Bias, confounding and causation
Introduit| English | Sens |
|---|---|
| selection bias/sɪˈlekʃn ˈbaɪəs/ | A systematic error in who gets into the study or who stays in it. |
| information bias | A systematic error in how exposure or outcome is measured. |
| recall bias | recall bias |
| confounder | A third variable linked to both the exposure and the outcome, and not on the path between them. |
| temporality | The one that cannot be missing. |
Does coffee cause lung cancer?
- A study finds coffee drinkers have twice the risk of lung cancer. Before blaming coffee, ask: do coffee drinkers also smoke more?
- Numbers can mislead in three ways: through bias, through a third variable, or by association without cause.
Selection bias
- Selection bias 选择偏倚 is a systematic error in who gets into the study or who stays in it.
- Hospital patients, volunteers or people who drop out can differ from the population the result is meant to describe.
Information bias
- Information bias 信息偏倚 is a systematic error in how exposure or outcome is measured.
- In recall bias 回忆偏倚, cases search their memory harder than controls, so past exposures seem more common among cases.
Mothers of babies with birth defects report more medicine use in pregnancy than other mothers. What is the main worry?
Difficulté 1/5 · 100 pièces + un petit bonus lors de la première réussite aujourd'hui
Mothers of affected babies may search their memory harder.
Confounding
- A confounder 混杂因素 is a third variable linked to both the exposure and the outcome, and not on the path between them.
- Smoking is linked to coffee drinking and causes lung cancer, so it can make coffee look harmful.
Match each problem to its name.
Difficulté 2/5 · 140 pièces + un petit bonus lors de la première réussite aujourd'hui
Who is studied, how it is measured, and a third variable.
The stratification test
- Split the data by the suspected confounder and compute the RR inside each group. If the association disappears in every stratum, it was confounding.
- If it stays, the exposure keeps its own effect. Randomisation prevents confounding in advance by balancing even unknown factors.
(Synthetic) Among smokers, exposed 18/120 and unexposed 6/40 had events. What is the RR in this stratum?
Difficulté 2/5 · 140 pièces + un petit bonus lors de la première réussite aujourd'hui
0.15 ÷ 0.15 = 1.0, so no effect inside the stratum.
Put the steps of a stratification check in order.
Difficulté 3/5 · 200 pièces + un petit bonus lors de la première réussite aujourd'hui
If the association vanishes inside the strata, it was confounded.
A very large sample removes confounding.
Difficulté 1/5 · 100 pièces + un petit bonus lors de la première réussite aujourd'hui
It only narrows the interval around a biased estimate.
From association to cause
- Bradford Hill's viewpoints help judge whether an association is causal: strength, consistency, dose-response, plausibility and others.
- Only one is essential: temporality 时间顺序. The cause must come before the effect.
Which Bradford Hill viewpoint must be met for a cause?
Difficulté 1/5 · 100 pièces + un petit bonus lors de la première réussite aujourd'hui
The cause must come before the effect; the rest support but are not required.
Worked example: confounding revealed
- (Synthetic) Crude: exposed 20/200 events against unexposed 10/200, RR 2.0. Exposed: 120 smokers (18 events) and 80 non-smokers (2). Unexposed: 40 smokers (6) and 160 non-smokers (4).
- Smokers 18/120 = 6/40 = 15%; non-smokers 2/80 = 4/160 = 2.5%. RR = 1.0 in each stratum: smoking confounded.
Worked example: three appraisal questions
- (Fictional abstract) "Coffee drinkers had twice the risk." Who was studied (selection)? How was coffee measured (recall)?
- What could confound it (something tied to both coffee and the outcome, such as smoking)? Three questions, before any belief.
Which method prevents confounding by unknown factors?
Difficulté 1/5 · 100 pièces + un petit bonus lors de la première réussite aujourd'hui
Only chance assignment balances factors nobody has measured.
Three common slips
- Thinking big data cures confounding. More data gives a more precise wrong answer.
- Treating Hill's viewpoints as a scorecard. Only temporality is required.
- Thinking adjustment fixes measurement bias. No analysis fully repairs badly measured data.
Recap
- Bias = systematic error: selection (who is studied), information or recall (how it is measured).
- Confounder = a third variable linked to both exposure and outcome; stratify to test for it, randomise to prevent it.
- Hill's viewpoints judge causation; temporality is the one that cannot be missing.