Berkson's Paradox & Collider Bias

Dr. R. Düsing · Osnabrück University

Help — Berkson's Paradox & Collider Bias

What does this tool show?

Two uncorrelated variables X and Y can show a strong negative correlation within a selected subsample — even though there is no relationship whatsoever in the overall population. This effect is called Berkson's paradox or collider bias. It arises whenever you condition on a variable that depends on both variables of interest.

The running example

An HR department rates applicants on talent (X, IQ test) and diligence (Y, self-report). In the population, talent and diligence are uncorrelated (ρ = 0). Career success Z depends on both: Z = X + Y + noise. Looking only at the most and least successful applicants, a negative correlation suddenly appears: highly talented people seem less diligent — and vice versa.

The mechanism (why?)

Selecting on Z restricts the value space: people in the upper Z range must have high X or high Y values. Whoever has low X must have high Y — and vice versa. Within the selected band, the sum X + Y ≈ constant forces a negative dependency. In DAG terms, Z is called a collider — it has two incoming arrows (X → Z ← Y). Conditioning on a collider opens a spurious path between X and Y.

The DAG

X → Z ← Y — Z is the collider. The dashed red line shows the conditioning (selection on extreme Z values). Without conditioning: X ⊥ Y. With conditioning: X ⊥̸ Y.

The visualization

Left panel: Overall population. Gray points = not selected, red points = selected extreme group. The orange diagonal lines show the selection boundaries (Z = X + Y = threshold). Note: the boundary lines have slope −1 — that is exactly the reason for the negative correlation in the selected group.
Right panel: Only the selected people. The red regression line shows the negative correlation. The dashed gray line is the overall population's regression line, shown for reference.

Controls

Selection mode: None / Top only / Bottom only / Both extremes. Selection fraction: what % on each side is selected. ρ_XY: true underlying correlation (normally 0). Noise in Z: how strongly Z still depends on other factors (weakens the effect).

📋 Example — Talent & diligence in job applications
—
Causal Structure (DAG)
no direct connection (X ⊥ Y) Talent X Success Z Diligence Y CONDITIONING Z is the collider — sample only at extreme Z
Basic Principle
Z = X + Y + ε (collider structure: Z depends on X and Y)
Corr(X, Y) = ρXY (overall population — no conditioning)
Corr(X, Y | Z extreme) ≠ ρXY ← Berkson's Paradox
Selection boundary: Y = zthreshold − X  →  slope −1  ·  r in selection: —
Overall Population vs. Selected Group
ρ_XY (population)
—
true underlying correlation
n selected
—
— % of the sample
ρ_XY (selection)
—
correlation in extreme group
Δρ (bias)
—
selection − population
Conclusion: —
Concepts
What is Berkson's paradox?
Joseph Berkson (1946) described it using hospitals as an example: two diseases A and B are uncorrelated in the population. In the hospital, however (where patients are selected who have at least one of the diseases), they appear negatively correlated — someone with disease A less often needs disease B as the reason for admission. The paradox occurs whenever you condition on a variable that has mixed causes.
Collider — the DAG term
In a directed acyclic graph (DAG), a collider is a node with two incoming arrows: X → Z ← Y. Normally a collider blocks the flow of information between X and Y (d-separation). As soon as you condition on Z — through selection, adjustment, or stratification — the path X — Z — Y is opened and a spurious correlation arises. This is the opposite of confounding.
Why slope −1?
The selection threshold corresponds to X + Y ≈ c (a constant). That's the equation of a line with slope −1. People in the selected band approximately satisfy X + Y = c, i.e. whoever has high X must have lower Y. The orange boundary lines in the left plot show exactly these anti-diagonals. As noise in Z increases, the boundaries blur — and the Berkson effect weakens.
One-sided vs. two-sided — two effects
One-sided selection (Preset B/D — top only or bottom only) shows the classic Berkson effect: within the group, X + Y ≈ constant → whoever has high X must have lower Y → negative spurious correlation. Two-sided selection (Preset C — top + bottom together) instead produces a positive spurious correlation: the upper group has high means on X and Y, the lower group low means — the between-group effect dominates the within-group Berkson effect. Berkson's hospital example (1946) is a one-sided case.
Consequences for research
Collider bias is common in empirical research and hard to spot. Known cases: hospitalization bias (hospital-based studies), survivorship bias (only survivors/successes surveyed), selective publication (only significant findings published), online surveys (participation depends on interest, which depends on the outcome). Wherever study participation depends on both X and Y, Berkson lurks.
Reversing a positive underlying correlation
With Preset D (strong one-sided selection, 10%) and the ρ_XY slider set to +0.40, you can show: even if X and Y are positively correlated in the population, strong selection on Z can reverse the direction. With a sufficiently narrow selection window and low noise in Z, the Berkson effect becomes strong enough to overwhelm the originally positive correlation — the selected group then shows a negative correlation.
Distinguishing it from confounding
Confounding: a common cause U → X and U → Y creates a spurious correlation; controlling for U removes it. Collider bias: a common effect X → Z ← Y; controlling for Z creates a spurious correlation. Both produce bias, but by opposite logic. In regression models, "control for all available variables" is therefore wrong — colliders should never be controlled for.
Distinguishing it from range restriction
In common: both are selection effects on the observed correlation — you don't see the whole population. Difference: Berkson conditions on a collider (a common effect X→Z←Y) and thereby creates a spurious correlation that also biases the regression slope — a structural problem. Range restriction, by contrast, selects directly on the predictor X: the correlation is merely dampened (or inflated for extreme groups), while the slope Y·X stays unbiased — a variance problem that can be corrected with the Thorndike formula. Rule of thumb: never control for a collider, but do correct for range restriction. → Range Restriction
Detecting and avoiding it
Steps: (1) Draw a DAG before starting the analysis. (2) Classify every variable as a confounder, mediator, or collider. (3) Adjust only for confounders — never for colliders. (4) Check the sample's selection mechanism: does study participation depend on both X and Y at once? (5) Run sensitivity analyses with other selection models. References: Berkson (1946); Pearl (2009) Causality; Hernán & Robins (2020) Causal Inference.