Correlation is not (necessarily) causation
We know HPV causes cervical cancer. It’s so well-accepted we developed a cancer-preventive vaccine based on this fact. But not so long ago, a different virus was considered the top contender among potential cervical cancer causes.
“I remember when I was in medical school, somebody standing up and giving a lecture and saying the herpes virus causes cervical cancer,” said Dr. Anne McTiernan, now an epidemiologist at Fred Hutchinson Cancer Research Center.
Spinning Science
Overhyped headlines, snarled statistics lead readers astray
Read our special series of articles with tips and tools to help you better understand health risks, scientific research, statistics and clinical studies
How did the lecturer get it wrong? By assuming causation based primarily on correlation — a common misstep seen in dramatic headlines warning about the latest health risks “discovered” by scientists.
Correlation, or association, means that two things — a disease and an environmental factor, say — occur together more often than you’d expect from chance alone. But that doesn’t tell you if one causes the other to occur. Often, both in the news media and in our own perception, we see causes where there are only correlates.
The most likely culprit
In the 1950s, when epidemiologists began digging into risk factors for cervical cancer, the correlations they saw pointed to a sexually transmitted infection as the cause, and herpes, or HSV, seemed the most likely culprit.
How did researchers definitively tie HPV, or human papillomavirus, to cervical cancer while clearing HSV of wrongdoing? The story of how scientists untangled these associations, and others, helps illustrate how research moves from correlation to causation — and why it can be so tricky.
The relationship between HPV, HSV and cervical cancer perfectly illustrates the difference between an association that’s due to happenstance and one that’s due to a causal link.
Cervical cancer, a sexually transmitted disease
What if HPV had never been discovered?
Epidemiologists would still have had enough evidence to make public health recommendations, McTiernan said.
They often take a more bird’s eye view of cause-and-effect than the average person. Yes, HPV causes cancer — but because HPV is transmitted via sexual contact, sex can also be thought of as a cause of cervical cancer, just one step removed.
Without discovery of HPV, “we could still make the case that cervical cancer was a sexually transmitted disease,” McTiernan said. “While we didn’t then know about the particular vector, we could say sex with multiple partners is a cause of cervical cancer, and advise to use condoms to reduce risk.”
It’s true that HSV infection is associated with cervical cancer. But so is a higher number of sexual partners — which causes an increase in the chance of contracting an STI, any STI.
“It turns out that anything that is related to a higher number of partners is going to be elevated and look like it's correlated with cervical cancer,” said Dr. Denise Galloway, who directs the Hutch’s Pathogen-Associated Malignancies Integrated Research Center and has studied virally caused cancers for decades.
She led the Fred Hutch-University of Washington collaboration that was recognized by the American Association for Cancer Research for their contributions to the cancer-preventive HPV vaccine.
Other diseases, including most types of cancer, are caused by a much more complex interplay of factors, which can include diet, exercise level, environmental exposures and genetics. (Some of these also probably come into play in cervical cancer too — though everyone who gets cervical cancer was infected with HPV at one point, not everyone who gets HPV will also get cervical cancer.) This makes it even more difficult to untangle cause from mere correlation, let alone determine how strongly each factor contributes to disease.
“Even for professionals there are many instances where we're not really sure,” said epidemiologist Dr. Noel Weiss, who arrived at the Hutch in the 1970s to help the National Cancer Institute establish the population-based cancer registry that’s part of the Surveillance, Epidemiology and End Results (SEER) Program.
“You can't give somebody a carcinogen and see what happens. So in most cases, we're relying on what people are exposed to and looking at who gets the disease. … [Or] we look at the people who have the disease and compare them to people who don't, in terms of their exposure. Then we talk about association because we've not caused the disease to occur.”
— Fred Hutch epidemiologist Dr. Anne McTiernan
What makes it so hard?
One of the main reasons is observational data. Randomized, controlled trials — the ultimate benchmark for determining causation — are rare, expensive, often small (which can somewhat undercut their findings), and, in certain cases, unethical.
“You can't give somebody a carcinogen and see what happens,” said McTiernan, who studies how lifestyle choices can prevent new or recurrent cancer, and, as part of the 2018 Physical Activity Guidelines Committee, helped put together the U.S. Department of Health and Human Service’s physical activity recommendations to reduce disease risk.
“So in most cases, we're relying on what people are exposed to and looking at who gets the disease. … [Or] we look at the people who have the disease and compare them to people who don't, in terms of their exposure. Then we talk about association because we've not caused the disease to occur,” she said.
One problem with observational evidence is that people may not report their behavior accurately. This may be because they don’t remember (such as their exact breakfast three months ago) or may underreport caloric intake and overreport exercise levels. This underreporting falsely reduces the size of the association.
“But the question is, when do you make the jump to say the word 'causal'?” McTiernan said.
Weighing the evidence
Epidemiology as a discipline hit its stride in the 1950s and ’60s when researchers began noting that the incidence of lung cancer in smokers dwarfed that of nonsmokers, Weiss said. But even then, scientists didn’t rely on just one study to make public health recommendations.
“There was the surgeon general's report basically saying [smoking] was harmful. But then what? Well, how do we know it's harmful, right?” he said.
The need to support their arguments prompted epidemiologists to come up with a set of criteria to consider as they assess the less-than-perfect observational evidence they usually work from. These criteria help guide scientists’ thinking, but they aren’t rules etched in stone or a checklist that can be ticked off to prove causation — each has caveats, may not apply to every case, and needs to be considered in context.
“These have evolved over the years as we’ve come to realize that some things are more important than others. Some things work for some scientific questions more than others,” Weiss said.
Hutch scientists walked through some of the top criteria they use to assess whether a correlation likely arises from causation:
- Strength of the association
The smoking and lung cancer link knocks this one out of the park. According to the Centers for Disease Control and Prevention, smokers are a whopping 15–30 times more likely to get lung cancer than nonsmokers.
“The bigger the difference [between those with and without the disease], the less likely it is that some other factor that also leads to disease could somehow be distorting the results,” Weiss said.
McTiernan pointed out that many established causes of cancer don’t affect cancer risk to the same degree seen for smoking and lung cancer. Both secondhand smoke and menopausal hormone therapy, known carcinogens, raise the risk of cancer (lung and breast, respectively) by about 30%.
And Dr. Holly Harris, a Hutch epidemiologist who studies how lifestyle factors and genetics influence women’s health, cautioned that after decades of epidemiological work on many health questions, sometimes an outsized effect gives her pause.
“Honestly, if I see an odds ratio of 10,” she said, referring to a high value of the statistic that quantifies the strength of an association, “I think this study was probably really poorly designed, because maybe you see that for smoking and lung cancer, but if you're seeing something like that these days — why haven't we figured that out already?”
- Biological plausibility and experimental evidence
Scientists also use their knowledge of biological mechanisms when assessing whether an association could be causal. In addition to the overwhelming strength of the association between lung cancer and smoking, scientists also took into account lab studies showing the carcinogenic effect of cigarette smoke to support their anti-smoking arguments.
Of course, mice, the animals usually used for these types of studies, are not always the best proxies for humans. And, these types of studies don’t necessarily lend themselves well to diseases that arise from a multitude of factors, such as cardiovascular disease and many cases of cancer, McTiernan said.
“They usually look at one exposure, one variable at a time. … So you’re just really taking pieces from biological models and then applying them to what you see in humans,” she said.
Scientists also think about how the pattern of the association matches with what they know of the biology of the disease, Weiss said: “When your observations correspond perfectly to that prediction, then that further enhances your ability to make an inference.”
- Consistency and replication
It’s also important to consider a study within the context of other studies — are they all reporting much the same results?
“If the first study looks at [an association] and they find a 20% difference in risk — you want to take that with a grain of salt,” Weiss said. “Whereas if you have five studies that are well done with a 20% increase consistently found across them, that means something else.”
It can be difficult to get this bigger picture from news reports, which often focus on new, unreplicated science instead of a growing body of work that’s showing the same consistent results, Harris said. In general, the researchers recommended looking to guidelines from groups like the U.S. Department of Health and Human Services, in which scientists have made a collective assessment of the existing evidence.
- Dose response
As scientists weigh the evidence for causation, a dose response can help tip the balance. This means, for example, that a higher exposure to the factor results in a higher incidence of disease. Cigarette smoking and lung cancer showed this relationship, with the risk of lung cancer increasing as the number of cigarettes smoked increased.
- Specificity of association
Weiss argued that looking at a question too broadly can obscure more specific associations. For example, researchers will study tea drinking “not just in relation to, let's say, stomach cancer,” Weiss said, “but they'll study tea drinking in relation to all-cause mortality. Well, all-cause mortality has got all these different conditions. And it's hard to imagine the tea is affecting them all. It could be affecting one or two. Its harmful effect or beneficial effects can be watered down by all these others.”
“If a first study looks at [an association] and they find a 20% difference in risk — you want to take that with a grain of salt. Whereas if you have five studies that are well done with a 20% increase consistently found across them, that means something else.”
— Fred Hutch epidemiologist Dr. Noel Weiss
Detangling HSV from cervical cancer
Initially, herpes seemed like a great candidate for the cause of cervical cancer. It ticked many of the boxes. There was a strong, consistent association between HSV infection and cervical cancer.
“You see higher antibodies to herpes [in women with cervical cancer] than you do in [women without cervical cancer],” Galloway said.
HSV also ticked the biological plausibility box: In the lab, scientists found the virus could turn normal cells cancerous — a process known as “transformation.”
But a limit of using biological plausibility to argue for or against a potential disease cause is that these arguments are only as good as the current scientific knowledge. When HSV was the top contender for the cause of cervical cancer, HPV had yet to be discovered.
One scientist showed that papillomaviruses caused tumors in rabbits. Other researchers wondered whether there was a human version that could do the same thing. Once HPV arrived on the scene, the case against HSV started falling apart. Those cells HSV transformed? Only mouse cells, never human, and only poorly.
In contrast, “You could actually find the [HPV] viral DNA within the tumor,” Galloway noted. Two specific HPV genes, E6 and E7, are sufficient to robustly transform human cells into immortal, cancer-like cells.
“That really became the basis for saying HPV caused cervical cancer because it could cause the precancers in tissue culture,” she said.
It would have been too unethical to follow HPV-infected people to see who developed full-blown cervical cancer, but further epidemiological studies strengthened the link between HPV and risk of developing precancerous lesions. By the mid- to late-'80s, HPV had been declared the cause of cervical cancer.
Seeking causation in complicated diseases
Unlike cervical cancer, most diseases don’t have a single dominant contributing factor. So how do researchers determine whether a single factor is contributing in an outsized way?
Weiss described work he contributed to in the 1970s showing that taking estrogen without accompanying progestin raised the risk of endometrial cancer. The determination was the result of piecing together several types of evidence.
At first, an early epidemiological study showed that women taking estrogen had more overgrowth of the endometrium, or lining of the uterus, a condition known as endometrial hyperplasia. Another study showed a higher incidence of endometrial cancer in women with endometrial hyperplasia.
“There was a lot of suspicion, but nobody had actually done a study directly testing the hypothesis [that estrogen raised the risk of endometrial cancer],” Weiss said.
A gynecologist at the University of Washington, Dr. Donald Smith, did a case-control study. He interviewed patients with endometrial cancer (the cases) and other types of gynecological cancer (not an ideal group of controls, “But it was all he had,” Weiss said) about their hormone use. He worked with Dr. Ross Prentice, a Hutch biostatistician.
They saw that the risk of endometrial cancer was 4.5 times higher for women taking estrogen. Weiss took advantage of his links to national cancer registries and looked to see if there had been an increase in endometrial cancer over the five to 10 years preceding the mid-1970s, a period during which hormone use was rapidly increasing.
“Sure enough, in every place you looked, there was an increase [in endometrial cancer]. … It was particularly large in places where estrogens were commonly used, like the western U.S. So that, along with the data from these case-control studies, was really quite compelling,” Weiss said.
The potential causal link also had biological plausibility to back it up: In lab studies, estrogen causes excess growth of the endometrium. Progestin, on the other hand, causes endometrial cells to stop growing and slough off; studies of women taking estrogen and progestin together did not show an increased incidence of endometrial cancer.
An abundance of evidence
One thing all the researchers agree on: A single study is never enough.
“It's really when you get an abundance of evidence, when you think, OK, this could be causing a disease,” McTiernan said.
Sabrina Richards, a staff writer at Fred Hutchinson Cancer Center, has written about scientific research and the environment for The Scientist and OnEarth Magazine. She has a PhD in immunology from the University of Washington, an MA in journalism and an advanced certificate from the Science, Health and Environmental Reporting Program at New York University. Reach her at srichar2@fredhutch.org.
Are you interested in reprinting or republishing this story? Be our guest! We want to help connect people with the information they need. We just ask that you link back to the original article, preserve the author’s byline and refrain from making edits that alter the original context. Questions? Email us at communications@fredhutch.org
Help Us Eliminate Cancer
Every dollar counts. Please support lifesaving research today.
For the Media