Berkson's Paradox: Why Your Dating Pool Lies to You About Reality

You have been dating for a while, and you have noticed something: the more attractive someone is, the less kind they tend to be. Your friends report the same pattern. It feels like a law of nature — as if the universe rations total human quality, distributing attractiveness and kindness as a trade-off. But there may be a completely different explanation: you are only dating people who clear a minimum threshold on at least one of those dimensions. That selection criterion, applied to an otherwise independent pair of traits, creates the illusion of a trade-off that does not exist in the broader population. This is Berkson's Paradox — and its implications stretch far beyond romantic disappointment.

The Statistical Structure

Berkson's Paradox was described by the American biostatistician Joseph Berkson in a 1946 paper examining why hospitalised patients seemed to show negative correlations between diseases that were actually independent in the general population. His original context involved hospital admission bias: if patients are admitted when they have disease A or disease B, then among admitted patients, having A is associated with not having B — because people who have neither are excluded from the sample. The correlation is artefactual.

The formal structure: suppose two variables, X and Y, are independent in the population — knowing someone's value on X tells you nothing about their value on Y. Now suppose we observe only individuals who clear some minimum threshold on a combined criterion: people enter our sample if X + Y ≥ threshold (or more generally, if they score high enough on some function of both). Within this selected sample, X and Y will be negatively correlated, even though they are independent in the full population.

The intuition: if someone is in our selected sample and has a low value on X, they almost certainly have a high value on Y — otherwise they would not have cleared the admission threshold. Within the selected group, low X predicts high Y and high X predicts low Y. A negative correlation has been manufactured from two independent variables by the selection mechanism.

The Dating Pool Example

Alice will go on a date with someone if they are at least reasonably attractive or at least reasonably kind — she excludes only people who are both unattractive and unkind. In the general population, suppose attractiveness and kindness are genuinely independent: knowing how attractive someone is tells you nothing about how kind they are.

Now consider who Alice actually dates. The unattractive-and-unkind people are excluded. The people she dates fall into three categories: (1) attractive and unkind, (2) unattractive and kind, (3) attractive and kind. Within this sample, kindness and attractiveness are negatively correlated: the most attractive people she dates include plenty of unkind ones (since their attractiveness alone qualified them), while the least attractive people she dates are disproportionately kind (since their kindness alone qualified them). She observes a negative correlation that does not exist in the broader population. The universe is not rationing total human quality; Alice's selection criterion is generating the illusion that it is.

This same structure applies to any selection process: job candidates screened on any two of skills and cultural fit; papers accepted when statistically significant or theoretically novel; restaurants chosen when cheap or highly rated. Wherever a "pass if at least one criterion is met" or "pass if sufficiently high on a combination" logic operates, the resulting sample exhibits correlations among the selection criteria that reverse or distort the population relationships.

The Collider in Causal Diagrams

In the framework of causal graphical models (directed acyclic graphs, or DAGs), Berkson's Paradox is an instance of "collider bias." A collider is a variable on a causal path that is caused by two other variables — it is a node where two arrows collide. Conditioning on a collider — selecting your sample based on it — opens a non-causal path between its parents and induces a spurious correlation.

In the dating example, the selection criterion (being dated) is caused by both attractiveness and kindness. It is their collider. When we restrict our analysis to the selected sample (people who are dated), we have conditioned on the collider, and attractiveness and kindness become correlated. This is formally equivalent to Berkson's original hospital admission example, and it generalises to any situation where an analysis is restricted to a subset defined by a variable that is influenced by two (or more) of the variables we want to study.

Collider bias is pervasive in medical research, epidemiology, and social science. Its recognition as a formal causal concept — rather than just an intuitive puzzle — has been one of the important contributions of the causal inference literature associated with Judea Pearl, James Robins, and others over the past three decades.

Medical Examples: Hospitalisation and Beyond

Berkson's original example remains instructive. If diabetes and cholecystitis (gallbladder disease) are both reasons for hospital admission but are independent in the general population, then among hospitalised patients, having one will appear to protect against the other: patients who are hospitalised for diabetes are less likely to also have cholecystitis (since their admission was already explained by diabetes), and vice versa. A clinician studying only hospitalised patients would conclude that diabetes protects against cholecystitis — an illusory negative association created by the admission selection mechanism.

More contemporary examples:

• COVID-19 severity and smoking. Early in the pandemic, several hospital-based studies found a negative association between smoking and COVID-19 hospitalisation — leading to speculation about protective effects of nicotine. Later analyses suggested collider bias: smokers are hospitalised more frequently for respiratory conditions generally, changing the composition of the hospitalised sample in ways that create artefactual negative correlations with COVID severity.
• Obesity and survival in certain conditions. The "obesity paradox" — where obesity appears protective in some hospitalised patient populations — may partly reflect collider bias: obese patients enter the hospital at lower disease severity thresholds than non-obese patients, creating a sample where the non-obese patients who are present are the most severely ill.
• Talent and work ethic in elite sports. Studies of elite athletes (already a selected population) sometimes find negative correlations between natural athleticism and training intensity — suggesting hard-working athletes are less gifted and gifted athletes work less hard. In the broader population, if these traits are independent or positively correlated, the negative correlation among elites may be a selection artefact: only athletes who are high on at least one dimension reach elite level.

Why This Error Is Hard to See

Berkson's Paradox is cognitively challenging because the selection process that generates the spurious correlation is often invisible or naturalised. We take for granted that our sample is "the world." We forget that we are only looking at people who are in our dating pool, or hospitals that admitted patients, or companies that survived to appear in a dataset. The selection event has already happened by the time we look at the data; the excluded observations are simply absent, leaving no trace of what generated the pattern we observe.

This connects to the broader problem of availability bias: the cases we reason from are the ones we can see, and the ones we can see are precisely those that cleared the selection criterion that creates the bias. The absent cases — the unattractive-and-unkind non-daters, the non-hospitalised patients, the failed startups that never appear in business school case studies — are systematically excluded from the inference we draw.

The base rate fallacy is adjacent: we fail to account for the population distribution of the traits we observe, reasoning instead from within the biased sample as if it were representative.

Survivorship Bias: The Famous Cousin

Berkson's Paradox is closely related to survivorship bias — a special case of selection bias where we condition on having survived some filtering process. The WWII example (examining returning bomber planes for damage patterns and reinforcing where they were hit, rather than where they were not hit — because the hit planes did not return) is the canonical illustration. Survivorship bias and Berkson's Paradox share the same causal structure: conditioning on having cleared a threshold introduces spurious correlations among the variables that determine threshold-clearance.

Startup mythology provides a rich source of survivorship bias examples: successful founders are described as risk-tolerant, visionary, and rule-breaking. Failed founders exhibit many of the same traits. The traits that appear associated with success in the visible sample of successful companies may be uncorrelated or even negatively correlated with success in the full population of attempted companies — we simply never see the failures.

Detecting and Correcting Collider Bias

The primary defences against Berkson's Paradox are awareness of selection mechanisms and explicit causal modelling:

• Draw the DAG. Before interpreting correlations in a dataset, identify the selection criterion that determines who is in the dataset, and ask whether that selection variable is a collider — caused by multiple variables in the analysis. If so, correlations among those variables within the selected sample cannot be taken at face value.
• Collect data on excluded cases. Wherever feasible, extending data collection to include individuals who did not clear the selection criterion allows estimation of population-level relationships. This is rarely fully achievable but partial data on non-selected cases can support sensitivity analyses.
• Use population-representative samples. Study designs that randomly sample from the full population, rather than from a filtered subset (hospital patients, app users, survey respondents), are not subject to Berkson's Paradox in the same way.
• Apply inverse probability weighting. Statistical methods that weight observations by the inverse of their selection probability can partially correct for selection bias when selection probabilities are known or estimable.

Sources & Further Reading

• Berkson, J. (1946). Limitations of the application of fourfold table analysis to hospital data. Biometrics Bulletin, 2(3), 47–53.
• Pearl, J. (2009). Causality: Models, Reasoning, and Inference. (2nd ed.) Cambridge University Press.
• Hernán, M.A., Hernández-Díaz, S., & Robins, J.M. (2004). A structural approach to selection bias. Epidemiology, 15(5), 615–625.
• Griffith, G.J. et al. (2020). Collider bias undermines our understanding of COVID-19 disease risk and severity. Nature Communications, 11, 5749.
• Wald, A. (1943). A Method of Estimating Plane Vulnerability Based on Damage of Survivors. Statistical Research Group, Columbia University. [The canonical survivorship bias analysis]