Confounding

What it is

A confounder is a variable that:

1. Affects the exposure (e.g. health-conscious people take vitamins).
2. Affects the outcome (e.g. health-conscious people also exercise, eat well, sleep enough — and live longer).

The apparent vitamin-taking → longer-life association is confounded by health-consciousness, not driven by vitamins themselves.

Classic longevity-research confounders

• Healthy-user effect: people who do one healthy thing tend to do many. Hard to isolate any single behaviour’s effect.
• Socioeconomic status: confounds nearly every health behaviour.
• Underlying disease: sicker people may avoid coffee, alcohol, exercise → making "abstinence" look bad in observational data.
• Recall bias: dietary self-report differs by health status.
• Selection into supplement use: same as healthy-user effect.

Adjustment is partial

Statistical adjustment for measured confounders helps but never fully removes confounding. Unmeasured confounders — especially those correlated with measured ones — persist.

Examples from history

• Hormone replacement and CVD: observational data suggested HRT protected against CVD. Healthy-user effect was a major contributor. WHI RCT inverted the apparent benefit in older women (timing-of- initiation issues then partially restored it).
• Vitamin E and CVD: observational benefit; RCTs (HOPE, others) showed null or harmful.
• Beta-carotene and lung cancer: observational protective; CARET + ATBC trials showed harm in smokers.

Tools that help

• Randomised controlled trials (gold standard).
• Mendelian randomization (genetic instruments).
• Target-trial emulation (designing observational analyses to mimic RCTs).
• Negative controls.
• Sensitivity analyses for unmeasured confounding.

Related entries

Survivorship bias, Mendelian randomization, Bradford-Hill criteria, NNT, ARR, RRR.