Collider

Definition

A collider is a variable affected by both the treatment (X) and the outcome (Y) (a common effect). In the structure X → C ← Y, C is a collider.

DAG representation:

  Treatment (X) → Collider (C) ← Outcome (Y)

Key properties:

• Default state: blocks the association between X and Y
• Conditioning on C: creates a spurious association between X and Y (collider bias)

Why “Collider”?

• Two arrows “collide” at the same variable

Intuitive Understanding

Core idea:

Controlling for a collider makes an X-Y relationship that was never there appear to exist

Example: Attractiveness and Niceness on a dating app

   Attractive → Date ← Nice

• Being attractive or being nice gets you selected as a date
• Whole population: attractiveness and niceness are unrelated
• Analyzing only the dates (conditioning on Date):
  • “A less attractive person must be nicer to be selected”
  • A spurious negative correlation arises!

Collider Bias Examples

1. Publication Bias (Meta-analysis)

  Methodological Rigor → Publication ← Innovativeness

• Analyzing only published papers (conditioning on Publication):
  • Even with low rigor, an innovative paper gets published
  • Even if not innovative, a rigorous paper gets published
• Result: rigor and innovativeness appear negatively correlated
• Reality: in fact unrelated, or positively correlated

2. Berkson’s Paradox (Hospital Sample)

  Disease A → Hospitalization ← Disease B

• Analyzing only the hospital sample:
  • Even without Disease A, one is hospitalized for Disease B
  • Even without Disease B, one is hospitalized for Disease A
• Result: Disease A and B appear negatively correlated
• Population: in fact unrelated

3. Nonresponse Bias (Survey)

  Variable X → Response ← Variable Y

• Analyzing only respondents (conditioning on Response):
  • X and Y affect whether one responds
• Result: the X-Y relationship is distorted

4. Attrition Bias (Longitudinal Study)

  Baseline X → Dropout ← Outcome Y

• Analyzing only the remaining participants (non-dropout):
  • X and Y affect whether one drops out
• Result: selection bias

5. Sample Selection Effect

  Variable X → Sample Selection ← Variable Y

• Analyzing only a particular sample:
  • e.g., only successful people, only college entrants
• Result: the X-Y relationship differs from the population

Why Does Collider Bias Occur?

Mathematical Intuition

$C = f(X, Y) + \epsilon$

Conditioning on C: $P(Y|X, C=c) \neq P(Y|X)$

• Fixing the value of C makes information about X informative about Y
• “If X is large, then for C=c to hold, Y must be small”

Information Flow

Without conditioning on C:
    X         Y     (no path, independent)

With conditioning on C:
    X → [C] ← Y     (path opened, dependent)

• Conditioning opens a “channel of information”

Identifying Colliders

Determining from a DAG

A variable C is a collider when:

1. $X \rightarrow C$ (X affects C)
2. $Y \rightarrow C$ (Y affects C)

Temporal Clue

Rule of thumb: a post-treatment variable can be a collider

• A variable that occurs after the treatment and the outcome
• e.g., a final result, a selection variable

Caution: Not Every Post-treatment Variable Is a Collider

X → M → Y    (M is a Mediator, not a collider)
X → C ← Y    (C is a Collider)

Do NOT Control for Colliders

A Mistaken Practice

“Let’s control for as many variables as possible” → dangerous!

The Correct Approach

1. Draw the DAG and grasp the causal structure
2. Identify colliders
3. Exclude colliders from the controls

Exception: Descendants of a Collider

X → C ← Y
    ↓
    D

• Controlling for D (a descendant of C) also induces collider bias
• Because it transmits partial information about C

Collider vs Confounder

Aspect	Confounder	Collider
DAG structure	X ← Z → Y	X → C ← Y
Role	Common cause	Common effect
Default state	Creates spurious association	Blocks association
Effect of control	Removes spurious association	Creates spurious association
Whether to control	Must control	Must not control

Related Concepts

• DAG - Visualizing causal structure
• Confounder - Common cause (must control)
• Mediator - A variable on the causal pathway
• Back-door Criterion - Conditions for causal identification
• Selection Bias - A form of collider bias
• v-structure - Unshielded collider, key to distinguishing MECs

References

• rohrerThinkingClearlyCorrelations - Explanation of collider bias
• Berkson, J. (1946). Limitations of the application of fourfold table analysis
• Pearl, J. (2009). Causality