next up previous index
Next: iposter.03 Up: Invited Posters (Number to Previous: iposter.01


Sponsoring Section/Society: Health Policy Statistics Section

Session Slot: Sunday, 4:30 - 7:00

AudioVisual Request: None

Invited Poster Title: The Clustered Encouragement Design

Frangakis, Constantine,   Harvard University

Address: Harvard University, Department of Statistics, Science Center, 1 Oxford St., Cambridge MA 02138

Phone: (617) 495 4888

Fax: (617) 496 8057


Rubin, Donald B., Harvard University

Zhou, Xiao-Hua (Andrew), Indiana University

Abstract: Comparative study designs where the encouragement, rather than direct receipt of treatment, is randomly assigned have attracted attention in settings where traditional randomization of receipt is not ethical or not feasible (e.g., Hirano, Imbens, Rubin, and Zhou, 1997). Combining this encouragement design with the practical constraint of clustered randomization (e.g., at the physician level), generates a relatively simple design with potential to evaluate the effectiveness of health-care-related treatment options. Unfortunately, the design does not prevent indirect effects, of no policy interest, from being present, which renders standard intention-to-treat or exclusion-restriction-based instrumental variables analyses (e.g., Angrist, Imbens, and Rubin, 1996) generally inappropriate or inapplicable. In this work, we develop an analytic method to address this situation by bringing together and extending ideas from recent work on encouragement designs (Hirano et al., 1997), while allowing for interference between units (Rubin, 1978) to accommodate the clustered randomization. We show that the resulting procedure has key theoretical advantages over existing ones and that, despite its multilevel structure, still it can be relatively easily implemented using current simulation techniques. We apply our methods to a study on the effectiveness of physician-patient discussions of Advanced Directive forms, and show how the theoretical results are reproducible in real studies, thereby supporting the clustered encouragement design as a practical tool.

Theme Session: Yes

Applied Session: Yes

next up previous index
Next: iposter.03 Up: Invited Posters (Number to Previous: iposter.01
David Scott