The oft-repeated mantra that correlation does not imply causation is not always effective. The media, the general public, and even scientists themselves, are often guilty of seeing cause-and-effect when there is only an association.
Now, two researchers think they might have a solution: when presenting correlational results, simply tilt the graph by 45 degrees (See 'Correlation or causation') .
The new graph, called a diamond plot, is the brainchild of evolutionary biologist Carl Bergstrom and information scientist Jevin West, at the University of Washington, Seattle. Together, the pair run a course on the way numbers are misused, titled Calling Bullshit.
They described the diamond plot in a paper posted to the preprint server arXiv last week1. They propose rotating scatterplots anti-clockwise by 45 degrees, so that they become diamond-shaped with symmetrical axes. They hope that the unusual orientation of the graph will have a “jarring” effect, forcing the viewer to explicitly consider — and hopefully dismiss — the issue of causality.
The researchers were inspired in part by the frequent erroneous claims of causality that they see in the popular press. “You could go through the last 2 weeks and you could find any number of examples of cases,” says Bergstrom. “Red wine causes this, chocolate causes that.”
But Bergstrom adds that it’s not just the media that’s at fault: scientists themselves are liable to make the same mistakes. In experimental methodology, scientists are taught that the independent variable — the one controlled by the researcher — belongs on the x-axis of a graph, while the dependent variable — the one being measured — belongs on the y-axis.
When it comes to observational data, scientists might still apply these rules instinctively, and mistakenly suppose that the variable on the x-axis is influencing the other. The symmetrical nature of the diamond plot means that neither axis takes precedence, hopefully reducing the risk of this misattribution.
“I thought it was a nice idea, and I understood the logic behind what they were doing,” says Catey Bunce, a medical statistician at King’s College London. “Do I think the use of those plots will lead to less people making this confusion about correlation and causation? No, I don’t.”
That's because, Bunce says, there are other factors that can lead scientists to assume causation, and these could undermine any positive impact from diamond plots.
“I think that publicity has something to do with that,” she says. “If you write a paper that says there’s causation, it’s quite a bold statement to make, and that will attract attention.”
She adds that diamond plots, like any new statistical method, might have hidden dangers that become clear only gradually, when people start misinterpreting them.
Bergstrom recognises that there is a way to go before advocating that diamond plots become the norm. The team now plans to investigate whether people can actually read the diamond plots — and, if so, whether these graphs do reduce misinterpretations. “Either this helps or it doesn’t,” he says. “Right now it’s a proposal, and the next step, of course, is to do the proper user-testing.”
Bergstrom, C. T. & West, J. D. Preprint at https://arxiv.org/abs/1809.09328 (2018).