cell phone coverage, as well as propensity of conflict, I include second order polynomials of distance to major town, distance to major road and distance to railway, town mean elevation, slope, aspect of the slope, and variance of elevation and slope. I also include distance to nearest cell phone tower from MPT and from another provider. District fixed effects are included to control for broader geographical trends.9 I cluster standard errors at the district level (one level above township) to account for possible contemporaneous correlation between neighboring areas. The identification relies on the assumption that predicted cell phone coverage is an exogenous determinant of social media use. In other words, after controlling for local population and geographic characteristics, and distances to transmitters, differences in cell phone coverage are due to the terrain between the location and nearby transmitters. Then, cell phone coverage by MPT affects conflict only through increased Facebook access. As long as CoverageFB is exogenous, β is equal to the causal effect of CoverageFB. As I do not observe individuals’ cell phone subscriptions or internet use, but only have a measure of availability (i.e. cell phone reception), the empirical approach is an encouragement design: I estimate the effect of availability of zero rated Facebook, instead of Facebook use per se (Duflo, Glennerster, and Kremer 2007). The intuition of the empirical strategy is the following. Offering zero rated content constitutes a negative price shock on internet use. Zero rated content is only available to consumers that have a SIM card from MPT, and cell phone reception from that provider is a prerequisite for accessing zero rated content. Moreover, better coverage from a mobile network provider in a given area is likely to be associated with higher probability that consumers obtain a mobile plan from that provider. Furthermore, having access to the zero rated content is expected to increase Facebook use (i.e. that the individual is exposed to the treatment). It is likely that both the outcome and the independent variable are measured with some error. First, the conflict data is based on monitoring the news, so there might be some reporting bias. For instance, particular types of events, or events occurring in particular areas, might be more likely to be reported. Second, because the data collection in GDELT is automated, there may be some duplicate reporting. Measurement error can bias the results if it is correlated with the treatment, i.e. cell phone coverage from MPT (conditional on observables). The direction of the bias would depend on the nature of the error. Better cell phone coverage could naturally lead to higher reporting of violence, which could drive up the estimates. However, if access to cell phone coverage lead to higher reporting of conflict incidents, both CovergaFB and Coverage should have a positive effect on probability of conflict. Alternatively, intensified violence could also lead to lower reporting mechanically. For example, if a township is subject to mass 9. The topographic variables are calculated from the SRTM data. Distances are calculated from township centroids using data from Myanmar Information Management Unit. Calculations are done using GIS software. 14