Section VIII: Cofounding & Control

What is a confounder?

A third variable associated with both the exposure and the outcome that distorts their observed relationship but is not on the causal pathway.

What are the three criteria for a confounder?

(1) Associated with the exposure (2) Associated with the outcome (3) Not caused by the exposure.

How does a confounder distort an association?

It hides, exaggerates, or reverses the true relationship between exposure and outcome.

What is positive confounding?

When a confounder exaggerates an effect, pushing the RR/OR away from 1.

What is negative confounding?

When a confounder masks an effect, pulling the RR/OR closer to 1.

What is qualitative confounding?

When confounding reverses the direction of an observed association (e.g., RR > 1 becomes RR < 1).

How is a confounder different from a mediator?

A confounder distorts the relationship; a mediator lies on the causal pathway and explains how exposure causes outcome.

When is confounding most common?

In observational studies that lack randomization.

What are common confounders in epidemiology?

Age, sex, SES, comorbidities, health behaviors (e.g., smoking, alcohol use).

How does randomization reduce confounding?

It balances known and unknown confounders across groups, minimizing bias.

What are the main ways to control confounding during study design?

Restriction, Matching, and Randomization.

What are the main ways to control confounding during analysis?

Stratification, Mantel–Haenszel adjustment, and Multivariable regression.

What is restriction in study design?

Limiting participants to one level of a potential confounder (e.g., only non-smokers).

What is matching in study design?

Pairing participants with similar confounder characteristics (e.g., age, sex).

What is stratification in analysis?

Dividing data into subgroups (strata) based on the confounder and calculating association measures separately.

What does the Mantel–Haenszel method do?

Produces a weighted average of stratum-specific RRs or ORs to control for confounding.

When is multivariable regression preferred?

When adjusting for several confounders or continuous variables simultaneously.

How do you detect confounding in data?

Compare crude RR/OR with stratified or adjusted estimates — large differences suggest confounding.

If stratum-specific RRs are similar but differ from the crude RR, what does that mean?

Confounding is present.

If stratum-specific RRs differ across strata, what does that indicate?

Effect modification (interaction).

How is confounding different from interaction?

Confounding is bias to remove; interaction is a true difference to report.

What is residual confounding?

The distortion that remains after adjustment due to imperfect measurement or unmeasured confounders.

What are examples of residual confounding?

Broad age categories, self-reported data, unmeasured diet or SES variables.

If crude RR > adjusted RR, what type of confounding is present?

Positive confounding (crude overestimates the true effect).

If crude RR < adjusted RR, what type of confounding is present?

Negative confounding (crude underestimates the true effect).

If crude RR ≠ adjusted RR but stratum-specific RRs are similar, what’s present?

Confounding.

If stratum-specific RRs vary widely, what’s present?

Interaction or effect modification.

Why can confounding never be fully eliminated?

Because some confounders are unknown or measured with error.

What does it mean when a study reports both crude and adjusted RRs?

The authors are showing the direction and extent of confounding.

How can you minimize confounding by age in a study?

Use restriction (e.g., one age group) or match participants by age.

What does a Mantel–Haenszel adjusted RR lower than the crude RR imply?

The crude association was inflated by confounding.

What’s the main difference between controlling confounding in design vs. analysis?

Design control prevents confounding before data collection; analysis control adjusts for it afterward.

What is the key takeaway about confounding vs. interaction?

Confounding hides or exaggerates an association; interaction changes it.