Which Potential Confounders Should You Identify Before Conducting a Study?

Which Variables Could Distort the Relationship You Want to Study?

Suppose you want to examine whether participation in an online learning program improves students' academic performance. Before collecting data, you identify prior achievement, socioeconomic circumstances, digital access, motivation, age, and several other characteristics that might influence the findings.

Should you measure all of them? Are they all confounders? And how can you determine which variables matter before you know what your data will show?

These questions are consequential because confounding is partly a research design problem. A variable that was never measured may be impossible to address adequately during analysis, while adjusting for the wrong variable can introduce rather than eliminate bias.

Identify Plausible Common Causes Before Collecting Data

How to Determine Which Potential Confounders Matter

Start with the causal relationship you intend to estimate

Confounder identification begins with a clearly specified exposure, outcome, target population, and causal question. Without these, it is difficult to determine whether another variable could distort the relationship of interest.

Consider two questions about online learning:

• Is voluntary participation in an online tutoring program associated with examination performance?
• What is the causal effect of participating in the program on examination performance?

The first question is associational. The second requires an argument about what examination performance would have been if comparable students had participated or not participated. Confounding is especially important for this second interpretation.

If students choose whether to participate, their prior achievement, access to resources, or academic motivation might influence both participation and subsequent performance. Consequently, an observed difference between participants and nonparticipants might not be attributable entirely to tutoring.

The task is to identify plausible alternative explanations for the observed exposure-outcome relationship before data collection begins.

Look for plausible common causes of exposure and outcome

A useful starting point is to identify variables that precede the exposure and could causally influence both exposure assignment and the outcome. Such variables may create noncausal pathways between exposure and outcome.

For example, prior academic achievement might influence whether a student seeks tutoring and independently predict subsequent examination performance. It is therefore a plausible confounder of the tutoring-performance relationship.

However, the familiar rule that a confounder must be associated with both exposure and outcome is only a preliminary heuristic. Statistical associations alone do not establish a variable's causal role. Some variables are associated with both because they are consequences of other variables, or because the data were selected in a particular way.

A variable may also be important for confounding control because it helps block a noncausal pathway when considered jointly with other variables, even if it does not independently satisfy a simple association-based screening rule.

Distinguish confounders from other types of variables

Not every variable related to an exposure or outcome should be treated as a confounder. Its position in the proposed causal structure matters.

These categories are not interchangeable. Their implications also depend on the causal effect being estimated and the relationships among the remaining variables. The distinction between effect modification and confounding is particularly useful when deciding whether a variable should be controlled, examined as an interaction, or both.

Likewise, understanding why adjustment for mediators or colliders can introduce bias helps prevent the mistaken assumption that including more variables necessarily produces a more credible causal estimate.

Use a causal diagram to make your assumptions explicit

A directed acyclic graph, or DAG, represents hypothesized causal relationships using arrows. It can help researchers identify noncausal pathways and evaluate which variables might need to be measured or controlled.

For example, a simple proposed structure might contain:

• Prior achievement → Tutoring participation
• Prior achievement → Final examination performance
• Tutoring participation → Final examination performance

Here, prior achievement creates a potential backdoor pathway between tutoring participation and examination performance. Controlling it may help address that pathway, assuming the diagram adequately represents the relevant causal structure.

A more realistic diagram might include household resources, digital access, prior learning opportunities, and institutional factors. Some variables may cause others, meaning that controlling every variable individually may be unnecessary.

A DAG is not proof that the proposed relationships are correct. It is a structured statement of assumptions, which should be justified using substantive knowledge and relevant evidence. Different plausible diagrams may imply different adjustment strategies.

Identify variables using theory and prior research, not statistical significance

Review studies investigating the same or closely related exposure-outcome relationships. Pay particular attention to variables identified as causes of exposure assignment, determinants of outcomes, or important contextual factors.

However, a published study's adjustment list should not automatically become yours. Previous researchers may have selected covariates through statistical screening, convenience, or assumptions that do not apply to your population.

For each candidate variable, ask:

• What is the plausible causal mechanism connecting it to exposure?
• Could it also affect the outcome independently of exposure?
• Did it exist before exposure, or could exposure have changed it?
• Does its causal role depend on another variable?
• Is the supporting evidence applicable to the target population and setting?

Statistical significance should not determine whether a variable qualifies as a confounder. A genuine confounding pathway may remain important even when an individual association is imprecisely estimated in a small sample.

Consider timing, measurement quality, and feasibility

Once candidate confounders have been identified, determine when and how they can be measured.

Pre-exposure measurements are often especially useful because they help establish temporal ordering and reduce the risk of inadvertently adjusting for consequences of exposure. Yet timing alone is not sufficient: a pre-exposure variable may still be a collider, an instrument, or an unnecessary adjustment variable.

Measurement quality matters as well. For example, asking students to recall their academic achievement from several years earlier may introduce substantial error. Where feasible and ethically appropriate, verified historical records may offer more reliable measurement.

Distinguish between a concept and its operational measure. Socioeconomic circumstances, for example, may be represented by household income, parental education, or an area-level deprivation indicator. These measures are not interchangeable and may capture different aspects of the underlying construct.

Researchers should also consider whether sensitive information is genuinely necessary. Collecting identifiable personal data simply because they might be useful in a future model creates avoidable privacy and data-management obligations.

This is why measuring a variable for possible future adjustment requires a substantive justification rather than an indiscriminate approach to data collection.

Identify a sufficient adjustment set rather than the largest possible set

A sufficient adjustment set is a collection of variables that, under a specified causal diagram and its assumptions, blocks relevant noncausal pathways without introducing inappropriate conditioning.

Several adjustment sets may be sufficient. Some may contain fewer variables, while others may be more practical to measure accurately.

For example, if an underlying variable causes both exposure and outcome through other measured variables, the appropriate adjustment strategy depends on the full causal structure. Simply counting the number of covariates cannot establish whether confounding has been addressed.

When selecting among defensible adjustment sets, consider measurement validity, missingness, statistical precision, and the risk of model instability.

It is equally important to recognize that even a theoretically sufficient set cannot remove confounding when key causal assumptions are incorrect or necessary variables are measured inadequately. Statistical adjustment has important limitations that must be considered during study planning.

Account for confounding that cannot be resolved through measurement alone

Some potential confounders may be difficult to measure, unavailable in existing datasets, or fundamentally unobservable. For instance, academic motivation is multidimensional and may not be adequately represented by a single questionnaire item.

When important confounders cannot be measured well, consider whether a different design, additional data source, or sensitivity analysis would provide a more credible basis for inference.

Randomization, when implemented appropriately, can reduce systematic baseline confounding in expectation. Observational designs instead depend more heavily on defensible causal assumptions and the availability of relevant information.

The objective is not to claim that every possible confounder has been identified. It is to develop a transparent account of the most plausible sources of confounding, how they will be addressed, and what uncertainty remains.

Identifying Confounders in an Online Tutoring Study

The resulting adjustment set is provisional, not automatically sufficient. Its adequacy depends on whether the proposed causal relationships are credible and whether additional pathways, including selection processes, have been overlooked.

The practical achievement is that the researchers now know which information is important to collect before exposure occurs, rather than discovering after data collection that a central baseline characteristic was never recorded.

Common Mistakes When Identifying Potential Confounders

Turn Confounder Identification into a Defensible Design Decision

Before finalizing your research instruments, create a candidate-variable inventory. For each variable, document its proposed causal role, supporting rationale, measurement timing, data source, and whether it is needed for the intended analysis.

Then distinguish variables that are essential for confounding control from those collected for descriptive purposes, effect-modification analyses, or improved precision.

Ideally, these decisions should be incorporated into a prespecified confounding-control strategy. This connects the causal question to recruitment, measurement, analysis, and eventual interpretation.

Remember that collecting relevant variables is only one part of confounding control. Participant eligibility, selection mechanisms, and the intended statistical analysis may also determine whether a credible causal estimate is possible.

Before Data Collection, Check Your Potential Confounders

Frequently Asked Questions About Identifying Confounders

Identify Confounders Through Causal Reasoning, Not a Covariate Checklist

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.