Confounding and effect modification answer different questions. Confounding concerns a distorted exposure–outcome comparison. Effect modification concerns whether an effect measure differs across groups. A factor can matter for both, and the distinction depends on the study question and the measure being reported.
For CPH review, compare crude, adjusted and stratum-specific estimates without treating a percentage cutoff or a subgroup difference as automatic proof. Use relative risk and odds ratio carefully: they are not interchangeable.
Confounding: an unfair comparison
In a simple example, a factor that precedes exposure is related to both exposure and outcome, but is not a consequence of exposure. The exposed and unexposed groups may then differ in their underlying outcome risk even without an exposure effect. The observed comparison mixes those differences with the association of interest.
Age is a candidate confounder if it affects or predicts both exposure patterns and outcome risk in the study setting. The label is not enough: use subject knowledge about timing and causal relationships. A variable on the pathway from exposure to outcome is a mediator, not a baseline confounder for that exposure’s total effect. See CDC’s discussion of confounding and stratified analysis.
Random assignment can address confounding in expectation in a trial. Restriction, matching with an appropriate analysis, stratification, standardization, regression and weighting can address measured confounding under assumptions. None guarantees control of unmeasured factors or repairs every selection and measurement problem.
Worked example: age composition changes a crude risk ratio
These numbers are fictional. Suppose outcomes are measured over the same follow-up period in two exposure groups. Assume age is measured before exposure and is not caused by it.
| Age group | Exposed: cases / people | Unexposed: cases / people | Stratum-specific risk ratio |
|---|---|---|---|
| Younger | 45 / 900 = 5% | 5 / 100 = 5% | 0.05 / 0.05 = 1.0 |
| Older | 20 / 100 = 20% | 180 / 900 = 20% | 0.20 / 0.20 = 1.0 |
| Combined | 65 / 1,000 = 6.5% | 185 / 1,000 = 18.5% | 0.065 / 0.185 ≈ 0.35 |
On smaller screens, scroll the table horizontally to read every column.
The crude comparison looks protective, although the observed risks are equal within each age group. The exposed group contains many more younger people. In this constructed example, age composition explains the crude difference.
To put both groups on the same hypothetical standard with 50% younger and 50% older people:
Standardized exposed risk = (0.5 × 0.05) + (0.5 × 0.20) = 0.125.Standardized unexposed risk = (0.5 × 0.05) + (0.5 × 0.20) = 0.125.Standardized risk ratio = 0.125 / 0.125 = 1.0.
This arithmetic shows why adjustment can change a comparison. It is not a claim about a real intervention, and an adjusted estimate in a real study is not automatically the true causal effect.
Why a 10% change does not prove confounding
A change-in-estimate rule is sometimes used as a screening heuristic. If the adjusted estimate is the chosen reference, one calculation is |crude − adjusted| / |adjusted| × 100%. The reference and threshold must be specified; “a difference greater than 10%” is not a universal mathematical definition of confounding.
Measurement error, sampling variability, model choices and a change in the target comparison can affect estimates. Odds ratios have an additional complication: their crude and conditional values can differ because of noncollapsibility, even without confounding. Do not diagnose confounding from an odds-ratio change alone. See Schuster and colleagues’ methodological analysis.
Effect modification: specify the scale
Effect modification means that the effect measure differs across levels of another variable. It is not, by itself, a bias to eliminate. Report the relevant subgroup estimates and their uncertainty, and explain the population, time period and scale.
A higher baseline outcome risk in one subgroup does not alone show effect modification. Compare exposed with unexposed people within each subgroup. Also distinguish differences in observed associations from differences in causal effects: a causal interpretation needs the study’s identification assumptions.
Worked example: equal risk ratios, different risk differences
In another fictional study, the following point estimates describe the same follow-up period:
| Subgroup | Unexposed risk | Exposed risk | Risk ratio | Risk difference |
|---|---|---|---|---|
| A | 20% | 10% | 0.10 / 0.20 = 0.5 | 10% − 20% = −10 percentage points |
| B | 40% | 20% | 0.20 / 0.40 = 0.5 | 20% − 40% = −20 percentage points |
On smaller screens, scroll the table horizontally to read every column.
The risk ratios are equal, but the risk differences are not. These point estimates show heterogeneity on the additive scale, not the risk-ratio scale. They do not establish a statistically detectable interaction in a population: that assessment also needs sample sizes, uncertainty and a suitable model.
This is why “the effect is different” should name the measure. Interaction terms depend on the model’s scale; they are not interchangeable with every description of effect modification. See chapters 4 and 6 of Hernán and Robins, Causal Inference: What If.
A practical analysis sequence
- Define the exposure, outcome, target population, follow-up and effect measure.
- Use causal knowledge to choose plausible confounders; avoid automatic adjustment for every recorded variable.
- Examine relevant subgroup estimates with their uncertainty. Unequal point estimates alone do not prove meaningful heterogeneity.
- Address confounding under a justified design or analysis. If effect modification is important, report appropriate subgroup comparisons rather than hiding them in one summary.
- State remaining assumptions, measurement limitations and selection concerns. Adjustment is not a certificate of validity.
Confounding and effect modification can coexist. Adjustment within subgroups and reporting subgroup-specific effects may both be needed. Equal estimates in two measured strata do not establish that every source of bias is absent.
Practice questions
Question 1: interpreting adjustment
In a fictional observational study, the crude risk ratio is 1.5 and the age-adjusted risk ratio is 1.0. Age precedes exposure and is a plausible common cause of exposure and outcome. Which conclusion is most defensible?
- The adjusted result proves that all bias has been eliminated
- The change is compatible with age confounding, but does not prove the adjusted estimate is the true causal effect
- The change proves effect modification on every scale
- Randomization occurred
Answer and explanation
B) The change is compatible with age confounding, but does not prove the adjusted estimate is the true causal effect. The causal role of age supports examining it as a confounder. Other assumptions and possible biases remain relevant.
Question 2: baseline risk or modification?
Older exposed participants have more outcomes than younger exposed participants. Is that fact alone evidence of effect modification?
Answer and explanation
No. Compare exposure groups within each age stratum on a specified scale. The difference could reflect baseline age-related risk rather than a different exposure effect.
Question 3: choose the scale
Subgroup A has risks of 20% without exposure and 10% with exposure. Subgroup B has risks of 40% and 20%. Compare risk ratios and risk differences.
Answer and explanation
Both risk ratios are 0.5; risk differences are −10 and −20 percentage points. The point estimates differ on the additive scale but not the risk-ratio scale. Their statistical uncertainty cannot be evaluated from these percentages alone.
Question 4: an odds-ratio change
A logistic model’s exposure odds ratio changes after a variable is added. Must that variable be a confounder?
Answer and explanation
No. Noncollapsibility and other modeling issues can change an odds ratio. Evaluate the variable’s causal role and the target estimand rather than applying a change threshold mechanically.
Sources and related reading
- CDC Field Epidemiology Manual: confounding and effect modification
- Schuster et al. (2021): noncollapsibility and confounding in logistic regression
- Hernán and Robins: Causal Inference: What If
- Methods to reduce bias
- Types of bias in epidemiology
Editorial update — October 11, 2026: consolidated the original and expanded guides; replaced the rigid 10% rule and “true effect” claims; added reproducible hypothetical calculations; and clarified scale, uncertainty and practice answers.
Humanities Moment
The featured image for this article is “‘Bloedberg’ in Antwerp” by Léon-Eugène-Auguste Abry. Léon-Eugène-Auguste Abry was a Belgian painter (1857–1905) best known for realist, often humorous and documentary depictions of military life and portraits rooted in academic training. He contributed to late 19th-century Belgian art by combining careful observation and sketching from actual troop maneuvers with polished academic technique, helping popularize authentic military genre scenes and informing public visual understanding of the army.
