14. Information Bias and Misclassification: Bias Toward the Null
After this lesson you can
- Define information bias and describe a measurement by its sensitivity and specificity
- Recompute a risk ratio under non-differential misclassification and explain why it moves toward the null
- Recompute the same truth under differential misclassification and explain why it can move either way
- Distinguish recall, interviewer, reporting and detection biases and instrument error
- State the direction of a bias before its size, including for a protective exposure
- Match each bias to a design defense such as blinded assessment or a validation substudy
Video coming soon
Scheduled for release on November 11, 2026.
The second lesson on bias, covering errors in measuring exposure and outcome, where the right people land in the wrong cells. One invented Oakhaven truth table with a risk ratio of 2.0 is recomputed under equal error, observed as about 1.73, and under differential recall, observed as about 2.63.
Questions and answers
Questions a learner would ask about this lesson. Think through your own answer before you open each one.
- What is information bias, and how does it differ from selection bias?
Show answer
Information bias is distortion from error in measuring or classifying exposure, outcome or covariates. The right people are in the study, but some are counted in the wrong cells. Selection bias is about who enters or stays in the study, so the table's edge is wrong; information bias is about how exposure and outcome were measured, so the table's cells are wrong. Both are systematic, and neither shrinks with sample size.
- What are the two leaks of a measurement instrument?
Show answer
Sensitivity of measurement is the share of truly exposed people classified as exposed, and specificity is the share of truly unexposed people classified as unexposed. One minus each is a leak across the table, and both leaks run at once. For example, with 90 percent for both, 100 truly exposed cases become 90 measured exposed, and 5 of 50 truly unexposed cases leak in, giving 95 measured exposed cases.
- Why does non-differential misclassification usually pull a risk ratio toward the null?
Show answer
When error rates are the same in every group, each measured group becomes a mixture of the two true groups, and mixtures are more alike than the pure groups. The two observed risks move toward each other, so their ratio moves toward 1.0. In Oakhaven, a true risk ratio of 2.0 measured with 90 percent sensitivity and specificity in everyone was observed as about 1.73. The rule holds for two categories with independent errors.
- Why can differential misclassification move a result in either direction?
Show answer
Its error rate depends on the other variable, so it can add people to whichever cell the group difference favors. When Oakhaven cases recalled exposure perfectly but non-cases only 80 percent of the time, the same truth of 2.0 was observed as about 2.63, away from the null. If non-cases were probed harder instead, the exposed non-case cell would grow and the ratio would fall.
- How is recall bias different from detection bias?
Show answer
Recall bias is about who remembers the exposure: an exposure error that depends on outcome, which threatens case-control studies. Detection bias is about who gets examined for the outcome: an outcome error that depends on exposure, which threatens cohort studies, for example when exposed workers are examined yearly and others are not. Both are differential, but they act on opposite variables.
Practice quiz
- Question 1Exposure is misclassified at the same rate among cases and non-cases. The misclassification is:
- Question 2Truth: 500 exposed with 100 cases, 500 unexposed with 50 cases. Exposure is measured with 90 percent sensitivity and 90 percent specificity in everyone. What is the observed exposed-case cell?
- Question 3In that same scenario, the observed risks are 95 of 500 and 55 of 500. What is the observed risk ratio?
- Question 4Cases remember past exposure better than controls. This is:
- Question 5Which design feature best prevents interviewer bias?
- Question 6Same truth table. Cases recall exposure with 100 percent sensitivity, non-cases with 80 percent, and specificity is 90 percent in everyone. The observed risk ratio is about 2.63. Which direction did the bias run?
- Question 7A protective exposure has a true risk ratio of 0.75. Non-differential misclassification pulls it toward the null. The observed ratio will be:
- Question 8Exposed workers are examined for the illness every year, while unexposed workers are not. Which bias is most likely?
- Question 9The outcome is found with 80 percent sensitivity and perfect specificity, in both exposure groups. True risks are 20 and 10 percent. What is the observed risk ratio?
- Question 10A study observes a risk ratio of 2.0, and the only suspected bias is non-differential misclassification of a two-category exposure. What can be concluded?
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.