Independent, Dependent, Confounding and Mediating Variables: What They Actually Do in a Study

Suppose your question is:

Is distance to an adolescent-friendly clinic associated with use of the service?

You might call:

Independent variable: clinic distance.

Dependent variable: service use.

In epidemiology and much of health research, it is often clearer to say:

Exposure: clinic distance.

Outcome: service use.

Why?

Because “independent” can be misleading. The exposure may be statistically and causally related to many other variables.

It is not necessarily independent of anything.

The labels are useful for introductory thinking.

The causal structure matters more.

Clinic distance.

Service use.

Perceived accessibility.

Residential setting is a common cause of clinic distance and service use.

Now let us unpack each one.

In causal inference, confounding is not adequately defined as:

“a variable associated with both the independent and dependent variable.”

That shortcut can mislead.

A more useful causal question is whether there are non-causal pathways creating differences between exposed and unexposed groups that prevent a valid comparison of the causal effect of interest.

In the example:

Residential setting → distance

and:

Residential setting → service use

Residential setting can create a backdoor path between distance and service use.

If it is appropriately measured and the causal assumptions are correct, adjustment may help block that pathway.

But the label depends on the question.

If your exposure changed, the relevant confounding structure could also change.

A mediator lies on a hypothesized causal pathway from exposure to outcome.

For example:

Distance → perceived accessibility → service use

Here, actual distance may affect whether services feel accessible, which may then influence use.

Perceived accessibility is therefore a plausible mediator.

Now ask what you want to estimate.

What is the total effect of clinic distance on service use?

If you adjust for perceived accessibility, you may block part of the pathway through which distance operates.

You have changed the estimand.

What is the effect of clinic distance on service use not operating through perceived accessibility?

Now mediation analysis becomes relevant, but estimating direct effects requires stronger assumptions and an analysis designed for that purpose.

Simply adding the mediator to a regression and comparing coefficients is not automatically a complete mediation analysis.

Consider:

Household income → clinic distance → service use

If household income influences residential location and transport options before clinic distance is experienced, it may be part of a confounding structure.

Now consider:

Clinic distance → transport burden → service use

Transport burden may be partly caused by distance.

That makes it a potential mediator.

The difference is causal position.

Not whether one variable has p < 0.05.

This article is not titled “colliders,” but you need to know they exist.

A collider is a variable caused by two other variables along a path.

For example:

Exposure → Clinic attendance ← Health-seeking tendency

If you condition on clinic attendance, for example, by studying only clinic attendees, you can induce an association between the exposure and health-seeking tendency even if none existed before conditioning.

Causal diagrams are useful partly because they make this problem visible.

The lesson is not:

“Never adjust for anything.”

It is:

Adjustment decisions are causal decisions.

An effect modifier is different again.

Suppose the association between clinic distance and service use is much stronger among younger adolescents than older adolescents.

Age may modify the effect.

That means the effect differs across levels of age.

This is a substantive finding about heterogeneity.

Effect modification is not something to “control away.”

It may be something to report.

For example:

The estimated effect of distance on service use differed by age group, with a larger reduction in use among younger adolescents.

Whether effect modification is plausible should ideally be considered before endless subgroup testing begins.

Yes.

Suppose income is the variable.

Question A:

What is the effect of education on healthcare use?

Income may partly mediate the effect of education on healthcare use:

Education → income → healthcare use

Question B:

What is the effect of insurance type on healthcare use?

Income may influence both insurance type and healthcare use.

Now it may be part of the confounding structure.

Same variable.

Different exposure-outcome question.

Different role.

This is why a textbook list of:

“Examples of confounding variables: age, sex, income...”

can only take you so far.

“Control variable” is often used loosely for anything placed in an adjusted model.

That phrase hides the reason for inclusion.

Methods Bench prefers naming the role where possible:

• confounder;
• design variable;
• precision variable;
• mediator;
• effect modifier;
• predictor;
• descriptive covariate.

Clear labels make the analysis easier to defend.

A directed acyclic graph, or DAG, represents assumed causal relationships using arrows.

A DAG does not discover causality from the data.

It makes your assumptions visible.

That is useful because it forces you to ask:

• What causes the exposure?
• What causes the outcome?
• What happens after exposure?
• Which pathways should be blocked?
• Which pathways should remain open?
• Could adjustment open a non-causal path?

Two researchers can disagree about a DAG.

Excellent.

Now they are disagreeing about explicit causal assumptions rather than hiding them inside a regression model.