BiostatisticsTier 1Medical Sciences concept

Types of bias in epidemiological studies (recall, selection, confounding, information)

Core concept

Bias = systematic error - the estimate is wrong in a predictable direction
- A larger sample does not fix bias - it only narrows random error, so a biased estimate becomes precisely wrong
Three families:
- Selection bias - who enters the study, and who stays in
- Information (measurement) bias - how exposure or outcome is ascertained
- Confounding - a third variable linked to both exposure and outcome, and not on the causal path
Confounding requires all three:
- assoc with the exposure, AND
- independently assoc with the outcome, AND
- not an intermediate step (adjusting for a mediator is over-adjustment, not confounder control)

Selection bias

Type	Mechanism
Berkson's	Hospital-based controls - both exposure and disease raise the chance of admission
Healthy worker	Employed cohorts are healthier than the general population
Attrition / loss to follow-up	Dropouts differ from completers
Immortal time	Person-time in which the outcome could not occur is assigned to the treated group -> spurious benefit
Volunteer / referral	Self-selected participants differ systematically

Information bias

Type	Mechanism
Recall	Cases remember exposures better than controls - the signature weakness of case-control design
Observer / interviewer	Knowledge of allocation shapes assessment - fixed by blinding
Misclassification	Non-differential -> bias toward the null; differential -> either direction
Hawthorne	Being observed changes behaviour
Surveillance / detection	The exposed are investigated more, so more disease is found

Controlling confounding

Design: randomisation (the only method that also balances unmeasured confounders), restriction, matching
Analysis: stratification, multivariable regression, propensity score, instrumental variable
Residual confounding persists whenever the confounder is measured imprecisely

Screening biases - the exam's favourite triad

Lead-time bias - screening moves the diagnosis earlier -> survival from diagnosis lengthens with no change in date of death
Length-time bias - screening preferentially catches slow-growing, indolent tumours; fast aggressive ones present between rounds
Overdiagnosis - the extreme of length-time: disease detected that would never have caused symptoms or death
- South Korean thyroid cancer: ultrasound screening -> ~15x rise in incidence, mortality flat, relative survival approaching/exceeding 100%
- Relative survival at or above 100% is a red flag for overdiagnosis, not for excellent treatment
Only disease-specific (or all-cause) mortality in a randomised screening trial escapes all three. 5-year survival in screened vs unscreened is uninterpretable
Effect modification is not bias - a real difference in effect across strata, to be reported, not adjusted away

3 more sections, plus exam facts

Premium unlocks every note across every specialty, and the full exam fact library behind it.