Can Adjusting for Mediators or Colliders Introduce Bias?

Why Can Controlling Another Variable Make a Causal Estimate Worse?

Suppose you are studying whether participation in an online tutoring program improves examination performance. You measure baseline achievement, academic motivation, study confidence after tutoring, and whether students complete an optional follow-up assessment.

You might reasonably assume that including more of these variables in a regression model would produce a more accurate estimate of tutoring's effect.

However, study confidence may be part of the mechanism through which tutoring affects performance. Assessment completion may depend on both tutoring participation and other determinants of academic outcomes. Adjusting for these variables could distort the relationship rather than clarify it.

The problem is not simply that some variables are measured after exposure. It is that variables occupy different positions in the causal structure, and statistical conditioning changes which pathways contribute to an estimated association.

Mediators and Colliders Can Introduce Different Forms of Bias

How Mediator Adjustment and Collider Conditioning Affect Causal Inference

What does statistical conditioning mean?

Conditioning means restricting or organizing comparisons according to information about another variable. It can occur through regression adjustment, matching, stratification, restriction, weighting, or selection into an analytical sample.

For example, a regression comparing tutoring participants and nonparticipants while holding baseline achievement constant conditions on baseline achievement.

Conditioning can help address confounding when it blocks relevant noncausal pathways. However, it can also block causal pathways or create associations that would otherwise be absent.

The appropriate decision depends on the variable's causal role, not simply whether it is associated with exposure or outcome.

What is a mediator?

A mediator is a variable through which an exposure exerts some of its causal effect on an outcome.

Consider the proposed causal sequence:

Tutoring participation → Academic confidence → Examination performance.

If tutoring increases confidence and confidence improves examination performance, confidence mediates part of tutoring's effect.

The total effect includes causal pathways operating through confidence and any other mechanisms. Adjusting for confidence may remove or alter the contribution of the mediated pathway.

However, simply observing that confidence is associated with tutoring and examination scores does not establish mediation. The proposed causal sequence requires substantive justification and appropriate temporal ordering.

Why can mediator adjustment bias the estimated total effect?

Suppose the objective is to estimate the total effect of tutoring participation on examination performance.

If confidence is a mediator, conditioning on confidence changes the comparison. Instead of allowing confidence to vary as it would under the different tutoring conditions, the analysis attempts to compare students at similar confidence levels.

This can remove part of the effect attributable to tutoring and therefore prevent the adjusted coefficient from representing the intended total effect.

Schisterman, Cole, and Platt describe this problem as overadjustment bias when adjustment is made for an intermediate variable, or certain descendants of an intermediate variable, along a causal pathway.

Importantly, the adjusted estimate is not necessarily closer to zero. The direction and magnitude of the change depend on the causal relationships, effect scales, and statistical model.

Nor does the difference between a model with and without mediator adjustment automatically equal the indirect causal effect. Identifying mediated effects generally requires additional assumptions and appropriate causal mediation methods.

When is adjusting for a mediator appropriate?

Mediator adjustment may be relevant when the scientific objective concerns a direct effect rather than the total effect, or when researchers want to investigate causal mechanisms.

For example, a researcher might ask how much tutoring affects examination performance through pathways other than academic confidence.

That is a different causal question from asking whether tutoring improves performance overall.

Direct-effect estimation requires specifying the relevant estimand. Controlled direct effects, natural direct effects, and interventional direct effects are not interchangeable and may require different identification assumptions.

Simple regression adjustment for a mediator may fail when there is unmeasured mediator-outcome confounding or when exposure affects variables that subsequently confound the mediator-outcome relationship.

Consequently, researchers should not interpret an exposure coefficient from a mediator-adjusted model as a valid direct effect without examining the relevant causal assumptions.

What is a collider?

A collider is a variable that receives causal arrows from two or more other variables in a directed acyclic graph.

For example:

Tutoring participation → Follow-up completion ← Academic motivation.

Here, follow-up completion is a collider because it is influenced by both tutoring participation and motivation.

Suppose tutoring participation and motivation are initially independent, while motivation independently influences examination performance. Conditioning on completion can create an association between tutoring participation and motivation among the selected students.

That induced association can then distort the relationship between tutoring and examination performance.

The important feature is the collider's position in the causal structure. A collider can be measured before or after exposure, depending on the relationships being studied.

How does collider conditioning create a noncausal association?

Consider two independent causes of selection into a study. Either cause may increase the likelihood that a participant is selected.

Among selected participants, knowing that one cause is absent may make the other cause more likely, even if the causes were independent in the original population.

This phenomenon is sometimes explained using the example of university admission. If both academic ability and another independent characteristic influence admission, conditioning on admission may create an association between those characteristics among admitted students.

In causal diagrams, conditioning on a collider can open a pathway that was previously blocked. Elwert and Winship describe this mechanism as a central source of endogenous selection bias.

The resulting association is not necessarily evidence that one cause affects the other. It may arise entirely from the selection mechanism.

How do mediator adjustment and collider conditioning differ?

These mechanisms are conceptually distinct, although they can coexist. A mediator may also be a collider when it has additional causes, and conditioning on it may both block a causal pathway and open another noncausal pathway.

Can a variable be both a mediator and a collider?

Yes. Causal roles are determined by relationships within a particular graph, and a variable can occupy more than one role.

Suppose tutoring affects academic confidence, while unmeasured psychological circumstances also affect confidence and examination performance.

The proposed structure includes:

• Tutoring participation → Academic confidence → Examination performance.
• Psychological circumstances → Academic confidence.
• Psychological circumstances → Examination performance.

Confidence lies on a causal pathway from tutoring to performance, making it a mediator. It is also a collider on the pathway between tutoring and psychological circumstances because both influence confidence.

Conditioning on confidence can therefore remove part of the mediated effect while introducing an association between tutoring and psychological circumstances.

This example illustrates why simple rules such as adjusting for every outcome predictor can fail.

Can selection criteria and complete-case analysis create collider bias?

Yes. Collider conditioning does not require explicitly adding the collider to a regression model.

Suppose participation in tutoring affects whether students complete a follow-up assessment, while motivation also affects completion and examination performance.

If researchers analyze only students who completed the assessment, they condition on completion. This can create selection bias even when completion is never included as a regression covariate.

Similar problems can arise when eligibility depends on a common effect of exposure-related and outcome-related causes, or when analyses are restricted to participants surviving or remaining enrolled until a particular time.

Elwert and Winship emphasize that conditioning on colliders can occur through selection and stratification as well as statistical control, and may involve variables measured before or after treatment.

Researchers should therefore examine the causal implications of recruitment, exclusion, missing-data handling, and analytical restrictions.

Are all post-exposure variables inappropriate for adjustment?

No. Post-exposure timing raises important questions, but it does not establish a universal prohibition.

Some post-exposure variables are mediators, some are colliders, and others may be time-varying confounders of later exposure-outcome relationships.

In longitudinal studies, a variable affected by earlier exposure may also influence subsequent exposure and the outcome. Conventional adjustment can then create difficulties because the variable simultaneously serves as a confounder for later exposure and a consequence of earlier exposure.

Specialized approaches, including certain g-methods, may be appropriate under suitable assumptions.

The correct strategy depends on the treatment history and causal estimand. Hernán and Robins provide a systematic account of these longitudinal identification problems in Causal Inference: What If.

Researchers should avoid both automatic post-exposure adjustment and the equally simplistic rule that no post-exposure variable may ever be used in causal analysis.

Can a collider be measured without introducing bias?

Yes. Measuring a collider does not itself create collider-stratification bias. The concern is conditioning on it, or on certain descendants of it, in a way that opens a noncausal pathway.

For example, follow-up completion may need to be recorded to describe retention, investigate missing-data mechanisms, and assess study implementation.

Collecting that information is often essential. Problems arise when the analysis selects or conditions on completion without considering its causal determinants.

This distinction is closely related to deciding which variables should be measured for possible adjustment. Measurement and conditioning are separate methodological decisions.

Can directed acyclic graphs identify problematic adjustment?

Directed acyclic graphs, or DAGs, are useful for representing proposed causal relationships and determining whether conditioning blocks or opens particular pathways.

A mediator is typically visible along a directed pathway from exposure to outcome. A collider is identified by two arrows pointing into the same variable.

However, a variable may have multiple roles, and whether conditioning is appropriate depends on the complete graph and the effect being estimated.

A DAG can help identify a sufficient adjustment set for a total-effect analysis or reveal why an apparently reasonable covariate would introduce bias.

It cannot establish the truth of the causal assumptions represented. Researchers must justify the graph using substantive knowledge and available evidence.

Nor can the diagram alone quantify the magnitude of bias. That generally requires additional assumptions, data, or sensitivity analyses.

How should researchers decide whether to adjust for a questionable variable?

Begin by specifying the exposure, outcome, time ordering, and causal estimand.

Next, identify the variable's plausible causes and consequences. Determine whether it lies on a causal pathway, serves as a common effect, or contributes to confounding.

Then examine whether conditioning on it would block a noncausal pathway, block part of the desired effect, or open a previously blocked pathway.

Where the causal role is uncertain, researchers may need to consider alternative plausible graphs and report sensitivity analyses. Merely including the variable in several models and choosing the most favorable estimate is not a defensible solution.

These decisions should be incorporated into the planned confounding-control strategy rather than improvised after examining outcomes.

How Conditioning on a Collider Can Create a Misleading Association

Before conditioning on completion, the tutoring and non-tutoring groups each contain equal proportions of highly motivated students. Both therefore have an expected examination pass probability of 60%.

Among follow-up completers, however, only 50% of tutoring participants are highly motivated, compared with approximately 83.3% of nonparticipants.

This is a deliberately constructed numerical illustration, not an empirical finding. It demonstrates how conditioning on a common effect can create an association that does not exist in the original population.

The same reasoning applies to regression adjustment, restriction, matching, or other procedures that condition on a collider, although the magnitude and direction of resulting bias depend on the causal structure and selection mechanism.

Common Misconceptions About Mediators and Colliders

Evaluate Causal Pathways Before Conditioning on Variables

Before including a variable in the primary adjustment model, determine whether it is a confounder, mediator, collider, or another type of covariate relative to the specified causal estimand.

Do not infer its role from statistical significance or measurement timing alone. Examine the proposed causal relationships and whether conditioning changes the pathways contributing to the estimate.

These decisions should be integrated into the research design's confounding-control strategy.

Researchers should also recognize that unnecessary statistical adjustment is broader than mediator and collider bias. Some covariates may simply contribute little information or reduce precision without introducing the same causal distortions.

The objective is to condition on variables that support the intended causal comparison, not every variable that appears relevant to the outcome.

Before Adjusting for a Mediator or Collider, Check the Causal Structure

Frequently Asked Questions About Mediator and Collider Bias

Adjust for Causal Relevance, Not Mere Statistical Association

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.