Questions: Information Bias and Misclassification Error

Question 1

In a case-control study of breast cancer, cases (women with breast cancer) are asked about past hormone replacement therapy use, as are matched controls. Cases report substantially higher rates of past HRT use. A critic suggests this may reflect recall bias. If the critic is correct, what type of misclassification is this, and in which direction does it bias the odds ratio?

• A. Non-differential misclassification; biases the odds ratio toward the null
• B. Differential misclassification; biases the odds ratio toward the null
• C. Differential misclassification; biases the odds ratio away from the null (inflates the apparent association)
• D. Non-differential misclassification; biases the odds ratio away from the null

Answer: (C)

Recall bias is a classic form of differential misclassification: cases are more motivated to recall past exposures (searching their memory for an explanation for their illness), so they over-report HRT use compared to controls. The misclassification rate differs between cases and controls — that's what makes it differential. Because cases over-report exposure, the apparent odds ratio is inflated above the true value — bias away from the null. Non-differential misclassification (same error rate in both groups) would instead bias toward the null.

Question 2

A cohort study of lung cancer measures smoking status at baseline with a questionnaire that has a 10% misclassification rate applied equally to exposed (smokers) and unexposed (non-smokers) participants. What is the expected effect on the observed risk ratio?

• A. The risk ratio is inflated — random errors amplify apparent associations
• B. The risk ratio is biased toward the null — the two groups are blurred together
• C. There is no systematic effect — random errors cancel out across the sample
• D. The direction of bias depends on the baseline prevalence of smoking in the cohort

Answer: (B)

Equal misclassification rates in both groups (non-differential) blur the two groups toward each other. Some true smokers are classified as non-smokers and vice versa, making the 'exposed' group less purely exposed and the 'unexposed' group less purely unexposed. The observed risk ratio moves toward 1.0 — the null — because the contrast between groups is diluted. This is attenuated, not amplified, association. The 'random errors cancel out' reasoning (Option C) is wrong for bias — it applies to random variation around an estimate, not to systematic misclassification.

Question 3

Non-differential misclassification of a binary exposure always biases the observed risk ratio or odds ratio toward the null value of 1.0.

• T. True
• F. False

Answer: True

Under non-differential misclassification of a binary exposure (same error rate in both outcome groups), the mathematical consequence is systematic attenuation of the apparent association toward 1.0 — 'bias toward the null.' This makes studies conservative: they underestimate true effect sizes. This is why studies showing associations despite likely non-differential misclassification are particularly compelling evidence — the true effect would be even larger. Note that with polytomous (more than two category) exposure, non-differential misclassification can occasionally bias away from the null, but for binary exposure the direction is consistent.

Question 4

Because non-differential misclassification involves random measurement error applied equally to both groups, it does not introduce systematic bias into study results.

• T. True
• F. False

Answer: False

This is the key misconception. 'Random' here means the errors are applied equally across groups — it does NOT mean the errors have no systematic effect on the estimate. Non-differential misclassification produces a highly predictable, systematic bias: attenuation toward the null. The risk ratio is consistently underestimated. Random individual errors can produce a systematic direction of bias at the aggregate level. The term 'random error' in this context refers to the mechanism of error generation, not its effect on the effect measure.

Question 5

Why is differential misclassification considered a more serious validity threat than non-differential misclassification?

Model answer: Non-differential misclassification has a predictable direction of bias (toward the null), so researchers can reason about its effect: the true association is at least as large as observed, and often larger. Differential misclassification can bias in either direction — toward or away from the null — depending on the specific mechanism. Because the direction is unpredictable without knowing the mechanism in detail, differential misclassification can both underestimate and overestimate associations, making it harder to reason about what the 'true' result might be.

Recall bias (a common form of differential misclassification in case-control studies) inflates associations and can create apparent effects where none exist. An unknown direction of bias is fundamentally harder to account for than a known one. This is why study designs that reduce differential misclassification — blinded outcome assessment, objective biomarker measurement — are considered methodologically stronger than those relying on self-report.