In Science, Popularity Means Inaccuracy

Who's more likely to start digging prematurely: one guy with a metal-detector looking for an old nail, or a field full of people with metal-detectors searching for buried treasure?

In any area of science, there will be some things which are more popular than others - maybe a certain gene, a protein, or a part of the brain. It's only natural and proper that some things get of lot of attention if they seem to be scientifically important. But Thomas Pfeiffer and Robert Hoffmann warn in a PLoS One paper that popularity can lead to inaccuracy - Large-Scale Assessment of the Effect of Popularity on the Reliability of Research.

They note two reasons for this. Firstly, popular topics tend to attract interest and money. This means that scientists have much to gain by publishing "positive results" as this allows them to get in on the action -

In highly competitive fields there might be stronger incentives to “manufacture” positive results by, for example, modifying data or statistical tests until formal statistical significance is obtained. This leads to inflated error rates for individual findings... We refer to this mechanism as “inflated error effect”.

Secondly, in fields where there is a lot of research being done, the chance that someone will, just by chance, come up with a positive finding increases -

The second effect results from multiple independent testing of the same hypotheses by competing research groups. The more often a hypothesis is tested, the more likely a positive result is obtained and published even if the hypothesis is false. ... We refer to this mechanism as “multiple testing effect”.

But does this happen in real life? The authors say yes, based on a review of research into protein-protein interactions in yeast. (Happily, you don't need to be a yeast expert to follow the argument.)

There are two ways of trying to find out whether two proteins interact with each other inside cells. You could do a small-scale experiment specifically looking for one particular interaction: say, Protein B with Protein X. Or you can do "high-throughput" screening of lots of proteins to see which ones interact: Does Protein A interact with B, C, D, E... Does Protein B interact with A, C, D, E... etc.

There have been tens of thousands of small-scale experiments into yeast proteins, and more recently, a few high-throughput studies. The authors looked at the small-scale studies and found that the more popular a certain protein was, the less likely it was that reported interactions involving it would be confirmed by high-throughput experiments.

The second and the third of the above graphs shows the effect. Increasing popularity leads to a falling % of confirmed results. The first graph shows that interactions which were replicated by lots of small-scale experiments tended to be confirmed, which is what you'd expect.

Pfeiffer and Hoffmann note that high-throughput studies have issues of their own, so using them as a yardstick to judge the truth of other results is a little problematic. However, they say that the overall trend remains valid.

This is an interesting paper which provides some welcome empirical support to the theoretical argument that popularity could lead to unreliability. Unfortunately, the problem is by no means confined to yeast. Any area of science in which researchers engage in a search for publishable "positive results" is vulnerable to the dangers of publication bias, data cherry-picking, and so forth. Even obscure topics are vulnerable but when researchers are falling over themselves to jump on the latest scientific bandwagon, the problems multiply exponentially.

A recent example may be the "depression gene", 5HTTLPR. Since a landmark paper in 2003 linked it to clinical depression, there has been an explosion of research into this genetic variant. Literally hundreds of papers appeared - it is by far the most studied gene in psychiatric genetics. But a lot of this research came from scientists with little experience or interest in genes. It's easy and cheap to collect a DNA sample and genotype it. People started routinely looking at 5HTTLPR whenever they did any research on depression - or anything related.

But wait - a recent meta-analysis reported that the gene is not in fact linked to depression at all. If that's true (it could well be), how did so many hundreds of papers appear which did find an effect? Pfeiffer and Hoffmann's paper provides a convincing explanation.

Link - Orac also blogged this paper and put a characteristic CAM angle on it.

Pfeiffer, T., & Hoffmann, R. (2009). Large-Scale Assessment of the Effect of Popularity on the Reliability of Research PLoS ONE, 4 (6) DOI: 10.1371/journal.pone.0005996