Table 1: Cross-sectional results for probability of conflict Conflict dummy, GDELT (1) (2) (3) CoverageFB Coverage Observations District dummies Spatial controls Population controls Conflict dummy, ACLED (4) (5) (6) −0.142∗ (0.080) 0.104 (0.084) −0.258∗∗ (0.101) 0.201∗∗ (0.098) −0.134∗ (0.079) 0.185∗∗ (0.075) 0.029 (0.040) −0.024 (0.042) −0.054 (0.048) 0.013 (0.058) −0.028 (0.053) 0.024 (0.058) 330 X 330 X X 330 X X X 330 X 330 X X 330 X X X ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01. Robust standard errors clustered at district level are reported in parentheses. The dependent variable is an indicator for conflict in a township between June 2016 until end of August 2017. Population controls: log population, log population density, dummy for below median urban rate, age (15–64 y.o.), population with no ID, population with electricity, mobile phone, landline phone, and internet at home. Spatial controls: 2nd order polynomials of distance to major town, distance to major road, distance to railway, distance to MPT transmitter, distance to other transmitter, mean elevation, slope and aspect of the slope, variance of elevation and slope. The predictors are standardized. 6. Results 6.1. Cross-sectional estimates Table 1 presents the OLS estimates of model (1). The dependent variable is an indicator for experiencing conflict in the treatment period, i.e. when Free Basics was available (from June, 2016 until end of August, 2017). The dependent variable in columns (1)–(3) is an indicator for an event of violent conflict from the GDELT, and in columns (4)–(6) an indicator for a violent event from the ACLED data. The estimates suggest that cell phone coverage provided by MPT, i.e. free access to Facebook, is associated with lower probability of conflict. The coefficient of CoverageFB in column (1) indicates that, after filtering out the district fixed effects and controlling for cell phone coverage from other providers, one standard deviation increase in MPT cell phone coverage is on average associated with 14.2 percentage point decrease in the probability of conflict. The positive coefficient on Coverage suggests that better cell phone coverage per se is correlated with higher probability of conflict. A placebo test (Table C.2 in the Appendix) shows that CoverageFB had no effect in the pre-treatment period, but the effects only appears in the treatment and posttreatment periods. The two sources of conflict data exhibit different patters—in GDELT, the estimate is marginally significant in the treatment period, and relative large although imprecise in the post.treatment period, whereas based on ACLED the effect appears only after the Facebook campaign has ended. 16

Select target paragraph3