Berkson's Fallacy Research Articles

Relation of alcohol/tobacco use with metastasis, hormonal (estrogen and progesterone) receptor status and c-erbB2 protein in mammary ductal carcinoma.

An association between alcohol/tobacco use and risk of metastasis in breast cancer has been clearly shown. The present study explored, in 48 samples of tissue from mammary ductal carcinoma (taken from Mexican women with an average age of 58.2±10.9 years), the association of risk of metastasis with the status of hormonal receptors and the c-erbB2 protein (by immunohistochemistry) as well as clinical, histopathological and sociodemographic factors. Of 48 patients, 41.6% (20/48) presented with metastasis, 43.8% were positive for the estrogen receptor (RE+), 31.3% for the progesterone receptor (RP+) and 47.7% for c-erbB2 (c-erbB2+). The following combinations were found: RE+/RP+/c-erbB2+ 8.3%, RE+/RP+ 22.9%, RE+/RP- 20.8%, RE-/RP+ 8.3%, RE-/RP-/c-erbB2- 22.9% and RE-/RP- 47.8%. There were 12 patients who used alcohol/tobacco, of which 91.6% did not present metastasis and 81.9% were RE-/RP-. Compared to the RE-/RP-/c-erbB2+, the RE+/RP+/c-erbB2+ group had a 15-fold greater risk for metastasis (95%CI, 0.9-228.8, p=0.05). The carriers of the double negative hormonal receptors had a 4.7 fold greater probability of being (or having been) smokers or drinkers (95%CI, 1.0-20.4, p = 0.03). There was a clear protective effect of using alcohol and/or tobacco, in the cases included in the present study of mammary ductal carcinoma, associated with double negative hormonal receptors. However, this association could be due to a protective factor not measured (Neyman bias) or to a bias inherent in the rate of hospitalization (Berkson fallacy). This question should be explored in a broad prospective longitudinal study.

Controlling asthma during pregnancy prevents asthma in children: a Berkson fallacy?

To the Editors: Assessment of risk factors is absolutely essential to understanding how and why asthma develops in children. We read with great interest the study by Martel et al. 1, which was recently published in the European Respiratory Journal , and considered the association between mothers' asthma in pregnancy and their children's risk of developing the disease 1. In that case–control study, children whose mothers experienced moderate-to-severe uncontrolled asthma during pregnancy were at higher risk of being diagnosed with asthma than children whose mothers had …

Treatment-naive active alcoholics have greater psychiatric comorbidity than normal controls but less than treated abstinent alcoholics

Background Most alcoholism research in the U.S. uses convenience samples of treated alcoholics. The findings from treated samples have traditionally been applied to all alcoholics, including the 75% of alcoholics who are untreated. Improper generalization from select samples to an entire population is called ‘Berkson's fallacy’. We compared untreated versus treated alcoholics, in order to ascertain whether both groups belonged to the same population with regard to psychiatric comorbidity. Methods We compared psychiatric comorbidity in 1) active treatment-naive alcoholics (TNA; n = 86) 2) treated long-term abstinent alcoholics (TAA; n = 52) and 3) non-alcoholic controls (NAC; n = 118). We examined lifetime and current diagnoses, lifetime symptom counts, and psychological measures in the anxiety, mood and externalizing disorder domains. Results TNA did not differ from NAC in psychiatric diagnosis rates, were abnormal compared to NAC on all psychological measures, had more externalizing symptoms than NAC, and showed a strong trend for men to have more symptoms in the mood and anxiety domains. TAA compared to TNA had higher diagnosis rates (all domains), symptom counts (all domains), and psychological measures of deviance proneness, but were comparable to TNA on anxiety and mood psychological measures. Conclusions The abnormal thinking (psychological measures) in TNA (versus NAC) does not extend to behavior (symptoms) to the degree that it does in TAA. These results underline the importance of the use of subdiagnostic measures of psychiatric comorbidity in studies of alcoholics. The finding of lesser comorbidity in TNA versus TAA confirms the presence of Berkson's fallacy in generalizing from treated samples to all alcoholics.

Treated and treatment-naive alcoholics come from different populations

In most research on alcoholism, convenience samples of individuals who have been in some type of treatment are used. Berkson's fallacy results when the associations found in studies of select samples are incorrectly presumed to apply to all alcoholics (i.e., including untreated alcoholics in the general population). In the current study, we examined whether treated and untreated alcoholics have similar early alcohol use histories by comparing abstinent alcoholics (treated and sober at least 6 months) with treatment-naive alcoholics (active drinkers). We studied 14 pairs of women and 25 pairs of men matched on the age at which they first met criteria for heavy alcohol use (women, 80 drinks per month; men, 100 drinks per month). The timeline follow-back interview method was used to gather retrospective alcohol use information. Alcohol dose and duration of use were subsequently computed for two intervals: (1) time between the person's first drink and date at which the person met criteria for heavy drinking and (2) period between when criteria for heavy drinking were met and current age of the treatment-naive person from each pair. During the period before the matching “heavy drinking” criteria were met, alcohol dose did not differ between groups. In the period after criteria for heavy alcohol use were met, in comparison with treatment-naive alcoholics, the treated alcoholics had higher average and peak alcohol doses. We rejected the hypothesis that the treatment-naive alcoholics and the treated alcoholics have similar alcohol use trajectories over time, with the treatment-naive sample simply being observed earlier in its alcohol use histories. Instead, we concluded that the two groups come from different populations with regard to alcohol use. In fact, the treated alcoholics had alcohol doses more than 50% higher than those of treatment-naive alcoholics in the years just after they began drinking heavily. This finding supports the suggestion that results from studies of alcoholics in treatment or after treatment (i.e., most studies of alcoholics) cannot be generalized to untreated individuals (who make up the majority of alcoholics).

High Incidence of Myocardial Ischemia during Postpartum Hemorrhage or Poor Management of Hemorrhagic Shock?

University of Alabama, Birmingham, Alabama. froelich@uab.eduWith interest I read the recent article by Karpati et al. 1in the January issue of Anesthesiology. I have several questions, mostly regarding the study design, and hope the authors will comment.It was the defined goal of the study to prospectively evaluate the incidence of myocardial ischemia in women with postpartum bleeding (introduction). In a prospective trial, one would expect to learn about admission criteria, predefined group selection criteria, and outcome measures. This did not seem to be the case in this study. The authors did not mention inclusion or exclusion criteria and, for the most part, seemed to use laboratory data from patient admission as outcome measures. In the Discussion, the authors assert, “myocardial ischemia–induced injury associated with hemorrhagic shock is likely to increase the incidence of [postpartum hemorrhage–associated] cardiac complications.” This statement implies that the authors observed cardiac complications. How did the authors define or measure this variable?The title of this study describes the incidence of ischemia, which implies that the authors identified new cases and the population at risk. The study cohort exhibited severe postpartum hemorrhage. Most subjects were troponin positive or negative at the beginning of the study. There was a good possibility of an admission rate bias (Berkson fallacy).There were 21 comparisons between troponin-positive and -negative groups, but there was no mention of an adjustment for multiple comparisons. With reference to the multiple logistic regression used, it is a rule of thumb to have approximately 10 outcomes for every independent variable in the model to achieve adequate power. This would call for a total sample size of at least 210 patients in this study. To illustrate this point, how can the authors assure that being catecholamine dependent, a cell with zero outcomes, is not predictive of ischemia?The authors should comment on their definition of design, direction (prospective vs. cross-sectional), and study factors. They describe a cohort of poorly resuscitated female patients in “severe hemorrhagic shock” with an apparent high prevalence of being cardiac troponin positive. Did the authors find that surprising?University of Alabama, Birmingham, Alabama. froelich@uab.edu

Cortical Gray Matter Loss in Treatment‐Naïve Alcohol Dependent Individuals

Most studies of the impact of alcohol dependence on the brain have examined individuals in treatment. Such samples represent a small proportion of alcoholics in the general population. Such samples may embody a bias (Berkson's fallacy) if the association between variables (for example, alcoholism and cortical gray matter loss) differs between the population of alcoholics in treatment and alcoholics in the general population. Our objective was to determine if treatment-naïve alcoholics show structural brain changes versus controls and to compare our findings with reports evaluating alcoholic samples drawn from treatment populations. Structural MRI was used to assess whole brain and regional volumes of cortical gray matter and white matter in 24 young to middle-aged treatment-naïve alcohol-dependent males versus 17 controls. Cortical gray matter volumes in alcohol-dependent individuals were negatively associated with age and lifetime duration of alcohol use (which were highly confounded). These subjects showed reduced whole brain (p < 0.05), prefrontal (p < 0.01), and parietal (p < 0.05) cortical gray matter compared with controls. White matter and temporal cortex, tissues that usually show volume reductions in samples drawn from treatment, did not differ between treatment-naïve alcoholics and controls (all p > 0.40). Our findings are consistent with the hypothesis that structural brain changes in treatment-naïve alcoholics are less severe than those reported in clinical samples of alcoholics, perhaps due to less concomitant psychopathology and a reduced severity of alcoholism in treatment-naïve alcoholics. However, caution must be taken when comparing our findings with results from clinical samples, as we did not directly compare treatment-naïve alcoholics with treated alcoholics and our treatment-naïve sample tended to be younger than the (clinical) samples reported in the literature. Nevertheless, we suggest that most of the reports of the central nervous system consequences of alcoholism may not accurately describe the majority of alcoholic-dependent individuals.

A fresh look at proportional mortality ratios

This is a review of the use of proportional mortality ratios (PMRs). District health authorities rarely concern themselves with searching for unknown causes of disease but do need to monitor the health of a population. PMRs are a good way to do this. Textbooks of epidemiology tend to see PMRs as biased measures of risk and give them little attention. This review shows that the bias is small and of no practical importance. PMR studies can be seen as case-control studies and the major bias in these studies as analogous to Berkson's fallacy. We can then use the same techniques to reduce bias in PMR studies that are used in case-control studies. These are the use of several controls, the use of positive and negative controls, and the use of only one type of death in the denominator, rather than all causes of death. This article reviews these means of minimising bias using examples to show that careful selection of controls can overcome many of the supposed problems of PMR studies. It also shows how PMRs can control confounding in a way that SMRs cannot. PMR studies should be more widely used to monitor the health of the population.

A fresh look at proportional mortality ratios

This is a review of the use of proportional mortality ratios (PMRs). District health authorities rarely concern themselves with searching for unknown causes of disease but do need to monitor the health of a population. PMRs are a good way to do this. Textbooks of epidemiology tend to see PMRs as biased measures of risk and give them little attention. This review shows that the bias is small and of no practical importance. PMR studies can be seen as case-control studies and the major bias in these studies as analogous to Berkson's fallacy. We can then use the same techniques to reduce bias in PMR studies that are used in case-control studies. These are the use of several controls, the use of positive and negative controls, and the use of only one type of death in the denominator, rather than all causes of death. This article reviews these means of minimising bias using examples to show that careful selection of controls can overcome many of the supposed problems of PMR studies. It also shows how PMRs can control confounding in a way that SMRs cannot. PMR studies should be more widely used to monitor the health of the population.

Disability-

Objectives: The medical definitions of impairment (lesion, disability, the functional consequence thereof), and handicap (the social consequence, sanctified by the World Health Organization), differ from legal concepts, in which impairment encompasses some of disability, while disability encompasses most of handicap. The medical testimony provides evidence; the designated legal representative (judge, arbitrator or panel) determines whether disability is present, its extent, and the likely duration, based on the medical evidence and other factors. This brief personal review addresses some of the shadings and conflicts the encounter between two disciplines produces. Findings: Contradictions in "science" are adjudged on the preponderance of evidence, often an exercise in biostatistics, and the law representative sometimes makes medical judgments, whilst physicians and other health care professionals, because of training and humanitarian aims, often create the very dilemma they want to forestall. In relating the impairment to a temporally antecedent event, to permit a decision of casual or causal, Berkson's fallacy comes into play. Conclusions: The World Health Organization sequence, breaking down the problem into its various components, permits the physician to appreciate the evidentiary portions of the problem and thus to define disability medically in such a way that the "safety net" works appropriately.

Constructing Some Categorical Anomalies

Abstract Three situations in which the interpretation of categorical data can appear anomalous are considered: (a) Berkson's fallacy, (b) an example involving retrospective and prospective epidemiological studies from Neyman, and (c) Simpson's paradox. The odds ratio is used as the statistic of interest. Methods are presented to determine whether an anomaly could exist for a given parameterization and to construct an anomaly where feasible. The methods may require at most the preparation of a simple graph and some simple hand calculator computations.

A matched-paired comparison of treated versus untreated schizophrenia spectrum cases. A high-risk population study.

As a part of the American-Danish prospective study of children of schizophrenic mothers, psychiatrically hospitalised and untreated cases of schizophrenia spectrum conditions, i.e. borderline schizophrenia, schizoid and paranoid personality disorders, were compared on a number of characteristics. This exploratory analysis revealed that the groups were similar to each other on childhood conditions, sociodemographic variables, premorbid IQ and levels of schizophrenia-related psychopathology. The hospitalised group, however, exhibited higher levels of concomitant psychopathology, such as substance abuse, affective symptoms and psychopathic tendencies. It is suggested that the clinical population may not be representative of the diagnostic category in question owing to co-existence of confounding symptomatology (Berkson's fallacy).

Towards a general model of non-random sampling and the impact on population correlation: generalizations of Berkson's Fallacy and restriction of range.

A long history of restriction of range research has developed in the psychometric literature under the assumption of strict truncation. A conceptualization of the general problem of non‐random sampling is proposed leading to a model of which both restriction of range and Berkson's Fallacy (Berkson, 1946) are special cases. Cast in the framework of unequal sampling probabilities operating separately on the variables of interest, two major forms of the model are investigated: an additive and an interactive (multiplicative) form. The influence of unequal selection probabilities on population values of the bivariate normal correlation (ρ) is described and findings from other research on restriction of range are shown to be special cases of these results. Implications for inferences from data and evaluation of research designs are discussed.

Causes of Death in Duodenal and Gastric Ulcer

An analysis has been made of 235 deaths that occurred among 1905 patients with peptic ulcer who constituted a random sample of the occurrence of ulcer disease in an area of Denmark comprising half a million inhabitants. The disease itself, according to the death certificate, was considered the primary cause of death in 10% of the cases; half of these had been operated on immediately before death. The other patients died more frequently than expected from the following causes: chronic bronchitis, pulmonary emphysema, cancer of the lung, cirrhosis of the liver, and cancer of the pancreas. Although the comorbidity with chronic bronchitis and emphysema was especially pronounced in patients with gastric ulcer, the association with liver cirrhosis and cancer of the pancreas occurred only in patients with duodenal ulcer. In women the mortality rate attributable to cardiac and vascular diseases was lower than expected. No excess coincidence of suicide was found. Berkson's fallacy is considered to be of much less importance as a possible explanation of the comorbidity found in the present study than in the majority of publications concerned with this question.

Berkson Fallacy and Studies on Reserpine and Breast Cancer-Reply

In Reply.—The survey that Dr Jick is defending1could have yielded a spurious association between reserpine use and breast cancer in women, if some uneven selection rates were taking place. The articles1.2give no assurance that precautions were taken to prevent the Berkson fallacy. Let me illustrate how a spurious association could have occurred, and could have yielded the precise data reported by Dr Jick and his associates. In describing the Boston Collaborative Drug Surveillance studies, the investigators stated2that the survey patients were served in 24 hospitals that provided 45% of all general hospital beds in the greater metropolitan Boston area. They estimated that approximately 1,300,000 people are served by these hospitals. The total population served might be estimated to be about 2.9 million. Half of these would be women: 1,450,000. The incidence of breast cancer in women in the United States is about

Berkson fallacy revisited. Spurious conclusions from patient surveys.

Clinical surveys of "selected" convenient patient samples (hospital or clinic) can lead to spurious conclusions unless the relationship of selected patients to the source population is carefully controlled. The possibility of biased conclusions from uneven selection rates of patients is called Berkson fallacy. Studies of cerebral palsy and low birth weight provide clear examples of Berkson fallacy. A fictitious clinical survey of cerebral palsy and low birth weight is presented to illustrate how Berkson fallacy may lead to interesting but spurious conclusions from patient surveys.

A Simple Test for Berkson's Fallacy

A Simple Test for Berkson's Fallacy

Epidemiology of ischemic heart disease among white males: I. Relationship between coronary atherosclerosis and cancer of various sites

1. 1. An epidemiologic study was made of the relationship between cancer and ischemic heart disease among white males, utilizing an autopsy population. The coronary arteries of patients with cancer of 11 different sites were compared with the coronary arteries of accident victims. The records were matched for age, sex, race, year of death within a 5 year limit, and other factors known to influence the prevalence of coronary atherosclerosis. Adjustments were made also for hypertension and body weight at time of death. Next, the coronary arteries of cancer patients were compared with those of “all other (noncancer) patients” in the autopsy population. 2. 2. There were no significant differences in the severity of coronary artery atherosclerotic blockage or coronary clot formation when cancer patients were compared with accident victims. However, cancer patients had significantly less severe coronary artery atherosclerotic blockage and fewer coronary clots when they were compared with all other (noncancer) patients in the autopsy population. These conflicting results constitute an example of Berkson's fallacy. It was reasoned that accident victims were the group in the autopsy population whose coronary arteries should be most representative of those of the general population. On the other hand, the coronary arteries of all other (noncancer) patients in the autopsy population were heavily weighted with atherosclerotic conditions such as ischemic heart disease, diabetes mellitus, and hypertension. 3. 3. It was concluded that there is no association—either positive or negative—between cancer and coronary artery atherosclerotic blockage or coronary artery clot formation. This finding appears to contradict recent epidemiologic and pathologic evidence which tended to link together cancer, especially lung cancer, and atherosclerosis.