What is Dr. Zubair Khalid's research focus?

Dr. Zubair Khalid specializes in molecular virology, mRNA vaccine development, and computational biology, with a focus on avian pathogens like IBDV and Avian Reovirus.

Where is Dr. Zubair Khalid currently working?

Dr. Zubair Khalid is a Postdoctoral Research Associate at the University of Maryland (UMD), specifically within the Department of Animal and Avian Sciences.

Confounding Factor: Definition, Examples, and Control

By Dr. Zubair Khalid, DVM, MS, PhD · 2026-09-19

A confounding factor is a variable that is associated with the exposure and independently associated with the outcome, but does not lie on the causal pathway between them. Because it is linked to both sides of the comparison, a confounder can create, inflate, shrink, or reverse an apparent association even when no causal relationship exists.

Confounding matters because it is the single most common reason a real-world observation about diet, housing, breed, or treatment turns out to be wrong. Unlike random error, which scatters around the truth, confounding pushes results in a consistent direction, and no amount of extra sample size removes it [1]. A study of 10,000 dogs can be just as biased as a study of 50 if the groups differ on a variable that predicts the outcome.

The Formal Definition and the Three Conditions

A variable must satisfy three conditions at the same time to be a confounder. If any one fails, it is something else.

It is associated with the exposure. Dogs on the suspect diet are more likely to be a particular breed, or older, or fed in a particular household.
It is a cause of the outcome, independent of the exposure. The variable raises disease risk by a route that does not run through the diet.
It is not on the causal pathway. The variable does not sit between the exposure and the outcome as a step in the mechanism.

Condition three is the one students miss most often. Age confounds the diet-pancreatitis relationship because age is tied to which dogs eat the diet and age independently raises pancreatitis risk. By contrast, a variable such as reduced appetite sits downstream of pancreatitis, so it is a consequence, not a confounder. Adjusting for it would remove part of the very effect you want to measure.

Why a Single Confounder Can Flip a Conclusion

Confounding is a structural problem, not a statistical nuisance. The crude comparison between exposed and unexposed groups mixes the effect of the exposure with the effect of every variable that differs between the groups. If those variables predict the outcome, the crude estimate is a blend, not a clean effect.

A vaccination example shows how far a confounder can move an estimate. In test-negative studies of influenza and COVID-19 vaccine effectiveness, correlated vaccination behaviors act as a confounder. When these behaviors are not addressed, studies can underestimate vaccine effectiveness, and the size of the error depends on how much the two vaccination behaviors move together [2]. The bias is directional and predictable, which is exactly what makes it dangerous: a reader sees a plausible number and has no way to know it is distorted.

Worked Example: Does Diet Cause Pancreatitis in Dogs?

Suppose a researcher wants to test whether a specific diet causes pancreatitis in dogs. The plan is simple: recruit dogs eating the diet, recruit dogs on other diets, and compare pancreatitis rates. The crude result shows a strong positive association. The relative risk is 3.0, meaning diet-fed dogs appear three times as likely to develop pancreatitis.

Now consider age. Suppose the diet is a premium product marketed for adult and senior dogs, so diet-fed dogs are older on average. Age is associated with the exposure. Age also independently raises pancreatitis risk. Age is not a step in the diet-to-pancreatitis mechanism. Age is a confounder.

If the researcher stratifies by age, the picture changes. Within each age band, the relative risk falls to about 1.1. The crude association was largely an artifact of comparing an older group to a younger group. Age did not cause the pancreatitis association, but it manufactured one.

Breed works the same way. If miniature schnauzers are overrepresented in the diet group and also have a higher baseline risk of pancreatitis, breed is a confounder. The crude estimate mixes diet effects with breed effects, and the reader cannot tell which is doing the work.

This is the core lesson. A confounder distorts results by being unevenly distributed across comparison groups and by independently predicting the outcome. The fix is to measure it and design around it, not to hope it averages out.

Confounding Is Not Effect Modification

These two terms confuse more students than any other pair in epidemiology. They are fundamentally different.

A confounder is a nuisance you want to remove. An effect modifier is a real finding you want to report. Effect modification occurs when the strength or direction of the exposure-outcome relationship genuinely differs across levels of a third variable.

Using the diet-pancreatitis example, suppose the diet elevates pancreatitis risk in miniature schnauzers by a relative risk of 4.0 but has no effect in Labrador retrievers, with a relative risk of 1.0. Breed is now an effect modifier. The diet really does behave differently in the two breeds. You do not adjust this away, because there is no single true effect to estimate. You report the diet effect separately for each breed.

The practical test is this. If you stratify and the stratum-specific estimates differ meaningfully from one another, you have effect modification and you should present the strata separately. If the stratum-specific estimates are similar and both differ from the crude estimate, you have confounding and you should pool the strata with an adjusted estimate. On the additive scale, effect modification is the presence of interaction, while confounding is a bias to be corrected.

Confounding Is Not Mediation

Mediation describes a causal mechanism, while confounding describes a spurious or mixed association. A mediator sits on the pathway between exposure and outcome.

If the pancreatitis diet causes obesity, and obesity in turn causes pancreatitis, then obesity is a mediator. The causal chain runs diet to obesity to pancreatitis. Adjusting for a mediator blocks part of the exposure's effect and usually underestimates the total effect of the diet. That is a serious error, and it happens whenever a researcher mechanically adjusts for every variable that shows a statistical association with the outcome.

The distinction comes down to causal ordering. A confounder sits before or beside the exposure and predicts the outcome by a separate route. A mediator sits between the exposure and the outcome. You control confounders. You never control mediators if you want the total effect of the exposure.

Confounding Is Not Selection Bias

Confounding and selection bias are distinct problems that frequently occur together. Confounding compromises internal validity, meaning the comparison within the study is distorted. Selection bias compromises external validity, meaning the study population is not representative of the target population, so the result may not generalize [3].

A good design controls both. Strong confounding control does nothing to fix a study that enrolled only referral-hospital dogs, and a representative sample cannot rescue a study whose groups differ on a strong risk factor. Researchers should select study animals only after careful consideration of the primary objective and the intended target population, and they should reduce information bias through standardized data collection and blinding [1].

How to Detect Confounding in Practice

The first step is to draw the causal relationships before collecting data. A directed acyclic graph (DAG) is a simple causal diagram in which arrows represent presumed causal effects. Any variable that has arrows pointing into both the exposure and the outcome, and no arrow from exposure into that variable, is a candidate confounder. Any variable on the path from exposure to outcome is a mediator and must not be adjusted.

The second step is to measure candidate confounders in every study group with the same method. If age is recorded more completely in the diet group than the control group, the confounding control is compromised before analysis begins.

The third step is to compare crude and adjusted estimates. If the adjusted estimate barely moves, confounding is limited. If it moves substantially, confounding was present.

A useful complementary tool is the negative control. A negative control outcome is an outcome that the exposure cannot plausibly cause but that is expected to be affected by the same confounders. Negative controls help identify confounding and other sources of noncausal association in observational studies [4]. If the exposure shows an association with a negative control outcome, confounding by an unmeasured factor is likely.

Control Methods: Design and Analysis

Confounding can be addressed before or after data collection. Design-phase methods are generally stronger because they reduce confounding by construction rather than by modeling.

Randomization

Randomization assigns animals to exposure groups by chance, which balances both measured and unmeasured confounders across groups on average. It is the strongest single method because it does not require knowing what the confounders are. When randomization is applied properly, it eliminates most noncausal associations [4]. The limitation is ethical or practical. You cannot randomize a dog to develop pancreatitis, and many diet exposures cannot be assigned.

Restriction

Restriction limits enrollment to animals within a single level of a confounder. If age confounds the diet association, you enroll only dogs in one age band. Restriction removes the confounder's variability entirely. The cost is a smaller sample and reduced generalization to other ages. In one breast cancer study, restriction to high-risk women changed a crude hazard ratio of 2.6, suggesting harm, to 1.1, suggesting no association [5].

Matching

Matching selects unexposed animals that share confounder values with exposed animals, often on a one-to-one basis. Matching is efficient when the confounder is easily measured. It can be cumbersome with many confounders, and it makes case-control analysis trickier because the matching must be accounted for in the analysis, typically with conditional methods.

Stratification

Stratification divides the sample into levels of the confounder and estimates the effect within each level. It is transparent and easy to explain. The limitation is a small sample within each stratum, and it becomes unwieldy with more than a few confounders.

Multivariable Regression

Multivariable regression estimates the exposure-outcome association while statistically holding other measured variables constant. It handles many confounders at once and uses the full sample. The limitation is that it controls only for variables you measured and do not know the correct functional form of, and it does nothing for unmeasured confounders. Residual confounding is expected in every observational study regardless of how many variables are adjusted for [6].

Propensity Scores

A propensity score is the probability of receiving the exposure given the measured confounders. It condenses many confounders into a single number that can be used for matching, stratifying, or weighting. Propensity scores can balance groups well when confounders are numerous. However, in one study of confounding by indication, propensity-score adjustment yielded results similar to multivariable regression and did not resolve the underlying problem, because unmeasured factors remained [5]. Like regression, propensity scores address only measured confounding.

Sensitivity Analysis

Sensitivity analysis tests how strong an unmeasured confounder would have to be to overturn the result. Simulation-based bias analysis can quantify the potential impact of unmeasured confounding at the design stage and can show how much adjustment for a correlated proxy variable reduces bias [7]. The limitation is that the inputs are assumptions, not observations, so the output is a judgment about robustness, not a corrected estimate.

Instrumental Variables and Quasi-Experimental Designs

When unmeasured confounding is suspected, design-based approaches can estimate causal effects under weaker assumptions. Instrumental variables, difference studies, interrupted time series, and regression-discontinuity designs use a source of variation that is plausibly unrelated to the confounders [8]. Regression discontinuity, for example, uses a cutoff on a continuously measured variable to create a comparison that is nearly random close to the cutoff, provided the variable is not manipulated [9]. These methods are powerful but require strong, testable assumptions and are easy to misuse.

Method	When to use	Main limitation
Randomization	You can assign the exposure ethically and practically	Cannot assign harmful exposures, not always feasible
Restriction	A strong confounder has a manageable number of levels	Reduces sample size and generalizability
Matching	The confounder is easy to measure and groups are comparable	Hard with many confounders, complicates case-control analysis
Stratification	Few confounders, moderate sample sizes	Small strata, unwieldy beyond a few variables
Multivariable regression	Many measured confounders, full sample needed	Residual and unmeasured confounding remain
Propensity scores	Numerous confounders, balancing groups is the goal	Does not control unmeasured confounding
Instrumental variables and quasi-experiments	Unmeasured confounding is likely	Strong assumptions, hard to verify
Sensitivity analysis	Any observational result needs a robustness check	Inputs are assumptions, not data

flowchart TD
    A[Ask your research question] --> B[Draw the causal diagram]
    B --> C{Variable linked to exposure and outcome}
    C -->|No| D[Not a confounder]
    C -->|Yes| E{On the causal pathway}
    E -->|Yes| F[Mediator - do not adjust]
    E -->|No| G[Confounder - control it]
    G --> H[Choose a design method]
    H --> I[Restriction or matching or randomization]
    I --> J[Choose an analysis method]
    J --> K[Stratification or regression or propensity score]
    K --> L[Run sensitivity analysis]

This flowchart summarizes the decision path from defining a research question to choosing and checking a control strategy.

The Collider Trap: Why Bad Adjustment Makes Things Worse

Adjusting for the wrong variable can introduce bias when none existed. A collider is a variable that is caused by both the exposure and the outcome, or by two variables that also affect the outcome. Conditioning on a collider opens a noncausal path between the exposure and the outcome.

Consider a study of diet and pancreatitis in which a third variable, whether the dog was referred to a specialty hospital, is caused by both the diet history and the pancreatitis diagnosis. Referral status is a collider. If you restrict your sample to referred dogs, you induce an association between diet and pancreatitis that was not there in the full population. This is known as collider bias or selection bias of the conditioning type, and it is one reason researchers must resist the urge to adjust for every available variable.

The same logic explains why adjusting for a mediator biases the total effect. Both errors arise from failing to use causal reasoning before choosing covariates. Statistical software will happily produce an adjusted estimate whether the adjustment is valid or not.

Common Mistakes and Limitations

Adjusting for variables without a causal rationale. Selecting covariates because they correlate with the outcome is a common shortcut. It invites adjustment for mediators and colliders, which biases results [3].

Treating any associated variable as a confounder. A variable must be associated with the exposure, predict the outcome, and lie off the causal pathway. Correlating with the outcome is not enough.

Confusing confounding with effect modification. If the effect genuinely differs across strata, reporting a single pooled estimate hides a real finding.

Assuming more adjustment is better. Over-adjustment can block causal pathways and can condition on colliders. Precision also falls as you add covariates.

Believing adjustment fixes everything. No analytic method controls unmeasured confounders. Conventional methods do not control for unmeasured factors, which often remain important, especially with confounding by indication [5]. Even after adjusting for many variables, residual confounding by unknown or unmeasured factors is likely [6].

Ignoring negative controls. Negative controls are underused, and they can reveal confounding that standard adjustment misses [4].

Underreporting in the literature. In a review of high-impact observational psychiatry studies, only 55.0% of articles explicitly mentioned confounding in the Abstract or Discussion, and only 20.8% acknowledged unadjusted confounders [10]. Readers should look for that discussion and treat its absence as a warning.

Underpowered strata. Fine stratification can leave each stratum too small to estimate anything useful.

Confounding by indication. When the reason for treatment is itself a risk factor for the outcome, treatment groups differ by severity. This needs to be handled at the design stage so that groups include the same range of condition severity [6].

A Checklist of Study-Design Controls

Use this at the protocol stage, before data collection.

Draw the causal diagram and list every arrow.
Label each variable as confounder, mediator, collider, or effect modifier.
Decide which confounders you will measure, and measure all of them in all groups with the same method.
Choose a design-phase control if one is feasible: randomization, restriction, or matching.
Plan the analysis-phase control: stratification, multivariable regression, or propensity scores.
Pre-specify the primary confounders. Do not decide after seeing the results.
Add a negative control outcome.
Plan a sensitivity analysis for unmeasured confounding.
Pre-specify any subgroup analysis that could reveal effect modification.
Document how unmeasured confounding and residual confounding limit the conclusions.

Frequently Asked Questions

What is a confounding factor in simple terms?

A confounding factor is a variable tied to both the exposure and the outcome but not part of the causal path between them. It makes two groups differ in a way that has nothing to do with the exposure being studied.

How is confounding different from effect modification?

Confounding is a bias that distorts a single true effect, so you adjust for it. Effect modification is a real difference in the effect across groups, so you report it separately rather than adjusting it away.

Can you adjust for a variable that is a mediator?

No. Adjusting for a mediator removes part of the exposure's effect and biases the total effect estimate. A mediator sits on the causal pathway, so it should not be treated as a confounder.

Why does randomization control confounding?

Randomization assigns exposure by chance, so measured and unmeasured confounders balance across groups on average. It is the only common method that addresses confounders you never measured.

Does adjusting for many variables remove all confounding?

No. Adjustment only addresses variables you measured and modeled correctly. Residual and unmeasured confounding remain possible in every observational study [6].

Can a confounder make a true effect look smaller or larger?

Both. Confounding can inflate, shrink, create, or reverse an apparent association depending on how the confounder distributes across groups and how it affects the outcome.

Quick Review

A confounder is associated with the exposure and the outcome, but not on the causal pathway.
Confounding is systematic and directional, so larger samples do not fix it [1].
Effect modification is a real finding. Confounding is a bias to remove.
Mediation is mechanism. Never adjust for a mediator when you want the total effect.
Randomization is the strongest control because it handles unmeasured confounders.
Restriction, matching, stratification, regression, and propensity scores all work only on measured confounders.
Adjusting for a collider or a mediator introduces bias rather than removing it.

Individual cases differ, so anyone applying these principles to a specific animal should work with a veterinarian rather than relying on general study-level reasoning.