Observational Studies

Observational epidemiology refers to the branch of epidemiology devoted to using nonexperimental [Page 758]studies to describe the health status of populations and generate evidence about determinants of health outcomes. Experimental designs in epidemiology, generally referred to as clinical trials, involve assignment of the principal independent variable, the ‘treatment,’ to subjects. Often this assignment is randomly allocated, which offers profound advantages for making causal inferences, but it can also be nonrandom. In observational studies, the investigators do not assign treatment to subjects. The principal independent variable is some endogenous or exogenous exposure observed as it naturally occurred. When observational epidemiologic studies are designed to draw inferences about health outcome across different exposure groups, they are considered ‘analytic.’ When they are intended only to describe the frequency of a risk factor or disease in a population, they are considered ‘descriptive.’ However, the line between descriptive and analytic observational epidemiologic studies is often blurred as descriptive studies typically contrast disease frequency endpoints across population subgroups, and analytic studies can also report on the absolute frequency of disease or exposure in the population sample under study. This entry describes cohort studies, case-control studies, crosssectional studies, and ecologic studies and discusses the problem of confounding and ways in which it can be addressed.

There are fundamental challenges in making causal inferences about associations between exposures and diseases based on evidence from observational epidemiologic studies. In any analytic epidemiologic study, the goal is to estimate average causal effects in groups (i.e., we do not scrutinize individual participants, subject by subject, to see if in each case there is biological evidence linking exposure and disease). Because there is no way to observe the average disease experience of the exposed group under conditions of no-exposure (the ‘counterfactual’ condition), estimating the average causal effect of exposure is done by comparing the disease experience of the exposed group with that of a different group of individuals without exposure. Consequently, the fundamental challenge to valid estimation of these effects is the similarity of these groups (or their ‘exchangeability’). Here, randomization is a great help because study groups that are randomly assigned to exposed versus unexposed status will, on average, be similar. However, when exposed and unexposed groups are merely observed as they occur in nature, it is extremely unlikely that they will be similar. If the groups differ on factors other than exposure that are associated with disease risk, estimated associations between exposure and disease will be confounded. Minimization of the influence of confounding is, consequently, critical to success in observational epidemiologic studies.