13 Research methods, statistics, critical review and evidence-based practice13.2 Epidemiological studies
13.2.1.2 Confounding and its control
4 min read2 figures
Written by Resident Doctor Committee · Published
Defining confounding
A confounder distorts an observed exposure-outcome association because it is linked to the exposure and to the outcome independently of the exposure. It does not lie on the causal pathway.
A spurious association is an apparent relationship that is not real. A confounder acting on both variables, or chance, produces it, and it disappears once the confounder is held constant.
A mediator lies on the causal pathway, so an indirect association through it is real. Antipsychotic weight gain leading to diabetes is an example, and adjusting for a mediator removes part of the total effect.
Population size does not confound a comparison of crude rates, because it is the denominator of each rate. Differences in age, sex, ethnic and socio-economic composition do confound it.
Confounding can create, mask, exaggerate or reverse an association.
Simpson's paradox is a reversal. Lithium augmentation can show higher remission in each severity stratum but lower crude remission when more lithium recipients have the harder-to-treat presentation.
Confounding in observational comparisons
In an observational comparison the groups differ in other characteristics that independently affect the outcome, so confounding may produce the association.
A significant difference between matched cases and controls does not establish causation while confounders remain unmeasured and unadjusted.
Confounding by indication arises when prognosis shapes which drug a patient is given. It is the main threat to naturalistic registry-based comparisons of psychotropic drugs.
Comparing outcomes by observed adherence within a randomised trial reintroduces confounding. Patients choose adherence rather than being allocated to it, and motivation, insight and support may lower risk independently.
A matched, adjusted nested case-control study built on prospectively recorded data guards against recall, information, selection and attrition bias. It is observational, so unmeasured or residual confounding remains possible.
Reading crude and adjusted estimates
Crude and adjusted result | Interpretation |
|---|---|
Crude significant, adjusted interval includes 1 | The adjusted confounders explain the apparent effect |
Adjusted interval still excludes 1 | The exposure is more likely to be a genuine independent predictor |
Small shift towards the null, adjusted estimate still well away from 1 | The measured confounders explain only part of the association |
Significance depends on the covariates in the model. An exposure can survive adjustment for pre-specified confounders yet lose significance once correlated co-exposures are added.
Residual confounding can remain after adjustment from unmeasured factors, coarse categories, measurement error or wrong timing, so adjustment does not prove that confounding is absent.
Effect modification
An effect modifier is a third variable across whose levels the strength of an association genuinely differs, so the stratum-specific estimates differ from each other.
Confounding shows as a crude estimate that differs from the adjusted estimate.
Stratified analysis can show both: confounding when the pooled stratum estimate differs from the crude, and effect modification when the stratum estimates differ from each other.
Confounding is a distortion to remove, whereas effect modification is a real finding to report.
Controlling confounding
Design-stage methods act before data collection. Analysis-stage methods need the collected data.
Stage | Method | How it works |
|---|---|---|
Design | Randomisation | Balances confounders between groups rather than eliminating them |
Design | Restriction | Enrols only people at one level of a suspected confounder, removing its variation from the sample |
Design | Matching | Selects comparable participants, such as cases and controls of similar age and sex |
Analysis | Stratification | Compares exposure groups within each level of the confounder |
Analysis | Multivariable adjustment | Regression or analysis of covariance (ANCOVA) adjusts for several measured variables at once |
Analysis | Standardisation | Compares rates using a shared population distribution, usually of age |
Randomisation works because a confounder distorts the estimate only when it is unevenly distributed between the groups compared.
Balance after randomisation holds on average, so chance baseline imbalance can occur, especially in small trials.
Matched data need a matched analysis, such as conditional logistic regression for matched case-control data.
The Mantel-Haenszel method calculates an odds ratio within each stratum of the confounder and pools them into one weighted, adjusted summary odds ratio.
Stratification fails when there are many confounders, because the strata become numerous and sparse.
Methods that do not control confounding
Method | What it does instead |
|---|---|
Blinding | Reduces observer and recall bias; blinded outcome raters guard against observer bias |
Enlarging the sample | Reduces random error only |
Agreement and reliability statistics | Measure repeatability |
Meta-analytic pooling model | Handles between-study heterogeneity |