Confounding in observational studies happens when a common cause of the exposure and the outcome distorts the association between them. Report confounding and bias by sorting each threat into selection bias, information bias or confounding, explaining how you chose adjustment variables, giving unadjusted and adjusted estimates with confidence intervals, and stating the likely direction of any residual bias.
Bias is systematic error, which a larger sample would not remove. STROBE (2007), the reporting guideline for cohort, case-control and cross-sectional studies listed by the EQUATOR Network, asks you to address it throughout the paper. If your sample is too small for an adjusted model, see our guide to decoding statistical reviewer comments.
Sort each threat into selection bias, information bias or confounding
Naming the kind of bias tells a reader what could have gone wrong and which way it would push the estimate.
- Selection bias comes from how people entered or left the analysed sample, such as restricting a cohort to complete records or losing the sickest patients to follow-up.
- Information bias comes from errors in measuring the exposure, outcome or covariates. Error that differs between groups, such as cases recalling past exposures more completely than controls, can push an estimate either way. Unblinded abstraction is another source; writing the Methods for a retrospective chart review covers how to report it.
- Confounding comes from a common cause of the exposure and the outcome that the analysis has not accounted for.
Decide whether each variable is a confounder, mediator or collider
A confounder is a common cause of the exposure and the outcome. A mediator is caused by the exposure and in turn causes the outcome, so it lies on the causal path between them. A collider is caused by two or more other variables, for example by the exposure and by a cause of the outcome.
Adjusting for a confounder that is well measured and appropriately modelled removes the bias it causes. Adjusting for a mediator, when you want the total effect, removes part of the effect you set out to measure: overadjustment bias. Adjusting for a collider, stratifying on it or selecting the sample on it can create an association that did not exist: collider bias. So "associated with both exposure and outcome" is not a safe test for a confounder, because mediators and colliders pass it too.
Choose confounders from causal knowledge or explicit assumptions, not from P values or from which variables differed "significantly" at baseline. The STROBE explanation and elaboration paper also asks you to list every potential confounder you considered and the criteria for including or excluding each.
The study below is invented to illustrate the structure; its numbers are not real data. A retrospective cohort of 1,240 adults asks whether pre-operative anaemia is associated with readmission within 30 days of hip fracture surgery. The roles are the investigators' assumptions.
| Variable | Assumed causal role | Adjust for it? |
|---|---|---|
| Age | Confounder: affects anaemia and readmission | Yes |
| Chronic kidney disease | Confounder: affects anaemia and readmission | Yes |
| Frailty, recorded only as pre-fracture mobility | Confounder, measured imperfectly | Yes, via the proxy; name the gap as a limitation |
| Blood transfusion during the admission | Mediator: caused by anaemia, assumed to affect readmission | No, because the aim is the total effect |
| Discharge destination | Collider: influenced by anaemia and by social circumstances that also affect readmission; possibly a mediator too | No, nor restrict the cohort by it |
Draw a directed acyclic graph before you choose the adjustment set
A directed acyclic graph (DAG) is a diagram of the assumed causal structure. Variables are nodes, and each one-way arrow is an assumed causal effect that says nothing about its size, sign or shape. No variable can cause itself, hence acyclic. A DAG records assumptions, not evidence.
In the invented study, age, chronic kidney disease and frailty each point to anaemia and to readmission. Anaemia points to transfusion, which points to readmission. Anaemia and social circumstances both point to discharge destination, and social circumstances also point to readmission. For the total effect, the implied adjustment set is age, chronic kidney disease and frailty. That set also rests on an omitted arrow: social circumstances are assumed not to affect anaemia; if they did, they would be a confounder too.
A 2021 review of DAG use in applied health research makes reporting recommendations, including these:
- State the estimand (the quantity you intend to estimate) in the aims.
- Make the DAG available, in the paper or its supplement.
- Include every plausible confounder, even unmeasured ones, and justify any arrow you leave out.
- Report the estimate from the DAG-implied adjustment set; justify any alternative set and report its estimates separately.
Name confounding by indication when treatment choice follows prognosis
Our guide to choosing a study design notes that the reason one group received a treatment may itself predict the outcome. That is confounding by indication: treatments are chosen for reasons, such as severity, that also shape outcomes. To report it:
- Describe how treatment decisions were made, and by whom.
- List the severity markers you adjusted for and those the records lacked.
- Where possible, compare two treatments used for the same indication rather than treated with untreated patients; this can reduce the problem but may not remove it.
- State the expected direction: if sicker patients were more often treated, residual confounding tends to make the treatment look worse than it is.
Matching, regression and propensity scores adjust only for measured confounders
Stratification, regression, matching and propensity scores share one limit: they handle only confounders that were measured, and measured well. A propensity score is the estimated probability of receiving the exposure given the measured baseline covariates, so matching on it need not balance a confounder missing from the dataset. Our decision guide to choosing a statistical test covers which regression model suits each outcome.
After propensity score matching, report balance with standardised differences, not significance tests; the matched sample is smaller, so a non-significant test may reflect only its size.
Give unadjusted and adjusted estimates, and show how far adjustment moved them
STROBE item 16(a) asks for unadjusted and, if applicable, confounder-adjusted estimates with their precision, and which confounders were adjusted for and why. The explanation and elaboration paper adds the number of people in each adjusted analysis and confidence intervals (CIs) for each estimate, so readers can judge how much, and in which direction, adjustment moved it.
In the invented study, the unadjusted odds ratio was 1.80 (95% CI 1.30 to 2.49) in all 1,240 patients. Adjusted for age, chronic kidney disease and pre-fracture mobility, it was 1.42 (1.01 to 2.00) in the 1,112 with complete covariates. An unadjusted estimate in those same 1,112 patients separates the effect of adjustment from that of the smaller sample.
Report the exposure's estimate as the result. Reading the coefficients of the other covariates as their causal effects, as if each were an independent risk factor, is the Table 2 fallacy: the adjustment set chosen for the exposure is not generally right for the other covariates.
State the direction and likely size of residual bias
Adjusting for measured confounders does not make an association causal; residual confounding, random error, selection bias and information bias can remain. Nor is it true that observational data can never inform a causal question: they can, under assumptions you state explicitly. STROBE item 19 asks for the direction and magnitude of any potential bias, and item 20 for a cautious overall interpretation.
Vague: Residual confounding cannot be excluded.
Specific (invented study): Frailty was recorded only as pre-fracture mobility. Because frail patients are assumed to be more often anaemic and more often readmitted, residual confounding by frailty would most likely make the adjusted odds ratio of 1.42 an overestimate.
Support the statement with a sensitivity analysis where you can, such as repeating the model with a second frailty measure. How to write a limitations section covers pairing each limitation with its consequence, and writing a conclusion that matches your data covers the final claim.
Copy this checklist for reporting confounding in observational studies
Each row maps to a STROBE item. Examples use the invented study, whose numbers are not real data.
| Where | STROBE item | What to write | Example sentence (invented study) |
|---|---|---|---|
| Methods: variables | 7 | Define exposure, outcome, potential confounders and effect modifiers | "Confounders (age, chronic kidney disease, frailty) were chosen from the DAG in the supplement." |
| Methods: bias | 9 | Efforts to address potential sources of bias | "Readmission was abstracted by a researcher blind to haemoglobin results." |
| Methods: statistics | 12(a) | Methods used to control for confounding | "The model included the DAG-implied adjustment set; transfusion, a mediator, was not included." |
| Results: descriptive data | 14(a), 14(b) | Characteristics and confounders by exposure group; missing data per variable | "Table 1 shows each covariate by anaemia status, with the number missing." |
| Results: estimates | 16(a) | Unadjusted and adjusted estimates with precision; which confounders and why | "Unadjusted odds ratio 1.80 (95% CI 1.30 to 2.49; n = 1,240); adjusted 1.42 (1.01 to 2.00; n = 1,112)." |
| Discussion: limitations | 19 | Direction and magnitude of potential bias | "Residual confounding by frailty would most likely make 1.42 an overestimate." |
| Discussion: interpretation | 20 | A cautious overall interpretation | "Anaemia was associated with readmission; causality is uncertain." |
Our author guidelines ask you to match the manuscript to the relevant EQUATOR checklist; for these designs, that is the STROBE checklist. If you spot an error in this guide, report the problem to us.
Reviewers who include an observational study in a systematic review judge these same threats with a structured tool; choosing a risk of bias tool explains when ROBINS-I or Newcastle-Ottawa applies and how judgements are reported by domain.