deportations or burning down villages, subsequent conflict events may become unlikely. Second, because I use crowdsourced data on cell phone towers, it is likely that not all towers are included in the data. There may also be some error in the exact locations of the towers, and many of the technical parameters need to approximated in the coverage prediction. Measurement error in the independent variable will bias the estimates towards zero. Furthermore, if some areas lost cell phone coverage as a result of increased conflict, it would cause the treatment effect to be underestimated. 5.2. Difference-in-differences In order to take advantage of the time variation in Facebook availability, I also conduct a difference-in-differences analysis. Because the information on population and spatial characteristics is constant over time, the analysis uses only within township variation to identify the effect of Facebook availability on conflict. I estimate the following model: Yit = β1 CoverageFBi · T reatt + β2 CoverageFBi · P ostt + τt + λi + T reatt + P ostt + εit (2) where Yit is indicator for conflict in township i in time t. The unit of observation is township-month. CoverageFB represents the treatment intensity. The time variables T reatt and P ostt indicate the treatment period (June, 2016–August, 2017), and posttreatment period (September, 2017–March, 2019). The time effect τt captures time specific effects that are common to all townships, and the township fixed effect λi captures all township specific time invariant characteristics. The coefficient β1 represent the effect of Facebook availability during the Free Basics campaign (the treatment period), and β2 is the post-treatment effect. I use three time periods to allow for possible time variation or persistence in the treatment effect. As the treatment period is relatively long and conflict events are observed both before, during and after the treatment, I am able to study whether the impact of Facebook availability is different during and after the treatment period. Social media use may take some time to influence users’ beliefs and behavior, and these effects may depend on the share of population using social media. The availability treatment may have taken time to affect Facebook use. It is likely that Facebook gained popularity during the Free Basics campaign, which lead to growth in its user base also after the campaign. The difference-in-differences approach allows estimating the causal effect of treatment even if the treatment itself is not randomly assigned, but instead determined by the unobservable characteristics captured by λi . Unfortunately, the fixed effects approach exacerbates measurement error in the regressor, which increases attenuation bias. 15