Types of Bias: Selection, Recall, Lead-Time & Length-Time

Bias is systematic error: a flaw in who is studied, how data are collected or how results are interpreted that pushes the answer away from the truth in one direction. Unlike random error, it does not shrink as the sample grows. Sort every stem into four bins: selection bias (who gets into or stays in the study: Berkson, attrition, healthy worker), information bias (recall, measurement, observer), confounding (a third factor tied to both exposure and outcome) and the screening biases, where lead-time bias lengthens survival by starting the clock earlier and length-time bias over-samples slow-growing disease. Mortality in a randomized screening trial, not survival after diagnosis, is the endpoint these biases cannot fake.

Bias types at a glance

Selection bias: who got in, who stayed in

Selection bias means the people studied differ systematically from the population the result should describe. Ask of every stem: who got in, and who stayed in?

• Berkson (admission rate) bias: hospital samples over-represent people with several illnesses, because each condition adds a reason to be admitted. Berkson's 1946 analysis showed how hospital clinic data could create a spurious association between gallbladder disease and diabetes. Fix: population-based controls and incident cases.
• Attrition bias (loss to follow-up): a threat to cohort studies and trials. If dropouts differ in prognosis and losses are unequal between groups, the groups that remain are no longer comparable. A common rule of thumb (Catalogue of Bias): under 5% loss causes little bias, over 20% seriously threatens validity. Fix: minimize losses, analyze everyone as randomized with intention-to-treat analysis, and test worst-case and best-case assumptions for the missing.
• Healthy worker effect: people who can work are healthier than the general population, which includes people too sick to work. Comparing a workforce's death rates with national rates therefore tends to understate, or even hide, an occupational hazard. Fix: compare with another working population.
• Volunteer and nonresponse bias: volunteers tend to be healthier, more educated and more adherent than nonvolunteers, and nonresponders differ systematically from responders. In screening, people who choose to be screened have lower mortality even from unrelated causes (healthy screenee bias).
• Prevalence-incidence (Neyman) bias: sampling existing cases misses patients who died quickly or recovered, a weakness of the prevalence snapshots in cross-sectional studies. Fix: incident (newly diagnosed) cases.

Trap: randomization cannot make an unrepresentative sample representative or undo dropout after randomization, and a bigger sample only shrinks random error, the territory of p-values and type I and type II errors.

Information bias: recall, measurement, observer

Information bias is systematic error in how exposure or outcome data are collected, remembered or recorded. Error that is similar in both groups (non-differential misclassification) typically pulls the estimate toward the null; error that differs between groups (differential misclassification) can push it either way.

• Recall bias: cases remember or report past exposures more (or less) completely than controls. It belongs to designs that ask about the past, above all case-control studies built on self-report. Classic stem: parents of children with cancer or a birth defect, asked about past infections or medications, search their memories harder than parents of healthy children. Fix: records made before the outcome (pharmacy, prenatal, employment), prospective exposure data (a cohort), newly diagnosed cases and the same neutral questions for everyone.
• Measurement bias: the instrument or method is systematically wrong (an uncalibrated blood pressure cuff, self-reported weight), or it is applied differently in the two groups. Fix: standardized, validated, calibrated methods chosen in advance and used identically in every group.
• Observer bias: the person assessing outcomes records differently because of expectations or knowledge of group assignment. Review books call the expectation-driven version observer-expectancy bias, or the Pygmalion effect: a self-fulfilling prophecy in which an investigator who expects a treatment to work tends to find improvement. Fix: blind the outcome assessors and prefer objective outcomes.
• Hawthorne effect: participants behave differently because they know they are being watched (hand-hygiene rates rise during an announced audit). Fix: unobtrusive observation.
• Performance (procedure) bias: the groups receive different care apart from the intervention because participants or staff know who got what. Fix: blinding and a placebo.

Confounding vs effect modification

A confounder is a third variable that is associated with the exposure, independently affects the outcome, and is not on the causal pathway between them. It can create a false association or hide a real one; if coffee drinkers smoke more, coffee can look like a lung cancer risk factor until you stratify by smoking. Effect modification is different: the exposure's effect genuinely differs across levels of a third variable. Stratify by that variable and compare:

A change of more than about 10% between crude and adjusted estimates is a common working definition of confounding. Only randomization balances unknown confounders; restriction, matching and adjustment handle only the variables you measured.

Screening biases: lead-time vs length-time

Survival is counted from diagnosis, $\text{survival} = t_{\text{death}} - t_{\text{diagnosis}}$, so anything that moves diagnosis earlier or fills the screened group with slow disease inflates it.

Lead-time bias, worked (NCI's example): a man whose lung cancer is diagnosed at 67 because of symptoms dies at 70, so he survives $70 - 67 = 3$ years and 5-year survival for patients like him is 0%. If screening finds the same cancer at 60 and he still dies at 70, he "survives" $70 - 60 = 10$ years and 5-year survival becomes 100%, yet his life has not been extended at all.

Length-time bias: slow-growing cancers spend longer in the detectable but silent phase, so a periodic screen is more likely to catch them, while fast-growing cancers surface with symptoms between screens (interval cancers). Screen-detected cases look better because of their biology, not because screening helped. Overdiagnosis is the extreme: screen-detected cancers that would never have caused symptoms or death. Studies cited by NCI estimate that 19% of screen-detected breast cancers and 20% to 50% of screen-detected prostate cancers are overdiagnosed.

The fix for all three: a randomized trial of screening versus no screening, analyzed by intention to screen (every cancer in each arm counts, interval cancers included), with disease-specific mortality (the most widely accepted endpoint) or all-cause mortality as the outcome. Rising 5-year survival alone never proves that screening saves lives.

Practice question

Vignette. A health system reviews women in its biennial screening mammography program. Cancers found at a scheduled mammogram are compared with cancers diagnosed after a lump appeared between two screens. Every woman in both groups attended the program, and follow-up is complete. The screen-detected tumors are more often low grade with a low proliferation index, and 10-year survival from diagnosis is higher in that group, even after adjustment for the estimated lead time. The investigators conclude that mammography improved survival. Which of the following best explains the survival difference?

A. Lead-time bias B. Length-time bias C. Healthy volunteer bias D. Recall bias E. Observer bias

Answer: B. Length-time bias. Slow-growing tumors linger in the preclinical window, so a scheduled mammogram is more likely to find them; aggressive tumors outrun the screening interval and present as interval cancers. The comparison is slow biology versus fast biology.

Why not the others?

• Lead-time bias inflates survival by starting the clock earlier, even when tumors are biologically identical. Here the gap persists after adjusting for lead time, and the clue is a difference in grade and growth rate. The two often coexist, which is why survival comparisons cannot evaluate screening.
• Healthy volunteer bias needs screened people compared with people who declined screening; both groups here attended.
• Recall bias needs self-reported past exposures; grade and survival come from records.
• Observer bias needs an assessor's judgment to shift the outcome; death is an objective endpoint.

Key takeaways

• Bias is systematic error: a bigger sample does not fix it.
• Selection bias = who got in and who stayed in: Berkson (hospital samples), attrition (use intention-to-treat), healthy worker effect, volunteers, Neyman (prevalent cases).
• Recall bias = case-control study plus memory; fix it with records or prospective data. Observer bias = unblinded assessor; fix it with blinding.
• Confounding: strata agree with each other but not with the crude, so report the adjusted estimate. Effect modification: strata disagree, so report each stratum.
• Lead-time bias: survival rises while the date of death stays put. Length-time bias: screening over-samples slow disease; overdiagnosis is its extreme.
• Judge screening by mortality in a randomized trial, never by survival after diagnosis.

Sources

• Admission Rate (Berkson) Bias — Catalogue of Bias, Oxford CEBM (2018)
• Attrition Bias — Catalogue of Bias, Oxford CEBM (2017)
• Recall Bias — Catalogue of Bias, Oxford CEBM (2017)
• Observer Bias — Catalogue of Bias, Oxford CEBM
• Information Bias — Catalogue of Bias, Oxford CEBM (2019)
• Length Time Bias — Catalogue of Bias, Oxford CEBM (2025)
• Crunching Numbers: What Cancer Screening Statistics Really Tell Us — National Cancer Institute (2018)
• Cancer Screening Overview (PDQ), Health Professional Version — National Cancer Institute (2023)
• Bovbjerg ML. Foundations of Epidemiology, Chapter 6: Bias — Oregon State University (2020)
• Bovbjerg ML. Foundations of Epidemiology, Chapter 7: Confounding — Oregon State University (2020)
• Bovbjerg ML. Foundations of Epidemiology, Chapter 8: Effect Modification — Oregon State University (2020)