With the exception of experimental studies, epidemiologic research is based on observational studies, and as such, in theory prone to bias by ''confounding.'' Confounders are ''extraneous factors'' that may lead to an apparent (or conceal a true) association between putative risk factors (or protective factors) and disease, due to their own association with both the former and the latter. For example, various lifestyle factors, such as dietary habits, physical activity, smoking and alcohol consumption, which are clearly related to a variety of health outcomes at old age, are often also interrelated. Therefore, when the impact of one of these factors on some health outcome is assessed, it is crucial that the other factors, as well as additional relevant factors, such as age or gender, are carefully measured and controlled for in the analysis. Control for confounding is typically done by means of multivariable analysis, such as multiple logistic regression or the Cox proportional hazards model (Hosmer and Lemeshow, 1999 and 2000). But even if this caveat is taken into account, one can rarely rule out confounding from still other, unmeasured factors. The only design to guarantee absence of this possibility is large-scale randomized trials. Obviously, for ethical reasons, such trials cannot be conducted to experimentally assess the impact of putative risk factors on health outcomes in humans. Therefore, carefully conducted observational studies will always be in the center of etiologic epidemiologic research.
Besides confounding, selection bias and information bias represent other sources of systematic errors and have to be carefully addressed in planning, conducting and interpreting epidemiologic research. In addition, random variability and the role of chance due to a finite number of observations have to be considered. Whereas random variability has an effect on the precision of the study result, systematic errors compromise the validity of a study.
Selection bias occurs when the association between exposure and disease differs for those who participate and those who do not participate in the study and it is mainly due to systematic differences in characteristics between participants and nonparticipants. For example, a hospital-based study on cases with myocardial infarction will exclude those cases who die before admission to hospital, and such a selection may invalidate conclusions and generalizations.
Information bias arises if the information collected about or from the study participants is less than perfect. It can apply to either exposure or disease status or both. Imperfect measurement is often referred to as misclassification if the variable is measured on a categorical scale.
Misclassification of subjects can be either differential or nondifferential. The distinction between these two terms refers to the question of whether misclassification of exposure (disease) relates to disease (exposure) status. A more thorough discussion about sources and consequences of information bias and misclassification can be found in standard textbooks (e.g., Rothman and Greenland, 1998).
Was this article helpful?