0.3 Prevalence of conflict events Average no. of conflict events 15 10 5 0 0.2 0.1 0.0 2014 2016 2018 2014 Time 2016 2018 Time Figure 1: Distribution of events during the sample period The vertical dashed lines show the beginning and end of the Facebook campaign. Data source: GDELT use conventional mass violence. Most of the events fall into the categories coerce and fight. Coercion includes, for example, arrests, detentions, seizing and damaging property, and imposing restrictions on rights of civilians. Fight consists of all uses of military force, fighting and killings, which usually take place between organized groups. Assault includes the less organized forms of violence, such as physical assaults, abductions, assassinations, and use of explosive devices.6 The number of events in the main categories is shown in the Appendix figure B.1. The dataset contains 101,891 events of violent conflict during January 1, 2014–March 31, 2019. Figure 1 shows the distribution of conflict events over the time period. The left panel shows the average monthly number of conflict events across townships. The increased activity in late 2017 marks the timing of the military’s anti-Rohingya “clearance operation” in Rakhine State (Human Rights Council 2018). The right panel of Figure 1 presents the monthly share of townships with conflict events. Every month on average 20% of townships experienced at least one conflict event. The figure shows that the share of townships experiencing conflict has decreased since 2016. Although the number of conflict events spiked at the end of 2017, the events were geographically very concentrated. Appendix Figure B.2 plots the frequency of different types of violent events. 4.2. ACLED Data To asses the robustness of GDELT data I also conduct analysis with the Armed Conflict Location & Event Data Project (ACLED 2019; Raleigh et al. 2010). Because the data is collected by researchers, it contains considerably less events than GDELT (5,416 conflict events in the time period January 1, 2014–March 31, 2019). However, because the data is reviewed and checked, there is also less incorrect reporting. ACLED collects data 6. For a more detailed description of the event types, see Event Data Project (2012). 8