The role of designs in partially controlled studies
Department of Biostatistics
Bloomberg School of Public Health
Johns Hopkins University
Friday, September 15, 2006, 12:00pm
The ability to evaluate causal effects of factors on outcomes is
increasingly important for studies that control some but not all of the factors. Work on designs for such studies has been relatively limited, due to the lack of an appropriate framework for separating assumptions on causal effects from assumptions on the design. We consider the following two types of such studies, and show how to use the framework of
"principal stratification" to design them and evaluate their goals.
(1) Studies that control the location of sites that offer treatments. The complication is that studies do not directly control who gets treatment or who provides outcomes (application to needle exchange programs). Here, we develop designs that maximize the benefit of the treatment to
(2) Studies that interview individuals after a critical event (e.g.,
injury) to measure an exposure (e.g., drug use) preceding the event, in order to relate exposure to severity of injury (e.g., mortality). The complication is that the past exposure is missing precisely for the individuals who cannot be interviewed because they die from the critical event. Here, we develop design methods that allow nonignorable missingness of exposure due to death, and that can still estimate objectively the relation of exposure and mortality.
The work is join with David Vlahov, Don Rubin and Ming Wen-An.