Confounder

Definition

A confounder is a variable that affects both the treatment (X) and the outcome (Y) (a common cause), creating a spurious (non-causal) association between X and Y.

DAG representation:

    Confounder (Z)
       ↙        ↘
  Treatment (X)   Outcome (Y)

Mathematical definition: A variable $Z$ is a confounder when:

1. $Z$ affects $X$ (or is associated with $X$)
2. $Z$ affects $Y$ (or is associated with $Y$)
3. $Z$ is not an effect of $X$ (not on the causal path)

Intuitive Understanding

Core idea:

A confounder is a “third variable” that makes X and Y appear associated even without a direct causal relationship

Example: correlation between ice cream sales (X) and drowning accidents (Y)

       Summer (Confounder)
          ↙        ↘
   Ice cream sales    Drowning accidents

• Ice cream does not cause drowning
• The common cause, summer, affects both
• Spurious association: correlation, not causation

Back-door Path

A confounder creates a back-door path:

X ← Z → Y

• A “back-door” path from X to Y
• This path transmits non-causal association
• Must be blocked: in order to identify the causal effect

Methods for Adjusting for Confounding

1. Statistical Control

Stratification:

# Analyze the X-Y relationship at each level of Z
for z_level in Z.unique():
    subset = data[data['Z'] == z_level]
    analyze(subset['X'], subset['Y'])

Regression: $Y = \beta_0 + \beta_1 X + \beta_2 Z + \epsilon$

$\beta_1$ is the effect of X after controlling for Z.

Propensity Score: $e(X) = P(X=1|Z)$

• Matching or weighting via the propensity score

2. Design-based Control

Randomization (RCT):

• Assign treatment at random
• Confounders become independent of treatment
• Back-door paths are blocked automatically

Natural Experiments:

• Instrumental Variables
• Regression Discontinuity
• Difference-in-Differences

3. Control by Design

Twin Studies:

• Monozygotic twins: share genes + family environment
• Within-pair analysis removes genetic confounding

Adoption Studies:

• Break the genetic link to remove genetic confounding

Measured vs Unmeasured Confounders

Measured Confounder

• Observable in the data
• Can be adjusted for via statistical control
• e.g., age, sex, education level

Unmeasured Confounder

• Unobservable (or not measured) in the data
• The causal effect cannot be identified
• Assess the impact with sensitivity analysis

       Unmeasured U
          ↙        ↘
         X          Y

• If U is not measured, the X→Y effect is biased

Confounding vs Collider vs Mediator

Variable type	DAG structure	Whether to control
Confounder	X ← Z → Y	Must control
Collider	X → Z ← Y	Must not control
Mediator	X → Z → Y	Depends on the goal

Rule of thumb: do not control for post-treatment variables

Examples

Example 1: Education and Income

      Intelligence
       ↙        ↘
  Education  →  Income

• Confounder: Intelligence
• Spurious path: Education ← Intelligence → Income
• Solution: control for Intelligence

Example 2: Smoking and Lung Cancer (Historical)

      Genetics?
       ↙        ↘
    Smoking   Lung Cancer

• Fisher’s argument: genetics could be a confounder
• Subsequent research established the causal effect of smoking

Example 3: Maternal Affection and Child Depression

      Shared Genes
       ↙        ↘
  Maternal       Child
  Affection    Depression

• Genetic confounding: shared genes between parent and child
• Solution: remove the genetic link with adoption studies

Measurement Error in Confounders

Measurement error in a confounder is a serious problem:

$Z^* = Z + \epsilon_Z$

• If $Z$ cannot be measured exactly, it is replaced by $Z^*$
• Residual confounding: the influence of Z is not fully removed
• False positive rate: can approach 100% in large samples (Westfall & Yarkoni, 2016)

Related Concepts

• DAG - Visualizing causal structure
• Back-door Criterion - Conditions for adjusting for confounding
• Collider - A variable that must not be controlled for
• Mediator - A variable on the causal pathway
• Propensity Score - A method for adjusting for confounding
• Unconfoundedness - The no-hidden-confounders assumption

References

• rohrerThinkingClearlyCorrelations - Confounding and DAGs
• Pearl, J. (2009). Causality