Alex Edmans
953,305 views • 17:47

Belle Gibson was a happy young Australian. She lived in Perth, and she loved skateboarding. But in 2009, Belle learned that she had brain cancer and four months to live. Two months of chemo and radiotherapy had no effect. But Belle was determined. She'd been a fighter her whole life. From age six, she had to cook for her brother, who had autism, and her mother, who had multiple sclerosis. Her father was out of the picture. So Belle fought, with exercise, with meditation and by ditching meat for fruit and vegetables. And she made a complete recovery.

Belle's story went viral. It was tweeted, blogged about, shared and reached millions of people. It showed the benefits of shunning traditional medicine for diet and exercise. In August 2013, Belle launched a healthy eating app, The Whole Pantry, downloaded 200,000 times in the first month.

But Belle's story was a lie. Belle never had cancer. People shared her story without ever checking if it was true. This is a classic example of confirmation bias. We accept a story uncritically if it confirms what we'd like to be true. And we reject any story that contradicts it. How often do we see this in the stories that we share and we ignore? In politics, in business, in health advice.

The Oxford Dictionary's word of 2016 was "post-truth." And the recognition that we now live in a post-truth world has led to a much needed emphasis on checking the facts. But the punch line of my talk is that just checking the facts is not enough. Even if Belle's story were true, it would be just as irrelevant. Why?

Well, let's look at one of the most fundamental techniques in statistics. It's called Bayesian inference. And the very simple version is this: We care about "does the data support the theory?" Does the data increase our belief that the theory is true? But instead, we end up asking, "Is the data consistent with the theory?" But being consistent with the theory does not mean that the data supports the theory. Why? Because of a crucial but forgotten third term — the data could also be consistent with rival theories. But due to confirmation bias, we never consider the rival theories, because we're so protective of our own pet theory.

Now, let's look at this for Belle's story. Well, we care about: Does Belle's story support the theory that diet cures cancer? But instead, we end up asking, "Is Belle's story consistent with diet curing cancer?" And the answer is yes. If diet did cure cancer, we'd see stories like Belle's. But even if diet did not cure cancer, we'd still see stories like Belle's. A single story in which a patient apparently self-cured just due to being misdiagnosed in the first place. Just like, even if smoking was bad for your health, you'd still see one smoker who lived until 100.

(Laughter)

Just like, even if education was good for your income, you'd still see one multimillionaire who didn't go to university.

(Laughter)

So the biggest problem with Belle's story is not that it was false. It's that it's only one story. There might be thousands of other stories where diet alone failed, but we never hear about them.

We share the outlier cases because they are new, and therefore they are news. We never share the ordinary cases. They're too ordinary, they're what normally happens. And that's the true 99 percent that we ignore. Just like in society, you can't just listen to the one percent, the outliers, and ignore the 99 percent, the ordinary.

Because that's the second example of confirmation bias. We accept a fact as data. The biggest problem is not that we live in a post-truth world; it's that we live in a post-data world. We prefer a single story to tons of data. Now, stories are powerful, they're vivid, they bring it to life. They tell you to start every talk with a story. I did. But a single story is meaningless and misleading unless it's backed up by large-scale data. But even if we had large-scale data, that might still not be enough. Because it could still be consistent with rival theories. Let me explain.

A classic study by psychologist Peter Wason gives you a set of three numbers and asks you to think of the rule that generated them. So if you're given two, four, six, what's the rule? Well, most people would think, it's successive even numbers. How would you test it? Well, you'd propose other sets of successive even numbers: 4, 6, 8 or 12, 14, 16. And Peter would say these sets also work. But knowing that these sets also work, knowing that perhaps hundreds of sets of successive even numbers also work, tells you nothing. Because this is still consistent with rival theories. Perhaps the rule is any three even numbers. Or any three increasing numbers.

And that's the third example of confirmation bias: accepting data as evidence, even if it's consistent with rival theories. Data is just a collection of facts. Evidence is data that supports one theory and rules out others. So the best way to support your theory is actually to try to disprove it, to play devil's advocate. So test something, like 4, 12, 26. If you got a yes to that, that would disprove your theory of successive even numbers. Yet this test is powerful, because if you got a no, it would rule out "any three even numbers" and "any three increasing numbers." It would rule out the rival theories, but not rule out yours. But most people are too afraid of testing the 4, 12, 26, because they don't want to get a yes and prove their pet theory to be wrong. Confirmation bias is not only about failing to search for new data, but it's also about misinterpreting data once you receive it.

And this applies outside the lab to important, real-world problems. Indeed, Thomas Edison famously said, "I have not failed, I have found 10,000 ways that won't work." Finding out that you're wrong is the only way to find out what's right.

Say you're a university admissions director and your theory is that only students with good grades from rich families do well. So you only let in such students. And they do well. But that's also consistent with the rival theory. Perhaps all students with good grades do well, rich or poor. But you never test that theory because you never let in poor students because you don't want to be proven wrong.

So, what have we learned? A story is not fact, because it may not be true. A fact is not data, it may not be representative if it's only one data point. And data is not evidence — it may not be supportive if it's consistent with rival theories. So, what do you do? When you're at the inflection points of life, deciding on a strategy for your business, a parenting technique for your child or a regimen for your health, how do you ensure that you don't have a story but you have evidence?

Let me give you three tips. The first is to actively seek other viewpoints. Read and listen to people you flagrantly disagree with. Ninety percent of what they say may be wrong, in your view. But what if 10 percent is right? As Aristotle said, "The mark of an educated man is the ability to entertain a thought without necessarily accepting it." Surround yourself with people who challenge you, and create a culture that actively encourages dissent. Some banks suffered from groupthink, where staff were too afraid to challenge management's lending decisions, contributing to the financial crisis. In a meeting, appoint someone to be devil's advocate against your pet idea. And don't just hear another viewpoint — listen to it, as well.

As psychologist Stephen Covey said, "Listen with the intent to understand, not the intent to reply." A dissenting viewpoint is something to learn from not to argue against. Which takes us to the other forgotten terms in Bayesian inference. Because data allows you to learn, but learning is only relative to a starting point. If you started with complete certainty that your pet theory must be true, then your view won't change — regardless of what data you see.

Only if you are truly open to the possibility of being wrong can you ever learn. As Leo Tolstoy wrote, "The most difficult subjects can be explained to the most slow-witted man if he has not formed any idea of them already. But the simplest thing cannot be made clear to the most intelligent man if he is firmly persuaded that he knows already." Tip number two is "listen to experts." Now, that's perhaps the most unpopular advice that I could give you.

(Laughter)

British politician Michael Gove famously said that people in this country have had enough of experts. A recent poll showed that more people would trust their hairdresser —

(Laughter)

or the man on the street than they would leaders of businesses, the health service and even charities. So we respect a teeth-whitening formula discovered by a mom, or we listen to an actress's view on vaccination. We like people who tell it like it is, who go with their gut, and we call them authentic. But gut feel can only get you so far. Gut feel would tell you never to give water to a baby with diarrhea, because it would just flow out the other end. Expertise tells you otherwise. You'd never trust your surgery to the man on the street. You'd want an expert who spent years doing surgery and knows the best techniques. But that should apply to every major decision. Politics, business, health advice require expertise, just like surgery.

So then, why are experts so mistrusted? Well, one reason is they're seen as out of touch. A millionaire CEO couldn't possibly speak for the man on the street. But true expertise is found on evidence. And evidence stands up for the man on the street and against the elites. Because evidence forces you to prove it. Evidence prevents the elites from imposing their own view without proof.

A second reason why experts are not trusted is that different experts say different things. For every expert who claimed that leaving the EU would be bad for Britain, another expert claimed it would be good. Half of these so-called experts will be wrong. And I have to admit that most papers written by experts are wrong. Or at best, make claims that the evidence doesn't actually support. So we can't just take an expert's word for it.

In November 2016, a study on executive pay hit national headlines. Even though none of the newspapers who covered the study had even seen the study. It wasn't even out yet. They just took the author's word for it, just like with Belle. Nor does it mean that we can just handpick any study that happens to support our viewpoint — that would, again, be confirmation bias. Nor does it mean that if seven studies show A and three show B, that A must be true. What matters is the quality, and not the quantity of expertise.

So we should do two things. First, we should critically examine the credentials of the authors. Just like you'd critically examine the credentials of a potential surgeon. Are they truly experts in the matter, or do they have a vested interest? Second, we should pay particular attention to papers published in the top academic journals. Now, academics are often accused of being detached from the real world. But this detachment gives you years to spend on a study. To really nail down a result, to rule out those rival theories, and to distinguish correlation from causation. And academic journals involve peer review, where a paper is rigorously scrutinized

(Laughter)

by the world's leading minds. The better the journal, the higher the standard. The most elite journals reject 95 percent of papers.

Now, academic evidence is not everything. Real-world experience is critical, also. And peer review is not perfect, mistakes are made. But it's better to go with something checked than something unchecked. If we latch onto a study because we like the findings, without considering who it's by or whether it's even been vetted, there is a massive chance that that study is misleading. And those of us who claim to be experts should recognize the limitations of our analysis. Very rarely is it possible to prove or predict something with certainty, yet it's so tempting to make a sweeping, unqualified statement. It's easier to turn into a headline or to be tweeted in 140 characters. But even evidence may not be proof. It may not be universal, it may not apply in every setting. So don't say, "Red wine causes longer life," when the evidence is only that red wine is correlated with longer life. And only then in people who exercise as well.

Tip number three is "pause before sharing anything." The Hippocratic oath says, "First, do no harm." What we share is potentially contagious, so be very careful about what we spread. Our goal should not be to get likes or retweets. Otherwise, we only share the consensus; we don't challenge anyone's thinking. Otherwise, we only share what sounds good, regardless of whether it's evidence.

Instead, we should ask the following: If it's a story, is it true? If it's true, is it backed up by large-scale evidence? If it is, who is it by, what are their credentials? Is it published, how rigorous is the journal? And ask yourself the million-dollar question: If the same study was written by the same authors with the same credentials but found the opposite results, would you still be willing to believe it and to share it?

Treating any problem — a nation's economic problem or an individual's health problem, is difficult. So we must ensure that we have the very best evidence to guide us. Only if it's true can it be fact. Only if it's representative can it be data. Only if it's supportive can it be evidence. And only with evidence can we move from a post-truth world to a pro-truth world.

Thank you very much.

(Applause)