These concepts appear simple. They are not. And they are frequently misunderstood in both exams and clinical research.
Confounding – The Classic Third Variable Trap
A confounder is a variable that is:
- Associated with the exposure
- Associated with the outcome
- Not on the causal pathway
Classic example:
- Coffee → MI?
- Actually: Coffee drinkers often smoke → smoking is the real culprit.
Confounding distorts effect estimates:
- Can hide true associations
- Can create false ones
- Can overestimate or underestimate effects
How to adjust:
- Stratification → Group by confounder (e.g. smokers vs non-smokers)
- Multivariable regression → Adjust statistically for multiple confounders
- DAGs (Directed Acyclic Graphs) → Visualise what to control for
Bias – Built-In Study Errors
Bias = a flaw in the study design or conduct.
Confounding is about what you study.
Bias is about how you study it.
⚠️ Types of Bias You Need to Know:
Selection bias
Occurs when there are systematic differences in how participants are selected, enrolled, or lost to follow-up. This distorts the comparability of groups.
- Neyman bias (incidence–prevalence bias): Misses early deaths or quickly resolved cases, underestimating disease incidence.
- Berkson bias: Using hospital controls leads to artificially high exposure rates because patients are more likely to have both disease and exposure.
- Diagnostic purity bias: Excluding patients with co-morbidities creates an unrealistic sample that does not reflect the real-world population.
Performance bias
This happens when the care provided differs between study groups, apart from the intervention itself. For example, more attention given to one group might influence outcomes regardless of treatment.
Measurement bias
Systematic differences in how outcomes or exposures are assessed or recorded.
- Observer bias: The researcher’s knowledge of treatment allocation affects how they record outcomes.
- Responder bias: The participant changes their answers or behaviour based on what group they are in.
- Recall bias is a subtype — e.g. cases may remember exposures more clearly than controls.
- Hawthorne effect: People behave differently just because they know they’re being studied. This only leads to bias if it happens unequally across groups.
Publication bias
Studies with positive results are more likely to be published. This skews the available evidence and can lead to wrong conclusions in reviews and meta-analyses.
Effect Modification – Not a Flaw, But a Finding
Also called interaction.
Effect modification occurs when the size or direction of an effect varies across different levels of another variable (e.g. sex, age, smoking status).
Example:
- Asbestos alone → 3x lung cancer risk
- Smoking alone → 10x
- Expected combined = 30x
- Observed combined = 60x → suggests effect modification
Key points:
- It’s real. It reflects a biological interaction.
- It’s not a problem like confounding—it can be clinically meaningful.
- Can also go the other way (weaker effect than expected).
- Sometimes, it even reverses direction across groups.
- Don’t over-interpret it unless it’s biologically plausible and replicated.
Key Exam Tips
Confounding = third variable creating a false link
Bias = systematic design error
Effect modification = real variation in effect size across groups
Confounding must be controlled for.
Bias should be prevented at the design stage.
Effect modification should be explored, not adjusted away.
Thank you to the joint editorial team of MRCEM Exam Prep for this ‘Exam Tips’ blog post.
