THE COLLIDER BIAS

Talent and luck are independent in the world. Look only at people who got hired — a common effect of both — and a phantom correlation appears out of nothing. Conditioning on a collider manufactures association. This is the opposite of a confounder.

source Berkson, J. — Limitations of the Application of Fourfold Table Analysis to Hospital Data, Biometrics Bulletin 2(3):47–53, 1946. Rendered, not quoted.

□ Blue Team · builds & defends
3
The Model

A linear-Gaussian collider

Two independent causes flow into one common effect — the graph X → C ← Y. No edge joins X and Y.

X ~ N(0,1) Y ~ N(0,1) [independent]
C = a·X + b·Y + ε, ε ~ N(0,σ²)

With a=b=σ=1 the covariance of (X,Y,C) is exact, so every correlation below is a closed form, not a fitted guess.

marginal ρ(X,Y)0 Var(C)3 Cov(X,Y | C)−1/3 partial ρ(X,Y | C)−1/2
5
The Lineage

The confounder’s mirror image

A confounder is a common cause (C → X, C → Y): you must condition on it to remove a real spurious link. A collider is a common effect: conditioning creates a spurious link where none existed.

Same word — “control for C” — opposite consequence. This is why “adjust for everything” is a trap, and why Berkson’s hospital paradox is selection bias in disguise: the hospital is the collider you conditioned on.

7
The Witness · live re-check

Does the collider invariant still hold?

Re-derives the covariance of the currently displayed graph and asserts the collider signature: marginal independence and negative induced partial correlation.

…
◉ The Machine
4
Data In in ▼

Two honest, unrelated causes

Structural coefficients enter the machine. X and Y are drawn independently; C is assembled from them. Nothing here links X to Y.

0
The Panel LIT

Selecting on C invents a correlation

Blue dots: the full, unconditioned cloud (X vs Y) — a shapeless blob, ρ≈0. Amber dots: the sub-sample where C≈0. Within that slice a real downward tilt appears.

all points (unconditioned) selected: C in band
marginal ρ(X,Y) — closed form0.000 partial ρ(X,Y | C) — closed form−0.500 sample ρ inside band (seeded N)−0.5 E[Y | do(X=x)] slope — causal0.000
booting…
8
Data Out out ▼

The verdict

Association ≠ causation, and selection can forge association. The causal effect of X on Y is exactly zero (do(X) leaves Y untouched), yet conditioning on the collider C yields ρ(X,Y|C) = −½. The correlation is real in the sample and false about the world.

□ Red Team · attacks & breaks
1
The Adversary WALL

“You found a correlation, so control for it.”

The reflex of the naive analyst: see two variables move together in the data, add the shared variable as a covariate, call it rigor. But if that shared variable is a common effect, the adjustment is what created the movement. The wall is the belief that more conditioning is always safer.

2
The Graveyard

Claims that died here

  • “Diabetes protects against gallbladder disease.” Berkson, 1946: an artifact of studying only hospitalized patients — conditioning on admission (the collider).
  • “X and Y correlate in my clean sample, so they are related.” Only if the sample was not selected on a descendant of both. Berkson’s slice is selected on C.
  • “Adjusting for more variables removes more bias.” Adjusting for a collider (or its descendant) adds bias. Direction, not count, decides.
  • “Attractive people are ruder — I’ve met so many.” You met them because you dated on looks-or-personality: conditioning on a collider again.
6
The Tamper

Swap the arrows: collider → confounder

Reverse the graph to C → X, C → Y and keep insisting “conditioning on C is harmful.” Now X and Y are genuinely correlated and conditioning removes it. The collider witness must catch that its signature is gone.

graph: X→C←Y (collider)