"Correlation Doesn't Imply Causation" - You Know the Phrase, but Do You Actually Get It?

A deeper look at why the gap between research and headlines keeps costing us.

Almost everyone who's ever been in a science-adjacent conversation has heard it: "correlation doesn't imply causation." It gets dropped like a trump card. Someone cites a study, someone else says the phrase, everyone nods, and the conversation moves on. But here's the thing; knowing the phrase and actually understanding what it means in practice are completely different things. And the ways this misunderstanding plays out in health media? Some of them are genuinely serious.

How a careful finding becomes a scary headline

The most common version of this problem goes something like this: a big cohort study finds that people who eat more red meat have higher rates of colorectal cancer. By the time that result travels from the journal abstract to the news article to the tweet, "associated with increased risk" has quietly turned into "causes cancer."

The epidemiologists who ran the study knew what they were measuring, an association. They tried to account for confounders like age, smoking, BMI, and exercise levels. But they also knew that no amount of statistical adjustment fully rules out the possibility that something else is driving the pattern. The journalists writing about it, and the people sharing it, rarely hold onto that uncertainty.

The sneaky problem you've probably never heard of: residual confounding

Okay, "media oversimplifies science", that's not exactly a hot take. But there's a deeper issue that even careful readers tend to miss, and it's called residual confounding.

Even in really well-designed observational studies with loads of statistical adjustment, there are always confounders that weren't measured, or couldn't be. Health behaviours cluster together in frustrating ways. People who exercise regularly also tend to sleep better, eat differently, have lower stress, and have better access to healthcare. When you find that regular exercisers have lower rates of some disease, you've adjusted for the things you thought to include. But the things you didn't think to include are still quietly skewing your results.

This is why so many nutritional epidemiology findings have failed to replicate, or have completely flipped, when finally tested in controlled trials.

The dietary fat saga and other cautionary tales

The history of nutrition research is honestly a wild ride. For decades, associations between dietary fat and cardiovascular disease looked pretty robust across massive population datasets. Then controlled trials that actually changed people's diets started rolling in, and the picture got a lot messier. Hormone replacement therapy is another classic example: observational studies suggested it protected against heart disease, then the Women's Health Initiative RCT pointed in the opposite direction, and the field spent years trying to work out why.

These aren't failures of science, they're science doing what it's supposed to do, correcting itself. But they're a pretty good illustration of why you shouldn't treat an observational finding like settled truth.

The flip side: "it's just a correlation" can be just as dangerous

Here's something that doesn't get talked about enough; dismissing associations can be just as harmful as overclaiming them. Tobacco companies spent decades hiding behind "it's just a correlation." The link between smoking and lung cancer showed up in observational data long before the full biological mechanism was understood, and that phrase was used deliberately and cynically to delay public health action.

So the lesson isn't to reflexively distrust all associations. It's to actually understand what an association does and doesn't tell you, and to apply that consistently, not just when it's convenient.

The tool that should be everywhere but isn't: the Bradford Hill criteria

If there's one framework that deserves way more airtime outside of epidemiology circles, it's the Bradford Hill criteria. Developed in the 1960s, they're a set of questions you can ask to evaluate whether an observed association is likely to actually be causal: things like the strength of the association, whether it shows up consistently across different populations, whether the biology actually makes sense, whether more exposure leads to more effect, and whether the cause genuinely comes before the outcome. No single criterion is a slam dunk on its own, but taken together they give you a structured way to reason about causality, rather than just defaulting to either "it's proven" or "it's just a correlation."

Why health media keeps getting this wrong

The incentive structure doesn't reward nuance. A headline like "new study suggests coffee might be associated with slightly lower all-cause mortality in some populations after adjustment for several confounders" isn't getting clicked. "Coffee drinkers live longer" absolutely is. So the translation layer between research and public understanding keeps producing the same distortions, and people keep making health decisions based on a fundamentally broken picture of the evidence.

So what's the actual antidote?

It's not cynicism about all health research, that's just as unhelpful as naive belief. The real goal is developing a working intuition for study design. Knowing the difference between what an observational study can and can't tell you. Understanding what an RCT controls for that an observational study doesn't. Recognising what even a well-run trial can miss when its sample is narrow or its follow-up period is short.

That's a higher bar than most health coverage sets for its readers. But it's the actual bar you need to clear if you want to think clearly about this stuff, and that seems worth working toward.