Research Literacy
Correlation vs. causation in nutrition science
"People who eat X live longer" and "X makes you live longer" are different claims. Most nutrition headlines quietly swap one for the other.
The classic illustration
Ice cream sales and drowning deaths rise and fall together across the year. Nobody thinks ice cream causes drowning — both are driven by a third factor, hot weather, that increases swimming and ice cream sales at the same time. That third factor is called a confounding variable: something linked to both the exposure and the outcome that creates an association between them without either one causing the other.
Nutrition research runs into this constantly, because unlike a drug trial, you usually can't randomly assign people to eat a certain way for 20 years and just watch what happens. Most of what we know about long-term diet and disease comes from observational studies — watching what people already choose to eat and tracking outcomes — and observational data is exactly where confounders hide.
Healthy-user bias: nutrition's most common confounder
People who take daily multivitamins, or who eat a lot of a specific "health food," also tend to exercise more, smoke less, see a doctor more regularly, and have higher incomes on average than people who don't. When a study finds that multivitamin users have lower rates of some disease, it's genuinely difficult to know how much of that gap is the vitamin and how much is everything else about being a person who proactively buys and takes a daily vitamin.
Reverse causation
Sometimes the arrow points the other way entirely. People in the early, undiagnosed stages of an illness often lose weight or change their eating habits because they're getting sick — not the other way around. A study that finds "low body weight is associated with higher mortality" can be partly capturing this: illness caused the weight change, not the reverse.
Why randomized trials fix this — and why nutrition often can't use them
A randomized controlled trial (RCT) solves the confounding problem by design: if you flip a coin to decide who gets the intervention, then on average the two groups start out identical on everything else — smoking rates, income, exercise habits, genetics. Any difference in outcomes is much more confidently attributable to the intervention itself.
The catch is that most interesting nutrition questions are hard or unethical to test this way. You can randomize people to take a pill for 6 weeks. You generally can't randomize people to eat a high-processed-food diet for 20 years and see who develops heart disease. That's why so much nutrition evidence stays observational — and why a single observational study, no matter how large, deserves more caution than its headline usually gets.
See how to read a nutrition study for the broader checklist this fits into, and how industry funding shapes nutrition research for a real historical case where this exact confusion was exploited deliberately.
Frequently asked questions
What is a confounding variable?
A confounding variable is a third factor linked to both the exposure and the outcome being studied, which can create an association between them even though neither one causes the other — like hot weather driving both ice cream sales and drowning rates.
What is healthy-user bias?
Healthy-user bias occurs when people who adopt one healthy behavior, like taking a daily vitamin, also tend to have other healthy habits — more exercise, less smoking, more regular healthcare — making it hard to isolate the effect of the one behavior being studied.
Why can't more nutrition questions be tested with randomized trials?
Randomizing people to eat a specific way for years or decades is often impractical, expensive, or unethical, so most evidence linking long-term diet to disease comes from observational studies, which are more vulnerable to confounding and reverse causation.