Berkson's Bias

Berkson's bias—also known as Berkson's fallacy, Berkson's paradox, or hospital admission rate bias—is a classic form of selection bias (specifically collider-stratification bias) in clinical epidemiology and biostatistics that arises when investigators conduct a case-control or cross-sectional study restricted exclusively to hospitalized inpatients (or clinic attendees, autopsy series, or tested individuals), and both the exposure (or first disease, D₁) and the outcome (or second disease, D₂) independently increase the probability that an individual is admitted to the hospital (H = 1). First proved mathematically in 1946 by Joseph Berkson, Chief of the Division of Biometry and Medical Statistics at the Mayo Clinic, and empirically confirmed in community vs. hospital comparisons by David L. Sackett (1979) and Roberts, Spitzer, Delmore, and Sackett (1978) before being generalized within the formal theory of Directed Acyclic Graphs (DAGs) by Judea Pearl (1995, 2000) and Miguel A. Hernán, Sonia Hernández-Díaz, and James M. Robins (2004), Berkson's bias demonstrates that conditioning on a common effect (a "collider" variable H in the causal graph D₁ → H ← D₂) induces a spurious statistical association between D₁ and D₂ even when the two conditions are completely independent (OR_population = 1.00) in the general community. Understanding Berkson's bias is essential for designing valid hospital-based case-control studies and avoiding spurious risk-factor conclusions in public health surveillance.

In 1929, the Johns Hopkins biometrician Raymond Pearl analyzed 7,500 consecutive autopsy records at Johns Hopkins Hospital and discovered a striking inverse (negative) association between active tuberculosis and malignant cancer: hospitalized patients who died with active tuberculosis had a much lower prevalence of malignancy than hospitalized patients who died without tuberculosis. Interpreting this hospital autopsy correlation as a biological antagonism, Pearl actually began treating terminal cancer patients with tuberculin injections!

Seventeen years later, Joseph Berkson (1946) published his landmark paper "Limitations of the Application of Fourfold Table Analysis to Hospital Data" in the Biometrics Bulletin. Berkson demonstrated mathematically that Pearl's finding—along with apparent hospital correlations between cholecystitis (gallbladder disease) and diabetes mellitus—required no biological interaction whatsoever. Instead, simply because every Condition (D₁ = cholecystitis, D₂ = diabetes, and D₃ = refractive error / control condition) had its own independent probability of sending a citizen from the general community into a hospital bed, sampling from inside the hospital walls systematically distorted the fourfold 2 × 2 contingency table and generated illusory odds ratios.

To see how Berkson's bias operates with complete mathematical clarity, follow Berkson's original formulation across a General Community Population of N = 10,000,000 people. Suppose we are testing whether Disease 1 (D₁: Cholecystitis, Exposed) is associated with Disease 2 (D₂: Diabetes Mellitus, Case), compared against a hospital control condition (D₃: Refractive Error, which is independent of D₁ and D₂).

Step 1: Community Independence and Differential Hospital Admission Probabilities

Assume that in the general community outside the hospital, Cholecystitis (D₁, prevalence p₁ = 0.03 = 3%) and Diabetes (D₂, prevalence p₂ = 0.01 = 1%) are statistically and biologically independent, so the probability of having D₂ given D₁ present (1.0%) is identical to the probability of having D₂ given D₁ absent (1.0%), yielding a True Community Odds Ratio (OR_population) = 1.00 (and true attributable risk AR = 0).

Now suppose that each condition independently carries a specific probability of causing a person to be admitted to the hospital:

• Probability of hospitalization due to Cholecystitis (D₁) alone: h₁ = 0.20 (20%).
• Probability of hospitalization due to Diabetes (D₂) alone: h₂ = 0.05 (5%).
• Baseline probability of hospitalization among individuals with neither D₁ nor D₂ (admitted for trauma, appendicitis, or control condition D₃): h₀ = 0.005 (0.5%).

What is the probability of hospitalization (h₁₂) for an individual who happens to have both Cholecystitis (D₁) AND Diabetes (D₂)? Because either disease alone is sufficient to trigger a hospital admission (and a patient fails to be hospitalized only if they escape hospitalization from D₁ and escape hospitalization from D₂ and escape baseline hospitalization h₀), under independent admission forces:

1 − h₁₂ = (1 − h₁)(1 − h₂)(1 − h₀)

h₁₂ = 1 − (1 − 0.20)(1 − 0.05)(1 − 0.005) = 1 − (0.80 × 0.95 × 0.995) = 1 − 0.7562 = 0.2438 (24.38%)

Step 2: Comparing the Community 2 × 2 Table vs. the Hospital Inpatient 2 × 2 Table

Let us multiply the exact cell counts in the community of 10,000,000 citizens by their respective hospital admission probabilities (h₁₂, h₁, h₂, and h₀) to see who enters the hospital cohort:

Now compute the Hospital-Based Odds Ratio (OR_hospital) that a clinical epidemiologist observes when comparing hospitalized Diabetes cases (a_hosp, c_hosp) against general hospitalized controls (b_hosp, d_hosp), or when comparing combinations of two diseases:

1. Comparing Inpatients With Cholecystitis (D₁⁺) vs. Inpatients Without Cholecystitis (D₁⁻) (The Paradox of Negative / Inverse Association):

Look at the last column of the table: among all hospitalized patients who have Cholecystitis (D₁⁺), only 1.20% (731 / 61,319) have Diabetes. Yet among all hospitalized patients who do not have Cholecystitis (D₁⁻), 9.96% (5,311 / 53,326) have Diabetes! Computing the inpatient odds ratio yields: OR_inpatient = (731 × 48,015) / (60,588 × 5,311) = 0.109 (a nearly 90% spurious "protective" reduction, p < 0.0001)! Why did this happen? Because every patient in the hospital had to have some reason to be admitted (H = 1). If an inpatient does not have Cholecystitis (D₁⁻), they are much more likely to have been admitted because they have Diabetes (D₂⁺, which has a 10-fold higher admission rate than baseline h₀). Conversely, if an inpatient already has Cholecystitis (D₁⁺), that alone explains why they are in the hospital, so their probability of also having Diabetes is close to the low 1% general population rate!

1. Berkson's Case-Control Bias When Controls Have a Lower Hospitalization Rate (h_control < h₂) (Spurious Positive Association):

Conversely, suppose an investigator selects a specific hospital control group D₃ (such as patients admitted once for a minor elective procedure with low admission rate h₃) and compares them against Diabetes cases (D₂) with respect to Cholecystitis (D₁). Because patients possessing both D₁ and D₂ experience two independent doors into the hospital (h₁₂ = h₁ + h₂ − h₁h₂), double-diagnosis patients (D₁⁺D₂⁺) are enriched in hospital wards relative to single-diagnosis patients whenever the baseline control admission rate differs, generating a spurious positive odds ratio (OR > 1.0) (confirmed empirically in community vs. hospital household surveys of 2,784 adults by Roberts, Spitzer, Delmore, and Sackett, 1978, where true community odds ratios of 1.06 between respiratory and musculoskeletal diseases jumped to 1.66 [p < 0.001] among hospitalized patients).

In modern causal inference formulated by Judea Pearl (1995, 2000) and Hernán, Hernández-Díaz, and Robins (2004), Berkson's bias is the canonical archetype of Collider-Stratification Bias:

• In a Directed Acyclic Graph (DAG), a node C is a fork (confounder) when arrows point out of it (E ← C → D), a chain (mediator) when arrows pass through it (E → M → D), and a collider when two causal arrows collide into it:

E (Exposure / Disease 1) → H (Hospitalization) ← D (Outcome / Disease 2)

• By the rules of d-separation:
  1. An unconditioned path through a collider (E → H ← D) is naturally BLOCKED. In the whole community (where we do not restrict or stratify on hospitalization H), E and D are marginally independent.
  2. Conditioning on a collider H (or any descendant of a collider) OPENS the path between its parents E and D! In probability terms, conditioning on the effect H = 1 makes its two independent causes compete to "explain away" the observed hospitalization (Pearl's "Explaining-Away" Effect).

Modern Manifestations Beyond Hospital Wards

Collider-stratification (Berkson-type) bias extends far beyond traditional hospital admission studies:

1. COVID-19 Voluntary Testing Bias (Griffith et al., 2020): During the SARS-CoV-2 pandemic, studies restricted strictly to individuals who received a clinical PCR test (T_tested = 1) observed a puzzling apparent protective association between active cigarette smoking and testing positive for SARS-CoV-2. Why? Because both chronic smoker's cough (S → T_tested) and true SARS-CoV-2 infection (I → T_tested) independently caused citizens to seek symptom-triggered PCR testing! Conditioning on the collider T_tested = 1 made non-COVID smokers flood the tested denominator, artificially depressing SARS-CoV-2 positive predictive value and odds ratios among smokers.
2. The Birth-Weight and Obesity Paradoxes (Hernández-Díaz, Wilcox, & Schisterman, 2006): Among all newborns, maternal smoking increases infant mortality. Yet when epidemiologists restrict analysis strictly to low-birth-weight infants (LBW = 1), maternal smoking appears to reduce neonatal mortality! Here, Low Birth Weight is a collider caused by both maternal smoking (S → LBW) and catastrophic congenital/placental anomalies (U → LBW → Death). If a low-birth-weight baby has a non-smoking mother, its low birth weight is much more likely caused by a lethal chromosomal or placental defect (U), creating a classic Berkson inverse paradox (identical in structure to the "obesity paradox" in patients hospitalized with heart failure and the healthy worker effect).

To prevent Berkson's bias in observational epidemiology, investigators apply four design safeguards:

1. Population-Based / Community Controls: Whenever feasible, sample both cases (via a comprehensive population disease registry) and controls (via random population sampling or primary-care rosters) from the defined source community rather than from hospital wards.
2. Selection of Hospital Controls Whose Admission Is Strictly Independent of the Exposure (h_control|E=1 = h_control|E=0): When a hospital-based case-control design is necessary (to ensure comparable recall accuracy and biospecimen collection), Wacholder, McLaughlin, Silverman, and Mandel (1992) proved that Berkson's bias is eliminated if controls are sampled from incident admissions for acute conditions that have zero causal or behavioral relationship with the exposure under study (for example, using acute non-alcohol-related orthopedic fractures or acute appendicitis as controls when studying dietary exposures and colon cancer, while excluding chronic smoking- or obesity-related respiratory/cardiovascular diagnoses).
3. Pooling Multiple Diverse Diagnostic Control Categories: Sampling hospital controls across a broad variety of unrelated diagnostic categories dilutes any single admission-rate anomaly that might affect one specific control diagnosis.
4. Inverse Probability of Selection Weighting (IPSW): When the probabilities of hospitalization or testing Pr(H = 1 | E, D, C) can be estimated from external community health surveys or electronic health record catchments, weighting each hospitalized subject by 1 / Pr(H = 1 | E, D, C) mathematically reconstructs the unbiased source population pseudo-table and eliminates collider-stratification bias.

1. Berkson, J. (1946). "Limitations of the application of fourfold table analysis to hospital data." Biometrics Bulletin, 2(3), 47–53. https://doi.org/10.2307/3002000
2. Roberts, R. S., Spitzer, W. O., Delmore, T., & Sackett, D. L. (1978). "An empirical demonstration of Berkson's bias." Journal of Chronic Diseases, 31(2), 119–128. https://doi.org/10.1016/0021-9681(78)90097-890097-8)
3. Hernán, M. A., Hernández-Díaz, S., & Robins, J. M. (2004). "A structural approach to selection bias." Epidemiology, 15(5), 615–625. https://doi.org/10.1097/01.ede.0000135174.63482.43
4. Griffith, G. J., Morris, T. T., Tudball, M. J., Herbert, A., Mancano, G., Pike, L., Sharp, G. C., Sterne, J., Palmer, T. M., Davey Smith, G., Tilling, K., Zuccolo, L., Davies, N. M., & Hemani, G. (2020). "Collider bias undermines our understanding of COVID-19 disease risk and severity." Nature Communications, 11(1), 5749. https://doi.org/10.1038/s41467-020-19478-2