ABSTRACT
The effects of exposures to agents, either ingested or environmental, on humans are estimated most reliably from randomized controlled trials. These effects have a direct causal interpretation since randomization accounts for both observed and unobserved confounding (Hernán and Robins, 2006). There are, however, numerous situations when observational studies are necessary to study exposure associations. Sometimes this occurs in the context of rare events, which would require extremely large trials to detect an association. Often, information on an adverse event may not be available in previous clinical trials and conducting new trials is both lengthy and expensive. In other cases, the exposure is not ethically amenable to experimentation and thus cannot be the object of randomization. Finally, the population studied in clinical trials is generally highly selected and researchers may wish to determine if the results in the trial population can be generalized to other portions of the population (Mills et al., 2006).
