Impact Evaluation of the Green Municipalities Program, and Imazon’S Support, in Pará, Brazil

Impact evaluation of the Green Municipalities Program, and Imazon’s support, in Pará, Brazil

Phase 3 Report delivered to Mercy Corps

Authors: Erin Sills, Justin Kirkpatrick, Alexander Pfaff, Luisa Young, and Rebecca Dickson April 2017 Version 1

Table of Contents Terms and acronyms ...... 3 Acknowledgements ...... 4 Executive Summary ...... 5 1.0 Introduction ...... 8 2.0 Framing the research questions ...... 9 2.1 Background and characterization of interventions (“treatments”) ...... 9 2.2 Expected impacts (“outcomes”) ...... 11 2.3 Selection into the Imazon and PMV interventions ...... 12 2.4 Questions we seek to answer ...... 15 3.0 Data ...... 16 3.1 Outcomes – Deforestation ...... 16 3.2 Outcomes ‐ Agricultural land and yields ...... 25 4. Methods ...... 26 5. Estimation and results ...... 30 5.1 Deforestation outcome ...... 30 5.1.1 Deforestation Impact of the bundled PMV + Imazon10 based on synthetic controls 31 5.1.2 Impact of the PMV estimated with SCM ...... 33 5.1.3 Reasons for changes in impact estimates ...... 34 5.2 Agricultural yield and share ...... 35 5.2.1 Two‐way fixed effects estimation of agricultural yields ...... 35 5.2.2 DiD estimation of agricultural yields in matched sample of PMV and control municipalities ...... 38 5.2.3 Impact of bundled PMV + Imazon10 on agricultural yield estimated with SCM 41 5.3 Administrative costs of the PMV ...... Error! Bookmark not defined. ANNEX 1...... 42 ANNEX 2...... 46 A1. Donor pool ...... 46 A.2 Treatment and moderator variables ...... 48 ANNEX 3...... 55 ANNEX 4...... 56

Terms and acronyms

APP: Área de Preservação Permanente, or Area of Permanent Preservation, including riparian zones, steep slopes, and hilltops designated for forest protection per the Brazilian Forest Code Blacklist: federal priority list, established in 2007, identifying municipalities with high levels of deforestation, administered by the MMA CAR: Cadastro Ambiental Rural, or Rural Environmental Cadaster DID: differences in differences Gross deforestation: conversion of original mature forest to a non‐forest land‐use, not to be confused with net deforestation which is total change in forest cover IBAMA: Brazilian Institute of Environment and Renewable Natural Resources, federal environmental agency IBGE: Instituto Brasileiro de Geografia e Estatística, or the Brazilian Institute of Geography and Statistics Imazon10: Ten municipalities receiving technical support from Imazon under the USAID/Skoll program, including Dom Eliseu, Tailândia, and Ulianópolis; Brasil Novo, Monte Alegre, Novo Progresso, Novo Repartimento, Santana do Araguaia and Santarém; and Moju (all are also participants in the PMV) INPE deforestation year t: as monitored by INPE, from August 1 of year t‐1 to July 31 of year t LULCC: Land use / land cover change MMA: Ministério do Meio Ambiente, or Ministry of the Environment MPF: Ministerio Público Federal, or Public Prosecutor’s Office PMV: Programa Municípios Verdes (GMP in English) PMV category: Categorization of municipalities in PA developed for prioritization and planning purposes, including (1) blacklisted, (2) consolidated or well developed old frontier, (3) forested or beyond the frontier, (4) deforestation monitored and under control, having exited the blacklist, and (5) under deforestation pressure or active frontier, at risk of entering the blacklist PMV on paper: Indicator that municipality has signed agreement with the federal prosecutor in Pará (MPF, or Ministério Público Federal) and/or directly with the PMV, with year of treatment defined as earlier of these two possible agreements based on INPE definition of a deforestation year (August of previous year to July of current year) (Note that we exclude Paragominas, because it received assistance from Imazon although it was not part of the Imazon10, and the 11 municipalities that joined the program after 2011, because effects are not expected within such a short time frame.) PMV implementation: Subset of municipalities in PMV on paper that by 2015 had achieved at least three of the following steps identified by key informants as most important for active engagement with the program: celebrated a pact against deforestation, created a working group to combat deforestation, registered at least the median percent of land in the CAR, established at least minimum structure for municipal environmental governance PA: State of Pará Prefeitura/prefeito: municipal government/ mayor PSDB: Partido da Social Democracia Brasileira, political party of current governor of Pará SCM: Synthetic control methodology

Acknowledgements

We are grateful for the continued collaboration and insights shared by Amintas Brandão and Carlos Souza of Imazon. Luíza Andrade was a key collaborator on the analysis of socioeconomic impacts. The budget data were compiled by Bruno Marianno, who also shared his detailed knowledge of how these programs function.

Executive Summary

This report documents an impact evaluation of two programs working in concert at the municipal level in the Brazilian state of Pará: (i) Programa Municípios Verdes (PMV); and (ii) an initiative by Imazon to provide technical assistance to 10 municipalities within the PMV (Imazon10). These programs were designed to help municipalities respond to a federal mandate to reduce deforestation, as implemented in part through a ‘blacklist’ of municipalities responsible for the most deforestation that was maintained by the federal government from 2007 to 2013. In this context, municipalities may have selected into the PMV and Imazon10 with the goal of reducing penalties under such mandates, getting off the blacklist, and/or increasing local economic output conditional on meeting the mandate.

The two programs (or “treatments”) are best considered as a bundle of complementary and interdependent activities, with the Imazon10 serving to assist municipalities in their participation in the PMV. In addition to Imazon, the PMV involves other civil society organizations and federal and state agencies, resulting in both diffuse costs and diffuse impacts. The official “on‐budget” cost of the program as administered by the government of Pará was roughly one million dollars per year in 2015 and 2016 (see Annex 5), including the cost of program administration through the governor’s office (casa civil) and through the program office (Núcleo Executor do Programa Municípios Verdes) jointly funded by the state and a grant from Fundo Amazônia. Both the PMV and Imazon’s support are on‐going and expanding: municipalities have joined the PMV as recently as March 2017, and Imazon has expanded their technical assistance to 40 more municipalities.

In this report, we use the same master sample of 519 municipalities in the Brazilian Amazon as in previous evaluations (in this case excluding the original green municipality that we considered in an earlier analysis, Paragominas, because it received technical assistance from Imazon similar to but separate from the program that we are evaluating). In 2015 ‐ 2016, Imazon expanded its technical assistance to an additional 40 municipalities and the state of Mato Grosso launched a similar program of sustainable municipalities. In future analyses of deforestation that include 2016, we will identify and exclude these municipalities from the ‘donor pool’ of comparison municipalities. For our master sample of municipalities, we use our previous database of geographic, biophysical, political, and economic factors (covariates) believed to influence both participation in the PMV and deforestation (obtained primarily from Brazilian government sources such as the Instituto Brasileiro de Geografia e Estatística (IBGE)). We have expanded this database with new information on socioeconomic outcomes, revised estimates of deforestation in 2013 and 2014 from the Instituto Nacional de Pesquisas Espacials (INPE), and new estimates of deforestation in 2015 from INPE. That is, we report impacts on deforestation through July 31, 2015. The revised estimates reflect both improved geo‐rectification of images and deforestation revealed in 2015 in areas that had been under cloud‐cover in 2014 and 2013. While we did not expect those changes to be systematically related to the selection into our treatment or control groups, these new data have affected our findings substantively.

The selection bias inherent to intentionally‐targeted and voluntary programs such as the PMV and Imazon10 can confound estimates of impacts, unless counterfactual outcomes are estimated using quasi‐experimental methods. In our previous report, we argued that the synthetic control method (SCM) is an appropriate way to estimate program impacts on deforestation, given the small sample size of the Imazon10 program, availability of a long time‐series of data on deforestation outcomes,

5 and the possibility of heterogenous impacts across municipalities. Thus, we applied SCM to our database of historical deforestation rates and characteristics of municipalities in order to identify weighted combinations of municipalities (the ‘synthetic control’) that best represent what would have happened had the PMV or Imazon10 not been implemented. We assessed the validity of the synthetic controls based on the match between deforestation in the treated (with PMV or Imazon10) and synthetic control (without PMV or Imazon10) during the pre‐treatment calibration period (from 2001 to beginning of treatment in 2011, 2012 or 2013). We do not make claims about results for municipalities for which we could not obtain a good match during that time period. The municipalities in the synthetic control are drawn from donor pools that exclude all other municipalities in the program being evaluated (either PMV or bundled PMV + Imazon10) and include only municipalities in the same blacklist status (e.g., for black‐listed treated municipalities, the donor pool for the counterfactual was restricted to similar municipalities that were blacklisted during the 2007 – 2013 period).

We estimate the impact of treatment as the difference between deforestation in a treated municipality and its matched synthetic control in 2013, 2014, and 2015, using new data from INPE. To assess the significance of these estimated impacts, we conduct “placebo tests” to establish the range of false “effects” estimated when un‐treated municipalities are evaluated as if they were treated municipalities. Any estimated actual effects that are above the 95th or below the 5th percentile of the distribution of these false effects estimates are interpreted to be statistically significant. In municipalities where deforestation rates are already low within this region (i.e., municipalities not on the blacklist), any possible reduction in deforestation is likely to fall within this distribution and thus not be statistically distinguishable from zero. For example, this is the case with the two municipalities in the Imazon10 that were not blacklisted (Monte Alegre and Santarem).

Of the ten municipalities receiving support from Imazon, eight were on the federal blacklist. As reported previously, we were able to construct good quality synthetic controls for five of them (Ulianopolis, Dom Eliseu, Tailandia, Novo Progresso and Santana do Araguaia). Using the deforestation data previously released by INPE, we found that this bundled PMV + Imazon10 treatment reduced deforestation (as a percent of the municipal area) by an average of 0.3% or 21 km2 in 2014. The newly released deforestation data generally show more deforestation in these treated municipalities but similar levels of deforestation in the municipalities in the synthetic control. As a result, we no longer find any statistically significant impact of the bundled PMV + Imazon10 treatment on deforestation in 2013 or 2014. The only partial exception is Santana do Araguaia, where we estimate increasing effect sizes from 2013 to 2015. In 2015, we estimate that the bundled PMV + Imazon10 treatment reduced deforestation (as a percent of the municipal area) by 0.0035%, which was more than 90% of the placebos. However, this estimate is less than a square kilometer, orders of magnitude smaller than the previous estimate, and not statistically different from zero at the 90% level.

In our previous report, we said that in most blacklisted municipalities, both the bundled PMV+Imazon10 and the PMV alone significantly reduced deforestation in 2014 but not in 2013. Furthermore, we found that there were most likely to be significant impacts in municipalities that started treatment earliest (e.g., signed on to the PMV in 2010, the first year of the program). We suggested this could mean that impacts were increasing over time. However, since our last report, political and economic challenges have effectively distracted the Brazilian federal government from efforts to control deforestation. The threat of blacklisting (or the incentive of getting off the 6 blacklist) no longer appears to be a significant concern for municipal governments in the Amazon. Without this external “stick,” it is not surprising that we find that the PMV does not have an effect on deforestation in most participating municipalities in 2015. It is quite possible that such a program can continue to exist in name but shift in implementation if the external pressure to which it was a rational response is perceived to be considerably reduced. For that matter, it is quite possible that by 2014 municipal governments perceived decreasing pressure to reduce deforestation, helping explain why we now find – with updated deforestation data ‐‐ that the PMV did not have a significant effect in any year in most municipalities.

The exception is Cumaru do Norte, where the effect of the PMV grew over time from 2013 to 2015, becoming statistically significant at the 90% level in 2015. Compared to other blacklisted municipalities in the PMV (16, of which 11 have acceptable synthetic controls), Cumaru do Norte ‐ like Santana do Araguaia ‐ is a large municipality (thus making the blacklist criterion of 40 km2 of deforestation more binding) that joined the PMV early (in 2010), had established a societal pact against deforestation by 2016, and now has more than 80% of eligible land registered in the rural environmental cadaster (more than 85% in Santana do Araguaia). (Both municipalities would be defined as implementing the PMV in practice based on current information, although in 2015, we classified Cumaru do Norte as participating in the PMV on paper only.) This suggests at least anecdotally that the program is more likely to be effective when locally embraced and more fully implemented.

While the stated goal of the PMV (and Imazon’s efforts in support of the PMV) was to reduce deforestation, consistent with the rational responses of reducing penalties and getting off the blacklist, municipalities may also have joined the PMV or Imazon10 to reduce the costs of meeting federal mandates. That may, in turn, help to sustain the reduction in deforestation over the long run. We focus on the economic costs of reducing deforestation, which these programs seek to mitigate by encouraging shifts from extensive production on newly deforested land to more intensive production on previously deforested land. Thus, if federal rules seek to close the deforestation frontier, the PMV+Imazon10 may make that more palatable by increasing the profitability of agricultural production in areas already deforested.

To explore this, we examine shares of municipal land allocated to cattle pasture, annual crops and perennial crops, as well as the intensity of their production as measured by pasture stocking ratio and indices of crop yields. While much of the discussion about deforestation has focused on cattle and soybeans, we do not find any effects of the PMV or Imazon10 on shares or yields of cattle or annual crops. We do find that participation in the PMV increased both the share of total area allocated to perennial crops, as well as the yield of those crops. This provides some preliminary evidence that the PMV could shift municipalities to a more sustainable economic trajectory based on more stable and higher‐value land use, which could in turn reduce future deforestation. For our next report, we will also examine the impact of the PMV and Imazon10 on overall economic production in a municipality, as measured by value‐added tax receipts. We have requested these data from each state in the Amazon through procedures established by the Brazilian Access to Information Act.

1.0 Introduction

The Programa Municípios Verdes (PMV) in the Brazilian state of Pará (PA) seeks to engage municipal governments in reducing deforestation and developing economic alternatives that do not require deforestation, thereby helping municipalities to comply with national climate change mitigation policies while reducing the economic cost of compliance (in the sense of improved economic outcomes compared to a scenario of simply halting the activities that require deforestation).

As part of its support for this state‐wide program, Imazon provides additional technical assistance to 50 participating municipalities. Imazon piloted this approach starting in 2011‐2012 in 10 municipalities (Imazon10) selected to represent the range of conditions influencing deforestation in PA.

PA was the first Brazilian state with a program like the PMV, and its experience has generated interest in developing similar programs in other Amazonian states, including MT (Mato Grosso) and AM (Amazonas). In order to inform decisions about whether and where to scale‐up this type of intervention in other states, we need a better understanding of its causal effects, or impacts. While reducing gross deforestation rates (loss of mature forest) was the immediate objective of the PMV administration and Imazon, these programs ultimately aim to generate sustainable economic development while stabilizing forest cover. From the municipalities’ perspective, the goal is to reduce the probability of fines, restrictions on market access, or other penalties for deforestation imposed by the federal government. Thus, we both (1) analyze program impacts on gross deforestation, represented as the percent of the municipality deforested (as measured by INPE from August of previous year to July of current year), and (2) search for evidence of any resulting shifts in economic activities that would indicate progress towards an economic development path compatible with forest conservation.

The challenges to evaluating impacts of the PMV and Imazon10 include: 1) municipalities were selected into these programs based on factors that also influence their future deforestation rates and associated economic activity, 2) the number of treated units in the Imazon10 is very small, limiting the statistical power of most quasi‐experimental methods, 3) while the program goals are long‐term, many participating municipalities have been part of the PMV or collaborating with Imazon for five years or less, and 4) the programs have been implemented against a backdrop of highly variable federal efforts to reduce deforestation, which likely generated highly variable – and unobserved ‐ expectations about the costs and benefits of reducing deforestation.

To overcome these challenges, we use matching methods to control for self‐selection, synthetic control matching and placebo tests to address the small sample size of Imazon10, a range of proxy indicators to assess progress towards sustainable economic pathways, and either a donor pool or covariate based on each municipality’s status on the federal blacklist to control for the background level of pressure by the federal government.

2.0 Framing the research questions

2.1 Background and characterization of interventions (“treatments”)

In 2008, Imazon began collaborating with the municipality of Paragominas to support efforts to improve municipal environmental governance, develop economic alternatives compatible with forest conservation, and reduce deforestation in order to get off the federal “blacklist.” The blacklist, or priority list of the Ministério do Meio Ambiente, or Ministry of the Environment (MMA), had been established by the federal government in 2007 in order to focus enforcement efforts, restrict market access, and encourage better governance in municipalities where the most deforestation was occurring. A total of 48 municipalities were blacklisted between 2007 and 2013 (Figure 1.1). Although the blacklist officially still exists, the last decree adding or removing municipalities from the list was issued in 2013.

Figure 1.1: Counts of municipalities considered in this analysis (excluding Paragominas and municipalities that joined the PMV between August 2011 and March 2017)

72 8 32 PMV Blacklist

2 0

Imazon10

In 2010, Paragominas became the first municipality to be taken off the federal blacklist. This accomplishment captured the attention of other municipalities in the region as well as the state government. As a result, Imazon sought funding to work with additional municipalities, obtaining a grant from the Fundo Amazônia (Amazon Fund) that allowed them to expand such collaboration to other municipalities in the same region in 2011. Subsequently, the governor established a special secretariat for a state‐wide green municipalities program (Programa Municípios Verdes, hereafter “PMV”), also in 2011.

In March of 2011, the state government formally launched the PMV with the two official goals of promoting sustainable economic activities and reducing deforestation in the state. Imazon closely collaborates with the state government on this program, participating in the management council and the “environmental observatory” that tracks progress of reducing deforestation and expanding

9 registration in the CAR (rural environmental cadaster), and providing deforestation alerts that the PMV passes on to the municipal governments. Initially, municipalities could join the PMV by signing an agreement either with the PMV or with the MPF (specifically, the federal prosecutor from the Ministério Público Federal assigned to PA). When the PMV was launched, the federal prosecutor assigned to PA was actively working towards more effective implementation of laws restricting deforestation. However, both the level of activity and the influence of the MPF in PA have declined, reflecting an overall decline in federal efforts to slow deforestation.

When the PMV was established, it immediately incorporated 39 municipalities that had already signed agreements with the MPF in 2010. By July 2011 (the end of the 2011 INPE deforestation year), 91 municipalities (including Paragominas) had officially joined the PMV (Figure 1.2). By July 2015 (the end of the 2015 INPE deforestation year), another 15 municipalities had joined. All but one of these 15 municipalities were in either the forested zone (beyond the frontier, without much deforestation) or the consolidated zone (the old frontier, without much remaining forest cover). They signed agreements with the PMV in order to demonstrate support for the program, e.g. through initiatives to improve their urban environment. For our analysis, we consider municipalities that joined the program by 1 August 2011 as part of “PMV on paper,” defined as a municipal government (or prefeitura, referred to in this report as the municipality) signing up for the PMV (based on government records of the dates that municipalities joined the PMV). We do not consider the fifteen municipalities that have joined the program since that time, because the intervention has been applied too recently to expect impacts, and because they are also not valid ‘controls’ for estimating the counterfactual outcomes of municipalities that joined by August 2011. Likewise, we exclude Paragominas, because it neither underwent the same treatment as the rest of the Imazon10 nor serves as a valid control (see Figure 1.1).

Figure 1.2: Number of municipalities that joined the PMV in each year

Number of municipalities joining the PMV 60

0 2010 2011 2012 2013 2014 2015 2016 2017

When they joined the PMV, the municipalities committed to a series of specific actions (or steps), but they have made variable progress on implementing these steps, as tracked by the PMV in their on‐line database. In order to identify the steps most indicative of active engagement with and implementation of the PMV, we elicited input from key informants including researchers at 10

Imazon, consultants working with the PMV, municipal secretaries of the environment, and representatives of NGOs that collaborate with the PMV. Based on 11 responses, we identified four steps recorded in PMV’s database as the best indicators that a municipality is implementing the program. Thus, we defined “PMV implementation” as having completed at least three out of the following four steps by 2015: celebrated a pact against deforestation, created a working group to combat deforestation, registered land in the CAR (specifically in the top half of municipalities in terms of percent of land registered), established at least a minimum structure for municipal environmental governance.1 In addition, in this report, we interpret our results in light of steps accomplished by the municipalities through March 2017 (Annex 1), but we do not re‐define “PMV Implementation” because of changes in how information is elicited for the PMV database.

Imazon, with funding from USAID‐Skoll, is partnering with three of the municipalities where they established technical cooperation agreements in 2011 (Don Eliseu, Tailândia, and Ulianópolis); six municipalities that signed technical cooperation agreements with Imazon by the end of 2012; Mojú, which replaced one of the originally selected municipalities that dropped out of the program in 2013; 37 municipalities in the PMV that signed agreements with Imazon in 2016; and three municipalities that joined the PMV and started working with Imazon simultaneously in 2016. We focus on the original 10 that started working with Imazon in 2011 and 2012 (plus the replacement municipality Moju) in order to allow a sufficient number of years to observe impact (through 2015). These “Imazon10” municipalities are all participants in the PMV and have accomplished at least three of the steps identified as the best indicators of implementation of the program. Imazon’s objective with the Imazon10 is essentially to replicate the “Paragominas model”, including developing detailed maps and municipal government capacity in order to facilitate the CAR and eventually environmental licensing. For actors in other states considering whether to pursue this type of program (i.e., the target audience for scaling up), the most relevant question is the impact of implementing the PMV with the level of support provided by Imazon. Thus, we define one “treatment” of interest as joining the PMV with support from Imazon, or the “bundled” effect of the PMV and Imazon10 interventions. We consider years through 2010 as the pre‐treatment calibration period, and years starting in 2013 as treatment, with variation in how we consider 2011 and 2012 across the different municipalities, given their staggered entry into the program.

2.2 Expected impacts (“outcomes”) PMV and Imazon agree on both the short‐term goal of reducing gross deforestation and the long‐ term goal of economic development with zero net deforestation in Pará (by 2020, according to Imazon’s log frame for the USAID‐Skoll project). Since we are evaluating program impact over only a few years, we first assessed impacts on gross deforestation: area of original mature forest in the municipality converted to other land uses. This does not distinguish illegal from legal deforestation (e.g., clearing up to 20% of private properties in municipalities not on the blacklist is

1 In Portuguese: 1. Possuir Acordo de Cooperação específico com o PMV 2. Pacto contra desmatamento celebrado 3. Grupo de combate ao desmatamento 4. % propriedades cadastradas no CAR 5. Habilitação para gestão ambiental municipal

11 legal), nor consider forest degradation, nor count plantations, restoration, or forest regeneration aas forest.

We further consider whether the Imazon10 and PMV have different impacts in areas under different land tenure, specifically considering deforestation outcomes in (i) the land eligible for registration in the CAR, and (ii) the land under municipal control. The areas evaluated exclude federal protected areas, indigenous resources, and land reform settlements. We hypothesize that the effects of PMV are likely to be different in areas where the CAR is actively being implemented and areas where the municipal government has more direct control over land use, as compared to federally owned areas that may be subject to illegal activity more difficult for the municipal government to control. However, it is also possible that the PMV or the Imazon10 are more effective in federal areas, as a result of better communication and coordination between different levels of government.

In the municipality as a whole or in these sub‐areas, annual gross deforestation can be represented by the (i) percent of the remaining original forest deforested, (ii) area deforested (in square kilometers), or (iii) share of the municipality deforested (square kilometers deforested divided by square kilometers of the municipality). The third option is our preferred outcome measure for the impact evaluation, because it reflects the primary objective of the program, requires relatively little data manipulation (annual deforestation in square kilometers derived from shape files provided by INPE divided by constant area of municipality in square kilometers), and can be summed across years to calculate cumulative impacts. Nearly all of the municipalities in our sample were originally 100% covered in forest, and thus gross deforestation as a percent of the municipality also represents the percent of the original forest cover lost in a given year.

While the PMV was created in response to the federal goal of reducing deforestation, the goal of the participating municipalities may be better characterized as maximization of income or economic activity conditional on compliance with federal rules. Compliance in turn is likely to further those economic goals because it is expected to lead to removal of federal constraints on access to markets. Thus, we also look for evidence of progress towards the long‐term goal of economic development compatible with forest conservation. Ultimately, we would like to estimate the impact of the PMV on total economic activity in a municipality, thus capturing any shifts across sectors (e.g. away from land‐based activities). We have determined that for purposes of relating annual changes in economic activity to the PMV, the best available data are tax receipts, described in more detail below. In this report, we estimate impacts on the share of municipal land allocated to different agricultural activities and the yield r(o stocking ratio as a proxy for yield) of those activities. Increasing yields is often said to be a prerequisite to achieving economic development while limiting deforestation, and thus increasing yields (or stocking ratio) may be an indicator of progress towards this goal.

2.3 Selection into the Imazon and PMV interventions

It is important to place the PMV and Imazon’s efforts in the context of changes in federal deforestation policy over the past decade in Brazil. In 2007, a presidential decree established the legal basis for identifying priority municipalities for preventing, monitoring, and controlling deforestation, including increased enforcement effort by IBAMA (Brazilian Institute of Environment and Renewable Natural Resources, federal environmental agency), restrictions on 12 commodity sales, and stricter requirements for land registration. In 2008, another presidential decree specified the administrative processes for dealing with environmental crimes, increasing the clarity, speed, and likelihood of punishment through fines, seizure of production inputs, and/or embargoes, including the publication on the internet of a list of embargoed properties. Further, transporting, processing, trading or commercializing products from embargoed areas also become illegal. During 2009, the MPF reached various agreements with industry, producers, and state government in Pará requiring them to stop trade from embargoed areas and implement an environmental licensing system within 12 months. In the same year, Pará created a provisional version of the CAR, which allowed landowners to register before proving definitive title or having a field check of their APPs (Área de Preservação Permanente, or Area of Permanent Preservation). However, the municipalities were not able to meet the MPF deadline for an environmental licensing system, and so in 2010 and 2011, the MPF signed another set of agreements with municipalities in Pará, as well as the state government, IBAMA, and producers represented by FAEPA (Federação da Agricultura do Estado do Pará). These agreements postponed the requirement for environmental licensing, conditional on the municipalities implementing a series of actions to combat deforestation, which in 2011 officially became the steps required to implement the PMV.

All municipalities in the state of PA are eligible to join the PMV. The administration of the program categorizes municipalities according to whether they are on the federal blacklist, previously but no longer on the blacklist (with deforestation “monitored and under control”), under strong deforestation pressure, already largely deforested (“consolidated”), or forested. They prioritize and actively encourage participation by municipalities that are on the federal blacklist and those under strong deforestation pressure. These municipalities, as well as the municipalities that have accomplished all of the goals ofV, the PM receive priority for program support such as equipment “kits” (vehicle, computers, GPS units, and cameras) financed through Fundo Amazônia.

While the program has set priorities for membership, it is actually the individual municipalities that decide whether to join the program. At least until 2013, the federal blacklist was key to participation decisions, because municipalities hoped that the PMV would help them either avoid or get off the blacklist, thereby ensuring access to credit and markets for their agricultural producers. This motivation was strengthened by the connection between the PMV and the MPF, specifically the federal prosecutor’s office in PA. That office had demonstrated that it could effectively cut off credit and market access (in particular access to large beef and soy processors who not only purchase outputs but also finance production) for producers in municipalities not complying with agreements to improve municipal environmental governance. In the past several years, both the blacklist and the MPF have declined in influence, but the PMV has been successful at obtaining new funding from Fundo Amazônia and has thus been able to offer new benefits to participating municipalities, including the equipment kits and more hands‐on support from regional PMV offices (núcleos).

By joining the PMV, municipalities commit to developing a social pact to reduce deforestation, establishing a working group, monitoring deforestation, registering 80% of the municipality (outside of protected areas) in the CAR, and staying off the blacklist (by maintaining deforestation below 40 square kilometers per year, regardless of the size of the municipality). These goals were initially reinforced by the MPF, which declared that failure to reach the goals would nullify the extension of the deadline for environmental licensing. In exchange, the MPF agreed to take all possible measures to guarantee participating municipalities access to the government agencies 13 necessary to register and license producers; to work with INCRA (Instituto Nacional de Colonização e Reforma Agrária) to ensure that producers who comply also are able to register their land in the CCIR (Certificado de Cadastro de Imóvel Rural) system; to work with financial institutions to guarantee producers access to financing and to assist the municipalities with financing infrastructure; and to work with the municipalities to help generate benefits in compensation for forest conservation. The federal environmental agency, IBAMA, also agreed not to embargo areas in municipalities that were making progress towards meeting these requirements.

The PMV and Imazon directly helped municipalities meet the requirements for getting off the blacklist, for example, by facilitating the process of CAR registration, by helping establish GIS capacity in the municipalities, encouraging NGOs to develop comprehensive base maps, and contracting firms to help smallholders register in the CAR. In its initial years, the PMV accomplished this largely by coordinating the actions and resource allocations of other state and federal agencies and NGOs, including both Imazon and others such as The Nature Conservancy (TNC). Since July 2014, the PMV has also administered its own budget, including a large grant from Fundo Amazônia. The PMV has developed advertising campaigns to encourage registration in the CAR, and supported establishment of a new legal mechanism (COTP or Certificado de Ocupação de Terra Pública) to certify that land owners are in the process of obtaining title to state land, based on registration in the CAR. Further, PMV provides a communication channel between municipalities and both the state police and IBAMA, who can in turn obtain support from the federal police for enforcement actions. Other municipalities in the forested and consolidated zones have a variety of motives for joining the PMV, including offering political support to the governor (if the prefeito and governor are from the same political party), obtaining technical assistance and supplies from the PMV (e.g. “kits” for monitoring deforestation, including motorcycles financed through a World Bank loan to the Pará Rural program, and vehicles financed by Fundo Amazônia), attaining priority status for state assistance (e.g. from SEPOF), and receiving a greater proportion of the state’s revenues from value added taxes (beginning in 2015).

Thus, the PMV offers a mix of incentives to participate, some appealing directly to municipalities (access to “kits” of equipment, higher VAT (value added tax) allocation, and support for the governor when they are from the same political party) and some more appealing to agricultural producers (access to markets for beef and soy through CAR registration, liberation of subsidized agricultural credit from the government, provisional land titles). Partnership with Imazon basically provides technical assistance that allows more rapid progress towards the PMV’s goals, including building the municipalities’ capacity to take responsibility for environmental licensing, which in turn benefits producers who will not need to travel as far or wait as long for licenses. While PMV and Imazon tout the multiple benefits of the program to municipalities, there are clearly also costs, both on the administrative side and in terms of the opportunity cost to producers of not deforesting. In blacklisted municipalities, all deforestation is already illegal, and hence these opportunity costs may not be a legitimate part of social cost accounting. Nonetheless, the demand for deforested land and the cost of foregoing that deforestation are likely to influence municipal decisions whether to join the PMV and/or partner with Imazon.

All municipalities in the PMV are supported by Imazon through its participation in the overall program, for example, providing monthly deforestation alerts, pursuing clarification on the division of environmental responsibilities between state and municipal governments, and setting standards for municipal environmental agencies. However, Imazon has selected 50 municipalities for more intensive support, seeking to replicate the “Paragominas model.” We focus on the first 10 14 municipalities to partner with Imazon, including eight on the blacklist who Imazon selected because they (1) had a large area of forest remaining, and (2) had demonstrated interest in reducing deforestation (e.g., by signing an agreement with the MPF or creating a “social pact” with stakeholders in the municipality to decrease deforestation). Imazon also partnered with two other municipalities (Monte Alegre and Santarém), in order to gain experience working with municipalities from different regions of the state facing different deforestation pressures (i.e., speculation vs. colonization vs. expansion of commercial ranching and agriculture). For example, they picked Monte Alegre because it faces the most deforestation pressure in the Calha Norte region, although it does not have a high deforestation rate relative to the rest of the state and was not on the federal blacklist. Imazon had offered similar support to the municipality of Paragominas starting in 2008, and then expanded to nearby municipalities hwit financing from Fundo Amazônia. Several of these nearby municipalities were also selected for the USAID/Skoll project, although Paragominas is not among the initial ten selected (and is therefore excluded from our sample for this analysis). Some of the municipalities that Imazon originally identified as potential participants in the program ultimately decided not to partner with Imazon. This decision of the municipality is likely related to both their level of interest in controlling deforestation and their support for the PMV, which is closely identified with both Imazon and the governor of PA (from the PSDB political party). Thus, both Imazon’s selection rule and the municipalities’ choices about whether to participate are inherently – although not solely ‐ related to deforestation and associated economic activity. This creates a fundamental challenge for identifying the causal impact of these interventions upon rates of deforestation.

2.4 Questions we seek to answer

Based on consultations with Imazon, PMV, Mercy Corps, Skoll and USAID, we identified the following questions that are relevant for scaling programs to other states in the Brazilian Amazon and that reflect the inter‐dependence of the Imazon10, PMV and federal blacklist: 1. What has been the impact on deforestation through ,July 31 2015 of joining the PMV by July 2011 and receiving additional technical cooperation and support from Imazon? This bundled “PMV + Imazon10” intervention was fully implemented starting in 2011, 2012, or 2013 in the ten pilot municipalities. 2. What has been the impact on deforestation through July 31, 2015 of joining PMV by July 2011 without additional support of the Imazon10 intervention? In other words, what has been the effect of a local government committing to a low‐deforestation development path by signing agreements with state and federal governments (called “PMV”)? 3. What has been the impact on deforestation through July 31, 2015 of joining PMV by signing a commitment with the MPF or PMV no later than July 2011 and implementing a core set of actions required by the PMV by 2015? All of the municipalities in the Imazon10 and all but one of the blacklisted municipalities in the PMV had completed the core actions by the end of 2014. We refer to this intervention as “PMV implementation.” 4. What has been the impact on the productivity of annual crops, perennial crops, and cattle pasture (measured as yield and stocking ratio) of the PMV alone and the bundled PMV + Imazon10?

5. For future analysis (pending data availability), what has been the impact on economic activity, as measured by tax receipts, of the PMV alone and the bundled PMV + Imazon10?

3.0 Data We compiled a rich panel data set on a large pool of municipalities. In this section, we focus on our outcome variables, which are revised or new for this analysis. Our treatment and moderator variables (covariates), and our donor pool ‐ or the sample of municipalities used to construct the counterfactual – ehav not changed from previous analyses (Annex 1). 3.1 Outcomes – Deforestation All deforestation data reported are from the INPE PRODES project. PRODES is the source of official deforestation information used by the Brazilian government, e.g. to determine which municipalities to include in the blacklist and to meet reporting obligations to donors of the Fundo Amazônia. Although the minimum mapping area of PRODES data is 6.25 hectares and higher resolution deforestation data are available, we use PRODES data to maintain consistency with officially reported deforestation rates. The PRODES record of deforestation begins in 1988, and spatially explicit information is available annually from 2000. Some of the early years of data suffer from poor image availability or quality because of clouds, but for most of the region annual data are available from 2000. Additional details about the PRODES project are available from INPE.2

For the training period used to construct the synthetic controls, PRODES deforestation data were extracted from shapefiles provided by Imazon for 2000‐2012. We downloaded both previous and updated data directly from the PRODES website for 2013 and 2014, as well as data for 2015. Data were processed independently for each state in ArcGIS and then combined after extraction in MS Excel. To manage for variability in perceived deforestation rates caused by cloud cover (i.e. due to non‐continuous coverage), we used PRODES attribute data to identify deforestation polygons with previous cloud cover. We then distributed the area deforested evenly across the time period when clouds were present, following the approach developed by Imazon. Deforestation polygons coded by INPE as ʹDSF_ANTʺ in the states of Amapá and Maranhão were treated as deforestation prior to 2000, following consultation with Imazon.3 Finally, annual deforestation polygons were overlain on 2013 municipal boundaries.

PRODES also provides estimates of deforestation by municipality, and those are highly correlated with our estimates (e.g. correlation coefficient of 99.9% in 2013). We work directly with the shape files of deforestation provided by PRODES so that we can calculate deforestation using Imazon’s recommended methods for distributing deforestation in areas previously under cloud‐cover and so that we can identify whether the deforestation occurred in areas eligible for the CAR by removing deforestation in protected areas (state and federal conservation units, and indigenous territories)4.

2 http://www.obt.inpe.br/prodes/index.php. 3 http://www.dpi.inpe.br/gilberto/teses/dissertacao_mario.pdf 4 We also remove urban areas from the calculation of the area eligible for CAR, although this has negligible effect on our measure of deforestation in CAR‐eligible areas. In our previous report, we also presented a third measure of deforestation in the area under control of the municipality, which excluded agrarian reform settlements, military areas, and other areas that the federal government claims (glebas federais). However, because the CAR is applicable in agrarian reform settlements, we only have access to the boundaries of military areas in Pará, and 16

Starting with the 2015 release, PRODES made significant changes in the deforestation data. Specifically, they improved the geolocation processing, or the way that the location of the satellite imagery is related to a ground location. Although this correction is an improvement in the locational accuracy of the data, it creates challenges for analysis of historical deforestation. In order to better explain the challenges of working with these new data, we discuss the underlying remote sensing process, the changes in that process, and the implications for the deforestation data.

Geo‐rectification Geo‐rectification, sometimes called geo‐referencing or geo‐location, is the process of matching imagery precisely to ground locations. The ability to accurately relate satellite or aerial imagery to ground locations has improved greatly since PRODES began. The change in their methodology reflects changes in the industry standards due to rapid technological progress over the past two decades. Initially geo‐rectification was an onerous process that required technicians to match locations between paper, scanned or digitized topographic maps at the 1: 100,000 or 1: 250,000 scale. These “control points” were used to warp and stretch the image to match the spatial dimensions of the base topographic map. At the PRODES project, this was the method used until 2000. Between 2000 and 2005 each image was geo‐referenced or “spatially matched” to the image from the previous year. In 2000, NASA sponsored the creation of a geo‐rectified Landsat 7 mosaic of all of Earth’s landmasses (with the exception of Antarctica) as part of the GeoCover 2000 project. By 2005, PRODES was using this GeoCover data as the base layer for all geo‐referencing. This improved the spatial accuracy of all layers. All Landsat images available from INPE have been “pre‐processed” using NASA’s official ground control points, which has improved, but not always perfected, the spatial accuracy of images acquired through USGS (the clearinghouse of all Landsat imagery). After 2013, with the launch of the Landsat 8 OLI satellite, the imagery released by USGS are much more precisely geo‐rectified.

Geo‐rectification correction In 2015 INPE improved the geo‐rectification of the PRODES data based on Landsat 8 imagery from the year 2013. INPE estimated errors in both X and Y of up to 240 meters between the PRODES shapefiles and the Landsat 8 images. To correct this error, INPE first calculated the dislocation factor between each Landsat 8 scene and the PRODES data and later applied this factor as described in an Official INPE Methodological Note5. What this means practically is that the newly released 2015 data do not align correctly with the previously released data (Figure 3.1.1).

the tenure status of glebas federais is ambiguous, we choose to report impacts only on the second measure of deforestation in areas eligible for the CAR. 5 Details of the processing and the shift for each Landsat Path/Row are available in this report: http://www.obt.inpe.br/Prodes/NT_deslocamentoMascara.pdf

Figure 3.1.1. The image on the left is from the border of Sao Felix Do Xingu and Altamira. It shows a deforestation polygon from the 2015 Prodes mosaic data (red) and the previous dataset (green). The change in the area of deforestation is different for each of the datasets based on the shift factor.

In addition to the new geo‐rectification, PRODES combined all deforestation through 2012 into one unique class in the 2015 data release. This means that the 2015 version of PRODES data has the following deforestation classes: (i) cumulative deforestation up to 2012 in one class; and (ii) annual deforestations increments for 2013, 2014, and 2015 separately. We explored the potential of repairing the original data from 1997 to 2012 to maintain consistency and resolution with the historical data, but determined that this would require substantial effort outside of the scope of this analysis. This means that we do not have the ESPROD 2015 data release (with the more accurate geo‐rectification) for our calibration period and therefore cannot use those data to re‐construct synthetic controls. There are also implications for how we handle deforestation that appears in areas that had been under cloud cover. Using the previous data series, we distributed this deforestation across all years that the area had been under cloud cover. For example, if an area were observed to be forested in 2009, under cloud cover in 2010, 2011, and 2012, and deforested in 2013, we assigned equal probability to that deforestation occurring in 2010, 2011, 2012, or 2013. With the new data series, we can only distribute deforestation that appears in 2013, 2014, or 2015 back as far as 2013, because only the previous data release is available for earlier years. This results in higher estimated deforestation in those years in the new data.

Changes in deforestation estimates To our surprise, we also discovered that the total deforestation reported over the entire time series and for the entire Amazon has changed. Figure 3.1.2 portrays the change by municipality, clearly demonstrating that it cannot be due solely to changing boundaries of the municipalities with the new geo‐rectification.

When we compared the total deforestation mapped up to 2014 by Prodes 2014 (745,819.2 sq km) with that mapped by Prodes 2015 (763,517.8 sq km) for the same period, we found a difference of 2% (17,698.6 sq km.). Figure 3.1.2 shows the total change in deforestation per municipality for the full time period from 1997 ‐2014, while Figure 3.1.3 highlights the municipalities where total deforestation prior to 2013 reported by PRODES changed by 100 sq km or more. Both maps show that increases in estimated deforestation are concentrated in Pará.

Figure 3.1.2. Total change by municipality 1997‐2014.

Figure 3.1.3. Municipalities with > 100 sq km of change up to 2013. 12 were outside the Amazon forest biome boundary (red), 15 inside (green) and 4 on the boundary line (yellow). The biggest change is in the Codo municipality (665.8 sq.km), outside the Amazon forest biome boundary.

There is no official explanation for the substantially higher deforestation reported in the new data from PRODES, but there appear to be multiple contributing factors including increased imagery analysis and new automated techniques for detection and delineation of deforestation polygons, which have resulted in an overall increase in the amount of historical deforestation reported.

Below is an image of all deforestation reported in both data sets. The data used in previous reports are labeled “old” data while the 2015, geo‐rectified data are “new.” In figure 3.1.4, the old data (in grey) covers the new data (in red) where there is coarse alignment. This shows the multiple regions where previously unreported deforestation has been detected in the new data.

Figure 3.1.4. Map of the Brazilian Amazon comparing the two deforestation data sets. Red areas show where deforestation from 2014 or earlier is reported in the new data but was previously reported ein th old data.

While there is no documentation, we speculated that some areas PRODES may have analyzed more historical imagery in order to reduce problems of cloud contamination. The northeastern Amazon (Amapá and Pará) is often clouded, and has a large amount of deforestation newly reported in the historical imagery. See figure 3.1.5 below showing some very large areas of deforestation in Laranjal do Jari in Amapá and Almeirim in Pará. The large red areas are reported in the new deforestation data as occurring prior to 2014. Previously, the PRODES would analyze only one

20 image for every Landsat scene per year. In regions with persistent cloud cover this approach could result in multiple years without imagery to assess. Since tropical forest can regenerate quickly to the point of appearing “spectrally similar” to intact forest, it is possible that as short a gap in available imagery as four years would have allowed for deforested areas to regenerate past the point of detection. Thus, it is possible that new detection methods have identified areas that were previously deforested and regenerated thereby escaping past detection as deforestation.c

Figure 3.1.5. Map of Northeastern Pará. Red areas show where deforestation from 2014 or earlier is reported in the new data that was not in the previously reported.

In other regions, newly reported deforestation has occurred in areas even though they had not been under cloud cover. There is a possibility that a gap in the time series occurred as explained above, or a different technique may have been employed to detect deforested areas. Figure 3.1.6 shows such an area in Mato Grosso, in and around the municipality of Campo Novo do Pareci, where large clearings (in red) that were not included in the old deforestation data are seen within a matrix of previously cleared areas.

Figure 3.1.6. An area of Mato Grosso with new areas of deforestation (in red) reported in PRODES 2015 data release.

Another possible explanation could be the type of detection method employed. Differences in the detection approach are evident in some areas. In previous data, some deforested areas appear to have been detected or adjusted manually and outlined. In these areas, the edges are smooth compared to other areas where edges of pixels are clearly evident. Figure 3.1.7 shows an area of deforestation reported in the 2007 data outlined on top of the new deforestation data in red. The shape of the eastern edge in the old polygon had been smoothed from the “stair step” effect of pixel boundaries which are seen in the new data. This area is a good demonstration of the overall shift between the two data sets that we expect to see, but surprisingly the shape of the polygon has also changed. Also note that the polygon to the northwest was not reported as deforestation in the old data.

Figure 3.1.7. The old deforestation polygon is shown in grey outline here on top of the new deforestation data in red. The shift from the geo‐rectification is evident here as well as a change in the shape and size of the polygon.

As illustrated in the figures above, there are many areas of newly reported historical deforestation in this data set. Although we have not seen official reports of a new methodology for the analysis of historical imagery or the inclusion of more scenes from the historical archive, clearly more areas of past deforestation have been detected in the newly released data. This in turn is likely to affect our estimates of program impacts using the PRODES 2015 release for 2013, 2014, and 2015, especially because the greatest increases in deforestation observed are in Pará, where the PMV program operates. This is illustrated in the next two figures. Larger municipalities tend to show larger absolute changes in total area deforested (Figures 3.1.2 – 3). Therefore, to examine annual changes, we consider annual increments of change as percentages of the total deforestation reported in the new data (Figures 3.1.8 – 9). In these figures, our two treatment groups are identified by cross‐hatching (PMV) and bold outlines (Imazon10 + PMV).

3.2 Outcomes ‐ Agricultural land and yields As noted above, the municipalities that join the PMV or Imazon10 may have been seeking to maximize local economic outcomes, conditional on compliance with federal regulatory constraints. One strategy that has been widely promoted is to intensify production on land that is already deforested, in order to absorb labor, supply the market, and generate income and tax revenues. We therefore examined patterns in land allocation and yields of major crops and livestock production systems. Specifically, we use the following outcome variables:

(1) The pasture stocking ratio (head of cattle per hectare of pasture) in each municipality in each year from 2008 to present. The stocking ratio is highest on average in blacklisted municipalities. We also examine as separate outcomes each element of that ratio, i.e., the head of cattle and the hectares in (or area of) pasture. Head of cattle are reported by IBGE in the annual Pesquisa Pecuária Municipal. Area of pasture is reported by the Mapbiomas project for each year from 2008 to 2015 (with fewer than 10% missing due to difficulties with interpretation of remote sensing). We use data only through 2014 for consistency with other models. Mapbiomas is working towards a longer time series on land use, which we expect to be available for our next report. (2) Average value of production per hectare planted in perennial crops, and average value of production per hectare planted in annual crops (dominated by cassava, rice, maize, soy and sugarcane). These are extracted from IBGE’s annual Pesquisa Agrícola Municipal for 2002 to 2014. The yield of perennial crops is highest on average in forested municipalities, while the yield of annual crops is highest in consolidated and “green municipalities” that have implemented all steps of the PMV. Again, we also examine as separate outcomes each element of those ratios, i.e., the value of production and the hectares in (or area of) annual crops or in perennial crops.

We have been pursuing data on the value‐added tax revenues generated in each municipality from the tax called the ICMS (Imposto Sobre Operações Relativas à Circulação de Mercadorias e Serviços de Transporte Interestadual de Intermunicipal e de Comunicações). Once these data are compiled, they will provide a summary metric for overall economic activity in a municipality that takes into account the whole suite of economic activities (as measured by market prices and quantities). Receipts from value added taxes are summarized in a measure called the VAF, or Valor Adicionado Fiscal. The state VAF is used to calculate federal transfers to the states. The municipal data are not available from the federal government, so we have requested from state finance ministries through the Access to Information Act. For a few states, we have already obtained data from 2000 to present. We also considered municipal GDP as a potential aggregate economic outcome, but we discovered that IBGE generates the data series for each municipality by dividing each state’s GDP across its municipalities based on fixed proportions defined using census data. Thus, this data series is not suitable for identifying impacts of treatments that vary across municipalities and time.

We have been pursuing data on the value‐added tax revenues generated in each municipality. Once these data are compiled, they will provide a metric for overall economic activity in a municipality, one which takes into account the whole suite of economic activities (as measured by market prices and quantities). Receipts from value added taxes are summarized in a measure called the VAF, or Valor Adicionado Fiscal. The state VAF is used to calculate ICMS transfers. The municipal data are not available from the federal government, so we have requested from the state finance ministries through the Access to Information Act. For a few states, we have already obtained data from 2000 25 to present. We also considered municipal GDP as a potential aggregate economic outcome, but we discovered that IBGE generates the data series for each municipality by dividing each state’s GDP across its municipalities based on fixed proportions defined using census data. Thus, this data series is not suitable for identifying impacts of treatments that vary across municipalities.

We also explored measures of municipal government costs and capacity building. For municipal government finance and structure, we can obtain data on the budget, staffing, and receipts generated by the municipal environmental agencies (SEMMAs) for the state of Pará but not for all municipalities in the Amazon biome that are rin ou donor pool. We therefore do not analyze PMV impacts on these administrative costs.

4. Methods There is a growing literature that applies “quasi‐experimental” methods to estimate the causal impacts of conservation interventions, i.e., to discern if a particular intervention (treatment, project, program) actually caused a particular observed outcome, such as reduced deforestation or increased productivity. To assess the causal impacts of a program, we must establish what would have happened in areas exposed to the program if they had not been exposed (i.e., the ʹcounterfactualʹ to treatment). This would be relatively easy if either the treated areas before treatment or non‐treated areas could be assumed similar in all aspects except treatment – justifying before/after or treated/not comparisons.

However, if temporal or spatial similarity cannot be assumed, then such comparisons are biased estimates of impacts. Temporal and spatial factors that differ systematically across treated and non‐treated areas make it considerably harder to causally relate a program to any change in outcome, like a decrease in deforestation rates. These factors confound impacts when they are correlated both with the program (or its lack) and with outcomes of interest, such that their own impacts can be misinterpreted as a program‐outcome link. One important feature of such potential confounders is that while sometimes they are observed (i.e., as variables in the dataset), in other settings they may be equally consequential and yet still be unobserved by the data analyst.

For observables, perhaps the most frequent approach for addressing such potential confounding is multivariate regressions that control for various factors while including a variable for the program being evaluated. As the workhorse of modern statistical inference, multivariate regressions are a logical starting point for any evaluation. When panel data are available, fixed effects can be employed to control for factors that are unobserved but time invariant, i.e. differences‐in‐ differences estimation, as well as factors that are constant across tunits bu vary with years, i.e. two‐ way fixed effects estimation. (We previously reported estimation results for two‐way fixed effects models of deforestation using the full sample, and in this report, we add estimation results for two‐ way fixed effects models of economic outcomes using the full sample.) However, if selection processes strongly influence where a program is applied, it could be the case that the values for important characteristics included in regression specifications do not overlap a great deal between the program sites and comparison (or control) sites. When the program and comparison sites are different in terms of such characteristics, then regression can fail to separate the impacts of programs from the influences of these characteristics. Selection bias can be compounded by low power when there are relatively few program sites, i.e. few treated units. 26

Quasi‐experimental methods have been developed to address selection bias due to both observables and unobservables. Matching methods have become the leading quasi‐experimental approach, usually pairing units (tracts, pixels or households) subject to interventions with ʹvery similarʹ comparison or control units in order to isolate impacts using statistical modeling. Various metrics of the similarity of control units to treated units based on observable characteristics can be employed, e.g., the Mahalanobis distance or probability of participation in a policy or program (the propensity score) as a summary statistic for the suite of ecological, socio‐economic, institutional and geographic factors judged by the analyst to be relevant. By explicitly searching for greatest similarity, however that is defined, the analyst can compare treated units to similar control units with all of the relevant observed characteristics – except the intervention ‐ in common.

Matching methods can be powerful, especially when it is possible to match or ‘balance’ all observed factors considered likely to influence selection into the treatment as well as the outcome. However, matching by itself does not help address any of the potential unobservable differences between the units being compared. One improvement is to analyze “differences‐in‐differences” (DiD) within the matched sample, in order to remove the effect of any unobservable factors that are time‐ invariant as well as balancing on observed factors. For economic outcomes, we first use matching to establish a balanced sample, i.e. with similar distribution of outcomes for municipalities in the PMV and comparison municipalities not in the PMV. Using that matched sample, we estimate two‐ way fixed effects models, which control both for idiosyncratic differences across municipalities and temporal trends. In addition to the fixed effects, we include covariates for each of blacklist status (1 for years that a given municipality was on the blacklist), PMV membership (1 for years since a given municipality joined the PMV), and Imazon10 (1 for years since a given municipality started collaborating with Imazon).

Another challenge with conventional matching methods is that establishing statistical similarity for observable factors typically relies upon quite large samples in order to find well‐balanced sub‐ samples of treated and control units with similar distributions. Thus, we apply matching to the PMV but not to the PMV bundled with the Imazon10. A potential solution for “low N” treatments is the ʹsynthetic control methodʹ or SCM (as described in the first publication of this project by Sills et al. 2015). When only a few units are treated, researchers often compare outcomes for a few selected ʹsimilarʹ units to those for a ʹtreatedʹ unit influenced by an intervention or exogenous event. Selecting similar units is critical to estimating impacts, because those unitsʹ outcomes are used to generate the counterfactual, i.e., the estimate of what would have happened within the treated unit had there not been any intervention. Thus, regardless of whether statistical or case study methods are employed, policy impact studies always involve some form of judgment concerning similarity across treated and untreated units. For example, ʹdifference in differenceʹ (DiD) estimates implicitly assume that comparison and treated units have similar time trends.

While conventional ʹmatchingʹ approaches search for similarity in the observed characteristics of treated and control units, SCM searches for a weighted blend of control units that – for the pre‐ treatment years − actually did have a similar time trend in termse of th outcome of interest. SCM makes explicit not only the weighting of those pre‐intervention outcomes and unitsʹ characteristics − defining the differences that are being minimized by searching for similarity − but also ultimately the definition of ʹthe synthetic controlʹ, i.e., the weights that are assigned to each of the potential comparison units in order to match the past outcomes for any given treated unit.

The next optimization process of SCM , which identifies a combination of control units in terms of both observed characteristics and pre‐treatment outcomes, blends the focus on characteristics from matching with the focus on unobservable trends from DiD approaches. For DiD, the analyst assembles a pool of comparison units (possibly through matching) assumed to have similar time trends and thus the treatment impact is estimated as the change in the outcome over time for the treated units minus the change over time for the controls (the latter identifying the common time trend to subtract it out). In SCM, the best fitting combination of units in terms of the pre‐treatment trends in the outcomes − i.e., the synthetic control – is assumed to continue to follow a time trend similar to that of the treated unit but for the effect of the intervention. If the pre‐treatment outcomes have been well‐matched in the creation of the synthetic control, the impact estimate for each post‐ treatment year is simply the difference between the observed outcome for the treated unit and the syntheticʹs weighted outcome.

When a long time series of data on the outcome variable is available, SCMʹs use of pre‐treatment outcomes can implicitly improve upon matching on only observed characteristics (though of course the pre‐treatment outcomes could also be called “observed characteristics”), thereby ʹmaking up forʹ the limited cross‐sectional sample size (due to few treated units) by employing intensive computation (feasible for each of a few treated )units to identify the optimal combination of controls. The observed driving characteristics in a data set will never be able to include absolutely all of the factors that influence the outcome of interest, e.g., deforestation. The pre‐treatment outcomes, in contrast, clearly must reflect all of the key factors that drove them (at least during the pre‐treatment period). Thus, by matching on pre‐treatmentoutcomes, we can in principle construct a more accurate counterfactual.

For example, consider a treated municipality whose deforestation is affected by its responsiveness to some exogenous factor that is not directly observed. That factor could be a shock in the price for a crop that can only be produced under particular agro‐climatic conditions, or it could be a regional shift in the relevant political dynamics that has local effects. For either example, among a large set of control units only some will be sensitive to this factor. Looking at outcomes across pre‐treatment years – and most usefully during years during which that factor shifted − and controlling for all of the observed characteristics that are believed to matter for deforestation, matching on pre‐treatment outcomes can help to identify which control units have the same sensitivities as the treatment unit.

In practical terms, we have a long‐enough time series and small enough sample to recommend using SCM to estimate the impact of the PMV on deforestation and on crops yields in the 16 blacklisted municipalities in our sample (8 in the Imazon10, and 8 only in the PMV). We implement SCM using the Synth package in R. The steps include defining a ʹdonor poolʹ, or a set of potential controls judged to have some underlying structural similarity in terms of the processes that generate the outcome of interest. One important set of factors to consider is the historical conditions that constrain the range of the outcome, e.g. how much forest was still standing in a municipality at the start of the study period. Because the nested optimization process in Synth may fail to find a solution when the donor pool is too large, we narrow down large donor pools to no more than the 80 most similar municipalities.

Given an outcome of interest and a donor pool, SCM selects weights (W) on potential control units to define a linear combination of controlsʹ outcomes, i.e., ʹthe synthetic controlʹ. These weights very directly determine the impact estimate which, for any post‐treatment time period, is the difference between the outcome for the treated unit and this weighted‐average or synthetic outcome. The W 28 is chosen based on the core idea that the average of pre‐treatment outcomes for control units − each weighted by its W value, which often is zero for any given donor pool unit − should be as close as possible to the pre‐treatment outcome in the treated unit. Recall that the outcome (Y) ‐ whether deforestation or crop yields‐ is affected by both observed (Z) and unobserved (U) factors (Y = βZ + U). Were we to confirm only similarity in Y (i.e., search for low WYcontrol ‐ Ytreated) for pre‐treatment years, then we might label as ʹsimilarʹ units with higher Z and lower U. The innovation of SCM is to select units based on similar Y as well as similar Z, implying similar U (although the U cannot be observed). Since one cannot minimize all differences at the same time, but only a combination of differences, another weighting vector V is required to assign weights to the variables in Z and to each year or averages of pre‐treatment Y. In the Synth package for R, V is selected to maximize predictive power for pre‐treatment outcomes.

The quality of the synthetic control is measured by how closely the weighted synthetic outcomes match the outcomes for the treated unit in the years prior to the intervention in question. The goal is to minimize the difference, as measured by the mean squared prediction error (MSPE). A ʹbestʹ fit is always a matter of judgment and, in particular, one might go beyond MSPE to judge whether the synthetic cohort appears to mirror the treated unit in terms of the turning points in a plot of the outcomes across pre‐treatment years (which could indicate common sensitivities as noted above). If the outcome plots look similar especially in the years immediately preceding treatment, it is more likely that the treated and the synthetic control will have similar responses – or similar sensitivity ‐ to common factors. Another possible check is the size of the error in the time periods immediately preceding treatment. A persistent error here could easily be replicated in the estimated impacts. In our previous report, we evaluated the quality of the synthetic controls obtained for the deforestation outcomes.

The final step is to assess the statistical significance of any post‐intervention divergence between the synthetic outcome and the treated outcome, Mi.e., the SC estimate of impact. In this report, we calculate this difference in deforestation based on the new PRODES data series, as well as crop yields. Standard statistical tests of the similarity between the treated unit and the constructed or weighted synthetic control − such as tests of ʹbalanceʹ on covariates or of diﬀerences in outcomes − are not possible here due to the small number of treated units. Thus, other methods must be employed to characterize the ʹnoiseʹ in such procedures, e.g. placebo tests for different years or municipalities, bootstrapping, or a combination of the two. Placebo effects are typically estimated for all units in the donor pool, i.e., by ‘pretending’ that they were treated and using SCM with a donor pool of all of the other control units to create synthetic controls for ʹthe treatedʹ and calculate the resulting ʹtreatment effectsʹ. As the units were not in fact treated, the effects should be zero and thus the false ‘treatment effects’ characterize the ʹnoiseʹ, i.e., what SCM yields where there is not really any effect. Estimated effects above the 95th percentile or below the 5th percentile of the placebo effects are significant at the 10% level.

However, because the optimization process in Synth works best with a donor pool of fewer than 100 units, we obtain fewer than 100 placebo effects, meaning that their distribution is not very smooth and the critical values for 5% and% 95 often must be extrapolated from the nearest values. We therefore bootstrap the placebo tests, applying the procedure developed in Sills et al. (2015), by taking multiple random draws of their donor pool, ensuring that in each case we drop a proportional number of donors that received large and small weights in the original synthetic control. For this report, we employed this approach to establish confidence intervals around our

29 new estimated impacts using INPE’s revised estimates for deforestation in 2013 and 2014 and new estimates for 2015.

5. Estimation and results

Results from the evaluations performed are detailed in this section. To assess impacts on deforestation, we use the synthetic controls constructed for the previous report using annual data on deforestation from the old PRODES data series (recalling that we do not have new PRODES data across the entire period). For the economic outcomes, we focus on DiD estimated within a matched sample, while also comparing to the results of synthetic control matching for perennial crop yields for the municipalities in the Imazon10 treatments.

5.1 Deforestation outcome

In order to evaluate the impacts of PMV and the Imazon10, we need to generate ʹapples to apples baselinesʹ, i.e., controls that have the same outcome‐relevant characteristics as treated units. In standard matching, characteristics that affect both the probability of participating in the program and the probability of deforestation ear balanced. In synthetic control matching, the goal is to balance both those characteristics and historical deforestation. Though they emphasize different balances across control and treated, these two types of matching ultimately require the same characteristics data, in addition to data about the deforestation outcome. In our last report, we concluded that synthetic control matching was most appropriate, given the sample sizes and data available, and that where possible to compare with conventional matching, the results were qualitatively similar. To implement synthetic control matching, we identified factors that could drive deforestation and selection into these programs and we then sought variables to represent those factors, including:

1. Factors in agricultural profitability, including slope, elevation and market access (e.g., % of municipality that is within 5 km of a road) 2. Descriptors of land tenure, including % under federal control, in protected areas, indigenous territories, and agricultural settlements 3. Indicators of social welfare and structure, such as education levels, per capita income, and income distribution 4. Indicators of demand for agricultural production: in sanitary zone free of hoof and mouth disease, whether slaughterhouses present 5. Political indicators such as party of the prefeito, and percent of municipal income generated by agricultural sector 6. Indicators of pressure for land speculation and invasion, including land tenure and # of conflicts recorded by the Catholic Church 7. Indicators of cost to implement the CAR, such as count of farms and furthest distance from municipal seat to municipal boundary 8. Other deforestation drivers like population density, cattle density, and possibly mine density Starting with the master sample of 519 municipalities either fully inside the Amazon Biome or adjacent/crossing the boundary of the biome and with more than 50% historical forest cover, we 30 then constructed synthetic controls for each municipality in the PMV and the Imazon10. In this report, we use those same synthetic controls and analyze their deforestation trends compared to the trends in the treated municipalities in 2013, 2014, and 2015, based on the new PRODES data series. We demonstrated that the synthetic controls were always more similar to the treated municipality than the mean of the donor pool in terms of the covariates. We further assessed the quality of the synthetic controls based on the historical deforestation plots. Based on the plots, we sorted the municipalities into three categories: (i) poor synthetic control, i.e., unable to identify a linear combination of donor municipalities with similar historical deforestation, in particular in the final pre‐ treatment years for which errors in fitting appear to imply correlated errors in impact estimates; (ii) deforestation always low or, even if not, then low immediately pre‐treatment, implying little potential for the intervention to reduce deforestation beyond the influences represented in the donor pool that have already maintained low deforestation; or (iii) synthetic controls with low MSPE (mean square prediction error) that are closely matched to the treated unit in the years immediately preceding treatment.

We followed the SCM literature and dropped group (i). For low‐pressure locations (ii), synthetic controls appeared to match well, although there was really not enough variation pre‐treatment to assess the fit. They also generally show no impact of the program. Because we find many such municipalities, category (ii) matters for overall impacts of interventions. For locations facing more pressure and where the synthetic control is a reasonable match (iii), we calculate treatment impacts as the treated’s percent deforestation minus the synthetic’s percent deforestation in 2013, 2014, and 2015 using the new PRODES data series.

5.1.1 Deforestation Impact of the bundled PMV + Imazon10 based on synthetic controls

For the bundled PMV + Imazon10 treatment, we drew from a pool of blacklisted municipalities to construct the synthetic control for the 8 municipalities that are in the blacklist, the PMV, and the Imazon10, and from a donor pool of non‐blacklisted municipalities for the 2 municipalities in the Imazon10 that are not blacklisted. We found that: (i) three municipalities have poorly fit synthetic controls: Brasil Novo, Moju, and Novo Repartimento; (ii) the deforestation rate has been low and steady in Monte Alegre and Santarém and this not surprisingly leads to low estimated impacts, but also makes it difficult to evaluate the quality of the synthetic control; and (iii) the synthetic control closely matches historical deforestation in Novo Progresso and Santana do Araguaia, and matches reasonably well in Dom Eliseu, Tailandia, and Ulianopolis. It is always a judgment call whether the match quality is ʺsufficiently good.ʺ In all cases, the construction of the synthetic control improves the match, or balance, of the covariates. However, the heart of SCM is use of historical data on the outcome to match on unobservables as well as observables, and thus we ultimately judged quality based on the plots of historical deforestation. In our previous estimations, we found that the bundled PMV + Imazon10 treatment reduced deforestation in the five blacklisted municipalities with good synthetic controls (by an average of 0.3% or 21 km2 in 2014 and by <0.1% or <0.7 km2 in 2013). However, as described previously, the deforestation data have since changed significantly, generally showing more deforestation in the Imazon10 (Figure 3.1.2). As a result, we no longer find significant differences between deforestation in the Imazon10 and their synthetic controls in 2013, 2014, and 2015, based on the 95% 31 confidence interval constructed with placebo tests (Table 5.1). Santana do Araguaia is a partial exception, as shown in Figure 5.1. For that municipality, we estimate increasing effect sizes from 2013 to 2015. In 2015, we estimate that the bundled PMV + Imazon10 treatment reduced deforestation (as a percent of the municipal area) by 0.0035%, which was more than 90% of the placebos, and thus statistically significant at the 80% but not the 90% level (using a two‐sided test). We hypothesized that the program may have more effect on deforestation in the areas most directly impacted, i.e. excluding protected areas and indigenous territories that are not under municipal control and not eligible for the CAR. However, again we found that although the point estimates consistently suggest that being in the Imazon10 lowers deforestation in blacklisted municipalities, none of the estimated effects are significantly different from zero.

Table 5.1. Effect of bundled Imazon10 + PMV intervention, estimated with SCM: first symbol (+ or ‐) indicates whether the point estimate of the effect is positive or negative, and second symbol indicates its statistical significance based on placebo tests ‐ all are zero (0) because they fall within the 90% confidence interval around zero established by the placebos.

Municipality Year Quality of Sign and significance of effect of joined Synth Control Imazon10+PMV 2013 2014 2015 Blacklist Brasil Novo 2012 Unacceptable Dom Eliseu 2011 Good ‐ 0 ‐ 0 ‐ 0 Moju 2013 Unacceptable Novo Progresso 2012 Excellent + 0 ‐ 0 + 0 Novo Repartimento 2012 Unacceptable Santana do Araguaia 2012 Excellent ‐ 0 ‐ 0 ‐ 0 Tailandia 2011 Good + 0 ‐ 0 ‐ 0 Ulianopolis 2011 Good ‐ 0 ‐ 0 ‐ 0 Non‐blacklist Monte Alegre 2012 Unknown‡ + 0 ‐ 0 + 0 Santarem 2012 Unknown‡ + 0 + 0 + 0 ‡ While the synthetic controls match annual deforestation in these municipalities prior to 2012, this is not a good test of whether they are valid counterfactuals, because annual deforestation was consistently very low and therefore very easy to match by simply picking other municipalities outside the active deforestation frontier.

Figure 5.1 Percent of municipality deforested in Santana do Araguaia (solid black line), average of the donor pool of blacklisted municipalities (green line), and synthetic control (dotted line)

5.1.2 Impact of the PMV estimated with SCM

As with the bundled PMV and Imazon10 intervention, we examine effects of the PMV on blacklisted municipalities and non‐blacklisted municipalities separately, recognizing different deforestation histories (creating different opportunities to reduce deforestation) as well as different sets of interventions designed to reduce deforestation. Previously, we found that the PMV had some impact in blacklisted municipalities that joined the program early. More specifically, in the nine blacklisted municipalities that joined the PMV in 2010 (excluding Paragominas), we found that treatment significantly reduced deforestation (below the noise observed in the placebos) in five of the municipalities. This was also observed in one out of the seven blacklisted municipalities that joined the PMV in 2011.

Similar to our findings for the Imazon10 with the new PRODES data, we no longer find a consistent effect of the PMV on blacklisted municipalities using the new PRODES data series. The point estimates of the effect of the PMV on blacklisted municipalities are almost all negative (with exceptions including very small and unexpectedly positive point estimates in Altamira and Pacajá), but as with the Imazon10, these are generally not significantly different from zero, except in a single municipality: Cumaru do Norte (Figure 5.2). In Cumaru do Norte, the effect ofV the PM grew over time from 2013 to 2015, becoming ‐0.0072% and statistically significant at the 95% level in 2015. Compared to other blacklisted municipalities in the PMV (16, of which 11 have acceptable synthetic controls), Cumaru do Norte ‐ like Santana do Araguaia ‐ is a large municipality (thus making the blacklist criterion of 40 km2 of deforestation more binding) that joined the PMV early (in 2010), had

33 established a working group to respond to deforestation and a societal pact against deforestation by 2016, and now has more than 80% of eligible land registered in the rural environmental cadaster (more than 85% in Santana do Araguaia). This suggests anecdotally that the program is more likely to be effective when locally embraced and more fully implemented.

Figure 5.2 Percent of municipality deforested in Cumaru do Norte (solid black line), average of the donor pool of blacklisted municipalities (green line), and synthetic control (dotted line)

Turning to the analysis of deforestation in the areas most directly affected by the PMV, .i.e excluding protected areas and indigenous territories that are not under municipal control and not eligible for the CAR, we again find that most point estimates are negative but not significantly different from zero. The three exceptions (with point estimates in parentheses) are Cumaru do Norte in 2014 (‐0.0105), Dom Eliseu in 2014 (‐0.00376), and São Felix de Xingu in 2015 (‐0.0052). Dom Eliseu and São Felix de Xingu were both implementing the PMV in practice (and not just on paper) by July 2015. For the larger pool of municipalities in the PMV but not on the blacklist (Figure 1.1), we also do not find significant effects on deforestation for any municipality, confirming our previous results.

5.1.3 Reasons for changes in impact estimates

The results for 2013 and 2014 differ from our previous results because of the change in the PRODES data series, which INPE has attributed to their new geo‐rectification, as discussed in section 3.1. Consider first the change in deforestation detected in the municipalities in the Imazon10. On

34 average across the eight blacklisted municipalities in the Imazon10, the new PRODES data series shows deforestation 2.5 km2 higher in 2013 and 4.2 km2 higher in 2014. These changes amount to about 0.05% of the average municipal area, which is substantial in comparison to the previously estimated effect sizes. Specifically, in the five municipalities with good quality synthetic controls, the new data series shows deforestation higher by 0.06% of the municipality area in 2013 (larger than the previously estimated effect size) and 0.05% of the municipality area in 2014 (about an eighth of the previously estimated effect size). The estimated effect size is the difference between deforestation in the treated municipalities (e.g. the Imazon10) and their synthetic controls. For the evaluation of the bundled Imazon10 + PMV intervention in blacklisted municipalities, we used the same eight synthetic controls constructed previously from a donor pool of other blacklisted municipalities not in the PMV. eWith th shift to the new data series, deforestation in 2013 in the average synthetic control for the blacklisted Imazon10 decreased by ‐0.002% of municipal area. Deforestation in 2014 increased on average in the synthetic controls, but only by 0.003%. Likewise, in the synthetic controls constructed to evaluate the impact of the PMV on blacklisted municipalities, there was on average no change in deforestation in 2013, while deforestation in 2014 in the new data series increased by just 0.007% of municipality area. These numbers confirm the general pattern shown in the maps in section 3: the change in the PRODES data series led to higher estimated deforestation in most municipalities in the Imazon10 and the PMV, that is, most treated municipalities. The synthetic controls draw on municipalities from across the Amazon, and thus on average, were not as affected by INPE’s change in data processing. The end result is that we no longer find statistically significant impacts of either the bundled Imazon10 + PMV or the PMV treatment on blacklisted municipalities. For non‐blacklisted municipalities, we generally did not find effects previously and the new data series does not change that finding. INPE has explained the new PRODES data series as a result of improvements in geo‐rectification. Because our new findings in this report are also based on those improved data, they should be more accurate. However, because the new data series is only available in annual increments starting in 2013, we were not able to apply our preferred method for distributing deforestation detected in areas that had previously been under cloud cover. We expect that annual data using the new geo‐rectification will eventually be made available, and analyzing that longer time series may lead to different conclusions.

5.2 Agricultural yield and share

5.2.1 Two‐way fixed effects estimation of agricultural yields To accomplish its ultimate goal of economic development compatible with stable net forest cover, the PMV must shift municipalities away from economic activities dependent on deforestation, most notably conversion of forest to cattle pasture and crops fields. One way to make this economically (and politically) viable is to intensify agricultural production in areas already deforested, in order to maintain agricultural employment and profits even as the expansion of agricultural land into forest is blocked. While there are concerns about the environmental externalities of intensive cattle production and the potential perverse effects of making cattle ranching more profitable, this intensification strategy has garnered broad support as part of climate change mitigation efforts in Brazil. For municipalities in the PMV, it represents a feasible short‐term strategy for maintaining economies heavily dependent on agriculture and thus unlikely to quickly shift to an entirely new forest‐based or service‐based economy. To assess whether the PMV or Imazon10 are facilitating 35 this transition, we examine the impacts of participation in these programs on crop yields and pasture stocking.

We report first on two‐way fixed effects panel estimation of pasture stocking ratio (head of cattle/Ha of pasture) using the master sample of 519 municipalities in the Brazilian Amazon. Table 5.2, first row, shows no effect of the PMV, Imazon10, or the blacklist on stocking ratios.

Table 5.2 Two‐way fixed effects estimation of stocking ratio and yields

Coefficient, Standard Error, Significance

PMV Imazon10 Blacklist Years Adj. R2

Cattle stocking density ‐0.974 2.296 0.527 2008 ‐ 2014 0.584

0.97 3.019 1.837

‐‐‐ ‐‐‐ ‐‐‐

Yield of temporary crops 0.01 ‐0.778 0.046 2002 ‐ 2014 0.514

0.104 0.293 0.12

‐‐‐ *** ‐‐‐

Yield of perennial crops 0.609 0.649 ‐0.408 2002 ‐ 2014 0.458

0.193 0.537 0.225

*** ‐‐‐ *

‐‐‐ not statistically significant, * p < 0.10, ** p < 0.05, *** p < .001

Turning to crops, we disaggregate annual (or temporary) crops and perennial crops. While annual crops occupy a much greater area than perennial crops, perennial crops are often advocated as a production strategy consistent with forest conservation, because they require more investment and produce more over a longer time period from each hectare of land. Again, we report first on two‐ way fixed effects estimation of yields and shares of municipalities in annual and perennial crops using the master sample of 519 municipalities in the Brazilian Amazon.

Table 5.2, second row, shows that the PMV had no effect on the index of temporary crop yield, while the Imazon10 had an unexpected negative effect. This may reflect some other idiosyncratic factors in those 10 municipalities, which is the risk of trying to estimating the effect on such a small treatment group using conventional methods. In clear contrast, Table 5.2, last row, shows that the PMV had a positive effect on perennial crop yield, after controlling for fixed municipal and time effects and a small negative effect of the blacklist.

Table 5.3 disaggregates the effect of the PMV by year since a municipality joined, showing that the effect rises over time to the strongest effect in the fourth year (the last possible in our data). That is consistent with rising efficacy of a rational local response (increasing the yield of crops grown year

36 after year in the same area) to a federal constraint (discouraging conversion of forest to additional cropping area).

Table 5.3 Two‐way fixed effects estimation of Perennial Crop Yields as a function of years in the PMV PMV ‐ PMV ‐ PMV ‐ PMV ‐ yr 1 yr 2 yr 3 yr 4 Imazon10 Blacklist Years Adj. R2

Coef. ‐0.034 ‐0.115 0.631 1.909 0.621 ‐0.387 2002 ‐ 2014 0.459

Std Err 0.319 0.331 0.631 0.342 0.536 0.225

Signif. ‐‐‐ ‐‐‐ * *** ‐‐‐ *

“PMV – yrN” is the municipalityʹs Nth year of participation in the PMV ‐‐‐ not statistically significant, * p < 0.10, ** p < 0.05, *** p < .001

If PMV motivates greater yields in perennial crops, it might also motivate a shift into perennial crops from temporary crops. Table 5.4 supports this notion, showing a consistent rise in the area used for perennial crops and consistent fall in the area used for annual crops as a result of participation in the PMV. Since Imazon supports the entire PMV, its influence is reflected in the coefficient on the PMV, but again ether is no additional effect of the Imazon10. With only 10 participating municipalities, there is limited statistical power to estimate the effect of the Imazon10 on top of the effect of the PMV on either yield or share or land in perennials, and thus in section 5.2.3 we re‐examine this question using the synthetic control method.

Table 5.4 Two‐way fixed effects estimation of share of land in crops and pasture

Coefficient, Standard Error, PMV ‐ PMV ‐ PMV ‐ PMV ‐ Significance yr 1 yr 2 yr 3 yr 4 Imazon10 Blacklist Years Share in perennial crops 0.004 0.002 0.002 0.002 0.0004 0.001 2002‐ 0.001 0.001 0.001 0.001 0.001 0.0004 2014 *** *** *** *** ‐‐‐ ‐‐‐ Share in temporary crops ‐0.008 ‐0.011 ‐0.012 ‐0.014 0.006 0.015 2002‐ 0.003 0.003 0.003 0.003 0.005 0.002 2014 *** *** *** *** ‐‐‐ *** Share in cattle pasture ‐3.842 ‐3.731 ‐3.197 ‐2.094 0.126 2.482 2002 ‐ 1.087 1.134 1.171 1.177 1.949 0.743 ‐‐‐ ‐‐‐ ‐‐‐ ‐‐‐ ‐‐‐ ‐‐‐ 2014 * p < 0.10, ** p < 0.05, *** p < .001

While Tables 5.2 ‐5.4 provide benchmark results, they are subject to concerns about confounding related to selection into the PMV. If the municipalities that selected into the PMV were on different trajectories than other municipalities, that could confound estimates of program impact, because we are only able to control for time‐invariant factors specific to municipalities and temporal factors that affect all municipalities with our two‐way fixed effects strategy. Thus, we next turn to matching to pre‐process the data.

5.2.2 DiD estimation of agricultural yields in matched sample of PMV and control municipalities Adding to our full‐sample panel regressions above, we used traditional matching on PMV status as of 2011 with the single nearest or the five nearest neighbors to pre‐process the sample before re‐ estimating the panel regressions. To check for robustness, we matched with large and small sets of factors. The results were reasonably robust, and we therefore focus on matching with the larger number of factors and larger number of neighbors. Annex 3 demonstrates that matching improves balance, while not fully eliminating statistically significant differences between the treated and control units. As expected, selecting only the single nearest neighbor into the matched sample results in the best balance, but also uses less of the sample. Thus, we report estimation results using both matched samples, and place the greatest confidence in results that are consistent across the two.

For cattle, Table 5.5 suggests that if anything, the PMV generated the opposite of higher density stocking. Table 5.6 shows that the average negative effect comes from a negative effect early on that fades to zero over time. For groups promoting intensification of cattle production, the PMV clearly has not had the hoped for effect, and there has been no different effect in the Imazon10. By looking at the two components of the stocking ratio, we can see how this happened. Table 5.6 shows that in the early years, PMV starts with a larger positive effect on pasture that gets smaller over time, and that the effect of PMV on cattle herds starts at zero and grows positive over time. Table 5.5 Two way fixed effects estimation of pasture stocking ratio in matched sample # Coefficient, SE, Significance Neighbors PMV Imazon10 Blacklist Years Adj. R2 Cattle stocking 1 ‐2.046 2.301 0.792 2008 ‐ 0.592 1.078 2.843 2.047 2014 * ‐‐‐ ‐‐‐ Cattle stocking 5 ‐1.839 2.323 0.667 2008 ‐ 0.601 0.836 2.206 1.496 2014 ** ‐‐‐ ‐‐‐ ‐‐‐ not statistically significant, * p < 0.10, ** p < 0.05, *** p < .001

Table 5.6 Two way fixed effects estimation of pasture stocking ratio, share in pasture, and total herd as a function of years in the PMV Coefficient, Standard Error, # Significance Neighbors PMV ‐ yr 1 PMV ‐ yr 2 PMV ‐ yr 3 PMV ‐ yr 4 Imazon10 Blacklist Years Cattle stocking 5 ‐3.407 ‐2.677 ‐0.706 0.819 2.023 0.793 2008 ‐ 1.15 1.219 1.294 1.31 2.204 1.494 2014 *** *** ‐‐‐ ‐‐‐ ‐‐‐ ‐‐‐ Share in pasture 5 0.028 0.031 ‐0.019 0.013 ‐0.006 0.002 2008 ‐ 0.007 0.007 0.008 0.008 0.0013 0.009 2014 *** *** ** ‐‐‐ ‐‐‐ ‐‐‐ Total head of cattle 5 7.892 9.93 23.508 31.988 14.652 108.447 2002‐ 6.139 6.484 6.755 6.796 9.487 4.454 2014 ‐‐‐ ‐‐‐ *** *** ‐‐‐ *** “PMV – yrN” is the municipalityʹs Nth year of participation in the PMV ‐‐‐ not statistically significant, * p < 0.10, ** p < 0.05, *** p < .001

For crop yields, Table 5.7 supports the findings above concerning the effect of PMV, i.e., a significant positive effect from PMV on the perennial yield but no significant effect on the temporary yield, but this is not confirmed by estimation results using the sample created using the five nearest matches (Table 5.7) that show no significant impact. Turning to the finding that PMV might facilitate a shift in area from temporary crops to perennial crops, this is confirmed using both the sample generated by selecting the single nearest match and the five nearest matches (Table 5.8). Thus, we conclude that the PMV has encouraged the allocation of more land to perennial crops, and there is some indication – not fully robust – that is has also increased yields.

Table 5.7 Two way fixed effects estimation of yields (perennial and temporary crops)

Coefficient, St Error, # Adj. Significance Neighbors PMV Imazon10 Blacklist Years R2 Perennial 1 0.463 0.943 ‐1.227 2002 – 2014 0.611 yield 0.195 0.456 0.224 ** ** *** Perennial 5 0.249 0.757 ‐0.715 2002 – 2014 0.581 yield 0.162 0.378 0.195 --- ** *** Temporary 1 0.247 ‐0.778 ‐0.025 2002 – 2014 0.449 yield 0.151 0.354 0.174 --- ** ‐‐‐ Temporary 5 0.017 ‐0.748 ‐0.087 2002 – 2014 0.567 yield 0.089 0.208 0.106 --- *** ‐‐‐ ‐‐‐ not statistically significant, * p < 0.10, ** p < 0.05, *** p < .001

Table 5.8 Two way fixed effects estimation of shares (% area) of municipality allocated to perennial and temporary crops Coefficient, St Error, # Significance Neighbors PMV Imazon10 Blacklist Years Adj. R2 2002 ‐ Perennial 1 0.001 0.001 ‐0.001 2014 0.738 Share 0.001 0.001 0.001 ** 2002 ‐ Perennial 5 0.002 0.0003 0.00001 2014 0.771 Share 0.0005 0.001 0.001 *** 2002 ‐ Temporary 1 ‐0.003 0.01 0.006 2014 0.912 Share 0.001 0.003 0.002 ** *** *** 2002 ‐ Temporary 5 ‐0.008 0.008 0.011 2014 0.873 Share 0.002 0.004 0.002 *** ** *** ‐‐‐ not statistically significant, * p < 0.10, ** p < 0.05, *** p < .001

5.2.3 Impact of bundled PMV + Imazon10 on agricultural yield estimated with SCM Given the finding that the PMV (but not the Imazon10) has an impact on perennial crops (yield and share of area) based on DiD estimation using both the full and matched samples, we tried SCM to estimate the impact of the bundled PMV + Imazon10. The Imazon10 is small for conventional matching and classical statistical tests, but there is a long enough time series on perennial crop yields to construct synthetic controls. Annex 4 presents plots of the synthetic controls, demonstrating that their yields in the calibration period are closer to those of the Imazon10 municipalities than a simple average of all municipalities in the donor pool, but nonetheless, in most cases, it is obvious that the fit is far from perfect. For example, the fit is particularly poor in Dom Eliseu, Novo Progresso, and Santana do Araguaia.

For the other municipalities with better quality synthetic controls, the apparent direction of Imazon10 impact on perennial yield is variable. For example, in Novo Repartimento and Santana do Araguaia, the Imazon10 appears to have increased yields. However, in Brasil Novo and Ulianopolis, it looks like the Imazon10 decreased perennial yields. We are not aware of any credible theory of change suggesting how participation in the Imazon10 could lead to decreased yields, and therefore we do not believe that much can be gained from interpreting these differences across municipalities.

In sum, while there are isolated cases suggesting effects, the additional support for municipalities in the Imazon10 does not appear to have a consistent effect on yield. We find the same for the share of a municipality allocated to perennial crops. Like yield, the share of municipalities allocated to perennial crops rises in both the Imazon10 municipalities and their synthetic controls, with no systematic difference apparent in the figures.

This mixed record of the Imazon10 effect on perennial crops revealed by SCM is broadly consistent with the regression results presented earlier. In those regressions, the estimated effects of the Imazon10 are generally not significantly different from zero, with some suggestion of a negative effect on the yield of temporary crops and a positive effect on the yield of perennial crops. When we allow the effects of the Imazon10 to vary across municipalities by applying SCM, we find some positive and some negative effects, consistent with an overall insignificant effect. One possible explanation is that in the 10 municipalities where Imazon initially concentrated its efforts, it was aiming to sustain the impact of the PMV, and any effect would be expected to emerge only over the long‐run.

ANNEX 1. Steps completed by municipalities in the PMV

Working group % eligible land Can issue Joined Pact against Município to combat registered in environmental PMV deforestation deforestation CAR licenses Água Azul do Norte 2010 não possui não possui 72.17 Não Almeirim 2010 em 25/01/2013 sem data 87.2 Sim Ananindeua 2010 não possui não possui 40.39 Sim Anapú 2010 em 05/02/2011 em 25/03/2011 79.69 Sim Brasil Novo 2010 em 26/04/2016 em 11/07/2013 86.57 Sim Brejo Grande do 2010 não possui não possui 67.91 Araguaia Não Canãa dos Carajás 2010 em 23/05/2015 em 16/05/2011 75.32 Sim Chaves 2010 não possui não possui 55.01 Não Cumarú do Norte 2010 em 16/07/2016 em 23/06/2015 80.91 Sim Eldorado dos 2010 em 19/05/2011 não possui 94.51 Carajás Sim Gurupá 2010 em 07/07/2016 não possui 81.01 Sim Igarapé-miri 2010 em 02/06/2016 não possui 62.77 Sim Irituia 2010 em 01/02/2012 não possui 55.62 Sim Marabá 2010 em 31/05/2016 não possui 86.57 Sim Medicilândia 2010 em 15/06/2016 em 06/06/2014 87.5 Sim Melgaço 2010 não possui não possui 54.25 Sim Novo Repartimento 2010 em 10/06/2016 em 15/07/2013 77.41 Sim Óbidos 2010 em 14/06/2016 sem data 85.93 Sim Ourilândia do Norte 2010 em 19/05/2011 não possui 87.72 Sim Pacajá 2010 em 27/06/2016 não possui 86.06 Sim Paragominas 2010 em 10/06/2016 em 23/03/2011 91.17 Sim Ponta de Pedras 2010 não possui não possui 55.72 Não Porto de Moz 2010 em 02/06/2016 em 17/08/2016 47.34 Sim Prainha 2010 em 20/02/2017 não possui 56.21 Sim Redenção 2010 em 13/06/2016 não possui 79.74 Sim Rio Maria 2010 não possui não possui 77.97 Sim Santa Maria das 2010 em 23/06/2016 em 26/04/2012 83.62 Barreiras Sim Santa Maria do 2010 não possui não possui 32.7 Pará Não Santana do 2010 em 20/07/2016 em 19/04/2012 85.43 Araguaia Sim Santarém 2010 em 20/03/2017 não possui 77,992.39 Sim 42

São Félix do Xingu 2010 em 26/08/2011 sem data 81.3 Sim São Geraldo do 2010 não possui não possui 79.15 Araguaia Sim Tucumã 2010 em 30/03/2012 sem data 98.01 Sim Ulianópolis 2010 em 27/04/2016 em 19/07/2012 90.11 Sim Uruará 2010 em 13/06/2016 em 18/12/2013 80.33 Sim Viseu 2010 não possui não possui 59.72 Não Vitória do Xingu 2010 em 19/09/2014 não possui 68.89 Sim Xinguara 2010 em 27/04/2011 em 29/05/2015 82.73 Sim Abel Figueiredo 2011 em 20/04/2011 não possui 45.04 Sim Alenquer 2011 em 16/08/2016 em 14/03/2017 78.28 Sim Altamira 2011 em 01/06/2016 em 22/06/2011 62.99 Sim Aurora do Pará 2011 não possui não possui 75.84 Não Aveiro 2011 em 28/06/2016 não possui 64.26 Sim Bannach 2011 em 01/07/2011 sem data 78.42 Sim Belterra 2011 em 16/07/2016 não possui 68.09 Sim Bom Jesus do 2011 em 02/06/2016 não possui 74.57 Tocantins Sim Breu Branco 2011 não possui não possui 71.31 Sim Cachoeira do Piriá 2011 não possui não possui 80.76 Não Conceição do 2011 não possui não possui 68.52 Araguaia Sim Curionopolis 2011 não possui não possui 84.54 Sim Dom Eliseu 2011 em 23/02/2016 em 03/09/2012 81.64 Sim Faro 2011 não possui não possui 79.76 Não Floresta do 2011 não possui não possui 59.88 Araguaia Sim Goianésia do Pará 2011 em 09/06/2016 não possui 74.48 Sim Igarapé-açu 2011 não possui não possui 44.14 Sim Ipixuna do Pará 2011 sem data não possui 78.23 Sim Itaituba 2011 em 29/07/2015 não possui 30.37 Sim Itupiranga 2011 em 06/06/2016 não possui 83.93 Sim Jacareacanga 2011 em 14/06/2016 em 21/09/2015 32.96 Sim Jacundá 2011 em 13/06/2016 em 25/08/2016 71.88 Sim Juruti 2011 não possui não possui 88.2 Sim Mãe do Rio 2011 não possui não possui 64.29 Sim Moju 2011 em 30/05/2016 em 23/08/2012 71.45 Sim Nova Ipixuna 2011 não possui não possui 69.83 Sim Novo Progresso 2011 em 07/06/2016 sem data 71.88 Sim Oriximiná 2011 em 18/05/2011 sem data 69.2 Sim Palestina do Pará 2011 não possui não possui 76.01 Não 43

Parauapebas 2011 em 01/10/2013 em 21/10/2015 69.61 Sim Pau d'arco 2011 não possui não possui 81.59 Não Peixe boi 2011 não possui não possui 36.59 Não Piçarra 2011 não possui não possui 76.97 Não Placas 2011 em 10/07/2014 em 22/09/2014 81.2 Sim Rondon do Pará 2011 em 22/02/2016 não possui 72.76 Sim Rurópolis 2011 em 01/03/2016 não possui 75.94 Sim Salinópolis 2011 não possui não possui 4.52 Sim Salvaterra 2011 não possui não possui 51.8 Sim Santa Izabel do 2011 não possui não possui 42.73 Pará Sim Santa Luzia do 2011 não possui não possui 60.33 Pará Não São Caetano de 2011 não possui não possui 18.68 Odivelas Sim São Domingos do 2011 não possui não possui 78.86 Araguaia Sim São Domingos do 2011 não possui não possui 62.77 Capim Não São João de 2011 não possui não possui 10.82 Pirabas Sim São Miguel do 2011 não possui não possui 38.84 Guamá Sim Sapucaia 2011 não possui não possui 88.88 Não Senador José 2011 em 14/06/2016 em 15/03/2012 78.78 Porfirio Sim Soure 2011 não possui não possui 69.21 Sim Tailândia 2011 em 09/10/2016 em 18/07/2013 74.47 Sim Terra Santa 2011 em 07/03/2016 não possui 53.89 Não Tome-açu 2011 não possui não possui 69.78 Sim Trairão 2011 em 02/06/2016 não possui 53.82 Sim Tucuruí 2011 em 28/06/2011 não possui 71.89 Sim Abaetetuba 2012 em 04/04/2016 em 22/10/2012 63.13 Sim Barcarena 2013 em 24/11/2015 em 02/07/2015 35.95 Sim Belém 2013 em 11/12/2015 em 10/12/2015 30.63 Sim Bragança 2013 em 14/05/2014 não possui 18.23 Sim Cachoeira do Arari 2013 não possui não possui 97.77 Não Capitão Poço 2013 não possui não possui 72.26 Sim Mojuí dos Campos 2013 em 02/06/2016 não possui s/i Não Ourém 2013 em 27/04/2013 não possui 51.87 Sim Terra Alta 2013 sem data não possui 42.01 Sim Augusto Corrêa 2014 não possui não possui 8.48 Sim Concórdia do Pará 2014 não possui não possui 65.07 Sim Muaná 2014 não possui não possui 45.75 Sim Curuá 2015 não possui não possui 81.5 Não 44

Maracanã 2015 não possui não possui 12.87 Sim Castanhal 2017 não possui não possui 48.63 Sim

ANNEX 2. Data description

A1. Donor pool

The donor pool was defined as all municipalities in the Amazon forest biome. Specifically, a municipality was included in the donor pool if (i) it was fully contained within the Amazon Biome boundary published by IBGE, or (ii) if it crossed or bordered the Amazon Biome boundary and originally at least half of its area was under forest cover. We referenced municipal boundaries from the 2013 IBGE file6, which we would further propose be held constant in future evaluations, to ensure that all future processing can be streamlined and will not require additional adjustments to historical boundaries. To define original, or historical, forest cover7 we used the INPE PRODES data to map all areas that were ever categorized as forest (Figure A.1). Based on these criteria, we identified an initial “master” donor pool of 519 municipalities (including treated and control) (Figure A.2).

6 IBGE is the data source, but the shapefle was provided by Imazon. 7 For historical forest cover, we used all areas that had been classified as forest (or deforested) in any year of the PRODES analysis from 1997 ‐; note that in 1997, the first year PRODES produced spatial data, areas that had been deforested prior to 1997 were also included in the deforestation category, thus allowing for a full reconstruction of historic forest cover. 46

Figure A.1: Historical (Original) Forest Cover

Figure A.2: Sample (519 municipalities considered in analysis)

A.2 Treatment and moderator variables

We identified whether each municipality was participating in each of our treatments (PMV and Imazon10) and the year that they began participating. In addition, we determined whether the municipality had ever been on the federal blacklist, including year that entered and (if relevant) year that exited. We estimate all treatment impacts conditional on blacklist status, e.g. the impact of the PMV on a municipality given that it is on the blacklist (or not on the blacklist). Because reducing deforestation below 40 km2 is a more stringent requirement for larger municipalities, we also consider municipality size as another moderator variable.

Blacklisted The official list of municipalities that have been blacklisted was obtained from the document “Lista de Municipios Prioritarios da Amazonia and Lista de Municipios com Desmatamento Monitorado e Sob Controle” 8 from the Ministério do Meio Ambiente. Those municipalities included on the list were

8 Source: MMA (http://www.mma.gov.br/) 48 identified in the database. Information on year entered and where relevant, year exited, was included in the database.

PMV Participation in the PMV was extracted from “Green Municipalities Program: Lessons Learned and Challenges for 2013/2014” (Appendices: Situation of Para municipalities in relation to goals of the GMP – up to January 2013) and where necessary, confirmed and updated using the PMV’s on‐line database.9 The year of treatment is the earlier of the “data de Assinatura do TC com MPF” or the “data de Assinatura do Termo de Cooperação com PMV” for each municipality (with years defined according to INPE’s convention for deforestation: August of the previous year through July of the current year). The fifteen municipalities that joined after our cut‐off date for treatment (1 August 2011) and that are therefore excluded from the analysis are: Barcarena, Bragança, Bethlehem, Cachoeira do Arari, Capitão Poço, Captain Wells, Concórdia do Pará, Curuá, Maracaña, Melgação, Muaná, Mojuí dos Campos, Ourem, São Sebastião da Boa Vista, and Terra Alta.

Municipalities were further broken down according to two levels of engagement in the PMV as explained in Section 2.1: “PMV on paper”, i.e. signing agreement with the PMV/MPF, and “PMV implementation”, i.e. implementing by 2015 three out of four identified key actions (celebrated a pact against deforestation, created a working group to combat deforestation, registered at least the median percent of land in the CAR, established at least a minimum structure for municipal environmental governance).

Imazon10 The treatment years for the ten municipalities in the Imazon10 are 2011 for Dom Eliseu, Tailândia, and Ulianópolis (based on the year that they started collaborating with Imazon under an initiative funded by the Fundo Amazônia); 2013 for Moju (when it substituted for another municipality); and 2012 for all others.

The covariates used in the analysis are described below, including information on the processing steps.

ENVIRONMENTAL VARIABLES

Slope and DEM Information on elevation and slope were calculated from a raster image of the shuttle radar topography mission (SRTM) which provides free high quality digital elevation maps at a spatial resolution of 90 m. The file was provided by Imazon. Zonal statistics were used to derive the min, max, range and mean for the slope and DEM for each municipality.

SOCIO‐ECONOMIC

9 Source: PMV http://www.municipiosverdes.com.br/

Roads Imazon provided us with a map of roads that they had manually digitized from satellite imagery. The data include information on unofficial roads in 2003, 2007, 2008 and 2010 across the Amazon Biome, excluding the states of Maranhão and Tocantins. The Estrada shapefile contained classes 1, 2 and 3 corresponding to the type of roads. Class 1 is for official roads, class 2 for unofficial roads (not in federal or state maps) and class 3 are roads in human settlements. The classes were split by year (2003, 2007, 2008 and 2010) and area was calculated as a 5 km buffer around all roads.

INCRA Settlements Imazon provided us with data on INCRA settlements. The INCRA shapefile contained information regarding the capacity for the number of families and the actual number of families resettled. They are contained in the attribute table under CAPAC and Benef. To calculate the capacity and benefit totals, the INCRA polygons were converted to points. This was done to reduce error and retain the count information from the polygons that crossed two or more municipalities. The information on number of families was assigned to the municipality identified in the attribute table. The area of INCRA settlements was calculated based on the area within each municipality using the 2013 boundaries. A total of 985 INCRA settlements crossed municipal boundaries and are also included in the database.

Indigenous Areas The shapefile with polygons for indigenous territories was intersected with the municipalities shapefile to derive the name of the municipality each indigenous polygon fell in. The shapefile was dissolved by the municipalities to calculate the area in sq.km for each municipality.

Protected Areas The shapefile with polygons for state and federal protected areas (including APAs) was intersected with the municipalities shapefile to derive the name of municipality each protected area polygon fell in. The shapefile was dissolved by the municipalities to calculate the area in sq.km for each municipality. All years were used in the calculations

CAR Registration Data on properties in the state of Pará that are CAR APPROVED and CAR PROVISIONAL were provided by Imazon. These files were intersected with the 2013 municipal boundary shapefile and the area in both categories was calculated in sq. km. for each municipality. To calculate the total CAR eligible area, the total municipal area in 2013 without the areas in federal and state protected areas, indigenous territories, or urban areas was calculated.

Embargoed Areas The area under embargo in each municipality was calculated. Only embargo areas from the years 2000‐2012 were included, although the data included a range of dates from 1963‐2013. This embargo shapefile was intersected with the municipality shapefile and dissolved by municipality and date. The resulting shapefile was broken down by year, from 2000 to 2012 and all other years including missing years in one shapefile. The area under embargo was calculated in sq km.

INDUSTRIES

Mines

A shapefile of the point location of mines was obtained from Imazon. The data had no information on quantity or type of extraction, so the count of total mines within each municipality was extracted.

Frigorificos (slaughterhouses) This shapefile shows the location of slaughter houses for cattle. The shapefile was spatially joined with the municipality shapefile to derive the count of slaughter houses per municipality and sum of cattle head. A scale from 1‐10 was created to provide some information on the size of the plant using the cattle head field. The scale was produced using the same logic as the polos madeireiro scale.

We also drew on a database compiled by Imazon to accompany the series O Estado da Amazônia, extracting the following variables:

Cattle Density of cattle in municipality (heads / km2). Percentage share of income going to top 10% of earners (percent of Income10 total income). PctHomesWithLand Percent of homes with own land and settled (percent). Per capita PIB (Produto Interno Bruto, or GNP) sourced from PIB_AG_PC agricultural enterprises ($R million). PIBPC PIB (Produto Interno Bruto, or GNP) per capita ($R) PopDen Population density (people / km2). Percent of PIB (Produto Interno Bruto, or GNP) sourced from PIB_PCT_AG agricultural enterprises (percent).

CONFLICT.PCT Percent of municipality identified as in conflict (percent). EDUCATIONHDI HDI for education using new methodology (HDI).

Finally, we gathered additional data from Brazilian government sources (including the Ministry of Agriculture and the Electoral Commission or Tribunal Superior Eleitoral) on the political party of the prefeito elected in 2008 (coded as binary variable indicating whether the same party as state governor, or PSDB); the density of rural properties (count of properties per km2); and whether the municipality was quarantined for foot and mouth disease (summarized as number of years free of foot and mouth disease in the decade of the 2000s).

All variables used in analysis are listed in table below:

Variable Name Description Years Source

MEANSLOPE Average slope (degrees). N/A

MEANDEM Mean elevation (meters). N/A Percent municipality area within 5km buffered area of class 1(official roads) and class 2 (unofficial roads) (percent of 2003, 2007, ROAD.class12.PCT municipality). 2008, 2010 Percent of municipality designated as an indigenous area (percent). PCT.INDIG.MUNI

Percent of municipality designated as a protected area (percent). PCT.PA.MUNI

Percent of municipality designated an INCRA area of INCRA.PCT resettlement 2000 ‐ 2012

PCT.CAR.PRO Percent of municipality provisionally designated CAR

Political party affiliation of preferito election winner (1 is POLITICAL.PARTY PSDB, 0 is all other) 2008, 2012

One or more slaughterhouses are present in municipality SLAUGHTERHOUSE (0 = none, 1 = one or more) 2005

MINES Density of mines in municipality (mines / km2).

Cattle Density of cattle in municipality (heads / km2). 2000‐2004 Imazon database

Density of rural properties in municipality (properties / NUMBER.OF.AG km2). 2006 2006 Ag Census

Number of years out of previous decade municipality was

FMD Foot and Mouth Disease‐free (years). 2000 ‐ 2012 http://www.agricultura.gov.br/legislacao/sislegis

Percentage share of income going to top 10% of earners Income10 (percent of total income). 2000 Imazon database

PctHomesWithLand Percent of homes with own land and settled (percent). 2000 Imazon database

Per capita PIB (Produto Interno Bruto, or GNP) sourced PIB_AG_PC from agricultural enterprises ($R million). 2000 ‐ 2003 Imazon database

PIBPC PIB (Produto Interno Bruto, or GNP) per capita ($R) 2000 ‐ 2003 Imazon database

PopDen Population density (people / km2). 2000 Imazon database

Percent of PIB (Produto Interno Bruto, or GNP) sourced PIB_PCT_AG from agricultural enterprises (percent). 2000 ‐ 2003 Derived from Imazon database

CONFLICT.PCT Percent of municipality identified as in conflict (percent). 2005 Imazon database

EDUCATIONHDI HDI for education using new methodology (HDI). 2000 Imazon database

Percentage of municipality area under IBAMA embargo PCT.EMBARGO (percent). 2000 ‐ 2012

ANNEX 3. Balance pre‐ and post‐ matching

2 Matching 2.1 Full set of variables 2.1.1 Nearest neighbour

Table 21: Pre-matching balance table

Mean treatment Mean control Std diﬀerence P-value blacklist ever 0.181 0.082 25.531 0.016 sharedefor2000 0.366 0.360 1.923 0.866 roadclass2008 0.215 0.074 70.500 0 pct area ind2009 0.038 0.030 6.188 0.573 pct area pa2009 0.165 0.213 −18.496 0.122 pct incra2009 0.006 0.004 22.446 0.050 polparty2008 0.076 0.060 5.971 0.582 numberofag2009 0.001 0.002 −111.629 0.004 fmd2009 0.419 0.471 −10.526 0.345 pibpc2003 4, 767.835 3, 776.471 23.012 0.032 popden2000 43.522 22.300 9.225 0.354 pibagpct2003 0.457 0.336 58.101 0.00000 education2009 0.747 0.745 3.168 0.791 pctembargo2009 5.957 0.402 17.960 0.069 cattleindex2008 574.608 2, 317.084 −51.553 0.065 N 105 365

21 Table 22: Post-matching balance table

Mean treatment Mean control Std diﬀerence P-value blacklist ever 0.181 0.181 0 1 sharedefor2000 0.366 0.323 14.773 0.003 roadclass2008 0.215 0.174 20.233 0.0003 pct area ind2009 0.038 0.033 3.547 0.312 pct area pa2009 0.165 0.177 −4.651 0.473 pct incra2009 0.006 0.004 25.121 0.0001 polparty2008 0.076 0.057 7.145 0.415 numberofag2009 0.001 0.001 −9.010 0.314 fmd2009 0.419 0.448 −5.763 0.257 pibpc2003 4, 767.835 4, 372.078 9.186 0.079 popden2000 43.522 31.333 5.298 0.220 pibagpct2003 0.457 0.421 17.535 0.002 education2009 0.747 0.754 −10.008 0.210 pctembargo2009 5.957 1.559 14.218 0.077 cattleindex2008 574.608 237.691 9.968 0.343 N 105 365

22 2.1.2 5 nearest neighbours

Table 33: Pre-matching balance table

Mean treatment Mean control Std diﬀerence P-value blacklist ever 0.181 0.082 25.531 0.016 sharedefor2000 0.366 0.360 1.923 0.866 roadclass2008 0.215 0.074 70.500 0 pct area ind2009 0.038 0.030 6.188 0.573 pct area pa2009 0.165 0.213 -18.496 0.122 pct incra2009 0.006 0.004 22.446 0.050 polparty2008 0.076 0.060 5.971 0.582 numberofag2009 0.001 0.002 -111.629 0.004 fmd2009 0.419 0.471 -10.526 0.345 pibpc2003 4, 767.835 3, 776.471 23.012 0.032 popden2000 43.522 22.300 9.225 0.354 pibagpct2003 0.457 0.336 58.101 0.00000 education2009 0.747 0.745 3.168 0.791 pctembargo2009 5.957 0.402 17.960 0.069 cattleindex2008 574.608 2, 317.084 -51.553 0.065 N 105 365

Table 34: Post-matching balance table

Mean treatment Mean control Std diﬀerence P-value blacklist ever 0.181 0.156 6.401 0.132 sharedefor2000 0.366 0.356 3.414 0.588 roadclass2008 0.215 0.175 19.816 0.002 pct area ind2009 0.038 0.029 7.780 0.163 pct area pa2009 0.165 0.154 4.410 0.533 pct incra2009 0.006 0.003 29.807 0.0001 polparty2008 0.076 0.051 9.289 0.158 numberofag2009 0.001 0.001 -25.986 0.021 fmd2009 0.419 0.442 -4.610 0.448 pibpc2003 4, 767.835 4, 192.437 13.356 0.027 popden2000 43.522 21.741 9.468 0.206 pibagpct2003 0.457 0.411 22.259 0.002 education2009 0.747 0.759 -17.083 0.072 pctembargo2009 5.957 0.842 16.536 0.070 cattleindex2008 574.608 393.004 5.373 0.654 N 105 196

33 ANNEX 4. Synthetic controls for impact of the Imazon10 on perennial yields

0.3 Yield: perennial crops

(a) Donors plot (b) SCM plot

Figure 17: Brasil Novo

(a) Donors plot (b) SCM plot

Figure 18: Dom Eliseu

10 (a) Donors plot (b) SCM plot

Figure 19: Moju

(a) Donors plot (b) SCM plot

Figure 20: Novo Progresso

11 (a) Donors plot (b) SCM plot

Figure 21: Novo Repartimento

(a) Donors plot (b) SCM plot

Figure 22: Santana do Araguaia

12 (a) Donors plot (b) SCM plot

Figure 23: Tailandia

(a) Donors plot (b) SCM plot

Figure 24: Ulianopolis

13 ANNEX 5. State budget for the PMV

When the PMV was first launched, program administrators had only a small budget under their direct control and therefore focused on influencing and coordinating other state agencies, the federal government, and civil society. Over the past three years, the PMV has established its own infrastructure and own budget, with funds from the state government and from Fundo Amazônia. The program is now administered through both the governor’s office (Casa Civil) and a program office (Núcleo Executor do Programa Municípios Verdes) jointly funded by the state and a grant from Fundo Amazônia. The official “on‐budget” cost of the program as administered by the state government was roughly one million dollars per year in 2015 and 2016, approximately half for purchases of equipment (vehicles, local offices, computers, GPS units, cameras) with funds from Fundo Amazônia. In order to calculate this total cost, we assumed that equipment will last for 5 years and therefore attributed 1/5th of the cost of all equipment purchased in 2015 and 2016 to each of those years. Other major assumptions include that (1) the state budget for personnel is roughly the same as the personnel budget funded by Fundo Amazônia (because the Casa Civil does not report disaggregated personnel costs), and (2)e th exchange rate was USD0.32/BRL in 2015 and USD0.31/BRL in 2016. Based on these assumptions, we estimated the total cost of administration and operations to be USD975,530 in 2015 and USD1,009,520 in 2016. If the equipment actually lasts longer than five years, without a substantial increase in maintenance costs, our budget numbers are overestimated. The PMV is currently conducting an inventory of its equipment, which will allow more accurate treatment of those costs. In the previous report, we also estimated the cost of the PMV to selected municipalities, and discussed costs borne by other state and federal agencies. However, we have concluded that the most relevant number for other states considering whether to scale up the program is the total cost of administration and implementation by the state government.