PES Learning Paper 2019-1 Impacts of conservation incentives in protected areas: The case of Bolsa Floresta, Brazil Elías Cisneros, Jan Börner, Stefano Pagiola, and Sven Wunder November 2019 Environment, Natural Resources & Blue Economy World Bank Washington DC, USA Abstract Incentive-based conservation is a promising approach to tropical forest conservation, including within multiple-use protected areas. Here we analyze the environmental impacts of Bolsa Floresta, a longstanding forest conservation program combining conditional household payments with livelihood- focused investments in 15 multiple-use reserves of Amazonas State, Brazil. We use grid-based data, nearest-neighbor matching, and panel data econometrics to compare forest-related program outcomes (deforestation, degradation, fires) with non-participating reserves. While post-treatment deforestation and degradation was negligible, this was already the case pre-treatment, since low-threat reserves had preferentially been targeted. We thus find only statistically insignificant additional conservation effects from implementation. No important heterogeneous treatment effects could be detected either. Our findings thus add to the growing evidence that spatial mis-targeting towards low-hanging fruits, that is disproportionally selecting low-threat forest conservation areas into programs, constitutes a prime cause for low additionality found in rigorous impact evaluations of incentive-based forest conservation initiatives. Authors Elías Cisneros is Research Associate at the Georg-August-Universität Göttingen; Jan Börner is Professor of Economics of Sustainable Land Use and Bioeconomy at the Rheinische Friedrich-Wilhelms-Universität Bonn; Stefano Pagiola is Senior Environmental Economist in the Environment, Natural Resources & Blue Economy Global Practice, World Bank; and Sven Wunder is Principal Scientist of the European Forest Institute and Senior Associate of the Center for International Forestry Research (CIFOR). Keywords Payments for Environmental Services (PES), incentives, protected areas, deforestation, impact evaluation, quasi-experimental methods, spatial matching, Amazon, Bolsa Floresta, Brazil Acknowledgements This paper is based in part on the first author's doctoral dissertation at the Rheinische Friedrich- Wilhelms-Universität Bonn, for which he received financial support from the Robert-Bosch foundation and the Interamerican Development Bank. We would like to thank Virgilio Viana, Gabriel Ribenboim, and Suelen Marostica (FAS) for valuable program-related information, comments, and support, and Peter May for various conceptual contributions. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors, and do not necessarily represent the views of their institutions. Cover photo Areas cleared for cultivation within rainforest areas of Amazonas (Neil Palmer, CIAT). PES Learning Papers PES Learning Papers draw on the World Bank’s extensive experience in supporting programs of Payments for Environmental Services (PES). They are part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. The PES Learning Paper series disseminates the findings of work in progress to encourage the exchange of ideas about PES. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. The boundaries, colors, denominations, and other information shown on any map in this work do not imply any judgment on the part of The World Bank concerning the legal status of any territory or the endorsement or acceptance of such boundaries. Impacts of conservation incentives in protected areas: The case of Bolsa Floresta, Brazil Elías Cisneros, Jan Börner, Stefano Pagiola, and Sven Wunder 1. Introduction Protected areas (PAs) play a key role in preserving biodiversity-rich landscapes, storing forest carbon, and generating other environmental services. Worldwide, PA cover 15 percent of land areas (UNEP-WCMC and IUCN, 2016). Evidence suggests that PAs have significantly reduced—but seldom totally halted—deforestation within their boundaries (Joppa and Pfaff, 2010; Nelson and Chomitz, 2011; Cuenca and others, 2016). In multiple-use PAs that explicitly permit human presence and residents’ environmentally benign activities, incentives have been used to improve both conservation and livelihoods of local populations. Some evidence suggests that multiple-use PAs have globally performed comparatively well (Nelson and Chomitz, 2011; Porter-Bolland and others, 2012), although selection biases can rarely be fully controlled for in assessing the impacts. In this paper, we examine a longstanding effort to use conservation incentives within multiple-use PAs, the Bolsa Floresta program in the State of Amazonas, Brazil. Bolsa Floresta offers payments for environmental services (PES) to households residing in several of the state’s sustainable-use PAs to induce them further to conserve forests and use them sustainably. To evaluate its impacts, we use spatial matching techniques to identify counterfactual sites and estimate treatment effects using panel data. We proceed as follows. Section 2 provides geographical and topical context for the Bolsa Floresta program, which is described in Section 3. Section 4 describes the impact evaluation methods, and Section 5 presents the results. In Section 6, we discuss and conclude. 2. Conservation policy mixes: Brazilian Amazon and beyond Annual deforestation in the Brazilian Amazon as a whole fell by some 80 percent (from 27,000 km2 to 4,500 km2) between 2004 and 2012, and by 60 percent in Amazonas State (1,200 km2 to 520 km2) (INPE, 2019). This fall has been attributed to both economic and political factors (Canova and Hickey, 2012; Hargrave and Kis-Katos, 2013; Cisneros and others, 2015; Pailler, 2018). The Plan to Combat Deforestation in the Amazon, launched in 2004, played an important role by setting up an effective satellite forest-cover monitoring system, increasing budgets of the prime environmental enforcement agency (Instituto Brasileiro do Meio Ambiente e dos Recursos Naturais Renováveis, IBAMA), expanding the PA system, and promoting the standardization of land registration cadasters (Maia and others, 2011; Assunção and others, 2012; Arima and others, 2014; BenYishay and others, 2017), as demonstrated by quasi-experimental evidence (Canova and Hickey, 2012; Hargrave and Kis-Katos, 2013). However, 3 deforestation partially rebounded after 2012 (INPE, 2019), mainly due to legal-political changes (Sengupta, 2018; Pereira and others, 2019). Brazil’s PA system covers 2.2 million km2, or 24 percent of the Legal Amazon1, in 2012—up from 13 percent in 2004. It has made an important, well-documented contribution to reducing deforestation in the Amazon (Soares-Filho and others, 2010; Pfaff and others, 2015a). PAs are particularly effective in areas with high deforestation pressure, close to roads and cities (Nolte and others, 2013; Pfaff and others, 2015a), especially right after their establishment (Pfaff and others, 2015b). Unsurprisingly, leakage effects are larger in high-pressure areas (Amin and others, 2019). In the Brazilian Amazon, outcomes appear to be less dependent on the quality of management (Nolte and Agrawal, 2013) than to PA types: strict PAs have been more effective than multiple-use ones (Nolte and others, 2013). However, multiple-use reserves in Acre state avoided more deforestation than strict PAs (Pfaff and others, 2014).2 Where PAs effectively curb deforestation, however, political pressure may threaten this outcome. Command-and-control policies, be it law enforcement or PA establishment, typically impose uncompensated economic costs on local development- oriented actors, who thus tend to oppose them. Adding some carrots (that is, incentives) to the pre-existing sticks (that is, disincentives) could curb local welfare losses, making conservation politically more acceptable (Börner and others, 2010; Nepstad and others, 2014; Santiago and others, 2018). Yet, policy mixes of carrots and sticks need to adequately balance environmental impacts with welfare and equity objectives (Börner and others, 2015). Among these incentives, payments for environmental services (PES) are conditional on land stewards adopting environmentally-friendly land uses, while compensating for income foregone from deforestation and degradation activities (Engel and others, 2008; Wunder, 2015). PES have been widely adopted, especially across Latin America (Alix-Garcia and Wolff, 2014; Börner and others, 2017). Brazil was a latecomer in PES development, but numerous applications have been recently developed, especially watershed schemes in the Atlantic Forest biome and carbon initiatives in the Amazon (Pagiola and others, 2013). One key challenge of using PES for conservation is adverse self-selection bias: those landholders who were likely to protect forests even in the absence of payments will be the most eager to participate, yet their enrollment would not result in additional conservation (Persson and Alpízar, 2013). A systematic review of rigorous forest-based PES evaluation studies worldwide by Samii and others (2014) found relatively low PES 1 The Legal Amazon includes the states of Acre, Amapá, Amazonas, Pará, Rondônia, Roraima, and Tocantins, and part of Mato Grosso and Maranhão; covering 5 million km2. 2 Brazil’s National System of Protected Areas includes strictly protected areas (Áreas de Proteção Integral) and multiple-use reserves (Áreas de Uso Sustentável). Multiple-use reserves allow resident populations to pursue livelihood strategies with benign environmental impacts. Subcategories include Sustainable Development Reserves (Reserva de Desenvolvimento Sustentável, RDS), Environmental Protection Areas (Área de Proteção Ambiental, APA), Extractive Reserves (Reserva Extrativista, RESEX), and State Reserves (Reserva Estadual, RE). 4 additionality, but their demanding methodological filters selected very few studies, all from Costa Rica and Mexico. Börner and others (2016) reviewed a larger, less geographically biased sample, and found slightly more evidence of additionality. The first-ever randomized-control PES trial in Uganda found that PES reduced deforestation by more than half (Jayachandran and others, 2017), which then rebounded once PES ended (World Bank, 2018). In some cases, PES have already been used inside multiple-use PAs. For example, PA residents are eligible to participate in the national PES programs in Costa Rica and Mexico, to compensate them for the costs imposed on them by PA restrictions (Pagiola, 2008). Other PES interventions were designed specifically for PAs, for example, in the Monarch Butterfly Reserve in Mexico (Honey-Rosés and others, 2011). Going beyond land-cover protection, in Cambodia a PES program paid local communities to protect the nests of threatened bird species (Clements and others, 2014). PES inside multiple- use PA may also motivate local people to report violations or encroachments by others, thus assisting command-and-control monitoring (Robinson and others, 2010). PES-cum-PA pilot impact evaluations have so far shown fairly encouraging results. In Mexico, PES schemes and PA show a complementary conservation effect in PA buffer regions (Sims and Alix-Garcia, 2017). Honey-Rosés and others (2011) found high impacts in the Monarch Butterfly Reserve, but could not separate the effects of the two components. Clements and Milner-Gulland (2014) found that establishing PAs in Cambodia reduced deforestation by about half; and that adding economic incentives (including PES) reduced it by another half, while increasing the wellbeing of participants. Montoya Zumaeta and Rojas (2017) found that a watershed PES in Moyobamba (Peruvian Amazon), implemented within part of a simultaneously created multiple-use PA, jointly had moderate conservation impacts, fairly equally attributable to each treatment; the incentives also clearly increased the household income of PES recipients. 3. The Bolsa Floresta program The Bolsa Floresta program was established in 2007, with the double aim of protecting multiple-use PAs in Amazonas State against deforestation pressures and increasing local residents' welfare (Viana and others, 2009, 2012; Börner and others, 2013). Bolsa Floresta is operated by a non-governmental organization, the Sustainable Amazonas Foundation (Fundação Amazonas Sustentável, FAS), which is co-financed by Amazonas State and the Amazon Fund3, and has also been supported by multiple domestic and foreign private donors over the years (including Bradesco Bank, Coca-Cola Brazil, Marriott, Samsung, Petrobras, and Lojas Americanas). 3 Created in 2008, the Amazon Fund for Forest Conservation and Climate is a REDD+ mechanism to finance forest protection projects and a commitment for with a results-based evaluation (Correa and others, 2019). It is managed by the Brazilian National Bank for Economic and Social Development (Banco Nacional de Desenvolvimento Economico e Social, BNDES). 5 Figure 1: Protected areas in Amazonas State and the Bolsa Floresta program Nine multiple-use reserves were enrolled in Bolsa Floresta in 2007-08; another five joined in 2009, and one more in 2010.4 The program currently covers more than 10 million ha, making it one of Latin America's largest PES programs in terms of area (Figure 1). Over 9,600 households are currently enrolled; participation rates within Bolsa Floresta reserves range from 70 to 100 percent of resident households (Newton and others, 2012; FAS, 2013, 2018). The main component of Bolsa Floresta, Bolsa Floresta Family (BF-F), makes monthly payments of BRL505 to households who have lived in a PA for at least two years and sign a zero-deforestation commitment. These payments are intended to be conditional on respecting the zero-deforestation commitment and on a few other conditions, such as enrolling children in school. The payments are made every month, to the female household head or wife. Land-use commitments are variable across reserves, but typically only marginally more restrictive than the PA rules (Börner and others, 2013). Average monthly household income is estimated at BRL410 to BRL560, so 4 In 2014, RDS Puranga Conquista split apart from APA Rio Negro, so FAS (2018) now reports 16 reserves within their portfolio. Below we only consider the original 15 reserves. 5 BRL50 equals EUR15.84 or USD13.69, based on average 2018 exchange rates (Federal Reserve Bank of St. Louis, 2019). 6 Bolsa Floresta payments represent relatively large income increases (Newton and others, 2012; Börner and others, 2013). In addition to direct payments to households, Bolsa Floresta also has three other components: § Bolsa Floresta Association (BF-A) supports local associations and collaborations among communities and partnerships with other organizations and local governments. The program promotes meetings within communities and reserves in order to build leadership capacity and promote participation and to secure social justice and the interests of all inhabitants. The annual grants amount to 10 percent of all BF-F payments, and can be used by the communities as they wish (Börner and others, 2013; FAS 2013; Newton and others, 2012). § Bolsa Floresta Income (BF-I) provides BRL350 per household annually to foster forest-friendly production systems chosen by participating communities, with Technical assistance provided by FAS staff. The most frequent investments include poultry, nuts, natural oil production, agroforestry, fruit production, and tourism (Newton and others, 2012). The hope is that increased productivity will shift families' income sources towards more forest-friendly activities.6 § Bolsa Floresta Social (BF-S) provides BRL350 per household annually for basic service infrastructure investments within communities, including investments, in electricity, water supply, sanitation and communications systems (Börner and others, 2013). The total support provided by Bolsa Floresta in 2013 in RDS Rio Negro was calculated at BRL1,406 per household per year (FAS, 2013). Newton and others (2012) measured an average annual Bolsa Floresta support to families in the multiple-use reserve of Uacari of BRL1,300. Bolsa Floresta thus combines PES with interventions typical of integrated conservation and development programs (ICDPs) in what might be described as a ‘PES+’ program. Participating households receive direct payments and supposedly longer-term gains from higher productivity, as well as greater collective benefits from community support. In addition, it is hoped that social pressures will lead participants to assist informally in PA monitoring by actively reporting violations of PA rules. 4. Evaluating the impacts of Bolsa Floresta We investigate these conservation effects across the 15 PAs enrolled in Bolsa Floresta, using panel data and quasi-experimental empirical methods. We focus on the program’s purely environmental impacts, asking whether the incentives it provided 6 Swartz (2015) analyzes the effects of the income component with data on over 200 households living in RDS Rio Negro and APA Rio Negro, on opposite banks of the Rio Negro. Households in both reserves participate in Bolsa Floresta, but at the time of the survey the BF-I component had only started in RDS Rio Negro. For the short period through which households benefited from the income component, the study could not find a statistically robust difference in income or asset levels between both groups. 7 have conserved additional forest, beyond what would have been saved solely from having these areas under multiple-use reserve protection (business-as-usual scenario). We consider the program's effects on three environmental outcomes: deforestation, forest degradation, and fire incidence. Our analysis considers the full program over 10 million ha, but can only measure the initial effects of the program: it started in 2008 and forest loss data for this analysis is available from 2004 to 2015. By the nature of our analysis, we cannot capture impacts on behaviors or attitudes of participants, nor on the program’s development targets. Nonetheless, we extend the analysis from the usual average changes at the reserve level and investigate forest changes within different economic contexts. Appendix 1 describes the data sources and processing. Table 1: Characteristics of Bolsa Floresta reserves Deforestation rate Before After Year Year of Size treatment treatment protected BF start (km2) (%) (%) APA Rio Negro 1995 2010 7,415 0.008 0.007 RDS do Rio Negro 2009 2009 1,584 0.025 0.076 FE de Maués 2003 2008 5,201 0.013 0.009 RDS Canumã 2005 2009 518 0.042 0.011 RDS Uacari 2005 2008 7,684 0.002 0.001 RDS Rio Amapá 2005 2009 2,890 0.000 0.000 RDS do Uatumã 2004 2008 5,632 0.005 0.013 RDS Amanã 1999 2009 25,179 0.004 0.003 RDS Rio Madeira 2006 2008 3,809 0.013 0.011 RDS do Juma 2006 2008 7,359 0.023 0.005 RDS Mamirauá 1990 2008 16,124 0.000 0.001 RDS Cujubim 2004 2008 26,665 0.000 0.001 RDS Piagaçu-Purus 2004 2008 10,759 0.002 0.004 RESEX do Rio Gregório 2007 2009 3,914 0.016 0.006 RESEX do Catuá-Ipixuna 2004 2008 3,048 0.007 0.005 All Bolsa Floresta reserves 2003 2008 8,518 0.011 0.010 Non-Bolsa Floresta reserves in 2001 5,343 0.017 0.014 Amazonas* Non-Bolsa Floresta reserves in the 1999 4,723 0.200 0.096 Brazilian Legal Amazon* Notes: We use the August to July window to define yearly cycles to adjust to the deforestation data structure from PRODES (2004-2015). *Average values of non-Bolsa Floresta reserves are built as averages before and after 2008. 8 Construction of counterfactual Bolsa Floresta is implemented in 15 of 53 multiple-use reserves in Amazonas (Table 1, Figure 1).7 We use the non-selected reserves as a base to construct our counterfactual scenario of what would have been the outcome without the intervention. However, Bolsa Floresta didn’t select participating reserves randomly; the rollout started in state-administered reserves, but other criteria—unknown to us— were also used. We thus have to assume that the treated reserves differ systematically from non-treated reserves. Bolsa Floresta reserves had lower average annual deforestation rates (0.010 percent) after treatment than non-Bolsa Floresta reserves (0.014 percent) and reserves in the entire Brazilian Amazon (0.090 percent), so a naïve comparison would have us conclude that Bolsa Floresta has been highly successful (Table 1). However, the selected reserves already had lower average deforestation (0.011 percent) before treatment than non-Bolsa Floresta reserves in Amazonas (0.017 percent) and in the Brazilian Amazon (0.20 percent). To address potential selection biases from unobservable characteristics, we use a three-step approach. First, we ‘slice’ all reserves into smaller 5x5 km spatial units,8 to capture the large variety of environmental and economic characteristics across and within reserves. Second, we eliminate from our pool of control cells all those dissimilar to the treated Bolsa Floresta cells, using nearest neighbor matching techniques based on pre-treatment deforestation trends, socio-economic indices, and geo-environmental characteristics. Matching is a prominent quasi-experimental method to overcome selection biases in spatial environmental applications (Andam and others, 2008; Gaveau and others, 2009; Honey-Rosés and others, 2011; Pfaff and others, 2015b). It reduces potential selection bias by finding the most similar untreated unit for each treated unit, considering observable pre-intervention characteristics. Pfaff and others (2015b) describe this strategy as comparing ‘apples-to-apples’ by leaving out the oranges. 9 Third, we construct a panel data set, exploiting the fact that Bolsa Floresta start years varied across reserves. This allows us to filter out potential confounding factors that are unobserved and invariant over time. Figure 2 shows that control and treated deforestation rates in the period prior to the start of Bolsa Floresta align well after matching, particularly in the years 2004-2007 (see also Figure A1-3). 7 There are 33 state and 20 federal administered multiple-use reserves in Amazonas. The 5 municipally administered reserves were excluded from this analysis. 8 The 5 to 5 km grid cells are artificial data containers, and do not affect the resolution of the spatial data: the 30 meters resolution of the deforestation data from PRODES remains, but each pixel is assigned to a cell. The spatial data aggregation to a lower resolution of 5 to 5 km grid cells serves to avoid biases from strong spatial autocorrelation (Avelino and others, 2016). 9 Full details on data processing, slicing, and matching procedures are provided in Appendix 1, with a summary on covariate balances and pre-treatment deforestation trends. 9 0.08 Bolsa Floresta implementation Bolsa Floresta reserves Non Bolsa Floresta reserves Deforestation rate (%) 0.06 Matched control 0.04 0.02 0.00 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 Year Note: Deforestation rates shown are averages of forest loss by total area across reserves. Figure 2: Deforestation in Bolsa Floresta and non-Bolsa Floresta multiple-use reserves Estimation procedure Our panel model is: ln &'() = ') + + &'( + ) + &'() The outcome variable ln &'() represents the logarithmic annual deforestation rate within each cell , in reserve , within district , in a given year . The treatment dummy is denoted as ') , and takes the value of one in year and all subsequent years when a reserve starts to participate in Bolsa Floresta. denotes the matrix of additional covariates, including yearly cloud cover, lagged district income, and a dummy indicating protection status from the beginning of reserve creation (see also Appendix 1). The latter effect is also controlled for in the matching process, which pairs treated reserves only to older control reserves. Importantly, the panel data allows us to include fixed effects to capture unobservable time-invariant characteristics (&'( ). Year-fixed effects () ) control for common macroeconomic shocks or regional changes in environmental policies and enforcement. The idiosyncratic error term is denoted as &'() . We proceed similarly to examine forest degradation and fire incidence. Information on annual degradation of forest areas is available since 2007 (INPE, 2008). This reduces the timeframe on our estimates to 2007-13, with observations for only one year before treatment started. Our estimations thus rely more strongly on the matching assumptions. As a measurement of fire incidence, we use the aggregate count of heat foci within a year from different satellite sources provided by INPE. We use Fixed-Effects estimations to assess the impact of Bolsa Floresta on our outcome variables. For robustness we also run First-Difference estimations, allowing 10 for path dependencies where we further control for initial biophysical, economic, and reserve characteristics (See Appendix 1). Errors are most likely correlated within each reserve, as they were selected based on reserve, not cell characteristics. We therefore cluster standard errors at the reserve level, following Angrist and Pischke (2009). Ignoring the violation of the standard independence assumption (E[εirdt,εjrdt]=0) usually results in overestimated and untrue significance levels (Cameron and Miller, 2015). 5. Results As mentioned, Bolsa Floresta reserves already had systematically lower average deforestation rates than other reserves before 2008 (Figure 2). This pattern continues following the establishment of Bolsa Floresta in 2007-08. Compared to the matched control group, the deforestation rates of the treated group are in some years higher and in some years lower; notably the Bolsa Floresta reserves avoided a forest-loss peak in 2010 that affected non-treated reserves. Average effects on forest loss and quality The estimates of Bolsa Floresta’s average effects on the logarithms of deforested area, degraded area, and fires are reported in Table 2. The effect of treatment on deforestation is associated with a small and statistically insignificant negative coefficient (column 1). Standard errors increase when clustering at the reserve level (column 2), underlining the importance of clustering when treatment is set for groups. Even if the coefficient had been significant, deforestation would only be reduced by 2.8 percent, corresponding to an area of 114 ha ([-290,480] ha 90 percent CI). We find very similar results for yearly forest degradation and fire incidences (columns 3-4). Both impact coefficients are close to zero, and remain insignificant. The estimates hold with First-Difference estimations and including path dependencies on initial biophysical, economic and reserve characteristics (see Table A2-1). To further test for robustness, we use weighted regressions, a variety of sample restrictions (including matching only with state reserves as controls), different matching procedures, and control for anticipated and lagged effects. However, our results are very robust throughout: none of these procedures systematically changes the impact estimates, which invariably remain statistically insignificant (Appendix 2). Bolsa Floresta was implemented in the first reserve at the end of 2007. It might have had a slow start, creating conservation effects only in later years. Alternatively, program planning could have produced a pre-program dip with higher forest losses, as residents of reserves anticipated the Bolsa Floresta and started to reduce forest- harming activities ex-ante. We test for delayed or anticipated effects by shifting the treatment dummy to years after and before program start (Figure A2-1). Yet, the results do not indicate that Bolsa Floresta had delayed impacts, nor does it appear that the Bolsa Floresta was anticipated in treated reserves, or that it produced lagged conservation effects. 11 Table 2: Bolsa Floresta effects on forest loss, degradation, and fires 1. 2. 3. Forest 4. Dependent: Logarithm of Forest loss Forest loss degradation Fires Treatment -0.028 -0.028 -0.022 0.014 (0.018) (0.057) (0.047) (0.012) Cloud cover -0.284*** -0.284*** -0.136*** -0.010 (0.018) (0.073) (0.043) (0.013) Protection status -0.016 -0.016 -0.349 -0.022 (0.038) (0.074) (0.233) (0.046) Lagged log of GDP p.c. 0.041 0.041 -0.017 0.019 (0.030) (0.089) (0.075) (0.014) Year and cell fixed effects Yes Yes Yes Yes N 112,752 112,752 65,772 103,356 Cluster Reserve (9,396) Reserve (48) Reserve (48) Reserve (48) Years 2004-2015 2004-2015 2007-2013 2004-2014 2 Adj. R 0.225 0.225 0.015 0.239 Notes: The dependent variable is the log of yearly newly deforested area, log of degraded forest area, and log of fire incidences. Regressions use a matched sample based on one-to-one nearest neighbour matching with replacement on the Mahalanobis distance. Time-varying conditions include yearly cloud coverage over remaining forest area, the log of yearly lagged district GDP per capita, and a dummy for protection status to control for the effect of reserve protection. Clustered standard errors are reported in parentheses. *,**,*** denote significance at the 10/5/1% level, respectively. Context-dependent heterogeneous impacts In a context with considerable heterogeneity, small estimated impacts of the Bolsa Floresta could be due to diverging positive and negative effects of the Bolsa Floresta offsetting each other. Where opportunity costs are low or zero, payments would be unlikely to induce a behavioral change, thus having no effect. Where opportunity costs are very high, payments may be insufficient to cover opportunity cost, so deforestation will continue (Persson and Alpízar, 2013). We would expect opportunity costs to be related to the degree of market integration. We use the pre- treatment deforestation level within a 20km buffer around each cell and each cell’s distance to the next major market as proxies for market integration, and divide treated and control units into low, medium, and high levels of market integration subsamples, redo the matching and re-estimate Bolsa Floresta treatment effects. The results are shown in Table 3 and illustrated in Figure 3. As expected, Bolsa Floresta impacts are minimal where pressure levels are low (panel a). They are also not statistically significant where pressure levels are high. In the medium range of pre-existing pressure levels, on the other hand, the Bolsa Floresta is estimated to significantly reduce deforestation by 5.7 percent. Stratifying observations by distance to the nearest city results in a similar pattern, but the estimates are not significant. 12 Table 3: Effects of Bolsa Floresta under heterogeneous market integration Deforestation pressure Market distance Low Medium High Low Medium High 1 2 3 4 5 6 Treatment 0.007 -0.057** 0.073 -0.025 -0.175 -0.003 (0.004) (0.026) (0.123) (0.074) (0.136) (0.012) Cloud cover 0.005 -0.145*** -0.875*** -0.492*** -0.192** -0.052*** (0.005) (0.053) (0.262) (0.131) (0.087) (0.019) Protection status 0.007 0.003 0.032 0.239 0.038 -0.117 (0.006) (0.068) (0.159) (0.243) (0.214) (0.089) Lagged log of GDP p.c. -0.008 -0.087 0.152 0.418 -0.100 -0.010 (0.007) (0.098) (0.208) (0.355) (0.117) (0.034) Year and cell fixed effects Yes Yes Yes Yes Yes Yes N 50,284 34,632 34,632 35,542 33,410 50,544 Cluster Reserve Reserve Reserve Reserve Reserve Reserve (31) (37) (44) (26) (35) (38) Years 2004-15 2004-15 2007-13 2004-14 2004-14 2004-14 2 Adj. R 0.225 0.225 0.015 0.239 0.239 0.239 Notes: The dependent variable is the log of yearly newly deforested area. Regressions use a matched sample based on one-to-one nearest neighbour matching with replacement on the Mahalanobis distance. Cloud cover is measured as cloud area over remaining forest area. Clustered standard errors are reported in parentheses. *,**,*** denote significance at the 10/5/1% level, respectively. (A) Pre−treatement deforestation level Low deforestation pressure ● Medium deforestation pressure ● High deforestation pressure ● −0.2 −0.1 0.0 0.1 0.2 (B) Travel distance to next city High distance to city ● Medium distance to city ● Low distance to city ● −0.2 −0.1 0.0 0.1 0.2 Impact coefficient Figure 3: Effects of Bolsa Floresta under heterogeneous market integration 13 6. Conclusions and discussion A naïve comparison of deforestation in reserves with and without Bolsa Floresta would come out in favor of Bolsa Floresta. However, this result is biased by sample selection: Bolsa Floresta-enrolled reserves already had lower deforestation prior to treatment—notably, within an Amazonas State context of much lower overall forest loss than elsewhere in the Brazilian Amazon. Matching similar treated and untreated areas and combining it with fixed effects regressions allow us to control for this selection bias. Once we do so, we find that the overall effect of Bolsa Floresta has been very small and statistically insignificant, whether measured in terms of deforestation, forest degradation, or fire incidence. However, Bolsa Floresta does appear to have had a significant effect on deforestation in areas with moderate market access. The usual caveats of quasi-experimental evaluation, including the potential influence of unobserved confounding variables (Rosenbaum, 2002), apply to this study. Research on the effectiveness of PA indicates, however, that selection bias (usually due to ‘high and far’ location) typically leads to conservation impacts being overestimated. In combination with our systematic and extensive robustness analyses, this makes us confident that our analysis reflects the actual lack of impacts of the program on the measured outcome variables up to the end of our observations in 2015. Our results are similar to those found in Mexico by Alix-Garcia and others (2019). There, the impact of the country’s forest PES program (Pagos por Servicios Ambientales del Bosque, PSAB) was statistically insignificant at the national level, but significant in areas of high deforestation risk (50 percent reduction). Similarly, Jayachandran and others (2017) analyzed a PES program implemented in an area of high deforestation pressure in Uganda and found large, significant reductions in forest loss. Why, then, was Bolsa Floresta’s impact so small, when PES has been able to reduce deforestation elsewhere? We discuss here two potential causes: program design and implementation. In Mexico, poor targeting was likely one reason for the poor overall PSAB performance: although deforestation risk was one of the prioritization criteria used to select participants, it had a very small weight, thus resulting in many areas at negligible risk being enrolled (Alix-Garcia and others, 2019). To some extent, the same applies to the long-standing Costa Rican national PES program (Hanauer and Canavire-Bacarreza, 2015), and to Peru’s pilot Amazon public PES program (PNCB) (Giudice and others, 2019). A similar pattern is observed for Bolsa Floresta. As shown in Table 1 and Figure 2, the reserves selected for Bolsa Floresta tended to have much lower ex-ante deforestation than other Amazonas reserves, let alone those in all of the Brazilian Legal Amazon. Of course, Bolsa Floresta implementers could not, ex ante, have foreseen the sustained decline in overall Brazilian deforestation. In fact, science-based projections for the Bolsa Floresta-enrolled Juma Reserve—site of Brazil’s first subnational REDD project—indicated significant imminent deforestation pressures, and thus expected 14 future additionality from conservation action (Soares-Filho and others, 2006). Had these fears been realized, the impact of Bolsa Floresta might have been larger. However, among the reserves initially available for participation, Bolsa Floresta implementers have, intentionally or not, selected those with lower internal pressure, on average. This focus on relatively conserved areas may, perversely, have resulted in a targeting strategy with little leverage to change deforestation behavior. This suggests that the program’s environmental impact could be increased by extending it to areas at higher risk of deforestation. As shown in Figure 3, the Bolsa Floresta does appear to have had some impact in areas of medium deforestation risk. Table 3 and Figure 3 also show, however, that Bolsa Floresta did not succeed in reducing deforestation in the areas with the highest pressure. This result could be due to design problems (offering payments that are insufficient to cover locally high opportunity costs of reducing deforestation), or to implementation problems, for example, unenforced conditionality: Bolsa Floresta uses a system of two consecutive ‘yellow-card’ warnings, but has not pursued a sanctioning strategy of excluding non- compliant participants. If conditionality was enforced, we would expect an insufficient payment offer to result in either low participation or in many enrolled participants dropping out or being expelled. As we observe relatively high participation rates (and no contract cancellations) even in reserves where deforestation is on-going, we suspect that enforcement of conditionality is low. More strictly enforcing conditionality might improve results and reduce expenses by suspending payments to non-compliers. In sum, pilot cases worldwide have shown that PES can complement PA interventions, improving conservation outcomes and lowering social cost. Bolsa Floresta as currently designed and implemented has likely strengthened communal institutions and had positive social and welfare outcomes, transferring significant resources to local households and communities (Hayes and others, 2017; Börner and others, 2013). Bolsa Floresta has also been an important proof-of-concept example for other PES and PES- like conservation initiatives in the region. However, due mainly to its counter-intuitive spatial targeting of areas with ex-ante negligible deforestation pressures, rigorously measured conservation outcomes have so far by design remained limited. 15 References Alix-Garcia, J., and H. Wolff. 2014. “Payment for ecosystem services from forests.” Annual Review of Resource Economics, 6(1):361–380. Alix-Garcia, J.M., K.R.E. Sims, V.H. Orozco-Olvera, L. Costica, J.D. Fernandez Medina, S. Romo- Monroy, and S. Pagiola. 2019. “Can environmental cash transfers reduce deforestation and improve social outcomes? a regression discontinuity analysis of Mexico's national Program (2011–2014).” Policy Research Working Paper No.8707. Washington: World Bank. Amin, A., J. Choumert-Nkolo, J.-L. Combes, P.C. Motel, E. Kéeré, J.-G. Ongono-Olinga, and S. Schwartz. 2019. “Neighborhood effects in the Brazilian Amazônia: Protected areas and deforestation.” Journal of Environmental Economics and Management, 93:272–288. Andam, K.S., P.J. Ferraro, A. Pfaff, G.A. Sanchez-Azofeifa, and J.A. Robalino 2008. “Measuring the effectiveness of protected area networks in reducing deforestation.” Proceedings of the National Academy of Sciences, 105(42):16089–16094. Angrist, J.D., and J.-S. Pischke. 2009. Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton: Princeton University Press. Arima, E.Y., P. Barreto, E. Araújo, and B. Soares-Filho. 2014. “Public policies can reduce tropical deforestation: Lessons and challenges from Brazil.” Land Use Policy, 41:465–473. Assunção, J., C. Gandour, and R. Rocha. 2012. “Deforestation slowdown in the legal Amazon: Prices or policies.” CPI Working Paper. Rio de Janeiro: Climate Policy Initiative. Avelino, A.F.T., K. Baylis, and J. Honey-Rosés. 2016. “Goldilocks and the raster grid: Selecting scale when evaluating conservation programs. PLoS ONE, 11(12):1–24. Bauch, S.C., E.O. Sills, and S.K. Pattanayak. 2014. “Have we managed to integrate conservation and development? ICDP impacts in the Brazilian Amazon.” World Development, 64:135–148. BenYishay, A., S. Heuser, D. Runfola, and R. Trichler. 2017. “Indigenous land rights and deforestation: Evidence from the Brazilian Amazon.” Journal of Environmental Economics and Management, 86:29-47. Börner, J., S. Wunder, S. Wertz-Kanounnikoff, M.R. Tito, L. Pereira, and N. Nascimento. 2010. “Direct conservation payments in the Brazilian Amazon: Scope and equity implications.” Ecological Economics, 69:1272-1282. Börner, J., S. Wunder, F. Reimer, K.R. Bakkegaard, V. Viana, J. Tezza, T. Pinto, L. Lima, and S. Marostica. 2013. Promoting Forest Stewardship in the Bolsa Floresta Programme: Local Livelihood strategies and Preliminary Impacts. Rio de Janeiro: CIFOR. Börner, J., E. Marinho, and S. Wunder. 2015. “Mixing carrots and sticks to conserve forests in the Brazilian Amazon: A spatial probabilistic modeling approach.” PLoS ONE, 10(2):1–20. Börner, J., K. Baylis, E. Corbera, D. Ezzine-de-Blas, P.J. Ferraro, J. Honey-Rosés, R. Lapeyre, U.M. Persson, and S. Wunder. 2016. “Emerging evidence on the effectiveness of tropical forest conservation.” PLoS ONE, 11(11):1–11. Börner, J., K. Baylis, E. Corbera, D. Ezzine-de-Blas, J. Honey-Rosés, U.M. Persson, and S. Wunder. 2017. “The effectiveness of payments for environmental services.” World Development, 96:359- 374. Bowman, M.S., B.S. Soares-Filho, F.D. Merry, D. Nepstad, H. Rodrigues, and O.T. Almeida. 2012. “Persistence of cattle ranching in the Brazilian Amazon: A spatial analysis of the rationale for beef production.” Land Use Policy, 29(3):558–568. Cameron, A.C., and D.L. Miller. 2015. “A practitioner’s guide to cluster-robust inference.” Journal of Human Resources, 50(2):317–372. Canova, N.P., and G.M. Hickey. 2012. “Understanding the impacts of the 2007-08 Global Financial Crisis on sustainable forest management in the Brazilian Amazon: A case study.” Ecological Economics, 83:19–31. 16 Correa, J., R. van der Hoff, and R. Rajao. 2019. “Amazon Fund 10 years later: Lessons from the world’s largest REDD+ Program.” Forests, 10(3):272. Cuenca, P., R. Arriagada, and C. Echeverría. 2016. “How much deforestation do protected areas avoid in tropical Andean landscapes?” Environmental Science Policy, 56:56–66. Cisneros, E., S.L. Zhou, and J. Börner. 2015. “Naming and shaming for conservation: Evidence from the Brazilian Amazon.” PLoS ONE, 10(9):1–24. Clements, T., and E.J. Milner-Gulland. 2014. “Impact of payments for environmental services and protected areas on local livelihoods and forest conservation in northern Cambodia.” Conservation Biology, 29:78–87. Engel, S., S. Pagiola, and S. Wunder. 2008. “Designing payments for environmental services in theory and practice: An overview of the issues.” Ecological Economics, 65:663-674. FAS (Fundação Amazonas Sustentável). 2013. “Results of a baseline study for a REDD+ pilot area in Brazil: The Rio Negro APA and the Bolsa Floresta Programme.” London: IIED. FAS (Fundação Amazonas Sustentável). 2018. “Relatório de atividades 2018.” Manaus: FAS. Federal Reserve Bank of St. Louis. 2019. “FRED Economic data.” St. Louis: Federal Reserve Bank of St. Louis. Gaveau, D.L.A., J. Epting, O. Lyne, M. Linkie, I. Kumara, M. Kanninen, and N. Leader-Williams.2009. “Evaluating whether protected areas reduce tropical deforestation in Sumatra.” Journal of Biogeography, 36(11):2165–2175. Giudice, R., J. Börner, S. Wunder, and E. Cisneros. 2019. “Selection biases and spillovers from collective conservation incentives in the Peruvian Amazon.” Environmental Research Letters, 14:045004. Hanauer, M.M., and G. Canavire-Bacarreza. 2015. “Implications of heterogeneous impacts of protected areas on deforestation and poverty.” Philosophical Transactions of the Royal Society B: Biological Sciences, 370:20140272. Hayes, T., F. Murtinho, and H. Wolff. 2017. “The impact of payments for environmental services on communal lands: An analysis of the factors driving household land-use behavior in Ecuador.” World Development, 93:427-446. Hargrave, J., and K. Kis-Katos. 2013. “Economic causes of deforestation in the Brazilian Amazon: A panel data analysis for the 2000s.” Environmental and Resource Economics, 54(4):471–494. Honey-Rosés, J., K. Baylis, and M.I. Ramírez. 2011. “A spatially explicit estimate of avoided forest loss.” Conservation Biology, 25(5):1032–1043. IBGE (Instituto Brasileiro de Geografia e Estatística). 2000. Censo demográfico. Accessed: 2016-10-14. IBGE (Instituto Brasileiro de Geografia e Estatística). 2006. Agropecuário IC. Accessed: 2016-10-14. IBGE (Instituto Brasileiro de Geografia e Estatística). 2014. Produção da extração vegetal e da silvicultura (PEVS). Accessed: 2016-10-14. INPE (Instituto Nacional de Pesquisas Espaciais). 2008. “Monitoramento da cobertura florestal da Amazônia por satélites sistemas PRODES, DETER, DEGRAD e queimadas 2007-2008.” São José dos Campos: INPE. INPE (Instituto Nacional de Pesquisas Espaciais). 2019. “PRODES—Monitoramento da floresta amazônica brasileira por satélite.” São José dos Campos: INPE. Accessed: 2019-07-22. Jayachandran, S., J. de Laat, E.F. Lambin, C.Y. Stanton, R. Audy, and N.E. Thomas. 2017. “Cash for carbon: A randomized trial of payments for ecosystem services to reduce deforestation.” Science, 357:267-273. Joppa, L.N., and A. Pfaff.2010. “Global protected area impacts.” Proceedings of the Royal Society of London B: Biological Sciences, 278:1633–1638 Ludewigs, T., A. de Oliveira D’Antona, E.S. Brondízio, and S. Hetrick. 2009. “Agrarian structure and land-cover change along the lifespan of three colonization areas in the Brazilian Amazon.” World Development, 37(8):1348–1359. 17 Maia, H., J. Hargrave, J.R. Gómez, and M. Röper. 2011. “Avaliação do Plano de Ação para Prevenção e Controle do Desmatamento na Amazônia Legal (PPCDAm) (2007-2010).” Paper presented at the Seminário de Avaliação do PPCDAm, October 2011. Montoya Zumaeta, J., and E. Rojas. 2017. “The effects of Moyobamba rewards for Hydrological Ecosystem Services Mechanism on plot land cover and household wellbeing.” LACEEP Working Paper No.98. Turrialba: Latin American and Caribbean Environmental Economics Program. Nelson, A., and K.M. Chomitz. 2011. “Effectiveness of strict vs. multiple use protected areas in reducing tropical forest fires: A global analysis using matching methods.” PLoS ONE, 6(8):e22722. Nepstad, D., D. McGrath, C. Stickler, A. Alencar, A. Azevedo, B. Swette, T. Bezerra, M. DiGiano, J. Shimada, R. Seroa da Motta, E. Armijo, L. Castello, P. Brando, M.C. Hansen, M. McGrath-Horn, O. Carvalho, and L. Hess. 2014. “Slowing Amazon deforestation through public policy and interventions in beef and soy supply chains.” Science, 344(6188):1118–1123. Newton, P., E.S. Nichols, W. Endo, and C.A. Peres. 2012. “Consequences of actor level livelihood heterogeneity for additionality in a tropical forest payment for environmental services programme with an undifferentiated reward structure.” Global Environmental Change, 22(1):127–136. Nolte, C., and A. Agrawal. 2013. “Linking management effectiveness indicators to observed effects of protected areas on fire occurrence in the Amazon rainforest.” Conservation Biology, 27:155-165. Nolte, C., A. Agrawal, K.M. Silvius, and B.S. Soares-Filho. 2013. “Governance regime and location influence avoided deforestation success of protected areas in the Brazilian Amazon.” Proceedings of the National Academy of Sciences, 110:4956-4961. Pailler, S. 2018. “Re-election incentives and deforestation cycles in the Brazilian Amazon.” Journal of Environmental Economics and Management, 88:345-365. Pagiola, S. 2008. “Payments for environmental services in Costa Rica.” Ecological Economics, 65:712- 724. Pagiola, S., H. Carrascosa von Glehn, and D. Taffarello. 2013. “Brazil’s Experience with Payments for Environmental Services. PES Learning Paper No.2013-1. Washington: World Bank. Pereira, E.P.J, Silveira Ferreira, L.C. De Santana Ribeiro, T. Sabadini Carvalho, H.B. de B. Pereira. 2019. “Policy in Brazil (2016–2019) threaten conservation of the Amazon rainforest.” Environmental Science & Policy, 100:8-12. Persson, U.M., and F. Alpízar. 2013. “Conditional cash transfers and payments for environmental services: A conceptual framework for explaining and judging differences in outcomes.” World Development, 43:124-137. Pfaff, A., J. Robalino, E. Lima, C. Sandoval, and L.D. Herrera. 2014. “Governance, location and avoided deforestation from protected areas: Greater restrictions can have lower impact, due to differences in location.” World Development, 55:7–20. Pfaff, A., J. Robalino, D. Herrera, and C. Sandoval. 2015a. “Protected areas’ impacts on Brazilian Amazon deforestation: examining conservation: Development interactions to inform planning.” PLoS ONE, 10(7):e0129460. Pfaff, A., J. Robalino, D. Herrera, and C. Sandoval. 2015b. “Protected area types, strategies and impacts in Brazil’s Amazon: public protected area strategies do not yield a consistent ranking of protected area types by impact.” Philosophical Transactions of the Royal Society B: Biological Sciences, 370:20140273. Porter-Bolland, L.. E.A. Ellis, M.R. Guariguata, I. Ruiz-Mallén, S. Negrete-Yankelevich, and V. Reyes- García. 2012. “Community managed forests and forest protected areas: An assessment of their conservation effectiveness across the tropics.” Forest Ecology and Management, 268:6-17. Robinson, E.J.Z., A.M. Kumar, and H.J. Albers. 2010. “Protecting Developing Countries' Forests: Enforcement in Theory and Practice.” Journal of Natural Resources Policy Research, 2:25-38. Robinson, E.J.Z., H.J. Albers, G. Ngeleza, and R.B. Lokina. 2014. “Insiders, outsiders, and the role of local enforcement in forest management: An example from Tanzania.” Ecological Economics, 107:242–248. 18 Rosenbaum, P.R. 2002. Observational Studies. New York: Springer New York. Samii, C., M. Lisiecki, P. Kulkarni, L. Paler, and L. Chavis. 2014. “Effects of payment for environmental services (PES) on deforestation and poverty in low and middle income countries: A systematic review.” Campbell Systematic Reviews, 10(11):1–95. Santiago, T.M.O., J. Caviglia-Harris, and J.L.P. de Rezende. 2018. “Carrots, sticks and the Brazilian Forest Code: the promising response of small landowners in the Amazon.” Journal of Forest Economics, 30:38-51. Sekhon, J.S. 2011. “Multivariate and propensity score matching software with automated balance optimization: The matching package for R.” Journal of Statistical Software, 42(7):1–52. Sengupta, S. 2018. “What Jair Bolsonaro’s victory could mean for the Amazon, and the planet.” New York: The New York Times. Sims, K.R., and J.M. Alix-Garcia. 2017). “Parks versus PES: Evaluating direct and incentive-based land conservation in Mexico.” Journal of Environmental Economics and Management, 86:8-28. Soares-Filho, B.S., D.C. Nepstad, L.M. Curran, G.C. Cerqueira, R.A. Garcia, C.A. Ramos, E. Voll, A. McDonald, P. Lefebvre, and P. Schlesinger. 2006. “Modelling conservation in the Amazon basin.” Nature, 440:520-523. Soares-Filho, B., P. Moutinho, D. Nepstad, A. Anderson, H. Rodrigues, R. Garcia, L. Dietzsch, F. Merry, M. Bowman, L. Hissa, R. Silvestrini, and C. Maretti. 2010. “Role of Brazilian Amazon protected areas in climate change mitigation.” Proceedings of the National Academy of Sciences, 107(24):10821–10826. Swartz, E. 2015. “The effect of integrated conservation and development programs in protected areas on human wellbeing: An empirical analysis of Brazil’s Bolsa Floresta programme.” Masters thesis. Uppsala: Swedish University of Agricultural Sciences. UNEP-WCMC and IUCN. 2016. Protected Planet Report 2016. Cambridge and Gland: UNEP-WCMC and IUCN. Viana, V., M. Grieg-Gran, R. Della Mea, and G. Ribenboim. 2009. “The costs of REDD: Lessons from Amazonas.” London: IIED. Viana, V., J. Tezza, V. Salviati, G. Ribenboim, T. Megid, and C. dos Santos. 2012. “O Programa Bolsa Floresta no Estado do Amazonas.” In: S. Pagiola, H. Carrascosa Von Glehn, and D. Taffarello (Eds.), Experiências de Pagamentos por Serviços Ambientais no Brasil. São Paulo: SMA. World Bank. 2018. “Evaluating the permanence of forest conservation following the end of payments for environmental services in Uganda.” Report No.AUS0000379. Washington: World Bank. Wunder, S. 2015. “Revisiting the concept of payments for environmental services.” Ecological Economics, 117:234-243. 19 Appendix 1: Data Data sources Our outcome variables are annual deforestation, forest degradation, and number of fires per cell. Deforestation, forest degradation, and fire incidence data are downloaded in shapefile format from the web-site of the National Institute of Space Research (Instituto Nacional de Pesquisas Espaciais, INPE).10 INPE's deforestation measurement is based on Landsat imagery, and deforestation patches are defined as clear-cut deforestation—the complete loss of tree cover at 30 m resolution. Reliable deforestation data are available from 2003. To find comparable control group we match on pre-treatment data from 2003 to 2007. For our estimations of the treatment effects we use data from 2004 to 2015, when deforestation rates started to decline in Brazil. INPE uses the August to July cycle for its yearly measurements, exploiting the relatively cloud-free dry period of the year.11 INPE started to record forest degradation data in 2007, which reduces our timeframe significantly to one year before the first treatment started. In this analysis we could only consider degradation data until 2013. As consequence, estimations rely increasingly on matching assumptions and less on FE assumptions. The annual fire outcome is measured as the sum of counts of fire foci detected by several satellites on a daily basis. Data comprises the years 2004 to 2014. All spatial data processing is conducted on a PostgresSQL 9.2.3 data server with a PostGIS 2.0.1 spatial extension. Covariates used for the matching procedure and the time series analysis are summarized in Table A1-1. Data are either constructed through spatial calculations or obtained from official secondary data sources (district, reserve). Pre-treatment outcomes (deforestation, degradation, and fires), data on forest coverage, cloud coverage, non-forest coverage and water coverage (rivers, lakes) are provided by INPE's yearly deforestation database. Coverage is measured as a ratio of land use over the total area of a cell. Land use classes are obtained from the 2008 revision of INPE's TerraClass project.12 INPE classifies land use on deforested land into five categories: agricultural land, mixed land occupation, secondary vegetation, pasture land, and urban land. We use these classes to construct coverages for each cell. INPE classified these lands in the same year in which Bolsa Floresta rolled out. As the data corresponds to the year, in which Bolsa Floresta was actually rolled out, it is reasonable to assume that the program has not yet affected land use decisions and we can treat these variables as unaffected by the program. 10 INPE-PRODES Instituto Nacional de Pesquisa Espaciais/ Projeto PRODES—Monitoramento da floresta amazônica brasileira por satélite, http://www.obt.inpe.br/prodes/index.php. 11 For a detailed description of the methodology used to construct yearly deforestation data, see Cisneros and others (2015). 12 Available at: http://www.inpe.br/cra/projetos_pesquisas/terraclass2010.php 20 Table A1-1: Summary statistics of matching covariates after matching Variable class, variable Data level Primary source Mean SD Min Max Conservation policy Protected area created before 2009 Reserve IBAMA 0.99 0.11 0.00 1.00 Protected area created before 2008 Reserve IBAMA 0.99 0.11 0.00 1.00 Protected area created before 2007 Reserve IBAMA 0.97 0.18 0.00 1.00 Protected area created before 2006 Reserve IBAMA 0.87 0.33 0.00 1.00 Protected area created before 2005 Reserve IBAMA 0.79 0.41 0.00 1.00 Percentage area from original grid cell (%) Cell 0.86 0.27 0.05 1.00 Distance to next indigenous reserve (km) Cell IBAMA 73.06 49.20 0.00 277.36 Land in settlement project (%) Cell INCRA 0.35 0.47 0.00 1.00 Neigh. land in settlement project (%) Cell 2.59 3.53 0.00 9.81 Economic Deforestation in 2003 (%) Cell INPE / PRODES 0.00 0.00 0.00 0.11 Deforestation in 2004 (%) Cell INPE / PRODES 0.00 0.00 0.00 0.07 Deforestation in 2005 (%) Cell INPE / PRODES 0.00 0.00 0.00 0.04 Deforestation in 2006 (%) Cell INPE / PRODES 0.00 0.00 0.00 0.02 Accumulated deforestation in 2006 (%) Cell INPE / PRODES 0.01 0.04 0.00 0.74 Pre-treatment deforestation within 20 km Cell INPE / PRODES -0.11 0.36 -0.25 3.60 (st.dev.) Forest land (%) Cell INPE / PRODES 0.94 0.15 0.01 1.00 Travel time to major cities (st.dev.) Cell IBGE 0.21 1.09 -1.56 3.16 Agricultural land use (%) Cell INPE's TerraClass 0.01 0.05 0.00 0.80 Urban land (%) Cell INPE's TerraClass 0.00 0.00 0.00 0.06 Land speculation coverage (%) Cell Bowman and others, 0.35 0.43 0.00 1.00 2012 District population density in 2007 District IBGE Demo. Census 5.90 26.10 0.09 144.38 District GDP per capita (BRL ‘000) District IBGE 6.34 6.15 1.77 34.98 District farm coverage (%) District IBGE Agr. Census 0.02 0.04 0.00 0.43 District share of small farms (%) District IBGE Agr. Census 0.78 0.21 0.23 0.99 District tractors per farm District IBGE Agr. Census 0.01 0.02 0.00 0.10 Neigh. pre-treatment deforestation (03-06) (%) Cell 0.01 0.02 0.00 0.33 Neigh. forest land (%) Cell 6.70 1.72 0.28 10.00 Neigh. agricultural land (%) Cell 0.01 0.03 0.00 0.37 Neigh. urban land (%) Cell 0.00 0.00 0.00 0.02 Bio-physical Av. cloud area (03-06) (sq. km) Cell INPE / PRODES 1.66 2.25 0.00 15.64 Non-forest land (%) Cell INPE / PRODES 0.02 0.09 0.00 0.99 Water bodies (%) Cell INPE / PRODES 0.03 0.09 0.00 0.99 Neigh. water bodies (%) Cell 0.21 0.42 0.00 4.21 Notes: Statistics refer to 4698 treated BOLSA FLORESTA cells and 4698 matched control cells. We include data on the location of federal agrarian settlement projects, using the shape file provided by the National Institute of Colonization and Agrarian Reform (Instituto Nacional de Colonização e Reforma Agrária, INCRA). A land speculation map is provided by Bowman and others (2012), indicating whether or not a particular plot of forest land can be considered as being potentially profitable if converted for cattle ranching. We intersect Bowman's layer with the cells to construct an index of land profitability that can capture the pre-existing deforestation pressures on the reserves. 21 Non-spatial attributes to our cells database are cross-linked via the spatial location of the cell centroids within administrative entities. Reserve characteristics on size and the foundation year are obtained from the spatial database on reserves. The district boundaries from 2007 are used for our analysis and are publicly available from Brazilian Institute of Geography and Statistics (INPE, 2019). District characteristics include population densities in 2007 from the IBGE Demographic Census (IBGE 2000), GDP per capita and agricultural GDP per capita in 2006 from IBGE Agricultural Census (IBGE 2006). Information on the farm coverage, the share of small farms and tractors per farm also come from the Agricultural census. Timber prices between 2003 and 2006 are constructed as the ratio between quantity and total value of timber produced and obtainable from the IBGE-PEVS report (IBGE 2014). Spatially lagged variables are used as covariates to control for dependencies of deforestation on close contexts. Spatially lagged covariates are constructed by “queen style”, defining neighbors as such when two cells share a point on the boundary line. A spatially lagged variable is thus a weighted average of the values in neighboring cells. Slicing reserves To capture the spatial diversity and deforestation pressures within reserves we create a higher spatial resolution for the analysis. We intersect a grid with the administrative boundaries of all multiple-use reserves, which gives us a multitude of cells in each reserve (see Figure A1-1). The size chosen for the grid is 0.045 by 0.045 degrees which correspond to 5 by 5 km rectangles at the equator. Cell size is a compromise between spatial precision and spatial autocorrelation. Using a lower spatial resolution would fail to capture the spatial diversity, whereas a higher spatial resolution risks creating redundant observations and spatial autocorrelation. Owing to our vectored data structure the resulting spatial units range between 0 and 25 km2. At the border of reserves, cell size may thus only be a fraction of the original cell size (Figure A1-1). By keeping both entire and fractional cells as observations, we avoid a potential bias from loss of information or misattribution of treatment and outcomes at reserve borders. We only exclude very small units that are covered by only 5 percent of the slicing grid units. The 61 reserves are thereby divided into 14,397 grid cells. Spatial information on outcomes and controls are intersected and attributed to these smaller units, making the analysis spatially explicit. 22 APA Rio Negro PE do Rio APA ME Negro Rio Setor Sul Negro-S. RDS RDS Rio Negro do Tupé Unprotected area Indigenous territories APA MD do Strictly protected areas R.Negro-Setor Multiple-use protected areas Paduari-Solimões Bolsa Floresta Program reserve Unit of analysis Excluded cells from analytical sample Hydrography Deforestation 2007-2012 0 10 20 km Roads The figure depicts the slicing of reserves into spatial units of 5 to 5 km. The dashed reserves are the RDS do Rio Negro (south) and APA Rio Negro (north). RDS (Sustainable Development Reserve) and APA (Environmental Protection Area) are two subtypes of the multiple-use reserves category in the Brazilian protected area system. Figure A1-1: The spatial slicing of protected areas Sample The sample used in this research differs from the full database. As explained below, we drop some observations owing to data characteristics or to analytical reasoning before the analysis. The cell structure is constructed with 70,607 rectangular grid cells of 5 to 5 km covering the full Brazilian Legal Amazon. The resulting 37,942 cells are the data shell in which we fill the data described above. All spatial data measurements are constructed with PostgreSQL~9.2.3 and the PostGIS~2.0. add on. Distances between individual cells and line objects such as rivers, roads, reserve boundaries, etc. are constructed as direct lines from the center point of a cell to the nearest line-fragment of the respective object. Distances between individual cells and point objects (for example, district capitals) are constructed as direct lines from the center point of a cell to the nearest point. To calculate distances 23 to polygons, the object has to be converted into lines first. Area calculations, like deforestation within cells are based on the intersection of the two layers. As we use neighboring characteristics for this analysis, we have to exclude 18 cells that do not have any neighbor (for example, singular cells on islands). For analytical reasons, we drop 1,460 units that cover less than 5 percent of the original grid cell, that is, dropping each cell that is roughly smaller than 1.25 sq. km We suspect that these small spatial units are prone to measurement errors and create problems of spatial interdependencies in the unobservables. More importantly, we exclude all observations that have no forest cover (3,077). These can result from extensive logging in previous years, but more likely are fully covered by swamps and water bodies. 727 cells intersect the outer line of the Amazonas biome and are therefore excluded. Due to missing data in our travel distance measurement and from missing data at the district level, we further exclude 297 observations. We restrict the sample to state and federal administered reserves, excluding 668 cells which reside in a reserve administered at the municipality level. We exclude one reserve with 8 cells, the APA Nhamund, which protects lakes and flooding areas of the Amazon river. To avoid an analytical bias from spatial leakage we exclude all untreated cells within a distance of 20 km to treated cells (994). Clouds introduce a bias from systematic measurement errors of yearly deforestation for small spatial units. We therefore exclude all cells that had experienced more than 85 percent cloud coverage during 3 years of our timeframe, dropping 947 observations. As we are focusing only on multiple-use reserves that reside within the Amazonas state, with 53 multiple-use reserves, the drop 18,329 control cells without Bolsa Floresta. Finally, we identify 14 outliers in our control group that have experienced exaggerated deforestation levels with more than 15 percent (375 ha) forest cleared in some year prior enrolment in Bolsa Floresta. For this analysis we assume that these occurrences are driven by local unobservable peculiarities which could be related to treatment status and bias our results. Our final sample consists of 11,411 cells, 6,664 control and 4,747 treatment cells covering over 53 reserves including the 15 reserves of Bolsa Floresta. Spatial Matching The second step of our analysis consists of finding potential counterfactual cells. Guided by our theoretical framework we want to find control cells that experienced equal internal and external deforestation pressures. Similarity of deforestation pressures is approximated with observable pre-treatment environmental and socio- economic characteristics. Pressures from within and from outside of reserves are considered using variables on three data levels: cell characteristics, reserve characteristics and district characteristics (see Figure A1-2). Table A1-1 lists the data class, level, and sources of each covariate. 24 The figure depicts the three data levels: cell, district, reserve. Forest conservation at a cell level depends on its own context and on the surrounding context. Cell, reserve and district characteristics serve to describe the internal and external pressures each cell experiences. Figure A1-2: Data levels Matching is implemented with Sekhon's (2011) ‘Matching’ package, using a 1 to 1 nearest neighbor matching technique with replacement using the Mahalanobis distance measure. We implement two additional non-standard restrictions to the matching algorithm. First, each cell can cover an area of 5 to 100 percent of its slicing grid cell (5 to 5 km). In order to avoid matches between observations with different sizes, we restrict the algorithm to find only pairs of similar sized cells—within a margin of 5 percentage points. Second, the age of treated and control reserves varies significantly and some reserves were founded at the very outset of Bolsa Floresta while others are much older. To avoid to match control reserves that were founded later than Bolsa Floresta started, we restrict the procedure to only find control matches from reserves founded before Bolsa Floresta started. A detailed summary of all covariates used in the matching procedure is provided in Table A1-1. The table describes the attribute class of each factor (bio-physical, economic or political) and reports the data level (cell, reserve or district). 25 Characteristics of the natural environment are considered at the cell level with pre-existing deforestation trends, pre-existing number of fire incidents, initial forest cover, secondary vegetation, non-forest area (swamps and bush land areas) and water bodies (lakes and rivers). The economic environment is considered at cell level with infrastructural indices (distances to roads, rivers, and district capitals). Economic activities are controlled for at cell level with remotely sensed land use classes (agricultural land, mixed occupation, secondary vegetation, pasture and urban land). Economic pressures on reserves are approximated with official statistics at the district level (population density, GDP per capita, GDP per capita from agriculture, the percentage area under farms, the share of small farms, the average tractors per farm and an average timber price). Furthermore, we include an index on land speculation potential at a cell level based on Bowman and others’ (2012) spatial model of extensive cattle profitability. The institutional conservation environment is measured with data on settlement projects and data on further protected areas surrounding each cell. Settlement projects are a major influence, as they often take on characteristics in opposition to conservation although more recent efforts to establish “sustainable settlements” in the Amazon may change this situation (Ludewigs and others, 2009). We use a binary indicator to determine whether or not a cell is covered by a settlement project. A favorable conservation environment is considered with distances to the next strictly protected reserve and the next indigenous area. The conservation quality of each reserve is partly controlled for by the size and years of existence of the respective reserve. Distances of each cell to its own reserve border capture the relative internal position and reflect the degree of exposure to external deforestation pressures. Finally, we include indices on the narrower spatial context using neighborhood characteristics of adjacent cells. Matching estimators assume unconfoundedness and common support. Unconfoundedness requires that the selection of the reserves was solely dependent on the observable characteristics. Common support is a requirement of overlapping distributions of the distance measure or propensity scores between the control group and the treatment group. After personal discussions with the head of FAS (Virgílio Viana) and implementing agents we are confident that our covariate set covers the factors most relevant in the selection process.13 The matching assumptions for causal interpretation are relaxed by the panel-data-based post-matching estimation of the average treatment effect on the treated (ATT), described below. The 1:1 nearest neighbor matching results in 4698 paired controls created out of 881 unique controls spread over 33 reserves. The procedure reduces covariate imbalances on average by 83 percent (Figure A1-3). More importantly, pre-treatment deforestation trends of the matched control approach the treated. The difference reduces significantly in the pre-treatment deforestation indices (2003-06)—by 56-98 percent. That is controls and treated units in the matched dataset now have fairly equal 13 Visits to the headquarters of FAS in Manaus were conducted in July and November 2013. 26 trends before the program was rolled out, which increases confidence in the parallel time trend assumption of our estimation procedure. Protected area created before 2009 ● ● Protected area created before 2008 ● Protected area created before 2007 ● ● Protected area created before 2006 ● ● Protected area created before 2005 ● ● Percentage area from original grid cell ● ● Perc. deforestation in 2003 ●● Perc. deforestation in 2004 ●● Perc. deforestation in 2005 ● ● Perc. deforestation in 2006 ● ● Accumulated deforestation in 2006 ● ● Pre−treatement deforestaiton within 20km ● ● Av. cloud area (03−06) ● ● Forest land ● ● Non−forest land ●● Covariates Water bodies ●● Traveltime to major cities ● ● Agricultrual land use ● Urban land ●● Land speculation coverage ● ● Distance to next indigenous reserve ●● Land in settlement project ● ● District population density in 2007 ● ● District GDP per capita ● ● District farm coverage ● ● District share of small farms ● ● District tractors per farm ● ● Neig. pre−treatment deforestation (03−06) ●● Neigh. forest land ●● Neigh. water bodies ● ● Neigh. land in settlement project ● ● Neigh. agricultural land ● ● Neigh. urban land ● ● −1.0 −0.5 0.0 0.5 1.0 1.5 Standardized difference in means ● Matched ● Unmatched The figure depicts the mean standardized differences between control observations and treated observations, before (black) and after (orange) matching. The distances decrease after matching significantly towards zero. At zero difference between matched controls and treated units, the selection bias converges to zero. Figure A1-3: Covariate balances before and after matching 27 Appendix 2: Robustness of empirical results First-Difference estimations Difference estimations show negative impact coefficients (Table A2-1). To capture cell specific and reserve specific deforestation trends as well as the economic pressure level outside of reserves we include initial characteristics on biophysical, economic and reserve conditions (columns 2-4). Controlling those characteristics only marginally raises the treatment coefficient. Table A2-1: Path dependency by initial conditions 1 2 3 4 Treatment -0.064 -0.072 -0.075 -0.070 (0.082) (0.088) (0.088) (0.089) Time varying conditions Yes Yes Yes Yes Biophysical conditions Yes Yes Yes Economic conditions Yes Yes Reserve characteristics Yes Year fixed effects Yes Yes Yes Yes Cell fixed effects Obs. 112752 112752 112752 112752 Cluster 48 48 48 48 Cluster level Reserve Reserve Reserve Reserve R2 (full model) 0.015 0.017 0.017 0.017 2 Adj. R (full model) 0.015 0.017 0.017 0.017 Notes: Results are estimated with First Difference modelling including initial conditions to allow for differential trends. The analytical sample is based on a 1 to 1 nearest neighbour matching with replacement based on the Mahalanobis distance metric. The dependent variable is the change in log of yearly newly deforested area. Time varying conditions include yearly cloud coverage over remaining forest area, the log of yearly lagged district GDP per capita, and a dummy for protection status to control for the effect of reserve protection. Biophysical conditions control for initial conditions (in logarithms) of the cell size, non- forest area, size of water bodies, travel distance to the closest city, distance to the border of the multiple-user reserve, distance to the next indigenous territory, distance to the next strictly protected reserve and the forest cover in 2006. Economic conditions refer to logarithm in population density at the district level in 2007, total farm area in the district, as the share of small holder farmers and the logarithm of the average tractors and farm in the district. Reserve characteristics control for the age of the according protected area in years in 2015. Clustered standard errors are reported in parentheses. *,**,*** denote significance at the 10/5/1% level, respectively. Robustness The distribution of our dependent variable is highly skewed due to the low deforestation rate within forest reserves of the Amazon state. In the matched sample, 98 percent of all observations, across cells and years, report zero deforestation. Only 9 percent of all cells experience some deforestation during our timeframe. This could bias the treatment estimates downward, as variations in the explanatory variables lack 28 a response in large parts of the dependent variable. We deal with this in several ways (see Figure A2-1). (A) Sample properties Observations weighted by def. probability ● Observations weighted by cell size ● Min. cell size: 5.0 sqkm ● Min. cell size: 10.0 sqkm ● Min. cell size: 15.0 sqkm ● Min. cell size: 20.0 sqkm ● Matched excluding federal reserves ● −0.2 −0.1 0.0 0.1 0.2 (B) Alternative matching techniques Matching with caliper: 3.0 ● Matching with caliper: 2.5 ● Matching with caliper: 2.0 ● Matching with caliper: 1.5 ● Matching with caliper: 1.0 ● Matching w/o caliper, controls: 1 ● Matched controls: 2 ● Matched controls: 3 ● Matched controls: 4 ● Matched controls: 5 ● −0.2 −0.1 0.0 0.1 0.2 (C) Anticipated and lagged treatment effects Bolsa Floresta start in t−1 ● Bolsa Floresta start in t−2 ● Bolsa Floresta start in t−3 ● Bolsa Floresta start in t−4 ● Bolsa Floresta start in t+1 ● Bolsa Floresta start in t+2 ● Bolsa Floresta start in t+3 ● Bolsa Floresta start in t+4 ● −0.2 −0.1 0.0 0.1 0.2 Impact coefficient Bars indicate confidence intervals at a 90% level. Standard errors are clustered at the reserve level. Figure A2-1: The Bolsa Floresta effect in different specifications. We estimate a weighted Fixed Effects estimation, using weights constructed by the inverse probability of the cell experiencing some deforestation before treatment (2003-2007). Weighting the sample by probabilities gives less influence to observations that would not have been deforested in any case. Further, we weight by the size of our cell units, giving lower importance to small observational areas. This method controls 29 for the probability of observations experiencing zero deforestation simply because small areas are less probable to be affected by deforestation. In both cases, impact coefficients remain insignificant (Figure A2-1). Our database includes smaller and larger cells due to slicing reserves into a 5x5 km grid. In our preferred estimation model, we keep all irregular cells that are not fully covered by the original grid cells to avoid biases from the loss of information or misattribution and restrict the matching process to find only pairs where cells are equal in size (with a tolerance of 5 percent). To test whether this data structure drives our results, we examine whether estimates change when cells smaller than 5, 10, 15, 20 km2 are successively excluded. As can be seen in Panel A of Figure A2-1, successively excluding cells below 5 to 20 km2 has no effect on the impact estimate. Multiple-use reserves in Brazil are managed under federal, state, or municipal administration. The Bolsa Floresta program is implemented within state-administered reserves. In our preferred matching procedure, we use all administration types, to maximize the pool of potentially matched control cells. The matching procedure allows including federal administered reserves to the control sample, because after the procedure observations approximate similarity along the observed dimensions. On average federal reserves have higher deforestation rates, therefore we expect impact coefficients to fall when we exclude federal reserves from the pool of controls. Nonetheless, a bias will occur if federal reserves have sharply changed their management quality after Bolsa Floresta’s start in 2007. For example, if federal reserves improved their protection capabilities significantly after 2007, they would not serve as good controls and lead to an under-estimation of Bolsa Floresta's effects. The last row in Panel A of Figure A2-1 shows an insignificant estimate when excluding federal reserves before matching. The coefficient turns positive to 0.01, potentially suggesting an over-estimation in our preferred specification. Although matching achieved a significant improvement of the covariate balance, remaining imbalance could bias our results. We use a variety of alternative matching procedures to test for misspecifications (Panel B of Figure A2-1). We increase the stringency of 'similarity' between matched pairs. We restrict paired matches to 3.0, 2.5, 2.0, 1.5, 1.0 and 0.5 calipers of standard deviation difference in their covariate values. Impact coefficients increase and become significant at a 5-10 percent level indicating a reduction of deforestation by 4 to 8 percent due to the intervention. However, the sample size reduces considerably (15 percent-50 percent) and a reduction of 8 percent corresponds to a negligible 154 ha of avoided forest loss. Furthermore, we test our matching procedure by increasing the number of matched control units from 2 to 5 nearest neighbor pairs, which has no considerable influence on our estimate (rows 6 to 9). These results are in line with the negative impacts in the subsample of mid- range deforestation pressure. 30