Berkson's paradox
About this app
The filter invents the relationship: a light wall sweeps through 3,000 independent applicants, the selected fly onto a stage — and there the regression line falls (r ≈ −0.67). Sum, or, and, causal diagram with collider.
When selection invents a correlation
Berkson’s paradox is a spurious association created by selection. If inclusion in a data set depends on two traits X and Y, for example “invited if the overall impression is strong enough”, then X and Y look negatively correlated among those selected, even though they are independent in the population. In a causal graph the filter S is a collider: X → S ← Y. Looking only at cases with S = 1 conditions on the collider and opens a spurious path. Joseph Berkson described the effect in 1946 using hospital data.
What the app shows
A simulated population of 300, 3,000 or 10,000 cases sits as a point cloud at the bottom. When you filter, a wall of light sweeps through it and the selected cases rise onto a stage, each cloud with its own regression line. You choose the filter type (sum, OR, AND), the selection rate and the true correlation ρ in the population. The app reports r for population and selection, the theoretical value, a 95 % confidence interval and the p-value. Four scenarios apply the pattern to hiring, start-ups, restaurants and Berkson’s hospital case.
Things to try
Keep ρ = 0 and filter by sum: at a 15 % selection rate, r among the selected drops to about −0.67. More data does not change this; only the confidence interval narrows. Switch to the AND filter: with independent traits, r stays near zero. A collider can open a spurious path, but it does not have to. The causal panel shows whether the path is open.
Frequently asked questions
- What is Berkson’s paradox?
- Berkson’s paradox is a spurious correlation caused by selection. When you only look at cases that passed a hurdle based on two traits, those traits appear negatively related even though they are independent overall. A familiar example: restaurants in prime locations seem to taste worse, because mediocre kitchens can survive there.
- What is a collider in a causal diagram?
- A collider is a variable that two arrows point into, such as X → S ← Y. Left alone, it blocks the path between X and Y. Conditioning on it, for example by analysing only selected cases, opens a spurious path and creates a correlation that does not exist in the population.
- What is the difference between a confounder and a collider?
- A confounder is a common cause (X ← Z → Y); you condition on it to remove bias. A collider is a common effect (X → S ← Y); conditioning on it creates bias. Analysing only a selected group means conditioning on S.
- What settings does the app offer?
- You set the filter type, the selection rate, the true correlation ρ and the sample size, and you can draw a new sample. A scenario panel maps the axes to everyday examples, and a causal panel shows the collider. All data are simulated.
Subject: Data & decisions
More from Data & decisions
- Anchoring effect
- Benford's law
- Cherry-picking: base period
- Cherry-picking: base period 2.0
- The three-cowboy duel (truel)
- The three-cowboy duel 2.0
- The three means
- The three means 2.0
A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps