How Do You Explain Which Confounders Your Research Design Can and Cannot Address?

How Can You Explain What Your Study Actually Controlled?

You have completed an observational study examining whether voluntary participation in an online tutoring program improves examination performance. Your statistical model adjusts for prior achievement, age, and access to learning resources. However, academic motivation and family support were not measured.

In the manuscript, should you state that confounding was controlled? Should you describe the unmeasured characteristics as limitations? How much detail is necessary to explain whether the adjusted association supports a causal interpretation?

These decisions matter because readers need to distinguish the confounding pathways your design plausibly addresses from those that remain unresolved. Simply reporting that a regression was adjusted for several variables does not establish that the comparison is causally valid.

Report Both the Confounding Addressed and the Confounding That May Remain

How to Describe Confounding Control Accurately and Transparently

Why is listing adjusted variables insufficient?

A statement such as "the analysis controlled for age, sex, and socioeconomic status" tells readers which covariates were included, but it does not explain why those variables were selected or whether they adequately address the relevant causal pathways.

Some measured variables may be appropriate confounders. Others may be outcome predictors, mediators, colliders, or unnecessary adjustment variables. Their causal roles cannot be determined from the regression table alone.

Readers also need to know whether important common causes were unavailable, measured imprecisely, or excluded from the analytical model.

The STROBE explanation and elaboration recommends describing the methods used to control confounding, explaining variable-selection procedures, and considering residual confounding when interpreting observational findings.

Accordingly, report the scientific rationale for adjustment alongside the statistical procedure. A longer covariate list is not a substitute for a defensible causal explanation.

First, specify the causal comparison the design was intended to support

Confounding can only be evaluated in relation to a particular causal question.

Specify the exposure, comparison condition, outcome, target population, and relevant time frame. Where possible, identify the causal estimand, such as the average total effect of receiving an intervention compared with not receiving it.

For example, estimating the total effect of voluntary tutoring participation on final examination scores requires a different adjustment rationale from investigating whether tutoring affects performance independently of changes in academic confidence.

The former generally should not condition indiscriminately on mediators. The latter requires a properly specified direct-effect estimand and additional identification assumptions.

Without clarifying the intended effect, it is difficult to determine whether the chosen covariates were appropriate or whether relevant confounding remains.

Distinguish what was measured, what was controlled, and what remains uncertain

These are related but different claims. Measuring a variable does not mean that it was included in the analysis, and including a variable does not guarantee that its confounding influence was eliminated.

This distinction helps prevent an important reporting error: describing a variable as controlled merely because its name appears in a regression equation.

How should you explain the confounders your design addresses?

For each major confounding pathway, explain the suspected causal relationship, the relevant measurement, and the procedure used to address it.

Suppose prior achievement influences both tutoring participation and final examination performance. Researchers may explain that pre-semester achievement records were obtained and included in the prespecified adjustment model to address this pathway.

That is more informative than simply stating that prior achievement was a covariate.

If participants were matched on baseline achievement, report the matching method, balance achieved, and the characteristics of unmatched observations.

If eligibility was restricted to a particular achievement category, explain how the restriction reduced variation and how it changed the population represented by the study.

If treatment was randomized, describe the assignment mechanism and its implementation. Proper randomization addresses systematic baseline confounding in expectation, although chance imbalance and other sources of bias may remain.

The discussion should connect each method to the causal problem it was intended to address rather than treating all control procedures as interchangeable.

How should unmeasured confounding be described?

Unmeasured confounding arises when an unavailable characteristic contributes to a noncausal pathway between exposure and outcome that is not otherwise adequately addressed.

For example, baseline academic motivation may influence both voluntary tutoring participation and subsequent examination performance. If motivation was not measured, the adjusted comparison may remain confounded.

However, researchers should distinguish a plausible source of confounding from a demonstrated source of bias. Without additional evidence, they may not know whether motivation actually differed between groups or how much it affected the estimate.

A defensible explanation identifies the proposed mechanism and its implications without inventing a magnitude or direction.

Where relevant, explain why the confounder was unavailable. It may have been omitted from the original data collection, inadequately represented in administrative records, or difficult to operationalize reliably.

The distinction between insufficient confounder measurement and adequate causal identification is central to interpreting such limitations.

What is residual confounding, and how should it be reported?

Residual confounding refers to confounding that remains after attempted control.

It may arise because a relevant variable was omitted, measured inaccurately, categorized too broadly, or represented inadequately in the statistical model.

Suppose researchers adjust for prior achievement using only three broad categories. Students within each category may still differ substantially in their actual baseline scores. If those differences affect both tutoring participation and examination performance, some confounding may remain.

Similarly, a single indicator of device ownership may inadequately represent access to learning resources.

Researchers should explain the particular measurement or modeling limitations that make residual confounding plausible.

STROBE emphasizes that measurement error in confounders can increase the risk of residual confounding and recommends describing measurement validity and reliability where relevant.

A general statement that residual confounding is possible may be accurate, but it is less useful than explaining which variables or pathways are most concerning.

Can you determine the direction of remaining confounding?

Sometimes a plausible direction can be proposed using substantive knowledge about how a confounder relates to exposure and outcome.

For example, if highly motivated students are more likely to participate in tutoring and motivation independently improves examination performance, unmeasured motivation could make tutoring appear more beneficial than it actually is.

However, that interpretation depends on the assumed relationships and the absence of competing mechanisms.

Other confounders may act in the opposite direction. Students experiencing academic difficulties, for example, may be more likely to seek tutoring and also have lower expected examination scores.

The combined direction of bias may therefore be uncertain.

Researchers should not claim that residual confounding necessarily overestimated or underestimated an effect without a defensible basis. When the direction cannot be established, say so.

How can you discuss the possible magnitude of confounding?

The magnitude of confounding is often more difficult to assess than its existence.

Where suitable external evidence is available, researchers may examine plausible associations between an unmeasured confounder, exposure, and outcome. Quantitative bias analysis can then explore how different assumptions would change the estimated effect.

For certain effect measures and assumptions, the E-value provides one way to summarize the minimum strength of association an unmeasured confounder would need with exposure and outcome to explain away an observed association.

Such calculations are sensitivity measures, not evidence that a confounder of the required strength actually exists or does not exist.

Researchers should also distinguish statistical uncertainty from uncertainty caused by bias. A narrow confidence interval does not account automatically for unmeasured confounding, selection bias, or measurement error.

Consequently, an estimate may be statistically precise while remaining causally uncertain.

What should you say when adjustment changes the effect estimate?

Researchers often compare crude and adjusted estimates to describe how accounting for measured variables changes the observed association.

For example, a crude difference of eight examination points might decrease to four points after adjustment for baseline characteristics.

This change is compatible with confounding by measured variables, but it does not independently prove that the adjusted estimate is unbiased.

Differences between crude and adjusted estimates can also reflect changes in estimands, noncollapsibility of certain effect measures, model specification, or inappropriate conditioning.

In particular, adjusted and unadjusted odds ratios may differ even in the absence of confounding because odds ratios are noncollapsible.

Researchers should therefore interpret changes in estimates in light of the effect measure and causal structure rather than treating every change as a direct measurement of confounding removed.

How should you explain inappropriate or potentially harmful adjustment?

Confounding-control reporting should address not only omitted variables but also variables that were deliberately excluded because conditioning on them could introduce bias.

For example, study confidence measured after tutoring begins may be a mediator of tutoring's effect on examination performance. If the intended estimand is the total effect, routine adjustment for confidence could remove part of that effect.

Similarly, selecting only students who complete a follow-up assessment may introduce collider-related selection bias when completion depends on exposure and outcome-related causes.

Researchers can explain that such variables were excluded from the primary adjustment set because of their hypothesized causal roles.

This demonstrates that the adjustment strategy was selective for methodological reasons rather than incomplete by accident.

The distinction is developed further in the discussion of bias introduced by conditioning on mediators and colliders.

How should sensitivity analyses be reported?

Sensitivity analyses should address specific uncertainties in the causal interpretation.

Examples include alternative defensible confounder measurements, different plausible adjustment sets, quantitative assumptions about unmeasured confounding, and approaches to missing data.

Report why each analysis was conducted, which assumption it examines, and how its results affect interpretation.

When findings change substantially under plausible alternatives, that instability should be discussed as a limitation of the causal conclusion.

When findings remain similar, the results may provide some reassurance about the particular assumptions examined. They do not establish robustness to every possible source of bias.

A sensitivity analysis should not be described as proving that confounding was eliminated.

Where should confounding-control explanations appear in a research article?

Different parts of the manuscript serve different reporting purposes.

The STROBE reporting guidance provides a useful framework for observational studies, including the description of bias-control methods, statistical analyses, limitations, and cautious interpretation.

Reporting guidelines support transparency but do not certify that a research design is causally valid.

How should the conclusion change when important confounding remains unresolved?

The language of the conclusion should correspond to the strength of the identification strategy.

When a study cannot adequately address important confounding pathways, describing an adjusted association may be more defensible than claiming that exposure caused the observed outcome difference.

For example, "tutoring participation was associated with higher examination scores after adjustment for measured baseline characteristics" accurately describes an adjusted association.

By contrast, "tutoring improved examination scores" makes a causal claim that requires stronger support.

This does not mean observational research can never support causal inference. Well-designed observational studies may identify causal effects under credible assumptions and suitable methods.

The issue is whether those assumptions are sufficiently defensible for the particular study.

Researchers should explain the confounding-control strategy established during research design and evaluate its remaining limitations rather than treating causal language as a matter of stylistic preference.

Explaining Confounding Control in a Published Research Report

A defensible discussion might state:

Participation in online tutoring was associated with higher examination scores after adjustment for measured baseline characteristics. However, tutoring was voluntary, and academic motivation and family support were not measured. These characteristics may influence both participation and performance, leaving the findings vulnerable to residual confounding. The adjusted association should therefore not be interpreted as establishing the causal effect of tutoring.

This statement identifies the variables addressed, the unresolved causal concern, and the resulting interpretive limitation.

The numerical estimate is hypothetical. A real manuscript should tailor its explanation to the actual study design, available evidence, and results of any relevant sensitivity analyses.

Common Mistakes When Reporting Confounding and Study Limitations

Make Your Confounding Claims Proportional to the Evidence

When preparing a manuscript, distinguish the methods used to address confounding from the assumptions required for those methods to identify the intended causal effect.

Explain why important variables were selected, how they were measured, and which causal pathways remain uncertain. Where appropriate, supplement the explanation with a causal diagram and sensitivity analyses.

When reporting observational research, consult the STROBE reporting guidance and any applicable discipline-specific or journal requirements.

Transparent limitations should not be treated as an admission that a study lacks value. They help readers understand what the evidence supports, where uncertainty remains, and which improvements would be most consequential for future research.

Before Submitting Your Manuscript, Check the Confounding Explanation

Frequently Asked Questions About Reporting Confounding Limitations

Explain What Was Addressed, What Remains, and Why It Matters

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.