Berkson's Paradox
Berkson's paradox is a statistical collider-bias artifact in which conditioning on a shared selection event such as hospitalization induces a spurious negative correlation between two independent risk factors.
Berkson's paradox (also called Berkson's bias or Berkson's fallacy), discovered by Mayo Clinic biostatistician Joseph Berkson in 1946, is a classic form of collider stratification bias in epidemiology and probability theory. It occurs when two exposures or diseases that are completely independent (or even positively correlated) in the general population appear negatively correlated inside a selected sampleโsuch as hospitalized patients, clinic registries, or deceased cohortsโbecause both factors independently increase the probability of being selected into the sample.
Causal Graph (DAG) and Mathematical Mechanism
In causal inference using directed acyclic graphs (DAGs), Berkson's paradox arises whenever an analyst conditions on a common effect (collider C) of two independent causes A and B:
A → C ← B
Consider Berkson's original hospital example: let A = diabetes and B = cholecystitis (gallbladder inflammation), which occur independently in the general population, and let C = hospital admission. Suppose either A or B can cause hospitalization. If a researcher analyzes only hospitalized patients (C = 1) and learns that a patient does not have diabetes (A = 0), that patient is far more likely to be in the hospital because of cholecystitis (B = 1). Mathematically, conditioning on the union A ∪ B induces a negative conditional dependence:
P(B = 1 | A = 0, C = 1) > P(B = 1 | A = 1, C = 1)
even though P(B = 1 | A = 0) = P(B = 1 | A = 1) in the unrestricted population!
Clinical Manifestations: The "Obesity Paradox" and COVID-19
Berkson's paradox frequently generates misleading clinical claims when observational analyses restrict samples to diagnosed patients, ICU admissions, or survivors (selection bias by death):
- The Obesity Paradox: Among patients already diagnosed with heart failure, diabetes, or acute cardiac arrest, obese individuals sometimes exhibit lower short-term mortality than normal-weight patients. However, collider-bias analyses (Schooling & Au Yeung, 2017; Horita et al., 2022) show that because both obesity and marked unmeasured genetic/smoking frailty cause cardiovascular disease (the collider), non-obese heart-failure patients in the clinic are disproportionately enriched for marked unmeasured frailtyโcreating an illusory survival advantage for obesity inside the diseased subgroup.
- Hospital-Based CaseโControl Studies: During the COVID-19 pandemic, early hospital-only studies paradoxically reported lower hospitalization odds among active smokers because testing/hospitalization selection differed sharply between smokers presenting with non-COVID respiratory symptoms and non-smokers.
Prevention and Analytical Remedies
To avoid Berkson's paradox, epidemiologists:
- Use Population-Based Controls: Draw caseโcontrol and cohort participants from random population registries rather than hospital-admission pools sharing an admission pathway with the exposure.
- Avoid Conditioning on Descendants of Outcomes: Do not stratify or restrict regression models by post-exposure mediators, disease diagnosis, or survival unless modeling selection explicitly via inverse probability of selection weighting (IPSW).
- Conduct Quantitative Bias Sensitivity Analyses: Bound the magnitude of unmeasured collider pathways using external population prevalence benchmarks.
Causal Directed Acyclic Graphs and Study Design Mitigation
In modern causal inference, Berkson's paradox is formalized using directed acyclic graphs (DAGs) as conditioning on a common effect, or collider. When two independent exposures or traits (A and B) both increase the probability of selection into a sample (S, such that A → S ← B), restricting analysis to S = 1 opens a non-causal backdoor path between A and B. Even if A and B are completely uncorrelated in the source population, conditioning on the collider induces a spurious negative association inside the selected cohort, distorting regression slopes, odds ratios, and genome-wide association effect estimates.
Preventing collider bias requires careful sampling and analytical corrections prior to causal interpretation. Whenever feasible, epidemiologists recruit population-based cohorts or random community controls rather than relying solely on hospitalized patients, volunteer biobanks, or clinical registries. When only selected data are available, researchers apply inverse probability of selection weighting (IPSW) using auxiliary demographic or healthcare-utilization probabilities, perform quantitative sensitivity analyses for selection odds ratios, and explicitly map selection nodes in causal diagrams before interpreting inverse correlations as protective effects.
In clinical genetics and electronic health record biobanks, Berkson's paradox frequently emerges when investigators study comorbidities among patients hospitalized for acute conditions or recruited through specialty clinics. Recognizing how healthcare-seeking behavior, referral thresholds, and diagnostic screening act as colliders prevents misinterpreting spurious negative correlations as biological antagonism between unrelated diseases.
Frequently Asked Questions
How does Berkson's paradox differ from confounding?
Confounding comes from a common cause (A ← Z → B), which creates bias if you fail to adjust for Z. Berkson's paradox (collider bias) comes from a common effect (A → C ← B), which creates bias precisely because you restricted or adjusted for C.
Does Berkson's paradox appear outside medicine?
Yes. In college admissions, if admission depends on academic grades plus athletic or artistic talent, grades and talent appear negatively correlated among admitted students even if they are uncorrelated or positively correlated among all applicants.
References
- Schooling, C. M., & Au Yeung, S. L. (Mar 2017). โSelection Bias by Deathโ and Other Ways Collider Bias May Cause the Obesity Paradox. Epidemiology. https://doi.org/10.1097/ede.0000000000000591
- Horita, N., et al. (Sep 15, 2022). Collider bias and the obesity paradox. Nutrition Reviews. https://doi.org/10.1093/nutrit/nuac077
- Kimbrell, J. M., & Moskowitz, A. (Oct 2025). The obesity paradox and collider bias, the elephant in the room. Resuscitation. https://doi.org/10.1016/j.resuscitation.2025.110736