Author: Andrew Gelman

Concern has been growing in the past decade about published scientific claims that other laboratories can’t successfully replicate. Some of these studies are pretty silly – for example, the claim that women’s political preferences change by 20 percentage points depending on the time of the month. Others were potentially useful but didn’t work out, like the one which says that holding your body in a “power pose” gives you a hormonal boost.

Related:Study delivers bleak verdict on validity of psychology experiment results

Then there are claims which may have policy relevance, such as the study that says early childhood interventions could increase young adults’ earnings by 40%. The claim came from a longitudinal study which would require at least 20 years to replicate, but, on the basis of statistics alone, we have good reason to be sceptical about the findings.

This replication problem has become a crisis in the sense that researchers, ordinary citizens and policymakers no longer know what or whom to trust. Even the most prestigious scientific journals are publishing papers that fail to replicate and which, in retrospect, are simply ridiculous.

One notorious example is a 2014 paper from the Proceedings of the National Academy of Sciences (Pnas), comparing the damage done by hurricanes with male or female names. The research was based on historical data and so could not be replicated, but featured the same sort of statistical errors that commonly appear in any work that fails the replication test.

And it’s not just journals that get sucked in. Some of our most trusted explainers and interpreters of science have been fooled by work with fatal statistical flaws. Science writer Malcolm Gladwell fell for a mathematician’s claim to be able to predict divorces with 94% accuracy and the Freakonomics team fell for the erroneous claim that beautiful parents are more likely than ordinary-looking parents to have female babies.

Statistical errors are unfortunate but unavoidable. Science is open to all, and we wouldn’t want strict gatekeeping even if it were possible. Speculative (even completely misguided) work can still indirectly advance scientific understanding. The problems come when entire fields are so shaky that outsiders – and even insiders – don’t know what to believe. This is the replication crisis, and we need to do something about it.

The problems come when entire fields are so shaky that outsiders, even insiders, don’t know what to believe

For starters, researchers need to stop making excuses and address attitudes that are getting in the way of progress. By progress, I mean moving towards a future where there are clearer links between research designs, data, analyses, criticisms and replications. The goal is not the elimination of errors, but a system with better feedback, so that dubious claims can be disputed and discussed at the point of publication, not years later when the findings have been used in news articles, TED talks, radio features and beyond.

So what’s getting in the way? Sunk cost fallacy – the error of throwing good money (or, in this case, scientific resources) after bad – certainly plays a role.

An example of this can be found in a recent New York Times op-ed by psychologist Jay Van Bavel, entitled Why Do So Many Studies Fail to Replicate?. Bavel doesn’t dodge the bad news that only 39 percent of the 100 psychological studies used had been successfully replicated – but he moves quickly to the position that the studies failed to replicate because it was difficult to recreate the exact conditions of the original.

Context certainly matters, but we should also be aware that a lot of published work is just noise. It’s always worth considering the possibility that a published finding was real and that it failed to replicate because of changing conditions, but that should not be the default assumption.

It’s natural to want to spare the feelings and reputations of hardworking researchers and it’s horrible to think that there could be hundreds of papers, published in leading journals, that are nothing but dead ends. I can see the appeal in trying to preserve some value in this mountain of published work. A paper can be seriously flawed and fail to replicate but still contain valuable insight. But our starting point has to be that any given finding can be spurious.

Related:Academics: you are going to fail, so learn how to do it better

Replications are often controlled, meaning that the researchers have chosen their data selection and analysis rules ahead of time. But published findings are almost always uncontrolled, meaning that researchers have degrees of freedom to come up with statistically significant findings. When a well-publicised study fails to replicate, this is typically consistent with a model in which the first study was merely capitalising on chance.

So how can we do better? As scientists, we have to recognise sunk cost fallacy. We need to be willing to cut our losses and accept when a research programme has not advanced, rather than grasping to explain variations that can easily be understood as mere chance.

Researchers should, of course, feel free to explore speculative routes. But we must also accept that failure is an option.

La imagen destacada es tomada de: http://www.theguardian.com/higher-education-network/2016/jun/08/scientists-arent-superheroes-failure-is-a-valid-result#img-1

Professor Gelman is giving the keynote lecture at the ESRC Research Methods Festival

Join the higher education network for more comment, analysis and job opportunities, direct to your inbox. Follow us on Twitter @gdnhighered. And if you have an idea for a story, please read ourguidelines and email your pitch to us athighereducationnetwork@theguardian.com