Treatment effect bounds

Jay Bhattacharya

"We implement alternative bounding strategies to reanalyze data from the observational study by Connors et al. (1996) on the impact of Swan-Ganz catheterization on mortality outcomes. We implement both traditional bounds, which exploit access to an instrumental variable but impose no other assumptions (Manski, 1990), and the new bounds of Shaikh and Vytlacil (2004), which impose additional relatively mild nonparametric structural assumptions. Both of these approaches require an instrumental variable that shifts the probability of catheterization, but that does not alter mortality risks. We propose and justify using indicators of weekday admission as an instrumental variable in this context.We find that, while the traditional instrumental variable bounds are almost entirely uninformative in our application, the Shaikh and Vytlacil (2004) bounds often produce a clear answer -...