic Review of Publication Bias in Studies on Publication Bias, with a right lovely funnel plot. And that’s what I call a punchline.