Direct comparison
Effect Modification vs. Confounding
A confounder distorts your estimate, so you adjust for it. An effect modifier changes the true effect by subgroup, so you report it — never average it away.
Ask about Effect Modification vs. Confounding
Answers are drawn from this comparison and the rest of the CASRAI corpus, with a link to every source.
Answers are AI-generated from CASRAI’s own published pages and can be wrong, so check the linked sources before relying on one; your question is logged without personal data — never sold, never used to train a third-party model — to show us what CASRAI is missing, so please do not type personal or confidential details. How we use this
How do Confounding, Effect Modification compare side by side?
The table below compares Confounding, Effect Modification across 11 procurement-relevant dimensions, from what it actually is through where it's covered elsewhere on casrai.
Side-by-side comparison
| Dimension | Confounding | Effect Modification |
|---|---|---|
| What it actually is | A nuisance association: the third variable is tied to both the exposure and the outcome, so part of the crude association is really the third variable's effect leaking through. A property of how the study sample was assembled. | A real finding: the size, or direction, of the exposure's true effect genuinely differs by level of the third variable. A property of the thing being studied, not an artifact of the sample. |
| The three-part test that defines it | Associated with the exposure in this sample; an independent predictor of the outcome among the unexposed; and not a step on the causal pathway between exposure and outcome (a pathway variable is a mediator, not a confounder). | None of the above have to hold. A modifier does not need to be associated with the exposure at all — age, sex, or genotype can modify an effect while being unrelated to who ended up exposed. |
| How you find it | Compare the crude estimate to the estimate after adjusting for the candidate variable. A common rule of thumb: roughly a 10% or greater change signals confounding in this analysis — a heuristic, not a formal test. | Compute the estimate separately within each stratum, then formally test whether the stratum-specific estimates differ more than chance predicts — an interaction term, or a homogeneity test such as Breslow-Day. |
| The correct response once confirmed | Remove it. Adjust in the model, stratify and pool, match at the design stage, or restrict the population — then report one overall adjusted estimate. | Report it. Present the stratum-specific estimates side by side. Do not fold them into one pooled number — an average describes neither stratum accurately. |
| The wrong response | Reporting the crude, unadjusted association as the exposure's effect. Confounding by indication is the classic version — sicker patients preferentially get the more aggressive treatment. | Adjusting the modifier into the model as if it were only a confounder, then reporting one pooled effect — a number that can look reassuringly average while being wrong for both groups it was averaged from. |
| Can it be both at once? | Yes. A variable can meet the confounding test and get adjusted for, independent of what the modification test says about it. | Yes. The same variable can simultaneously show effect modification. The two tests are asked separately — passing or failing one is not evidence either way for the other. |
| Effect on a single summary number | A biased summary number — the crude estimate is systematically off from the true effect, in a direction the confounder's own associations determine. | No single summary number is right. Reporting one pooled effect, even an unbiased one, is a real information loss whenever the modifier changes the effect meaningfully. |
| Scale sensitivity | Not scale-dependent in the same way — a true confounder distorts the association on essentially any comparable scale. | Scale-dependent. A variable can show modification on the multiplicative scale (risk ratio) while showing little on the additive scale (risk difference), or the reverse. |
| Direction of the fix | Statistical: matching, stratification, restriction, multivariable adjustment, or, for unmeasured confounders, an instrumental-variable or negative-control design. | Presentational and analytic: stratified reporting or a pre-specified interaction term kept in the model — not a technique for removing bias, because there was none to remove. |
| What a causal diagram shows | A backdoor path: an arrow from the third variable into both the exposure and the outcome, open unless blocked by adjustment. | No backdoor path is implied. The modifier can sit entirely outside the exposure-outcome causal structure and still change the effect's size within its own strata. |
| Where it's covered elsewhere on CASRAI | See confounding variable, and, when a third variable's own path into the outcome is the real question, endogeneity: sources and remedies. | See mediator vs. moderator for the related interaction-term mechanics, and regression analysis for building the model that carries the interaction term. |
Common questions
Common questions about Confounding vs Effect Modification
Can a variable be both a confounder and an effect modifier at the same time?
+
Yes, and it happens often. The confounding test (does adjusting for it move the crude estimate?) and the modification test (does the effect differ meaningfully across its strata?) ask two separate questions about the same variable. A variable can pass both, one, or neither. When it passes both, the correct handling is to adjust for it as a confounder to get an unbiased estimate within each stratum, and still report the stratum-specific results separately rather than collapsing them, because the modification finding survives the adjustment.
What is the actual statistical test for effect modification?
+
Fit the model with a product (interaction) term between the exposure and the candidate modifier and test whether that term is significant, or, for stratified categorical data, run a formal homogeneity test such as Breslow-Day across the stratum-specific odds ratios. This is a different calculation from the change-in-estimate comparison used to check for confounding — one compares crude to adjusted, the other compares stratum to stratum.
Is effect modification the same thing as statistical interaction?
+
They're closely related but not defined identically. "Interaction" describes a model term — the coefficient testing whether the exposure's effect on the outcome depends on the level of another variable in that specific model. "Effect modification" is the substantive epidemiologic claim that a variable genuinely changes the exposure's effect, independent of which model or scale is used to detect it. In practice a significant interaction term is usually treated as evidence of effect modification, but the modification claim is the one that gets reported and interpreted; the interaction term is the mechanism that tests it.
Should an effect modifier ever be entered into the model as a plain covariate?
+
Only alongside an interaction term or stratified reporting, never instead of one. Adding a modifier as a plain covariate adjusts the model for it as though it were a confounder to be controlled for, which produces one pooled effect estimate and quietly discards the fact that the effect isn't uniform. If a variable has passed a modification test, the results have to show the stratum-specific estimates, not just a pooled number that happens to include the modifier as a covariate.
What is confounding by indication?
+
It's the pattern where clinicians preferentially prescribe a treatment based on a patient's underlying prognosis — sicker patients get the more aggressive option, or, in the opposite direction, a treatment is reserved for patients judged healthy enough to tolerate it. Either way, the patients who received the treatment differ systematically, in ways tied to prognosis, from those who didn't, and the crude treatment-outcome comparison is confounded by that underlying indication rather than reflecting the treatment's actual effect. It's one of the most common real-world sources of confounding in observational comparative-effectiveness research specifically because it's driven by clinical judgment that a simple demographic adjustment won't fully capture.
Does a non-significant interaction test mean there's no effect modification?
+
No, and treating it that way is a common overreach. Interaction tests are typically lower-powered than the corresponding main-effect test, because the product term is estimated from the variance left over after both main effects, so a study sized to detect the exposure's overall effect is routinely too small to detect a real difference in that effect across subgroups. A null interaction test is weak evidence of homogeneity, not proof of it, particularly when a subgroup is small.
Is the 10% change-in-estimate rule for confounding reliable on its own?
+
It's a useful screening heuristic, not a formal test, and it shouldn't replace reasoning through the causal structure first. The threshold is arbitrary, sensitive to sample size and to which other covariates are already in the model, and it can miss confounding that happens to leave the point estimate stable while still biasing the confidence interval. Identify candidate confounders from subject-matter knowledge and the causal diagram, then use the change-in-estimate comparison to confirm, not to discover, what to adjust for.
Going deeper
