PUBH5010 · Epidemiology Methods And Uses
Confounding, Effect Modification and Standardisation
Confounding is the third explanation for an association: a variable that is (1) associated with the exposure, (2) an independent risk factor for the outcome, and (3) not on the causal pathway between them, which mixes its own effect into the crude estimate. The exam wants more than the label — it wants the three criteria applied to a specific variable and, crucially, the direction the confounder pushes the estimate (toward or away from the null), which you reason from how it associates with exposure and outcome. Because a confounder was (in principle) measured, it can be controlled: in the design by restriction or matching, and in the analysis by stratification and Mantel–Haenszel pooling, or by regression. Effect modification is a different phenomenon and must not be confused with it: the exposure’s true effect genuinely differs across strata of a third variable, so you report the stratum-specific estimates rather than adjusting them away. Standardisation closes the chapter: direct and indirect methods remove the confounding effect of differing age (or other) structures so that two populations’ rates can be compared fairly — the source of the standardised mortality ratio.
What this chapter covers
- 01The three criteria that define a confounder
- 02Reasoning the direction of confounding bias
- 03Controlling confounding by design: restriction and matching
- 04Controlling confounding in analysis: stratification and Mantel–Haenszel
- 05Effect modification vs confounding — report vs adjust
- 06Direct standardisation and the standardised rate
- 07Indirect standardisation and the standardised mortality ratio (SMR)
Worked example: is age a confounder, and which way does it bias?
- +2(a) Three criteria. (1) Age is associated with the exposure (drinkers are older). (2) Age is an independent risk factor for heart disease. (3) Age is not on the causal pathway from coffee to heart disease. All three hold → age is a confounder.
- +2(b) Direction. Age is positively linked to both coffee and heart disease, so it inflates the crude association — the bias is away from the null. Adjusting for age should move the estimate back toward 1.
- +1(c) Control. Design: restrict to one age band, or match on age. Analysis: stratify by age and pool with Mantel–Haenszel (or adjust by regression).
Key terms
- Confounder
- A variable associated with the exposure, an independent risk factor for the outcome, and not on the causal pathway between them. It mixes its own effect into the crude association and must be controlled by design or analysis; the exam expects the three criteria applied to a named variable plus the direction of bias.
- Direction of confounding
- Which way a confounder pushes the crude estimate. A confounder associated in the same direction with both exposure and outcome biases away from the null (exaggerates); associated in opposite directions, it biases toward the null. Stating the direction, reasoned from the associations, is the marked skill.
- Effect modification
- When the true effect of the exposure genuinely differs across strata of a third variable (the stratum-specific estimates differ from each other). Unlike confounding it is a real finding to report — you present the stratum-specific measures rather than adjusting them into a single pooled number.
- Mantel–Haenszel estimate
- A method that pools stratum-specific measures of association into one adjusted summary, weighting strata to control a confounder while preserving efficiency. Used after stratification when the effect is consistent across strata (no effect modification).
- Standardisation
- Removing the confounding effect of a different population structure (usually age) so rates can be compared fairly. Direct standardisation applies each population's rates to a common standard structure; indirect standardisation applies standard rates to each population's structure, yielding the standardised mortality ratio (SMR).
Confounding, Effect Modification and Standardisation FAQ
What are the three criteria for a confounder?
A confounder must be (1) associated with the exposure, (2) an independent risk factor for the outcome, and (3) not on the causal pathway between exposure and outcome. If a variable lies on the pathway (a mediator), adjusting for it would wrongly remove part of the real effect; that distinction is a favourite exam trap. Apply all three to the specific variable named in the question.
How do I work out the direction of confounding?
Look at how the confounder associates with the exposure and with the outcome. If it is associated in the same direction with both (e.g. positively with each), it inflates the crude estimate — bias away from the null. If associated in opposite directions, it pulls the estimate toward the null. Stating and justifying this direction is worth more marks than naming confounding itself.
What's the difference between confounding and effect modification?
Confounding is a distortion you want to remove: a third variable contaminates the estimate, so you adjust it away to a single corrected number. Effect modification is a genuine feature you want to report: the exposure's real effect differs across strata of the third variable, so you present the stratum-specific estimates rather than collapsing them. Adjusting away a true effect modification would hide a real finding.
What does standardisation do?
It lets you compare two populations with different structures (typically different age distributions) fairly, by removing age as a confounder of the crude rate comparison. Direct standardisation applies each population's age-specific rates to one common standard population; indirect standardisation applies a standard set of rates to each population's age structure, producing the standardised mortality ratio (observed/expected deaths).
Exam move
For confounding, always do two things: apply the three criteria to the specific variable the question names, and state the direction of the bias, reasoned from how the confounder associates with exposure and outcome (same-direction → away from the null; opposite → toward it). Keep effect modification firmly separate — report stratum-specific estimates, never adjust them away — because conflating the two is a classic error. Know the control toolkit by where it acts: restriction and matching in design; stratification with Mantel–Haenszel, regression, and standardisation (direct vs indirect, and the SMR) in analysis.
