The Perils of Estimating Disengagement Effects of Deadly Terrorist Attacks Utilizing Social Media Data

Abstract

This comment discusses the impact of social media rule enforcement protocols on research on online data sources. It argues that the conclusions of the article ‘Do Islamic State's Deadly Attacks Disengage, Deter, or Mobilize Supporters?’ concerning the recruitment effects of deadly attacks cannot be assumed to hold when considering the timing of Twitter account suspensions. It highlights four ways in which suspensions can confound evidence of demobilization despite the introduction of control variables and fixed-effects model specifications. All change the composition of the sample in four non-random ways. First, suspending connected Islamic State accounts may result in follower loss. Secondly, Twitter suspension procedures may be tied to account characteristics, such as follower accrual rates. Thirdly, suspended accounts that re-emerge introduce replication bias. Fourthly, account closure may reflect user movement to other platforms in response to changing security environments following deadly attacks. In conclusion, caution is advised when platform-introduced variation risks altering the sample composition in non-random ways.

Islamic State's Terrorist Attacks Disengage Their Supporters: Robust Evidence from Twitter

“…Abstract This article responds to Hansen's (2022) comment on the use of social media data to evaluate the effects of terrorist attacks on related online behavior. Hansen casts doubts on our previous finding that terrorist attacks disengage supporters of terrorist groups.…”

“…In her comment, Hansen (2022) argues that Twitter suspension activities could still be a confounder in our models through four channels: (1) follower loss from suspended accounts; (2) heterogeneous Twitter suspension efforts based on account characteristics; (3) replication bias due to the resurgence of suspended accounts; and (4) replatforming to other social media. However, not all four channels are relevant to our empirical design.…”

“…Contradictorily, Hansen rules out the second and third pathways deeper in her comment. More specifically, she acknowledges that replication bias “is only problematic in random effects models,” not in the fixed-effects models, “as fixed effects models measure effects at a within-unit level” (Hansen 2022, fn. 5, page 7).…”