Examqo › Examqo library › Public health › Epidemiological methods · IMLE
Bias, confounding and effect modification
An association found in a study may arise from chance, from bias (a systematic error in selection or measurement) or from a confounder – a factor linked to both exposure and outcome that is not on the causal path between them; effect modification is a real finding in which the strength of the association differs between strata.
Why it matters in the exam
This is one of the most frequent topics in the research part of the exam. A critique of a study is presented and you are asked which bias it describes: recall, selection, loss to follow-up, the "healthy worker effect" or misclassification. Other items give a crude and an adjusted OR and ask whether confounding is present, or a stratified table in which the association differs between men and women – effect modification. Also tested are the ways to handle confounders at the design stage (randomization, restriction, matching) and at the analysis stage (stratification, multivariable analysis), and reading a DAG: adjust for a confounder, but not for a mediator or a collider. Understanding the difference between an error that analysis can fix and one created during data collection also solves the confusingly worded items.
Key points
- Bias is a systematic error in selection or measurement and is prevented mainly by design.
- A confounder is linked to exposure and outcome and is not on the causal path.
- Confounders are handled by randomization, restriction, matching, stratification and multivariable models.
- Effect modification is a finding that is reported separately for each stratum.
- In a DAG, adjust for a confounder but not for a mediator or a collider.
Summary table
| Type | What happens | Example | How to handle it |
|---|---|---|---|
| Selection bias | Groups are chosen so that the association is distorted | "Healthy worker effect", hospital controls, loss to follow-up | Controls from the same source population, complete follow-up |
| Information bias | Exposure or outcome is measured differently between groups | Recall bias in mothers of infants with a malformation | Data recorded beforehand, blinding, a uniform questionnaire |
| Non-differential misclassification | The same measurement error in both groups | An imprecise questionnaire for everyone | Usually biases the association towards the null |
| Confounder | A factor associated with both exposure and outcome that does not lie on the causal path | Age in the link between coffee and infarction, when smoking was not measured | Randomization, restriction, matching, stratification, multivariable model |
| Effect modification | The strength of the association differs between strata | A treatment that helps women but not men | Report each stratum separately rather than pooling |
| Screening biases | Lead time, length time, overdiagnosis | "Longer" survival with no change in mortality | A randomized trial with disease-specific mortality as the outcome |
In the full version for subscribers
- A simple algorithm
- Memory aid
- Exam traps
- 5 review cards
- Enough for IMLE
The full version is open to subscribers of: Public health, Medical licensing exam (IMLE), Family medicine.
Subscriber? Read the full topic Plans and prices 25 free questions
Sources
- Rothman, Epidemiology: An Introduction, 2nd ed., 2012 — Biases in study design; Dealing with confounding
- Gordis, Epidemiology, 6th ed., 2019 — More on causal inference: bias, confounding and interaction
- Hernán & Robins, Causal Inference: What If, 2020 — Graphical representation of causal effects
Sources are given for further reading; the text is written afresh and not copied from them.
Reviewed by the Examqo AI Professor · 2026-11-06. The text may contain errors; when in doubt, check the source.