Identification and sensitivity analysis of contagion effects in randomized placebo‐controlled trials. (12th November 2019)

• Record Type: Journal Article
• Authors: Imai, Kosuke; Jiang, Zhichao
• Is Part Of: Journal of the Royal Statistical Society. Volume 183:Number 4(2020)
• Page Start: 1637
• Page End: 1657
• Publication Date: 2019-11-12
• Subjects: Contagion effect -- Infectious effect -- Interference between units -- Non‐compliance -- Spillover effect; Social sciences -- Statistical methods -- Periodicals; Statistics -- Periodicals; 300.15195
• DOI: 10.1111/rssa.12528
• Languages: English
• ISSNs: 0964-1998

Abstract

Summary: In social science research, interference between units is the rule rather than the exception. Contagion represents one key causal mechanism of such spillover effects, where one's treatment affects the outcome of another individual indirectly by changing the treated unit's own outcome. Alternatively, the treatment of one individual can affect the outcome of another person through other mechanisms. We consider the identification and sensitivity analysis of contagion effects. We analyse a randomized placebo‐controlled trial of the get out the vote campaign, in which canvassers were sent to randomly selected households with two registered voters but encouraged only one voter within each household to turn out in an upcoming election. To address the problem of non‐compliance, the experiment includes a placebo arm, in which canvassers encourage voters to recycle. We show how to identify and estimate the average contagion and direct effects by decomposing the average spillover effect. Our analysis examines whether canvassing increases the turnout of a non‐contacted voter by altering the vote intention of a contacted voter or through other mechanisms. To address the potential violation of key identification assumptions, we propose non‐parametric and parametric sensitivity analyses. We find robust contagion effects among some households.