Public Attention Reduced Forest Fires in the Brazilian Amazon

10 International frameworks and agreements to reduce anthropogenic environmental disasters rely on international pressure driving local action. Although environmental catastrophes can occasionally capture international attention, it is unclear if focused media and increased public outcry can reduce environmental damage. We study the unusual and concentrated increase in international scrutiny on forest ﬁres in the Brazilian Amazon in August 2019. Comparing active ﬁres in the Brazilian Amazon versus those in the Peruvian and Bolivian Amazon before and after a surge in public attention on the Brazilian Amazon, we ﬁnd that increased public attention reduced ﬁres by 22% (93,607 avoided pixel-days of active ﬁre) avoiding 24.81 million MtCO 2 in emissions. Our results highlight the power of international pressure to compel governments to act on pressing environmental issues, even in political contexts hostile to environmental priorities.

Bolivia before and after the rise in international attention to the fires in Brazil 11 . Figure 2C shows the differential number of 39 fire outbreaks per km 2 in the Brazilian Amazon every two weeks relative to the Peruvian and Bolivian Amazon. We observe 40 no differential fires in Brazil prior to the international spotlight, supporting the common trends assumption that underlies a 41 causal interpretation of difference-in-differences estimates. We observe that fires decline by 0.004 days of fire per km 2 after 42 the spike in international attention. The effect is most prominent in September but persists through early November even as 43 the fire season comes to an end in Bolivia and Peru. This corresponds to 36% of the average fires during the same period in 44 the Peruvian and Bolivian Amazon. Altogether, our estimates indicate that international scrutiny resulted in 93,607 fewer 45 pixels-days of active fires than would have occurred in the absence of the same attention. Our estimates are robust to controlling 46 for precipitation and average fires from 2016-2018 at the pixel level (SI Appendix, Tab. S1, col. 1-4). Our estimates are robust 47 to a triple-differences strategy where we compare Brazil with its neighbors before and after August in 2019 versus 2016-2018 48 (SI Appendix, Tab. S1, col. 5). We find that reductions in fires were at least two times larger in areas with denser forest cover 49 (SI Appendix, Tab. S2, col. 1-4), suggesting that wildfires were the focus of the actions. 50 We find no evidence to suggest that these fires were simply displaced elsewhere. First, the map of fire intensity shows that 51 the main locations of active fires in August 2019 were not close to the border (SI Appendix, Fig. S2). Second, we compare 52 occurrences of fires around the Brazilian border with Bolivia and Peru using a regression discontinuity design 12 . SI Appendix 53 Figure S3 shows the average number of fire outbreaks within 200km from the border in August and September. We see that the 54 reduction of fires between August and September within 200km on the Brazilian side of the border was not followed by an 55 equivalent increase in fires in the neighboring countries close to the border.

57
Forest fires in the Amazon are an annual phenomenon, but a marked increase in international attention resulted in nearly 58 22% fewer pixel-days of active fire in 2019. The reduction on fires can be broadly attributed to greater engagement of the 59 civil society and to government actions taken after the outcry. The two main actions taken by the federal government were: 60 (i) fire control actions by recruiting and dispatching fire brigades to specific areas -some under the military "Green Brazil 61 Operation"; and (ii) a 60-day ban on the use of agriculture fires inside the Legal Amazon. As the timeline of events shows (see 62 SI Appendix), both actions were initiated after the marked increase in public attention. We estimate heterogeneous effects of the 63 international pressure on fires in municipalities that received external fire brigades or that received funds to recruit firefighters 64 (under PREVFOGO/IBAMA program). Our estimates imply that fire brigades partially contributed to reducing fires following the international outcry (SI Appendix, Tab. S2, col. 5), suggesting that actions mediated by non-government organizations 66 and local governments also contributed to the fire control. It is worth noting that the "Green Brazil Operation" launched in 67 2019 was a military operation that included additional fire-brigades over and above those sent by non-military agencies, such 68 as IBAMA. Unfortunately, data from the military operations are not available and therefore excluded from our analysis. As 69 such, our result is an existence result in that we show that increased public attention reduced fires in the Brazilian Amazon 70 through a combination of various government and civil society actions. However, disentangling the relative magnitude of each 71 government action in reducing fires is beyond the scope of this paper.

72
Our study has three caveats. First, we cannot determine the single element that led the global community to pay such close attention to fires in 2019 as opposed to previous years. Although the fire season arrived earlier in 2019 than it did in 74 2016-2018 13 , the peak in 2019 was similar to that in 2018 and lower than that in 2017 (Fig. 1A) world's rainforest.

97
The Amazon remains one of the major strands of tropical forests in the world and preserving its integrity is crucial to  Difference-in-differences is a statistical design where two groups, denoted treatment and control groups, are observed in 117 time periods before an event affects the treatment group but not the control group. In broad terms, the estimated treatment 118 effect takes the difference between time periods for each treatment and control groups (before and after the treatment period 119 for each group) and then takes the differences between these two differences. When treatment is not randomly assigned, this 120 method recovers causal estimates under the assumption that in the absence of treatment, the two groups would have been on 121 similar trajectories after the treatment, as they did before the treatment. Since this assumption is axiomatically untestable, a 122 commonly used alternative is to demonstrate parallel trends between the two groups prior to the treatment group receiving 123 treatment. In our case, we follow best practices in difference-in-differences designs and show that fires in Brazil have followed 124 similar trends as those in Peru and Bolivia before the increase in public attention (see Fig. 1B-C).

4/9
More precisely, we estimate the differential Amazon fires in Brazil relative to Bolivia and Peru over 2019 as captured by γ t from equation: where f ire i,r,w is the number of fires in pixel i in country r in week w, BrAm i is a dummy for the Brazilian Amazon region, α i are 126 pixels and δ w are week fixed effects (weeks staring in the date indicated) -pixel fixed effects control for time-invariant factors 127 that are specific to each pixel such as geography, regulatory context etc. X i,w is a vector of pixel-week controls (contemporary 128 precipitation and average fire outbreaks between 2016-2018). We allow η r to differ across countries r. ε i,r,w is the idiosyncratic 129 error, which we cluster at 25km × 25km grids to account for serial and spatial correlation. 130 Under the assumption that forest fires in the Brazilian Amazon, absent media and public attention on fires here, would have 131 followed a similar trend as the fires in Bolivia and Peru, γ t estimates the average treatment effect of international pressure 132 on fires over the remaining of the 2019 fire season. Figure 2B shows that the number of fire outbreaks followed a common 133 trend from 2016 up to August 2019. Because we have over 120 million pixels, for computational reasons, we estimate the 134 equation (1) using a random sample of 50% of pixels. We provide robustness checks and estimate heterogeneous effects in the 135 SI Appendix.  The R-squared of the regression is 0.746, larger than some regressions presented in GFED's regional estimates. Next, we 140 predict emissions between September and mid-November using these estimates and the fire count in this period. We predict 141 that the total emissions in the Brazilian Amazon in this period was 112 million tons of CO 2 . Thus, we calculate the reduction of 142 22% of fire days caused by international pressure (SI Appendix, Tab. S2 col. 2) helped preventing the release of 24.81 million 143 tons of CO 2 to the atmosphere. 144 We benchmark this number with the difference between Brazil's current emissions 19            against fires in Peru and Bolivia before and after the spike in international attention.

217
To generate Table S1, we estimate the average effect of international pressure on fires in the Brazilian Amazon between September and mid-November as captured by γ from equation: following a similar notation as in equation (1). 218 We also run a triple-difference estimate comparing fires in each pixel and fire-week in 2019 with its historical average between 2016-2018 for the same fire-week using the following equation: where AvgFire i,r,w,2016−2018 is the average fire outbreak in pixel i in fire week w between 2016 and 2018. γ in this equation 219 captures the triple-difference estimate. 220 D. Were forest fires displaced to other regions or across national borders? One immediate concern could be that at least 224 Second, Figure S3 shows the average number of fire outbreaks within 27km from the border in August (a) and September 225 (b); 27km is the optimal bandwidth for August 21 . Each point shows the average number of fire outbreaks by 1km bins of 226 distance to the border; positive distance represent pixels in Brazil and negative distances represent pixels in Bolivia or Peru. 227 We see that the reduction of fires between August and September on the Brazilian side of the border was not followed by 228 an equivalent increase in fires in the neighboring countries close to the border, on the contrary. Performing a regression 229 discontinuity estimation 21 , we find no discontinuous fires outbreaks at the border. Point estimates (bias corrected p-values) for 230 August is 0.003 (0.550), and for September-November is 0.003 (0.485); point estimates are smaller and less precise when we 231 control for distance to water, access to cities, and above ground biomass. 232 E. Heterogeneous effects on forest areas. Whether the international pressure was more effective to prevent fires in areas with forest cover than in other areas, for example pastureland, is important to understand its impact on the amount of carbon released by the fires. We assess this issue by estimating heterogeneous effects of the international pressure on fires in areas with greater forest cover, as captured by γ 2 from equation: where Forest i is an index equal to one for pixels with greater forest cover.  pixels with more than 50% of forest cover, pixels with more than 75% of forest cover, and pixels with more than 90% of forest 235 cover in 2015. 4 We see that international pressure was at least twice more effective to reduce forest fires than non-forest fires.

F. How did government actions curtail fires?
Although we cannot test for the extent to which each government action in 237 the aftermath of international coverage of fires affected subsequent fires, we assess whether the fire brigades dispatched or 238 recruited to control the forest can fully explain the effects we estimate. We use government records to identify the municipalities 239 that received external fire brigades (data from the Registry of Fire Incidents from the Environmental Regulatory Agency, 240 ROI/Ibama) or that received funds to recruit fire fighters (from the Ministerial Ordinance # 3020/2019). We regress equation 4 241 using a a dummy for the municipalities that received fire brigades instead of the Forest i variable. The coefficients in Table S2 242 column 5 show a differential reduction of fires in areas that received fire brigades, however we also see statistically significant 243 and meaningful effects on areas that did not receive the direct assistance captured on the records. This suggests that other 244 actions, such as the 60-day ban on the use of fire in the field, also contributed to the reduction. A caveat of this analysis is that 245 we could not obtain data on all missions under the Operation Green Brazil. Fig. S5 shows evidence that air pollution in major cities, for example, São Portuguese) and searches for the combination of the words "São Paulo" and "fire", "smoke", or "Amazon". We can see that the 255 focus of public attention was the Amazon fires, not São Paulo. Figure S4 plots the accumulated number of fire outbreaks for the Brazilian Amazon for   # Clusters 9,320 9,320 9,320 9,320 9,320 a This table presents the results of the difference-in-differences approach. The table shows the coefficient of the interaction term of a pixel belonging to the Brazilian Amazon with a dummy indicating the period after the week of the rise in international attention (coefficient γ from expression (2)). All specifications include pixel fixed-effect and week fixed-effect. Units of observation are 1km 2 pixels in a bi-week period. From columns (1) to (4) we vary the controls included. The results are robust to including precipitation at the pixel-bi-week unit and the average fire count of each pixel in the equivalent bi-week of the years 2016-2018. Column (4) presents our preferred estimates. In column (5) we consider only pixels near the border (distance from the border with Peru and Bolivia < 27km). Column (5) shows the estimates of a triple-difference estimate as represented in equation 3.illustrates that the result we have found in columns (1)-(4) is not a displacement effect from Brazil to Peru and Bolivia. Number of observations (and clusters) from the main specifications: 72,893,026 (10,532). Standard errors clustered at 100km 2 grids in parentheses. Significance levels: *10%, **5%, ***1%.. Significance levels: *10%, **5%, ***1%. # Clusters 9,320 9,320 9,320 9,320 9,320 a This table presents the results of the difference-in-differences approach with heterogeneous effects. Columns (1) to (4) show the coefficient of the interaction term of a pixel belonging to the Brazilian Amazon with the a dummy indicating periods after the bi-week of the rise in international attention interacted with a pixel's forest cover in 2015 (coefficient γ 2 from equation (4)). All specifications include pixel fixed-effect, week fixed-effect, and controls for precipitation and average fires from 2016-2018 at the pixel-bi-week level. Units of observation are 1km 2 pixels in a bi-week period. In column (1) the forest cover variable is the share of the forest cover in that pixel. From columns (2) to (4) we create dummy variables that equals to one when the forest cover of a pixel is above a threshold (50%. 75%, and 90% respectively). In these columns we see that the effect of the reduction on fires was stronger in areas with greater forest cover. In column (5) we consider the interaction term of a pixel belonging to the Brazilian Amazon with a dummy indicating the period after the bi-week of the fire ban with a dummy variable that indicates if a fire brigade or a special budget to combat fire was sent to the municipality that the pixel belongs, after the fire ban. We observe a stronger effect of fire reductions on municipalities that receive such help. Nonetheless, it does not explain all the reduction on fires. Standard errors clustered at 100km 2 grids in parentheses. Significance levels: *10%, **5%, ***1%.    Figure S6. Searches on fires in Brazil compared with searches related with São Paulo A shows Google searches in English about Amazon fire compared with searches about Amazon fire and the dark sky day in São Paulo. For this we used the words "smoke", "fire", "sao paulo", and "amazon". B shows the same Google searches results in Portuguese. For this we used the words "fumaça", "fogo", "são paulo", and "amazonia". 9/9