companies can be used to control for the general effect of cell phone coverage. Information on cell phone towers is obtained from OpenCellID, which is a crowdsourced project to collect cell phone tower locations. To measure the outcome, ethnic violence, I use a georeferenced data on conflict events. I take advantage of two datasets. The first data source is the GDELT Project (2019), which uses an automated system to extract information on conflict events from news media, by using natural language and data mining algorithms. It is the most comprehensive database on conflict events. In order to examine the reliability of this data, I conduct a comparative analysis with the Armed Conflict Location & Event Data (ACLED 2019). It is a widely used manually compiled data source, which makes it considerably narrower, but means there is less misreporting. I conduct both cross-sectional and a difference-in-differences analysis in order to explore different sources of variation. I use the cross-sectional analysis to compare conflict outcomes in areas that are similar in terms of socio-economic and geographical characteristics, but that have different cell phone coverage due to terrain between the location and cell towers. The results show that on average, higher Facebook availability had a decreasing effect on conflict occurrence. Previous literature has shown that disseminating propaganda tends to influence beliefs and behaviour in the intended way (e.g. Yanagizawa-Drott 2014; Adena et al. 2015; Peisakhin and Rozenas 2018), whereas enhancing the ability to communicate may either mitigate or exacerbate conflict (Pierskalla and Hollenbach 2013; Shapiro and Weidmann 2015). Therefore, this result suggest that inflammatory content on social media might play less of a role than the ability to communicate. When I focus on Rakhine State, a region which is central in the military’s crackdown on the Rohingya people (a predominantly Muslim ethnic minority), the results suggest that availability of Facebook led to a small increase in probability of conflict. The result indicates that the role of social media varies in different regional contexts, and that the anti-Muslim hate speech and other inflammatory content may have escalated conflict in deeply fractured and particularly volatile areas. To further study the impact of social media on conflict, I also use time variation in the Facebook campaign in a difference-in-differences set-up. In this setting, the information on population and spatial characteristics stays constant and I only examine whether there is systematic time variation in conflict depending on the cell phone coverage. This analysis indicates that there is no systematic variation in conflict within locations. However, the panel approach exacerbates attenuation bias due to measurement error in cell phone coverage. The remainder of the paper is organized as follows. Section 2 briefly presents the related literature. Section 3 describes the conflict situation in Myanmar and the details of the Facebook campaign. Section 4 provides a description of the data. Section 5 3