An empirical demonstration of Berkson's bias

doi:10.1016/0021-9681(78)90097-8

Journal of Chronic Diseases

Volume 31, Issue 2, February 1978, Pages 119-128

Journal of Chronic D...

https://doi.org/10.1016/0021-9681(78)90097-8 Get rights and content

Abstract

The importance of Berkson's bias (a systematic bias in the distribution of disease amongst hospitalized patients) in certain types of hospital based research has not been substantiated with empirical data. This paper firstly reviews the theoretical origins of the bias and secondly attempts to provide supportive empirical evidence. The basis of the analysis has been data generated in household surveys designed to capture health utilization information. Sufficient data were available for eight broad clinical conditions and six medications. All possible pairs of associations have been examined in both the entire data set and the hospitalized subgroups; statistically significant differences in relative risk were then identified. Berkson's theory provides a relatively weak yet statistically significant predictive validity for the change in relative risk from community to hospital.

References (18)

H.E.S. Pearson et al.
Stress and occlusive coronary-artery disease
Lancet
(1963)
H.K. Armenian et al.
Relation between prostatic hyperplasia and cancer of the prostate
Lancet
(1974)
B. Armstrong et al.
Retrospective study of the association between use of Rauwolfia derivatives and breast cancer in English women
Lancet
(1974)
O.P. Heinonen et al.
Reserpine use in relation to breast cancer
Lancet
(1974)
L. Garfinkel et al.
Breast cancer and hypertension
Lancet
(1974)
J.H. Medalie et al.
Association between blood-pressure and peptic-ulcer incidence
Lancet
(1970)
J Berkson
Limitations of the application of 4-fold tables to hospital data
Biomet Bull
(1946)
J Cornfield
A method of estimating comparative rates from clinical data. Applications to cancer of the lung, breast and cervix
J Nat Canc Inst
(1951)
United States Department of Health, Education, and Welfare
International classification of diseases, adapted for indexing hospital records by diseases and operations

There are more references available in the full text version of this article.

Cited by (111)

A short history of the concept of bias in epidemiology
2021, Revue d'Epidemiologie et de Sante Publique
La question des biais constitue un enjeu méthodologique central pour l’épidémiologie. Cependant, l’histoire de ce concept n’a fait l’objet que de rares études. De plus, la définition classique du biais comme déviation systématique des résultats ou des inférences par rapport à la vérité, telle qu’elle peut apparaître dans les dictionnaires d’épidémiologie, ne semble correspondre ni à la manière dont les épidémiologistes l’utilisent en pratique, ni aux différentes définitions qui jalonnent l’histoire de l’épidémiologie. Il s’agit donc d’expliquer ce paradoxe.
Cet article historique et épistémologique propose d’analyser, depuis les articles classiques des années 1950 sur le lien entre tabagisme et cancer du poumon jusqu’aux manuels d’épidémiologie les plus récents, les diverses occurrences du mot « biais » ainsi que la manière dont les épidémiologistes utilisent, définissent et rectifient ce concept au fil du temps.
Nous montrons que l’article de D.L. Sackett de 1979, consacré aux biais dans la recherche étiologique, vient clore, au moins provisoirement, une réflexion amorcée trente années auparavant en épidémiologie populationnelle. Plus précisément, nous montrons que la définition donnée par Sackett du biais se construit moins en relation avec les besoins et les buts de l’épidémiologie populationnelle que de ceux de l’épidémiologie clinique. Au même moment, le concept de biais est redéfini au sein de l’épidémiologie populationnelle, notamment par K.J. Rohtman, comme une menace à la validité interne d’une étude, dans le cadre d’une redéfinition de l’étude épidémiologique comme exercice de mesure.
Ainsi, il convient de distinguer entre deux concepts de biais : un concept épidémiologique de biais, conçu comme un défaut de validité interne d’une étude observationnelle ; et un concept médical de biais, conçu comme un écart par rapport à la vérité. Le premier porte sur le plan et la méthodologie des études épidémiologiques ; le second est plus général et invite les épidémiologistes et les médecins à une certaine forme de scepticisme, voire de criticisme, envers leurs propres inférences.
Bias is a major methodological issue for epidemiology. However, only a few studies have been dedicated to the past and present formulations of the concept of bias. Moreover, the classical definition of bias as systematic deviation from the truth of results or inferences, definition which can be found in dictionaries of epidemiology, does not seem to either match the way epidemiologists use it in practice, or correspond to the different definitions given throughout its history. It is consequently important to elucidate this paradox.
In this historical and conceptual article, we study the different uses of the word “bias” in epidemiological literature, from classic articles in the 1950's about the link between smoking and lung cancer to the most recent epidemiology textbooks, the objective being to analyze the ways in which epidemiologists have defined, applied and modified this concept over time.
We show that D.L. Sackett's article on bias in analytic research, published in 1979, put an end, at least temporarily, to reflection in populational epidemiology that started thirty years before. More precisely, we show that Sackett's definition of bias corresponds more to the needs and goals of clinical epidemiology than to those of populational epidemiology. Concomitantly, populational epidemiologists such as K.J. Rothman redefined bias as a threat to the internal validity of a study, and epidemiological study as an “exercise in measurement of an effect rather than as a criterion-guided process for deciding whether an effect is present or not”.
It is thereby important to draw a distinction between two notions pertaining to bias: an epidemiological concept of bias, viewed as the lack of internal validity of an observational study; and a medical concept of bias, defined as deviation from the truth. The former concerns the design and methodology of epidemiological studies; the latter is more general and impels epidemiologists and physicians to be skeptical, and even critical, towards their own inferences.
Computing disease incidence, prevalence and comorbidity from electronic medical records
2016, Journal of Biomedical Informatics
Electronic medical records (EMR) represent a convenient source of coded medical data, but disease patterns found in EMRs may be biased when compared to surveys based on sampling. In this communication we draw attention to complications that arise when using EMR data to calculate disease prevalence, incidence, age of onset, and disease comorbidity. We review known solutions to these problems and identify challenges for future work.
The ENPP1 K121Q polymorphism is associated with type 2 diabetes and related metabolic phenotypes in a Taiwanese population
2016, Molecular and Cellular Endocrinology
Citation Excerpt :
Due to a lack of heterogeneity, we were able to pool data from the two study cohorts and thereby considerably increase the statistical power of the analysis (De Cosmo et al., 2009). Furthermore, the hospital-setting for recruitment used in our study could be prone to selection bias, known as Berkson’s bias (Roberts et al., 1978), so that the observed prevalence and distribution of T2D may not be extrapolated to the entire population. Additionally, hip circumference data were unavailable, so we were unable to calculate waist-hip ratios to give more information regarding the association between the explored SNP and phenotypes of T2D and obesity.
Increased risk of developing type 2 diabetes (T2D) has been associated with a single nucleotide polymorphism (SNP), rs1044498 (K121Q), in the ectonucleotide pyrophosphatase/phosphodiesterase 1 (ENPP1) gene, but this association is unclear among Asians. In this replication study, we reassessed whether the ENPP1 rs1044498 SNP is associated with T2D, obesity, and T2D/obesity-related metabolic traits in a Taiwanese population. A total of 1513 Taiwanese subjects were assessed in this study. The ENPP1 rs1044498 SNP was genotyped by the Taqman assay. T2D/Obesity-related quantitative traits, such as waist circumference and fasting glucose, were measured. Our data showed a significant association of the ENPP1 rs1044498 SNP with T2D (P < 0.001) among the subjects. Moreover, the ENPP1 rs1044498 SNP was significantly associated with T2D/obesity-related metabolic traits, such as waist circumference (P = 0.002) and fasting glucose (P < 0.001), among the subjects. However, we found no association of ENPP1 rs1044498 with obesity (BMI ≧ 27 kg/m²). Our study indicates that the ENPP1 rs1044498 SNP is associated with T2D, waist circumference, and fasting glucose in Taiwanese subjects.
Pain intensity and smoking behavior among treatment seeking smokers
2016, Psychiatry Research
Citation Excerpt :
Given that gender has been shown to influence both pain ratings and smoking behavior (Jamner et al., 1998; Unrod et al., 2004), it is important to determine the unique contribution of both pain and gender when examining smoking behavior. Second, past work has focused almost exclusively on treatment-seeking pain patients, which may have resulted in a severity range restriction governing or influencing the strength of past observed associations (i.e., a Berskon’s bias, wherein the most severe cases are studied, (Roberts et al., 1978). In a recent example of research on pain-smoking relationship among treatment seeking smokers, Bastian et al. (2015) examined the role of smoking behavior on pain intensity.
Empirical evidence supporting the interplay between pain intensity and tobacco smoking has been growing. The current investigation advances this work in three important ways: (1) controlling for negative affectivity and gender; (2) examining pain intensity in smokers from a community sample, rather than specialized pain treatment centers; and, (3) studying smokers who are highly motivated to quit. Participants were adult smokers (N=112; 35% female; M_age=41.4, SD=13.1) participating in a larger study examining barriers to cessation during a self-guided quit attempt. At baseline, participants completed self-report measures on pain intensity and smoking severity outcomes. As hypothesized, more intense pain was significantly associated with all four smoking severity variables: years as a daily smoker, current cigarettes per day, cigarettes per day during the heaviest lifetime smoking period, and current level of nicotine dependence. These associations remained when taking into account the variance accounted for by gender and negative affectivity. These data provide evidence that more intense pain is related to more severe smoking behavior and nicotine dependence. Pain reduction could be an important target in regard to smokers with chronic pain.
Bias in clinical epidemiological study designs
2013, Allergologia et Immunopathologia
Citation Excerpt :
Roberts et al. published the first empirical study demonstrating Berkson bias, affecting the relationship between respiratory diseases and diseases of the locomotor system. In the study published by Roberts, involving a sample of 2784 subjects, 257 individuals had been admitted to hospital in the six months prior to the study.9 The data in Table 3 correspond to the global population and the subgroup of hospitalised individuals with respiratory diseases and diseases of the locomotor system in the study of Roberts.
Systematic error, or bias, is error that occurs in each measurement made and which has a direction, i.e., the measured value is always either greater or smaller than the true value. The presence of systematic error directly affects the internal validity of the study, and indirectly affects the external validity of the results obtained. In general, such error can be classified as selection bias, classification bias or confounding bias.
It is essential to deal with possible bias in the research design phase, since only confounding bias can be controlled in the phase corresponding to analysis of the results.
History of the Rochester epidemiology project: Half a century of medical records linkage in a US population
2012, Mayo Clinic Proceedings
Citation Excerpt :
The risk of selection bias when describing the clinical spectrum of severity and the outcomes of diseases was finally recognized and addressed.19-21 In 1946, Joseph Berkson described the selection bias that may occur in case-control studies based on hospital admissions (known as Berkson bias).22,23 Stimulated by the publication of a study on the incidence of multiple sclerosis in Olmsted County,9 Leonard T. Kurland came to Mayo Clinic for a fellowship in neurology and returned a decade later to become head of what is now called the Department of Health Sciences Research (Figure 3, A).24
The Rochester Epidemiology Project (REP) has maintained a comprehensive medical records linkage system for nearly half a century for almost all persons residing in Olmsted County, Minnesota. Herein, we provide a brief history of the REP before and after 1966, the year in which the REP was officially established. The key protagonists before 1966 were Henry Plummer, Mabel Root, and Joseph Berkson, who developed a medical records linkage system at Mayo Clinic. In 1966, Leonard Kurland established collaborative agreements with other local health care providers (hospitals, physician groups, and clinics [primarily Olmsted Medical Center]) to develop a medical records linkage system that covered the entire population of Olmsted County, and he obtained funding from the National Institutes of Health to support the new system. In 1997, L. Joseph Melton III addressed emerging concerns about the confidentiality of medical record information by introducing a broad patient research authorization as per Minnesota state law. We describe how the key protagonists of the REP have responded to challenges posed by evolving medical knowledge, information technology, and public expectation and policy. In addition, we provide a general description of the system; discuss issues of data quality, reliability, and validity; describe the research team structure; provide information about funding; and compare the REP with other medical information systems. The REP can serve as a model for the development of similar research infrastructures in the United States and worldwide.

View all citing articles on Scopus

^☆: supported by National Health Grant 606-21-95, Data generated through Ontario Ministry of Health Grants, Burlington-DM34, Smithville-DM35, Hamilton-DM53.

View full text

An empirical demonstration of Berkson's bias☆

Abstract

Lancet

Lancet

Lancet

Lancet

Lancet

Lancet

Limitations of the application of 4-fold tables to hospital data

Biomet Bull

A method of estimating comparative rates from clinical data. Applications to cancer of the lung, breast and cervix

J Nat Canc Inst

International classification of diseases, adapted for indexing hospital records by diseases and operations