Explainer · Plain-language
Selection Bias: Definition, Meaning & Examples | CASRAI
Selection bias is a systematic error introduced by the way participants are selected into a study or retained within it, so that the people analysed differ systematically from the population of interest. It distorts results and threatens both internal and external validity.
The step most authors miss
Doing CRediT right? Don’t stop at the statement.
A CRediT statement credits you inside one paper. The recognition CRediT was built for happens when those roles are tied to you, persistently. Sign in with your ORCID — free — and claim your CRediT contributions on casrai.org, the home of the standard. They become a verified, portable part of your identity, not a line that disappears into one PDF.
Free: claim your contributions, then export a journal-ready CRediT statement, schema.org structured data, JATS XML, CSV or BibTeX — and preview your public profile. A membership publishes that profile publicly and verifies the journals you serve.
How selection distorts results
Selection bias arises whenever the mechanism that determines who is included in — or excluded from — a study is related to the variables under investigation. If a treatment group and a control group differ systematically at the outset, any difference in outcomes is confounded with that initial difference. Because the distortion is baked into the composition of the sample, it persists no matter how carefully the data are later analysed; a larger but equally skewed sample is just as biased.
Recruitment versus attrition
Selection bias has two main entry points. At recruitment, the people who enrol may differ from those who do not — volunteers may be healthier, more motivated, or more affluent (volunteer or self-selection bias). During the study, attrition (loss to follow-up) can reintroduce bias if those who drop out differ systematically from those who stay — for example, if people doing poorly leave a treatment trial. Both routes break the comparability that valid inference depends on.
Common varieties
Selection bias takes many named forms. The healthy-worker effect makes employed populations look healthier than the general public. Berkson’s bias distorts associations studied only among hospital patients. Survivorship bias draws conclusions only from cases that "survived" some process, ignoring those that did not. Sampling bias — a closely related idea — arises when the sampling frame or method systematically over- or under-represents parts of the population. Recognising the specific mechanism is the first step to mitigating it.
Reducing selection bias
The strongest defence in experiments is randomisation, which makes groups comparable in expectation on both measured and unmeasured factors. In observational work, probability sampling, a well-defined sampling frame, broad and active recruitment, and strategies to maximise retention all help. Where bias cannot be prevented, analysts may adjust for it — for instance with weighting or sensitivity analyses — and should report response and attrition rates transparently so readers can judge how far the findings may be skewed.
Key facts
At a glance
- Definition: Systematic error from how participants are selected or retained
- Two sources: Recruitment (who enters) and attrition (who drops out)
- Threatens: Both internal validity and generalisability
- Examples: Volunteer, healthy-worker, Berkson’s, survivorship bias
- Not fixed by: Simply collecting a larger but equally skewed sample
- Defences: Randomisation, probability sampling, retention strategies
Common misconceptions
What people often get wrong
Often heard: A bigger sample size cancels out selection bias.
Actually: No — if the sample is systematically skewed, a larger sample is just as biased. Bias is about how participants are chosen, not how many there are.
Often heard: Selection bias only happens at recruitment.
Actually: No — it also arises through attrition, when those who drop out differ systematically from those who remain.
Often heard: Selection bias is the same as random sampling error.
Actually: No — random error averages out and shrinks with sample size; selection bias is systematic and does not.
Going deeper
