Content area
Full text
In many important social scientific settings, researchers are interested in estimating the effect of behavior on behavior. For instance, What is the effect of an increase in protests or violence by antigovernment groups on government policy, be it concessions or repression (Collins and Margo 2007; Dell 2012; Gould and Klor 2010; Henderson and Brooks 2016; Hendrix and Salehyan 2012; Huet-Vaughn 2013; Madestam et al. 2013; Ritter and Conrad 2016)? What is the effect of a politician issuing press releases on voter behavior (Grimmer, Messing, and Westwood 2012)? What is the effect of one female candidate running for office on another female candidate’s decision to run for office (Baskaran and Hessami 2018; Broockman 2014; Ladam, Harden, and Windett 2018)? Does indiscriminate violence by a counterinsurgent increase or decrease violence by rebels (Benmelech, Berrebi, and Klor 2014; Condra and Shapiro 2012; Dell and Querubín 2017; Jaeger et al. 2012; Lyall 2009)?
Although such questions are central, understanding what exactly we mean when we talk of the effect of one person’s behavior on another person’s behavior is something of a conceptual muddle, requiring careful analysis. This article is an attempt to contribute to that enterprise by articulating a framework and addressing some conceptual issues.
Despite the careful attention paid to identification issues by experimental methodologists, the conceptual difficulties for theoretical interpretability that arise in applications can be hard to spot precisely because it appears straightforward to describe such questions within the potential outcomes framework. We are trying to learn the effect of some agents’ actions; call them the treatment agents. Call the agents whose actions are being affected the outcome agents. The set of treatments corresponds to the set of actions available to the treatment agents. And the potential outcomes correspond to the action each outcome agent would take under each possible action by the treatment agents. The causal effect of the treatment agents’ actions on the actions of outcome agents is the difference in these potential outcomes under the different actions by the treatment agents.
With treatments and potential outcomes so defined, as emphasized by Angrist and Pischke (2009, chap. 1), we can then get clarity on the estimand of an empirical strategy by articulating an “ideal experiment.” The ideal experiment again appears straightforward. One would...