Bias in case-control studies. A review.

Abstract

It has been widely accepted that one reason for inconsistent or contradictory results of epidemiologic studies is bias.Therefore, an appreciation of potential sources of bias has become a critical issue in epidemiology.A large number of different sources and possible mechanisms of bias have been described but a consistent terminology has not evolved.Different authors have used different terms when referring to essentially the same type of bias.For example, Kleinbaum et al' and Rothinan2 classify biases into selection bias, information bias and confounding.Sackett offered a more detailed classification, with five major types of sampling bias (eg, admission rate bias, membership bias) and four types of measurement bias (eg, diagnostic suspicion bias, recall bias).3Feinstein distinguishes, inter alia, among susceptibility bias, performance bias, transfer bias and detection bias.4Last's Dictionary of epidemiology5 gives definitions of 26 biases but fails to mention many of the terms proposed by other authors.It is not our objective to provide a comprehensive list of biases, or to propose a new, improved taxonomy.Our purpose is to review the most common threats to validity in the design of case-control studies-commonly regarded as being particularly susceptible to bias-and to discuss some useful strategies in dealing with these problems.Certain issues will not be addressed, notably methods of controlling confounding, and the statistical analysis of case- control studies.Greenland and Morgenstern, case-control studies are distinguished from cohort studies by their use of "outcome-selective sampling".1oMiettinen rejects the case-control versus cohort duality entirely.In his view, the "case-referent strategy" consists in: (1) identifying and classifying all ofthe cases in the "study base"a, defined as the "population experience captured in a study", and (2) drawing a sample of the base to obtain information on the exposure.12 A good illustration of this concept is a "nested", or "within cohort" case-control study, which can be regarded as an efficient form of the cohort design, rather than an entirely different type of study.'3It may be noted that the two aspects (or "axes") of the design, ie, order of measurement and method of sampling, are interdependent.The reason for determining exposure status for all subjects in most cohort studies is to obtain as valid information as possible for the relatively few cases-to-be.Therefore, in cohort studies all potential cases are classified as exposed or unexposed.The logic behind the case-control method is to make the design more efficient by classifying only the actual cases, since information about the source population can be obtained by sampling.This, however, implies a particular order of measurement (directionality), namely ascertaining the outcome first.Measuring exposure for all members of the study population after the cases have been ascertained is unnecessary and impractical, if the disease is rare.