Definition of Confounding in Research and Ways in Which It Can Be Managed

A lighter doesn’t cause cancer, but let’s say a study shows that people who carry lighters are more likely to get lung cancer. This is confounding. In research, confounding happens when a third variable affects the relationship between an exposure and an outcome.

This is common in observational and epidemiological studies, because, unlike randomized controlled studies, researchers don’t always get to design what participants do. As such, what appears to be a cause could actually be a harmless exposure, and what is a true cause could be hidden. In this class, you’ll learn how to uncover these causes.

What is Confounding in Research?

Confounding happens when a third variable is associated with both the exposure and the outcome. That third variable is known as the confounding variable. In observational studies, the difference in outcomes is likely due to confounding bias.

Three components are always present in a research study:

Exposure: What is being studied?

Outcome: What is measured?

Confounder: The third variable which is related to exposure and the outcome.

Smoking and cardiovascular disease is an example where older people do have more cardiovascular disease. Smoking behavior is different for older people vs. younger people. So if your smokers are older people than your nonsmokers, then part of the increased risk that smokers have for cardiovascular disease is due to age, and not smoking.

Of the different types of observational studies (cohort, case-control, and cross-sectional) case-control studies and cross-sectional studies are more susceptible to confounding bias because in these two study designs, exposure is not assigned, and for the reasons people smoke and the reasons for people’s use of medications.

Randomized trials guard against confounding because chance, not circumstance, decides who is exposed. Observational designs have no such protection. Cohort studies, case-control studies and cross-sectional studies all take exposure as they find it, and the reasons people smoke or take a medication are often the same reasons they fall ill.

What are the Criteria for a Confounding Variable?

Traditionally, a variable has to pass three tests to count as a confounder:

1. It’s associated with the exposure, so exposed and unexposed groups carry it in unequal amounts.
2. It’s an independent risk factor for the outcome, so it matters even in people never exposed.
3. It doesn’t sit on the causal pathway. If the exposure causes it on the way to the outcome, it isn’t a confounder.

Aging affects all three in the smoking example: it relates to smoking behaviors, raises the risk of cardiovascular diseases in nonsmokers, and smoking does not cause it.

Those tests are a good first step, but by themselves, numbers can’t indicate what to adjust for. That’s where modern causal inference relies on causal diagrams, specifically directed acyclic graphs (DAGs). You draw how the variables relate to one another, utilizing your clinical expertise, and the graph provides a suggestion as to what to adjust for.

Examples of Confounding in Epidemiology and Clinical Research

In the following examples, I’ll show how confounding changes the result of a study.

Example 1: Age and cardiovascular disease

Think of a study comparing myocardial infarction rates in active and inactive adults. If the inactive group is older, age is doing part of the damage. If you adjust for age, then the active group’s younger average age makes exercise appear more protective than it is.

Example 2: Smoking and lung cancer

Smoking causes lung cancer, but underestimating the true effect in a study is inevitable. Sociocultural factors and occupational factors must also be considered when studying smoking and health. People working in certain occupations might smoke more and be also be exposed to asbestos at work. If you don’t account for occupational factors, you might misattribute certain cancers to smoking.

Example 3: BMI and chronic kidney disease

Higher BMI is associated with chronic kidney disease. High blood pressure could be the mechanism through which this happens. Therefore, in this case, high blood pressure is a mediator. Once you adjust for a mediator, part of the total effect is removed, and the total risk becomes less than it really is. Overadjustment bias is the term used to describe this error.

Confounding vs Selection Bias vs Recall Bias

All three are systematic errors, but each one gets into a study by a different door.

Type	Meaning	Example
Confounding bias	A third variable distorts an exposure-outcome association	Age affecting smoking and cardiovascular risk
Selection bias	Participant selection creates a distorted association	Selecting hospital controls who differ systematically from the source population
Recall bias	Exposure information is remembered or reported differently between groups	Cancer patients recalling past exposures differently from controls

Confounding depends on the relationship between variables in the real world. Selection bias depends on who participates in the study, and recall bias is about how exposures are remembered. It’s possible for a single study to include all three. Fixing one doesn’t affect the other two. Adjustment can help deal with measured confounding, but it doesn’t address sampling or exposure issues.

Appraisal tools check for all three. Risk of bias assessment in systematic reviews covers how ROBINS-I treats confounding as its first domain and how the Newcastle-Ottawa Scale scores comparability.

How to Identify Confounding Factors in Clinical Studies

This largely depends on what’s previously known about the disease and about the exposure. The other step is to:

1. Determine your research question and your exposure-outcome relationship
2. Review the literature and clinical knowledge
3. Come up with potential common causes of exposure and outcome
4. Make a directed acyclic graph (DAG)
5. Decide on an adjustment set. Ideally, this should be the smallest set possible.
6. Check that the confounders you’ve identified are adequate.

The first item on that list is only as good as your question. The PICO framework and its observational variants cover how to define population, exposure and outcome precisely enough to draw the diagram.

A DAG is just a drawing. You show causal relationships by drawing arrows. If you have confounder C pointing to both exposure X and outcome Y (X ← C → Y), this creates a backdoor path and you need to block it. A mediation relationship looks different (X → M → Y) and adjusting for a mediator creates overadjustment. There are other relationships that are similar to a collider relationship where adjusting for creates collider bias.

How to Control for Confounding Variables in Research

You have two opportunities to deal with confounding variables: during study design and data analysis. Well designed studies do both.

A. Control during Study Design

 Randomization: Groups are balanced and confounders are approximately even. Randomization is effective and the strongest method but it may not always be practical or ethical.

Restriction: When enrollment is restricted to a single level of a confounder (e.g. enrolling only participants in the 50-59 age group). Confounding is minimized but the sample size is restricted and the findings may not be generalizable.

Matching: Enrolling a balanced number of participants with the exposure and a confounding variable. Participants must also be matched for other possible confounders (e.g. age and sex).

Matching has its own rules in case-control studies, where matching on the wrong variable can hide a risk factor you wanted to study. Matching in case-control studies covers individual and frequency matching, overmatching and the analysis that goes with each.

B. Control during Data Analysis

 Stratification: Separates subjects by levels of a confounder and analyzes exposed and unexposed separately within each level of the confounder.

 Standardization: Distributes confounders equally within the exposure and comparison groups so that the exposure rates in the groups can be meaningfully compared.

Multivariable regression: Estimates effect of exposure while controlling for potential confounding variables.

 Propensity score analysis: The idea is similar to multivariable regression but focuses on reducing the many potential confounding variables to a single score for each study participant.

When designing a study, know what variables you’ll adjust for, provide a rationale for each and name the variables. There are common pitfalls when designing a study. For instance, it is common to adjust for everything and anything. This gives opportunity for mediators and colliders to creep into the model. Measure what you think may be confounders of the study because no analysis can fix an unmeasured confounder.

In a chart review that sentence is a hard limit, because you can only adjust for what the record contains. How to do a retrospective chart review covers building a data collection form that captures likely confounders before extraction starts.

Statistical Methods to Adjust for Confounding

Statistical adjustment asks what the exposure-outcome association would look like if the groups were alike on measured confounders. Which method fits depends on your study design, your sample size, and how many confounders you have.

Method	How it works	Main limitation
Stratification	Divides participants into groups based on confounder levels	Hard to manage with many confounders
Multivariable regression	Adjusts for multiple covariates within a statistical model	Depends on a correctly specified model
Standardization	Estimates outcomes using a common distribution of confounders	Becomes difficult with many confounders
Propensity score matching	Matches participants with similar probabilities of receiving an exposure or treatment	Unmatched participants may be dropped
Inverse probability weighting	Uses propensity-score-based weights to create a weighted population	Extreme weights can make estimates unstable

Propensity score matching comes up a lot in observational treatment-effect studies, where sicker patients may be likelier to get a certain treatment. Watch for three problems:

• Poor overlap: Treated and untreated patients are too different to compare.
• Inadequate covariate balance: The groups still differ after matching or weighting.
• Unmeasured confounding: The score can only use what was recorded.

Whatever distortion survives adjustment is called residual confounding. Sensitivity analyses can show how strong an unmeasured confounder would have to be to change your conclusion.

What is Simpson’s Paradox in Confounding?

Simpson’s paradox is what you get when an overall association reverses or changes direction once the data are split by a third variable. A classic kidney stone dataset shows how:

• All patients: percutaneous nephrolithotomy 83%, open surgery 78% success
• Small stones: open surgery 93%, percutaneous nephrolithotomy 87%
• Large stones: open surgery 73%, percutaneous nephrolithotomy 69%

Open surgery won in both subgroups yet lost overall. Why? It was used more often for large stones, which are harder to treat. Stone size is the confounder here, so always check subgroup distributions before trusting observational data.

How to Report Confounding in a Clinical Research Paper

The STROBE Statement is a reporting guideline for observational studies, not a statistical method for controlling confounding. It tells authors what to include so readers can judge how well confounding was handled. In your paper:

• Define your potential confounders and say why you chose them.
• Explain how you controlled confounding.
• Give both crude and adjusted estimates, so readers can see what adjustment changed.
• Include confidence intervals.
• Describe any sensitivity analyses.
• Admit what’s still uncertain, including possible residual confounding.

Frequently Asked Questions

Q1. What is an example of confounding in medical research?

Think of age in a study on smoking and cardiovascular disease. It’s linked to both, so it can distort the result.

Q2. What is the difference between a confounder and a mediator?

A confounder affects both exposure and outcome. A mediator sits between them and carries part of the exposure’s effect. Effect modification is different: the exposure’s effect truly differs between subgroups.

Q3. Can regression analysis eliminate all confounding?

No. Regression only handles variables you measured, so unmeasured and residual confounding can stay behind.

Q4. What is the best statistical method to control for confounding?

There isn’t one. The right choice depends on your study design, sample size, and which confounders you measured.

Spot confounders early, be honest about what remains uncertain, and readers can trust your findings.

Building your research profile for the US residency match? IMG Helping Hands guides international medical graduates through the USMLE pathway and residency applications. Visit IMG Helping Hands to plan your next step.

Sources

1. Jager KJ, Zoccali C, MacLeod A, Dekker FW. Confounding: what it is and how to deal with it. Kidney Int. 2008;73(3):256-260.
2. Kahlert J, Gribsholt SB, Gammelager H, Dekkers OM, Luta G. Control of confounding in the analysis phase: an overview for clinicians. Clin Epidemiol. 2017;9:195-204.
3. Feeney T, Hartwig FP, Davies NM. How to use directed acyclic graphs: guide for clinical researchers. BMJ. 2025;388:e078226.
4. Desai RJ, Franklin JM. Alternative approaches for confounding adjustment in observational studies using weighting based on the propensity score: a primer for practitioners. BMJ. 2019;367:l5657.
5. Hajian-Tilaki K. Methodological issues of confounding in analytical epidemiologic studies. Caspian J Intern Med. 2012;3(3):488-495.
6. Julious SA, Mullee MA. Confounding and Simpson’s paradox. BMJ. 1994;309(6967):1480-1481.
7. Vandenbroucke JP, von Elm E, Altman DG, et al. Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): explanation and elaboration. Epidemiology. 2007;18(6):805-835.
8. Rothman KJ, Greenland S, Lash TL. Modern Epidemiology. 3rd ed. Lippincott Williams & Wilkins; 2008.

Disclaimer:

Articles published by American Academy of Research & Academics are prepared by our team using information from direct experience, publicly available resources, and educational references. AI tools may be used to assist with drafting, proofreading, and formatting; however, all content undergoes review and approval before publication.
The information provided is intended for educational purposes only. Requirements, policies, and processes may change over time. Readers should consult official sources for the most current information.