13. Selection Bias in Epidemiology: Mechanism, Examples, Fixes
After this lesson you can
- Distinguish bias from chance and from confounding, and internal from external validity
- State the selection bias mechanism and show it as one leaking cell in a table
- Recompute an odds ratio under differential participation
- File volunteer, non-response, loss to follow-up, healthy worker, referral and prevalence-incidence biases under one mechanism
- Reason out whether a bias runs toward, away from or across the null
- Choose design defenses such as a single source population and a record of non-participants by group
Video coming soon
Scheduled for release on November 10, 2026.
The first of two lessons on bias, defining bias as systematic error and explaining selection bias through a single mechanism: selection that depends on exposure and outcome together. It recomputes the invented Oakhaven case-control study when exposed controls stay away, moving the odds ratio from 3.0 to 5.0.
Questions and answers
Questions a learner would ask about this lesson. Think through your own answer before you open each one.
- What is the mechanism behind selection bias?
Show answer
Selection bias arises when getting into or staying in a study depends on exposure and outcome together. The relation between exposure and outcome among those studied then differs from the relation in the source population. Selection by exposure alone shrinks rows and selection by outcome alone shrinks columns, and in both cases the ratio holds; only joint dependence distorts it.
- How did the repeated Oakhaven study turn an odds ratio of 3.0 into 5.0?
Show answer
In the true data, cases were 75 exposed and 25 unexposed and controls were 50 and 50, an odds ratio of 3.0. When controls were recruited by a newspaper notice, only 60 percent of exposed controls took part, 30 of 50, while everyone else did. The observed odds ratio became (75 x 50) / (25 x 30) = 5.0, because exposure looked rarer among controls than it really was, a bias away from the null.
- Is a sample biased just because it is not representative?
Show answer
No. The test is whether the exposure-outcome relation inside the sample matches the source population, not whether the sample resembles it. A narrow sample can be unbiased and a broad sample can be biased. A limit on generalizability is an external validity question, so the lesson advises saying "narrow," not "biased."
- What is the healthy worker effect?
Show answer
Employed people are healthier than the general population before any exposure acts, because seriously ill people are rarely hired and rarely stay employed. In Oakhaven the factory cohort's all-cause death rate of 5 per 1,000 against 8 per 1,000 for district adults gave a ratio of about 0.63, making the factory look protective when it was not. It is a selection problem because it concerns who entered the cohort.
- Does a larger sample fix selection bias?
Show answer
No. Size fixes random error, and only design fixes systematic error. Had Oakhaven recruited 1,000 newspaper controls instead of 100 with the same 60 percent participation among the exposed, the odds ratio would still have been 5.0, now with a narrow confidence interval that made the wrong answer look certain.
Practice quiz
- Question 1True data: cases 75 exposed and 25 unexposed, controls 50 and 50. Only 60 percent of exposed controls participate; everyone else does. What is the observed odds ratio?
- Question 2That distortion arose because participation depended on:
- Question 3An occupational cohort has a lower death rate than the general population. What is the most likely explanation?
- Question 4Forty of 400 exposed workers leave a cohort before the outcome is assessed. What is the main threat?
- Question 5Which defense best prevents selection bias in a case-control study?
- Question 6In Oakhaven, clinic controls were 75 of 100 exposed instead of the true 50 of 100. The odds ratio moved from 3.0 to 1.0. Which direction is this bias?
- Question 7A study enrolls prevalent cases instead of incident cases, and the exposure affects how long people survive with the condition. What bias threatens it?
- Question 8Where does the selection leak usually enter in each design?
- Question 9A case-control study has 50 percent participation among cases and 50 percent among controls, with no difference by exposure. What happens to the odds ratio?
- Question 10Which statement correctly separates the three alternative explanations?
Saved in your browser only: no account, no server.
Principles of Epidemiology is education in epidemiologic reasoning, not clinical, screening, or policy guidance. Every example is set in the fictional Oakhaven Health District, and every number is invented for teaching.