Why studying hospital patients can make two unrelated diseases look linked
The idea
In 1946 the statistician Joseph Berkson published a short paper with a dry title about fourfold tables in hospital data. Its argument was that hospital patients are a poor place to look for links between diseases, even when the diagnoses are perfectly accurate.
The reason is the gate people pass through to be in the sample. If having condition A makes someone more likely to be admitted, and having condition B does too, then the inpatient population is not a fair slice of the public. People with both conditions have two separate routes in. Depending on the admission rates for each condition, the hospital sample can show an association between A and B that is absent, weaker or reversed in the population at large. Berkson’s own illustration involved gallbladder disease and diabetes.
A simple thought experimentthought experiment An imagined scenario used to test an idea by asking what we would judge or expect if it were real. shows the logic. Suppose a programme admits anyone who is either very good at maths or very good at music, and the two skills are unrelated in the population. Inside the programme, a student who is weak at maths was almost certainly admitted for music, so the two skills will look negatively related among the admitted, even though nothing connects them outside. The selection created the pattern.
Epidemiologists now call this Berkson’s bias, or admission rate biasBerkson's bias A false link between two conditions that appears when a sample, such as hospital patients, is chosen in a way that both conditions make more likely., and treat it as one example of collider biascollider bias A misleading pattern created when a study selects or adjusts for something that two other factors both influence.: when a variable is affected by two others, restricting or adjusting for it can manufacture a link between them.
A concrete case comes from data summarised by the Catalogue of Bias from David Sackett’s 1979 review. Comparing a random sample of the general population with the subset who had been in hospital in the previous six months, the relative odds of locomotor disease among people with respiratory disease was 4.06 in the hospital group and 1.06 in the population. Outside the hospital the two were essentially unrelated.
Why it might change a decision
The idea produces a specific, testable prediction: when both variables in a comparison influence who ends up in a sample, the association inside that sample should differ from the association in a sample drawn without that filter. That gives a researcher or reader something concrete to check, rather than a vague worry. One practical response in medical research is to draw controls from the general population rather than from other hospital patients, and the Catalogue of Bias lists this alongside using newly diagnosed cases and excluding patients admitted for other diseases.
The same structure shows up well outside hospitals. Hernández-Díaz, Wilcox, Schisterman and Hernán used causal diagrams to explain the long-standing birth weight paradox: among low birth weight babies, those whose mothers smoked appear to die less often than those whose mothers did not. Their argument is that comparing only small babies conditions on birth weight, which smoking affects and which shares causes such as congenital abnormalities with infant death. Among small babies of non-smokers, those other, deadlier causes are more common, so smoking can look protective without being so.
How solid is this?
The core claim is a mathematical consequence of selection, not a finding from a single experiment, so the question is less whether it can happen than how big it is in real data.
The most direct test is a 1978 study by R. S. Roberts, W. O. Spitzer, T. Delmore and D. L. Sackett. They used a household survey of health care use covering eight broad clinical conditions and six medications, looked at every pairwise association in the whole sample and in the hospitalised subgroup, and found statistically significant shifts in relative riskrelative risk How many times more likely an outcome is in one group than in another, where 1 means no difference.. But they described Berkson’s theory as having “relatively weak yet statistically significant predictive validity” for those shifts. In other words, the bias is real, but predicting its size and direction for a particular pair of conditions is hard, because it depends on admission rates that are rarely known.
Later hospital studies, such as a 2016 analysis of malaria and salmonella infection by Krumkamp and colleagues, have shown how much an estimate can move depending on which control groupcontrol group The people in a study who do not receive the treatment, so that the treated group can be compared against them. is used, which supports the general concern without pinning down a typical size.
The strongest competing explanation for any particular hospital-based association is that it is real: some diseases do cluster, through shared causes or one causing the other. Berkson’s bias does not mean every inpatient finding is spurious; it means the inpatient sample alone cannot tell the two apart.
Popular writing often uses Berkson’s paradox to explain everyday impressions, such as the idea that attractive people one dates tend to be unkind. These are tidy illustrations of the logic, but they are usually offered without data showing that selection, rather than a real pattern, explains them.
Sources
- Berkson J. (1946). Limitations of the application of fourfold table analysis to hospital data. Biometrics Bulletin (now Biometrics), 2(3), 47-53. PubMed 21001024
- Roberts RS, Spitzer WO, Delmore T, Sackett DL. (1978). An empirical demonstration of Berkson’s bias. Journal of Chronic Diseases (now Journal of Clinical Epidemiology), 31(2), 119-128. doi:10.1016/0021-9681(78)90097-8. McMaster Experts record
- Hernández-Díaz S, Wilcox AJ, Schisterman EF, Hernán MA. (2008). From causal diagrams to birth weight-specific curves of infant mortality. European Journal of Epidemiology, 23(3), 163-166. doi:10.1007/s10654-007-9220-4. PMC2722961
- Catalogue of Bias Collaboration: Spencer EA, Aronson JK, Nunan D, Heneghan C. (2018). Berkson’s bias (admission rate bias). Catalogue of Bias, Centre for Evidence-Based Medicine, Oxford. Source of the Sackett (1979) figures and the Krumkamp et al. (2016) example. catalogofbias.org