Types of bias: selection, recall, lead-time and confounding
Bias is systematic error that pushes a result in one direction. Learn the three families, why screening can raise survival without changing deaths, and the step that reduces each bias.
Chapters (10)
Identify common biases (selection, recall, lead-time, length-time, observer) and confounding, and name the design methods that reduce each.
The short version
- Bias is systematic error, and it comes in three families: selection bias, information bias and confounding.
- Selection bias is about who is analysed; information bias, such as recall and observer bias, is about how things are measured.
- Lead-time and length-time bias are screening biases that inflate survival counted from diagnosis; compare mortality instead.
- Confounding is a third factor tied to both exposure and outcome and not on the causal pathway.
- Randomise, restrict or match in the design, then stratify or adjust in the analysis; only randomisation also handles unknown confounders.
The puzzle: survival up, deaths the same
Imagine a screening programme announces that five-year survival for a cancer rose from 40% to 70%. But the number of people dying of that cancer has not changed at all. (A hypothetical example.)
Survival can go up while deaths stay the same when something is biasing the numbers. The types of bias below, in particular lead-time and length-time bias, explain this puzzle.
Types of bias: three families
Bias is systematic error: it pushes a result in one direction. It is commonly divided into three families: selection bias, information bias and confounding.[1]
| Family | The problem | Examples in this lesson |
|---|---|---|
| Selection bias | The people analysed are not comparable | Different source populations, healthier volunteers, unequal loss to follow-up |
| Information bias | Exposure or outcome is measured or recorded with error | Recall bias, observer bias |
| Confounding | A third factor mixes its effect with the exposure's | Coffee, smoking and lung cancer |
Lead-time and length-time bias are screening biases. Review books file them under selection bias or measurement (information) bias; either way, screening changes which cases are counted and when their clock starts.
Selection bias: who is analysed
Selection bias means the people analysed are not comparable, because of how they entered the study or who dropped out.[4][5]
- Controls drawn from a different population than the cases.
- Volunteers who are healthier than everyone else.
- More people lost to follow-up in one group.
Reduce it by choosing cases and controls from the same source population, keeping follow-up complete, and concealing allocation in trials.[4][5][6]
Information bias: recall bias and observer bias
Information bias means exposure or outcome is measured or recorded with error. In the cases taught here, the error differs between groups.[6]
| Bias | What happens | How to reduce it |
|---|---|---|
| Recall bias | In a case-control study, patients search their memory harder than healthy controls, so they report more exposures.[4] | Use objective records instead of memory |
| Observer bias | An investigator who knows the group status nudges the measurement or judgment.[6] | Blinding and standardised measures |
Lead-time bias: the clock starts earlier
Lead-time bias: screening finds disease earlier, so survival counted from diagnosis looks longer, even if death comes at the same time.[2][3]
- A patient dies at 72.
- Diagnosed from symptoms at 68: survival = 72 − 68 = 4 years.
- Screened at 65: survival = 72 − 65 = 7 years.
- Death came at the same age. Only the start of the clock moved.
Length-time bias, and why mortality is the fix
Length-time bias: screening more easily catches slow-growing cancers, which stay detectable for longer before symptoms. Those cancers also tend to have a better prognosis, so screen-detected cases appear to survive longer.[3]
The fix for both screening biases is to compare death rates in screened and unscreened groups, ideally randomised, not survival after diagnosis.[3]
That explains the opening puzzle: survival rose, but mortality did not.
Confounding: a third factor
A confounder is a third factor linked to both the exposure and the outcome, and not a step on the causal pathway between them. It can create an association that is not real, or hide one that is.[1][7]
Say coffee drinkers get more lung cancer. But smokers drink more coffee, and smoking causes lung cancer. If coffee does nothing, then within smokers, and within non-smokers, the apparent effect disappears. This is a hypothetical teaching example.
| Stage | Method |
|---|---|
| Design | Randomisation, restriction or matching |
| Analysis | Stratification or multivariable adjustment |
Practice question
Original · USMLE styleResearchers conduct a case-control study to examine whether medication use in the first trimester of pregnancy is associated with a congenital malformation. Cases are mothers of infants born with the malformation, and controls are mothers of infants born without it; both groups are selected from the same regional birth registry. Telephone interviewers who are unaware of each mother's group use the same structured questionnaire with 150 case mothers and 150 control mothers. Mothers of affected infants report significantly more first-trimester medication use than mothers of unaffected infants. Pharmacy dispensing records for the same women show similar rates of medication use in the two groups.
Common questions
What are the three main types of bias?
Selection bias, information bias and confounding. Selection bias is about who is analysed, and information bias is about how exposure or outcome is measured.
What is the difference between lead-time bias and length-time bias?
Lead-time bias means screening starts the survival clock earlier without changing the time of death. Length-time bias means screening more easily catches slow-growing cancers, which have a better prognosis anyway.
How do you reduce lead-time and length-time bias?
Compare death rates in screened and unscreened groups, ideally randomised, rather than survival after diagnosis.
How do you reduce recall bias and observer bias?
Use objective records instead of memory for recall bias. Use blinding and standardised measures for observer bias.
What is a confounder and how do you control for it?
A confounder is a third factor linked to both exposure and outcome that is not on the causal pathway. Use randomisation, restriction or matching in the design, and stratification or multivariable adjustment in the analysis.
Transcript
Read the full narration, by chapter686 words
0:00The puzzle: survival up, deaths the same
A screening programme announces its results: 5-year survival for a cancer rose from 40% to 70%. But the number of people dying of that cancer has not changed at all. How can survival go up while deaths stay the same? Something is biasing the numbers. Today: recognise selection, recall, observer, lead-time and length-time bias, and confounding, and name the step that reduces each.
0:28Bias is systematic error: three families
Bias is systematic error: it pushes a result in one direction. It comes in three families: selection bias, information bias and confounding.
0:39Selection bias: who is analysed, and how to reduce it
Selection bias means the people analysed are not comparable, because of how they entered the study or who dropped out. Controls drawn from a different population than the cases is one example. Volunteers who are healthier than everyone else is another. More people lost to follow-up in one group is a third. Reduce it by choosing cases and controls from the same source population, keeping follow-up complete, and concealing allocation in trials.
1:08Information bias: recall bias and observer bias
Information bias means exposure or outcome is measured or recorded with error. In the cases we study here, the error differs between groups. Recall bias: in a case-control study, patients search their memory harder than healthy controls, so they report more exposures. Observer bias: an investigator who knows the group status nudges the measurement or judgment. Reduce them with blinding and standardised measures. For recall bias, use objective records instead of memory.
1:42Lead-time bias: the clock starts earlier (65 vs 68, death at 72)
Back to the screening puzzle. Lead-time bias: screening finds disease earlier, so survival counted from diagnosis looks longer, even if death comes at the same time. A patient dies at 72. Diagnosed from symptoms at 68, survival is 4 years. Screened at 65, survival is seven. Only the start of the clock moved.
2:08Length-time bias: slow cancers, and why we compare mortality
Length-time bias: screening more easily catches slow-growing cancers, which stay detectable for longer before symptoms. Those cancers also tend to have a better prognosis. Both are screening biases. Books file them under selection or measurement bias; either way, screening changes which cases are counted and when their clock starts. The fix for both is to compare death rates in screened and unscreened groups, ideally randomised, not survival after diagnosis. That explains our puzzle: survival rose, but mortality did not.
2:47Confounding: a third factor, and design and analysis fixes
Last, confounding. A confounder is a third factor linked to both the exposure and the outcome, and not a step on the causal pathway between them. It can create an association that is not real, or hide one that is. Say coffee drinkers get more lung cancer. But smokers drink more coffee, and smoking causes lung cancer. If coffee does nothing, then within smokers, and within non-smokers, the apparent effect disappears. In the design, use randomisation, restriction or matching. In the analysis, use stratification or multivariable adjustment. Only randomisation also handles unknown confounders.
3:30Your turn: a practice question
Your turn. Researchers conduct a case-control study of medication use in early pregnancy and a birth defect, using the same regional birth registry for both groups. Telephone interviewers, unaware of group status, use one questionnaire with 150 mothers of affected infants and 150 mothers of healthy infants. Mothers of affected infants report far more medication use. Pharmacy records show similar use in both groups. Which best explains the difference? A, confounding. B, lead-time bias. C, observer bias. D, recall bias. E, selection bias. Pause here and decide.
4:19The answer, one option at a time
A is out: confounding needs a third factor linked to both medication and birth defect, and actual use was similar. B is out: lead-time bias belongs to screening and survival. C is out: the interviewers did not know who was a case. E is out: both groups came from the same registry. So the answer is D. Mothers of affected infants search their memories harder, so they report exposures more completely. That is recall bias.
4:51Recap pocket card and what is next
Three things to keep. One: selection bias is about who is analysed; information bias, such as recall and observer bias, is about how things are measured. Two: lead-time and length-time bias are screening biases that inflate survival; compare mortality instead. Three: confounding is a third factor tied to exposure and outcome; randomise, restrict or match, then stratify or adjust. Next, type one and type two errors, power and p values.
Sources
Numbers in the text, e.g. [1], point here- Assessing bias: the importance of considering confounding. Evid Based Spine Care J 2012. pmc.ncbi.nlm.nih.gov/articles/PMC3503514/
- Lead time bias. pmc.ncbi.nlm.nih.gov/articles/PMC11951120/
- Deceptive Measures of Success in Early Cancer Detection (length bias, survival vs mortality). Curr Oncol 2024. pmc.ncbi.nlm.nih.gov/articles/PMC11431433/
- Methodology Series Module 2: case-control studies (selection and recall bias). pmc.ncbi.nlm.nih.gov/articles/PMC4817437/
- Methodology Series Module 1: cohort studies (follow-up). pmc.ncbi.nlm.nih.gov/articles/PMC4763690/
- Observer bias and blinding. pmc.ncbi.nlm.nih.gov/articles/PMC9709863/ · pmc.ncbi.nlm.nih.gov/articles/PMC4862344/
- Confounding. pmc.ncbi.nlm.nih.gov/articles/PMC4966397/ · en.wikipedia.org/wiki/Confounding
- Recall bias; selection bias; lead time bias; screening (medicine). en.wikipedia.org/wiki/Recall_bias · en.wikipedia.org/wiki/Selection_bias · en.wikipedia.org/wiki/Lead_time_bias · en.wikipedia.org/wiki/Screening_(medicine)
The practice question is original and its patient is synthetic. Every fact is checked against the sources above in independent accuracy reviews. Always follow your local guidelines and your own laboratory’s ranges.
Image credits. Illustrations: Servier Medical Art (smart.servier.com), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), adapted (recoloured, outlined, resized): scanner, lung cancer, lock, woman, eye, elderly man, clock, cup, cigarettes, pregnancy, hospital, man. Crowd, database and phone icons: Health Icons (CC0). The bars, dials, timelines, trees, cards and all handwriting are drawn in code (original). Hook figures, the coffee and smoking example and the practice vignette are invented teaching examples, not real data. No AI-generated teaching images.
Corrections. None so far. Spotted an error? Email info@medservatory.com with the lesson code MS-S1-BIOS-006.
For exam study only, not medical advice. Read the medical disclaimer.
In this series
Playlist on YouTube- EP 01Sensitivity vs specificity: the 2×2 table, SnNout and SpPin5:34
- EP 02PPV, NPV and prevalence explained5:27
- EP 03Relative risk vs odds ratio: the 2x2 table explained5:15
- EP 04Number needed to treat: ARR, RRR and NNH from trial data5:00
- EP 05Study designs explained: case-control, cohort, cross-sectional and RCT5:22
- EP 06Types of bias: selection, recall, lead-time and confounding5:30
- EP 07Type I and type II errors: alpha, beta, power and the p-value5:16
- EP 08Confidence intervals and significance: is the null value inside?5:12