Colliders and Bad Controls
You'll learn how to recognize collider variables—those caused by both treatment and outcome—and understand why controlling for them can introduce bias rather than remove it.
What Is a Collider?
A collider is a variable that sits at the convergence of two causal arrows. Imagine both your treatment (X) and your outcome (Y) independently cause a third variable (C). That third variable is a collider:
X → C ← Y
Unlike confounders (which cause both X and Y), colliders should not be controlled for in causal analysis.
Why Colliders Are "Bad Controls"
When you condition on (or control for) a collider, you artificially create a spurious association between X and Y, even if none exists naturally. This is called collider bias or selection bias.
Example: The Talented Jerk Paradox
Suppose companies hire employees (C) based on either talent (X) or social skills (Y). Among hired employees only, you might observe a negative correlation: talented people seem less social. But this is an artifact! You've selected on a collider (being hired). In the general population, talent and social skills might be independent.
How to Spot Colliders
Look for variables that are:
- Caused by both your treatment and outcome
- Descendants of both pathways (e.g., intermediate outcomes influenced by X and Y)
Common bad controls include post-treatment variables like "job performance" when studying education's effect on salary—both education and inherent ability affect performance.
Key Takeaway: Colliders block causal paths naturally; controlling for them opens biased pathways. Only control for confounders (common causes), never colliders (common effects).