[ JESS CRAIG | HEALTH REPORTER ]
The news is full of headlines about officials drawing bold lines between two things happening at once: Americans are eating more ultra-processed foods while the number of people with obesity and chronic diseases is rising. More children are being vaccinated while autism and ADHD diagnoses are climbing. People spend more time on social media as mental health declines.
The implication is often the same: if two trends rise together, one must be causing the other. Often, these notions are based on data that show timely correlations between two events.
But correlation can be misleading.
For example, data from the Department of Agriculture and the Centers for Disease Control and Prevention show a strong correlation between national margarine consumption and the divorce rate in Maine. As margarine consumption steeply declined from 2000 to 2005, divorce rates fell. From 2005 to 2008, people across the country started to eat more margarine and lo-and-behold, more people in Maine divorced. Then from 2008 to 2009, both margarine consumption and divorce rates fell together again.
As convincing as that graph can be, margarine isn’t spreading divorce. The graph is simply showing a correlation.
[ ELIZABETH STUART, PROFESSOR, BIOSTATISTICS, JOHNS HOPKINS BLOOMBERG SCHOOL OF PUBLIC HEALTH ]
“Correlation really just means two things are related to each other, that they kind of move together.
[ CRAIG ]
Two other things that move together: Ice cream sales and drowning. Both increase seasonally in the summer, peaking around August.
[ STUART ]
It’s not that eating ice cream causes drowning necessarily, but they are correlated. They just, both happened in the summer, you know, when people are sort of, it’s hot and people are swimming more. So the distinction there is like, again, correlation is sort of just things that kind of move together in a sense. Causation is really this concept or idea that something leads to something else.
[ CRAIG ]
While some relationships reflect true biological links, others — like ice cream and drownings or margarine consumption and divorces in Maine — are coincidences.
Confusing correlation and causation is easy to do, especially in complicated, high-stakes situations. Public discourse and political agendas can easily distort the two. And science is very seldom black-and-white. Researchers report different, sometimes conflicting findings. But major health and science decisions — such as the approval of a new drug or the reversal of longstanding health guidelines — require that scientists and politicians follow careful procedures to sort out correlation from causation.
To do that, scientists run specific controlled experiments.
In the early 1970s, doctors blamed stomach ulcers on stress. But then, in the 1980s, physicians Dr. Barry Marshall and Dr. Robin Warren, noticed a certain bacteria, H. pylori consistently colonized the stomach lining of stomach ulcer patients. H. pylori is now known to spread through contaminated food or water or through direct contact with a contaminated person’s feces.
The doctors’ observation was a correlation. In 1984, Marshall, who later won the Nobel Prize, famously drank a culture of H. Pylori bacteria. A few weeks later, he developed gastritis — an inflammation of the stomach lining that can cause a burning pain or ache in the upper belly, vomiting or a loss of appetite. This proved causation, albeit in a not so ethical or controlled way. Marshall then cured himself with antibiotics. This was the first step towards proving that a bacteria caused peptic ulcers. Scientists no longer drink bacteria to prove causation, but they do turn to certain experiments.
[ Stuart ]
“So the best case scenario, and this is sort of very common in the medical and like food and drug association world, is a randomized trial where we might have some exposure of interest, some, again, maybe a new program, a new medication, and we literally enroll people in a study and randomly assign them to get it or not get it.
[ CRAIG ]
Let’s say scientists think a new medication can cure cancer.
[ Stuart ]
We might again randomly pick among a set of people, some people get it and some people don’t, and then we compare outcomes over time. And the beauty of randomization is that we sort of know because we picked people randomly, if we see a difference in outcomes, it must be due to this treatment that we gave them. It can’t be due to sort of anything else. It can’t be due to the fact that, all the people eating ice cream, it was also summer.
[ CRAIG ]
Randomized controlled trials are important, but they’re not always possible. For instance, in the early 1900s, physicians started observing higher levels of lung cancer among patients who smoked cigarettes. It would have been unethical to randomly assign people to smoke cigarettes while others abstained. Instead, they followed a group of people who were already smoking and a group who did not smoke. Ultimately, they found higher rates of lung cancer among those who smoked.
With that correlation established, scientists turned to better understanding how smoking caused lung cancer. They now know that the hundreds of carcinogens in cigarette smoke penetrate deep into the lungs damaging DNA and causing mutations that disrupt normal cell growth and division. When certain genes are damaged, cells can grow uncontrollably, causing cancer.
In many situations, scientists cannot perform the kinds of controlled experiments needed to prove causation. What do they do in that case?
[ STUART ]
There’s a number of approaches that people can use. Basically a lot of the time, in the absence of randomization, people try to sort of at least find a comparison group, like a group of people who didn’t get the new medication but who look as similar as possible to those who did. And so the idea is to, again, sort of in some sense try to have data that looks like it could have been randomized. It wasn’t, and we have to acknowledge that. But where we at least say, well, maybe we want to compare what is the effect of eating ultra-processed foods, sort of a topic of interest right now.
We know that people who eat a lot of ultra processed foods are quite different from those who don’t in many ways. They live in different environments. They might have different exercise strategies. They might have other health conditions, sort of preexisting. So what a statistician might do or epidemiologist might do is sort of say, well, can we sort of look at our data and find the people who kind of look like the ones who eat a lot of ultra processed foods, but who didn’t?
Step one. I would say, is just sort of like, can we find a group of people who look like they could have gotten this exposure and look like the people who did? But they didn’t for some reason — you know, maybe they they live somewhere where they have better access to a grocery store or something.
[ CRAIG ]
Whether it’s ultra processed foods or autism risk, new studies are constantly published around the world. Some studies agree, some conflict. At what point do scientists and policymakers determine we know enough to form a conclusion?
[ STUART ]
There is sort of a spectrum here. And certainly for a new drug approval by the Food and Drug Administration, there is a very formal process. And they require randomized trials and sort of a very thorough process to sort of establish that causal link and in some sense, like statistical proof that some new medication or treatment really does improve outcomes.
In other scenarios where, it’s maybe more recommendations like the ultra processed foods. And honestly, also those are often murkier. There’s just sort of, there’s so many factors that relate to health broadly. I think that’s where it really is about a body of evidence.
[ CRAIG ]
In the early 2000s, nutritional epidemiologist Carlos Augusto Monteiro noticed a shift in Brazil: As traditional diets of rice and beans were replaced by soda and cookies, obesity rates surged.
He hypothesized that eating ultra processed foods might be causing the rise of obesity and chronic diseases. But Monteiro had not yet proven causation; he had only established correlation.
In the two decades since Monteiro’s work, many studies have found links between consumption of ultra processed foods and a range of diseases from obesity, diabetes and high blood pressure to depression, anxiety, inflammatory bowel disease, autoimmune disorders and even cancer.
Nutrition science is challenging because health is shaped by intertwined factors, including genetics, income, environment and lifelong eating habits, making controlled experiments difficult.
Researchers rely largely on observational studies that track dietary patterns, including ultra-processed food consumption. Then, scientists look for consistent patterns across many studies. Today, there is a large body of evidence showing strong and repeated correlations between higher consumption of ultra-processed foods and poor health outcomes.
Researchers have not yet proven how exactly consuming these foods causes these various conditions. Some scientists hypothesize that ultra-processed foods may harm health because they are engineered to be highly addictive, easy to overeat and low in filling nutrients like protein and fiber, leading to excess intake of salt, sugar and saturated fat. Other researchers have suggested that ultra-processed foods lead to poor health because they alter the gut microbiome or impact hormone levels.
[ STUART ]
So much of science is about building sort of bodies of research over time. Like one individual study is rarely sort of the definitive sometimes, but not, you know, pretty rare, especially again, for these sort of nuanced health questions. And so it’s sort of, how do we triangulate and sort of look at these different lenses and kind of come up with an overall conclusion? So I think kind of keeping that lens of looking at the details, moving past the headline to look at the details and see how it fits into a broader literature is really, important.
[ Craig ]
Late last year, President Donald Trump said that pregnant women who take acetaminophen, commonly known as Tylenol in the U.S., may increase their child’s risk of developing autism.
“Any association between acetaminophen and autism is based on limited, conflicting and inconsistent science and is premature,” Alycia Halladay, the chief science officer at the Autism Science Foundation, said in a press release following Trump’s remarks.
[ Craig ]
Last year, U.S. researchers reviewed 46 studies examining links between acetaminophen use in pregnancy and autism or ADHD. Of those, 27 reported an increased risk, nine found no association, and four suggested a possible protective effect. Several studies also observed a dose-response relationship, with higher use tied to higher risk.
In mid-January, European researchers published a separate review of 43 studies and pooled data from 17 of the highest quality papers.
They found statistically insignificant increases in the odds of autism and intellectual disability, about 3% and 11% higher odds, respectively, but a decrease of 3% in the odds of ADHD .
The researchers also examined sibling-comparison studies that help control for factors — such as shared genetics and family environment. Those analyses found no increased risk of autism, ADHD or intellectual disability associated with acetaminophen use during pregnancy.
Hugh Taylor, chief of Obstetrics and Gynecology at Yale New Haven Hospital, said the study was well done.
[ Taylor ]
“We can’t say that any one is the definitive study but it’s certainly reassuring. Certainly if there was a huge signal that this was really a common consequence of using acetaminophen, Tylenol, it would have shown up in all the studies.”
[ Craig ]
All of these prior studies are observational studies that describe a correlation; none of them describe causation.
Thinking back to the correlation between shark attacks and ice cream sales: There was a third factor at play there that offered an explanation: Temperatures. In the summer, as temperatures rise, more Americans go on beach vacations and swim in the water; they also eat more ice cream.
Similarly, other factors might explain sporadic correlations between Tylenol use and autism. Genetics could be one. Fever may be another. Fever during pregnancy, especially the first trimester, is linked to an increased risk of autism. Because pregnant women might take Tylenol to treat fever, the increased risk of autism and ADHD could be due to the fever and not the medicine.
While the correlation between Tylenol and autism has been well researched, there has yet to be a single study that proves how Tylenol causes neurodevelopmental disorders. This might be difficult, if not impossible, because it would be unethical to experiment on pregnant women and their fetuses.
Science and health research is far from black and white. Studies often report conflicting findings, and because of ethical limitations, proving true causation is not always possible. As science and health headlines continue to shape public debate, the most important question news consumers can ask is not whether two trends appear to rise together, but how researchers studied the connection — and what remains uncertain. Single studies, dramatic graphs or political statements rarely tell the full story. Strong scientific conclusions are built slowly, through multiple lines of evidence and careful methods. In an era of fast-moving information and high-stakes decisions, understanding the difference between correlation and causation is not just a scientific skill — it is an essential tool for making sense of the world.
