Watching, not intervening

In an observational study nobody is given anything. Researchers watch what people already do and track what happens. Two common shapes:

  • Cohort study. Follow a large group forward for years, noting who takes what and who develops what. (Think of the big diet-and-health surveys.)
  • Case-control study. Start from an outcome — people who got a disease vs people who didn't — and look backward at what differed.

These can involve tens of thousands of people over decades, which makes them sound authoritative. But no coin was flipped, and that missing coin is everything.

The confounder: a hidden third cause

A confounder is a lurking factor that influences both the thing you're studying and the outcome — creating an apparent link that isn't cause and effect. The classic supplement version:

A confounder causing both supplement use and better health, faking a direct link THE CONFOUNDER (Z) Sunlight, activity, income, better health habits Takes vitamin D supplement (X) Lives longer, healthier (Y) causes causes looks causal?
The link you see (dashed) vs the cause (solid). People who supplement also tend to be wealthier, more active and more sun-exposed — and those things drive the better health. The study sees "supplement users live longer" and the headline drops the caveats. This exact pattern is called healthy-user bias.

Two more traps in the same family

  • Reverse causation. Maybe the outcome causes the behaviour, not the other way round. "People who use more painkillers have more pain" — obviously the pain came first. With supplements: sick people may start (or stop) taking things because they're unwell.
  • Selection bias. Who ends up in the study isn't random. Health-conscious volunteers, or people who fill in long food questionnaires, differ systematically from everyone else.

So why run them at all?

Because observational studies are genuinely valuable when used for the right job:

  • Generating hypotheses worth testing in an RCT later.
  • Questions you can't ethically randomise — you can't assign people to smoke for 30 years, so the smoking–cancer link was nailed observationally, using careful criteria (the Bradford Hill viewpoints: strength, consistency, dose–response, plausibility).
  • Very long-term or rare outcomes that no trial can practically follow.

The rule of thumb: observational data is strong for "this is worth investigating" and weak for "so take this pill".

Worked example The supplement graveyard

Over and over, an ingredient looked like a hero in observational data and then flopped when someone finally ran the RCT:

Beta-carotene: people eating more of it had less cancer → trials of beta-carotene supplements found higher lung-cancer rates in smokers. The confounder was "eats lots of vegetables".

Vitamin E: observational promise for heart disease → large RCTs found no benefit, some signals of harm.

Vitamin D: low blood levels track with almost every illness → but supplementing healthy, non-deficient people has mostly failed to move hard outcomes in RCTs. Low vitamin D is often a marker of being unwell, indoors and inactive — not the cause. The observational headline and the trial verdict pointed in opposite directions.

Next time you read "a study of 50,000 people found that taking X…", finish the sentence in your head with "…and those people also differed in a hundred other ways." Then wait for the RCT. That single habit would have saved a lot of people a lot of money on the ingredients above.

What to remember

“Linked to” is not “causes”. Finish the sentence with the confounder.

  • Observational studies watch, they don't intervene. No coin flip means the groups differ from the start — hello, confounding.
  • Healthy-user bias, reverse causation, selection bias all manufacture links that vanish under an RCT. The vitamin graveyard (beta-carotene, E, D) is built from exactly these.
  • They're for hypotheses, not prescriptions. Great at "worth investigating", weak at "so take this pill". Wait for the trial.
Try it · ~1 min

Spot the confounder in a headline.

  1. Find any "people who take/eat X have less Y" claim.
  2. Ask: what kind of person takes more X? Wealthier, more active, more health-conscious?
  3. Check whether those same traits could explain Y on their own. If yes, you've found the confounder — and the reason to wait for an RCT.

Next in this path

Lesson 05In-vitro & animal studies — the “in mice” problem Lesson 02RCTs — how randomisation defeats confounding

References

1
The Alpha-Tocopherol, Beta-Carotene Cancer Prevention Study Group. The effect of vitamin E and beta carotene on the incidence of lung cancer… N Engl J Med. 1994;330:1029–1035. doi:10.1056/NEJM199404143301501
2
Hill AB. The environment and disease: association or causation? Proc R Soc Med. 1965;58(5):295–300 — the Bradford Hill viewpoints on causal inference.
3
Autier P, Boniol M, Pizot C, Mullie P. Vitamin D status and ill health: a systematic review. Lancet Diabetes Endocrinol. 2014;2(1):76–89 — low vitamin D as marker vs cause.

Progress saves automatically in this browser · skip ahead anytime · no streak.