Scientific data-mining is contaminating serious study


February 12, 2018   5 mins

You may not have heard of the most surprising scientific result of all time. It was published in the journal Psychological Science1, and it found that listening to ‘When I’m 64’ by the Beatles literally made you younger by nearly 17 months.

Not “made you feel younger”, or “made your brain younger”: the study demonstrated that listening to certain songs made you younger. And these surprising findings met the standard of “statistical significance” which most scientific journals demand.

In fact, the three authors of the study had set out to deliberately make a point. They wanted to show that, by using routine statistical methods2, you could find apparently solid results that were not merely unlikely but patently absurd.

Science isn’t broken. But it is profoundly flawed. And the wrong results can have serious implications
-
And they were doing so to show that science is – well. Not broken. But profoundly flawed. And the wrong results can have serious implications; we build our lives on the results of these tests: from the efficacy of the drugs we take to the policy decisions our governments make.

The problem is that the standard of “statistical significance” is too easily subverted. And to prove it, the three authors used tricks to find relevant-looking patterns out of pure, random fluke results. We might call their techniques “cheating”, but they are routinely used in almost all walks of science.

Say you’re running an experiment into whether or not a coin is biased towards heads. You flip the coin three times; it comes up heads three times. Does that prove your hypothesis? No: sometimes you get three heads in a row just by chance.

That’d be called a “false positive”. Your chance of getting three heads in a row on a fair coin is 1 in 8, or as a scientist would write it, p=0.125.

In 1925, the great statistician RA Fisher arbitrarily decided3 that “significance” should be defined as p=0.05, or a 1 in 20 chance of a false positive. That definition is the standard that most research is still held to. Your three heads in a row wouldn’t do it, but five would: that’d be a 1 in 32 chance, or p<0.03.

The “When I’m 64” study had p=0.04, and thus met the threshold.

To prove the thesis, they did a simple test: they took 20 subjects, randomly divided them into two groups, and then played one group the Beatles and one group a control song, “Kalimba” by Mr Scruff. In theory, randomly dividing the two group means that they should be, on average, the same; the only difference should be the intervention, which song they listened to. And that means that, in theory, if you find a significant difference between the two groups, it has to be caused by that intervention. It’s a randomised controlled trial, just the same as is used to test cancer drugs.

And the study found that the “When I’m 64” group was, on average, 1.4 years younger than the control group, and found that this difference was statistically significant. So they had proved – to the standard most journals demand – that “When I’m 64” causes you to get younger.

A deliberately rigged study proved that listening to the Beatles could actually make you younger. The research was intended to draw attention to bad practice known as HARKing.

Essentially, they gave themselves lots of chances to get that 1 in 20 fluke, then hid all their failed attempts at getting five heads in a row.

For instance, as well as making their 20 participants listen to the Beatles and Mr Scruff, they also asked them to listen to “Hot Potato” by the Wiggles. And they also asked 11 other questions, including how old they felt, their political orientation, whether they referred to the past as “the good old days”, and so on.

This allowed them to look at lots of different possibilities which might demonstrate their conclusion. Did the people who listened to “Hot Potato” happen, by chance, to be more politically conservative? Did people who listened to the Beatles say “the good old days” more often? No? Then the scientists didn’t report on that. There were dozens of ways to analyse the data, easily enough to get a 1-in-20 result just by fluke.

This behaviour is called “hypothesising after results are known”, or HARKing4. It means waiting until you’ve got your data, then sifting through it in imaginative ways until you find something that looks like a result. It gives you a much higher chance of (falsely) reaching a “significant” finding: using tricks like this, collectively known as “p-hacking”, Simmons, Nelson and Simonsohn say that the chances of a false positive can rise above 60%.

And this happens all the time, in real papers. A study by the Oxford University Centre for Evidence-Based Medicine5 found that the top five medical journals in the world regularly publish studies which change what they’re measuring after the trial begins, a practice related to HARKing.

It isn’t that scientists are frauds; it’s that the incentive system in science is flawed
-
It isn’t that scientists are frauds; it’s that the incentive system in science is flawed. As a community, we want science to build up a reliable body of knowledge. But as an individual scientist, you are incentivised to publish lots of papers showing exciting, novel results. And when the individual’s incentives don’t align with wider society’s, the individual incentives will win out almost every time. The trouble is, this wastes resources, skews results and creates fake narratives.

Chris Chambers6, a professor of psychology at Cardiff, tweeted recently that he’d had a paper rejected for having conclusions that were “not necessarily novel”.

“We could have been more novel,” he wrote. “We could have HARKed this paper up the wazoo or p-hacked it. We didn’t.”

The trouble is a deep-seated flaw in the process of “peer review”. “They judge the work,” Chambers said, “not just on how important the question is and how well the study is designed, but also whether the results are novel and exciting.”

This leads to a fundamental problem.When you select what to publish based on results, the knowledge base only reflects some of the truth. According to Chambers: “Positive studies get into journals but equally high-quality studies that don’t show results are basically censored.”

We could have been more novel We could have HARKed this paper up the wazoo or p-hacked it. We didn’t.
-
This has major implications. Take big drug studies. If journals only publish the papers  that show a big effect and ignore ones that don’t, then the scientific record will overstate how effective the drug is.

This sort of “publication bias” is well-documented 7. “Publication bias is catastrophic for medicine,” says Chambers. It also has profound implications for government policy: in the US, the federal government set up a $22 million (£16 million) programme called “Smarter Lunchrooms” on the basis of studies that contained flawed or missing data, several of which have since been retracted or corrected.

Luckily, there are people trying to do things about it, such as ensuring the preregistration of hypotheses, to prevent HARKing. Chambers is the founder of an initiative called Registered Reports: journals who sign up for it review a study on the basis of the study methods alone, and agree to publish regardless of the results.

This would take publication bias out of the equation and equally removes the incentives for researchers to data mine for significant findings, because the novelty or otherwise of the results is unimportant.

Such initiatives won’t make science perfect. But they almost certainly would improve things by aligning scientists’ incentives with those of society. They would stop fruitless research, prevent ineffective policy incentives and encourage straight-forward scientific innovation. Our lives would be made measurably better.

FOOTNOTES
  1.  ‘False-Positive Psychology: Undisclosed Flexibility in Data Collection and Analysis Allows Presenting Anything as Significant’, 2011. Joseph P. Simmons, Leif D. Nelson, and Uri Simonsohn
  2.  Scott Alexander, ‘Two dark side statistics papers‘, Slate Star Codex, 2 January 2014
  3.  Statistical Methods for Research Workers, RA Fisher, 1925
  4.  Norbert L. Kerr, ‘HARKing: Hypothesizing after the results are known‘, Personality and Social Psychology Review, 1 August, 1998
  5.  The Centre for Evidence Based Medicine Outcome Monitoring Project (COMPare), Ben Goldacre, Henry Drysdale, Carl Heneghan, 18 May, 2016
  6.  Chris Chambers, ‘Psychology’s registration revolution‘, The Guardian
  7.  Ikhlaaq Ahmed, Alexander J Sutton, Richard D Riley, ‘Assessment of publication bias, selection bias, and unavailable data in meta-analyses using individual participant data: a database survey‘, BMJ 3 January 2012.

Tom Chivers is a science writer. His second book, How to Read Numbers, is out now.

TomChivers