Commentary: Berkson's Bias reviewed

Abstract

Listen

Translate article

Epidemiologists now routinely collect biological samples from study participants. Because blood and other samples are relatively easy to collect from hospital patients, hospital-based case–control studies are becoming increasingly popular after years of disfavor. Adding to their popularity is the fact that response rates from hospital patients are generally higher than they are from the community [1]. In the current issue of the European Journal of Epidemiology, Sadetzki et al. present an example of the possible occurrence of Berkson’s Bias in a hospital-based case–control study of cigarette smoking and bladder cancer [2]. To design and interpret hospital-based case–control studies it is important to understand Berkson’s description of the selection bias that may characterize these studies [3–8]. A hospital admission probability is the probability with which members of a group in the community will be admitted to the hospital. For example, people with traumatic brain injury are more likely to be admitted to the hospital than are people with allergic rhinitis. Berkson’s selection bias results from the greater probability of hospital admission for people with two or more conditions than for people with one condition. For example, if breast cancer is the ‘exposure’ of interest (diseases are often treated as exposures in hospital-based case–control studies) and meningioma is the case disease then people with both breast cancer and meningioma could be hospitalized for either breast cancer or meningioma or both. However, people with breast cancer only or meningioma only could be hospitalized because of one of these diseases. Therefore, a greater proportion of people in the community with both breast cancer and meningioma would be admitted to the hospital than would people with meningioma only. To estimate the degree of Berkson’s Bias the addition law of probability is applied to hospital admission probabilities of exposure (pr(He)), case (pr(Hd)), and control conditions (pr(Hc)). For example, if the hospital admission probabilities for breast cancer (the exposure), meningioma, and the control disease are 0.95, 0.50, and 0.99 respectively and there is no association among these three conditions in the community (OR 1⁄4 1.00) then people in the community with both breast cancer and meningioma (but not the control conditions) have a 97.5% probability of being hospitalized. (This percentage is based on: pr(He) + pr(Hd) ) pr(He)pr(Hd) 1⁄4 (0.95) + (0.50) ) (0.95)(0.50) 1⁄4 0.975). Probabilities for each combination of attributes in a 2 · 2 table are calculated and applied to the number of people in the community in each attribute group (e.g., breast cancer and control diagnosis, see Kleinbaum et al. [7] for details). As Boyd shows [5], the effect of Berkson’s Bias is to multiply the community odds ratio by the following odds ratio: