Commodity Prices and Robust Environmental Regulation: Evidence from Deforestation in Brazil

Increasing international agricultural commodity prices create pressure on tropical forests. We study the effectiveness of three regulatory policies implemented by Brazil in reducing this pressure: blacklisting of municipalities, the Soy Moratorium, and conservation zones. We use a triple difference approach that combines international agricultural commodity prices with the policies across three million km in the Brazilian Amazon. We find that the blacklisting program is effective, as it reduces deforestation related to the prices by 40%. The Soy Moratorium made deforestation in exposed municipalities more sensitive to nonsoy prices, in line with crop substitution. Conservation zones amplify the effect of prices on deforestation on the remaining unprotected land, consistent with reduced land supply. Our results highlight that the effect of environmental regulation depends on the economic pressure to use natural resources.


Introduction
Land-use change, largely due to tropical deforestation (Mitchard, 2018), is estimated to account for about 10-12% of anthropogenic CO 2 emissions in the years 2000-2015(Le Quéré et al., 2016Edenhofer et al., 2014). The backdrop of high deforestation rates has been strong global economic growth, high global energy prices, subsidies for biofuels and a doubling of the real price of agricultural commodities like grains (Mitchell, 2008;Alexandratos, 2008). Large scale agriculture accounted for about two-thirds of deforestation in Latin America and one-third in Africa and Asia in the period 2000-2012 (Kissinger et al., 2012). Around half of such deforestation can again be attributed to the cultivation of crops for export markets like the EU, China and North America (Lawson, 2014). In response, countries such as Brazil have committed to an array of command and control policies to reduce deforestation. The question addressed in this paper is whether such policies are effective in curbing deforestation related to higher commodity prices.
We evaluate the effectiveness of three central policy measures implemented in Brazil. The policies vary in terms of the deforestation they target. Blacklisting of municipalities (PM) targets municipalities with high deforestation rates by the means of increased monitoring and law enforcement as well as by more stringent conditions for subsidized rural credit. 1 This policy focuses on the total extent of deforestation at the municipality level. The Soy Moratorium (SM) is an industry-driven initiative that aims to keep the commodity supply chain clean of soybeans that come from recently deforested land. Hence it focuses on deforestation caused by soy cultivation. Conservation zones (CZ) impose regulation on certain geographic areas. In this paper, we include three broad categories of protected areas in what we call Conservation Zones, namely indigenous lands, sustainable use conservations zones and strictly protected conservations zones. We study the deforestation frontier in the Brazilian legal Amazon. This is the part of the Amazon, the largest forest left on earth, that is likely to have experienced the most intense deforestation pressure to date. Our main dataset is a balanced panel of 486 municipalities covering the years 2002-2013 and about three million km 2 . The main analysis focuses on deforestation outside of the protected areas.
We begin our analysis by estimating the direct effect of agricultural commodity prices on deforestation. We construct a municipality-level price index based on international real prices.
We use weights based on each municipality's cultivated area of the different crops in 2002, the initial year of our sample. Consistent with the finding of Hargrave and Kis-Katos (2013), we find that higher agricultural commodity prices are associated with higher deforestation.
We estimate that a 100% increase in the prices leads to an increase in deforestation of about 40%. The average 56% higher level of the price index over 2004-2013 compared to 2003 then contributes with 1,700 km 2 of additional deforestation each year. This adds up to about 19% of the total deforestation of 91,000 km 2 in our sample over the ten-year period 2004-2013.
Next, we estimate how the effect of international agricultural commodity prices varies with the policies, which represents the main contribution of the paper. We use the municipalityspecific index of prices interacted with policy exposure in a triple difference model (DDD). This model essentially compares price effects in municipalities exposed to a given policy with price effects in municipalities not exposed to the policy. Exposure to a policy varies both across municipalities and over time. We cannot reject common differential trends in deforestation in the pre-policy period, suggesting that our design effectively nets out potentially confounding factors driving both deforestation and the policy-roll out.
We find that the policy of blacklisting municipalities reduced the impact of commodity prices on deforestation by about 40%, saving 35 km 2 forest per treated municipality per year. In our sample, the total saved forest due to this effect is 9,000 km 2 . This is consistent with the expected effect that the policy increases the costs of deforestation. Previous studies have also suggested that this policy reduced deforestation. 2 For the Soy Moratorium, we do not find a robust statistically significant effect for the agricultural commodity price index. This overall ineffectiveness masks two effects working in opposite directions: the soy price has a lower effect on deforestation under the Soy Moratorium, while the prices of other crops have a higher effect. This is consistent with the Soy Moratorium reducing deforestation related to soy cultivation, while the production of alternative crops is moved to or expanded on newly deforested areas. Corn may be a case in point. We find 2 Arima et al. (2014) find that 10,653 km 2 of deforestation or 0.123 PgC of emissions were avoided over 2009-2011 in the targeted municipalities. Andrade and Chagas (2016) study spill overs of the blacklisting policy on non-targeted neighbouring municipalities and find a decrease of 15% to 36% in deforestation in the non-listed neighbours. Koch et al. (2018) also find reduced deforestation in priority municipalities, but no effect on dairy production or crop production.  find that the policy reduced deforestation by 40%, in period 2009-2010, and cut emissions by 39.5 million tons of carbon. PgC (petagrams of carbon) is the same as gigatonnes of carbon (GtC). The weight of CO2 is equal to 3.67 times the weight of Carbon, assuming that all the carbon is emitted. For more information on details of conversion of emissions measured in terms of carbon dioxide equivalent into carbon, see section 7. deforestation to be more sensitive to the price of corn due to the Soy Moratorium, potentially explaining some of the remarkable increase in corn production seen in the Brazilian Legal Amazon since 2006. We find that leakage to corn can explain about 20% of the leakage to non-soy crops. Our results suggest that studies of the Soy Moratorium that have not allowed for substitution across crops may have overestimated its effect on deforestation. 3 Finally, we find that conservation zones amplify the effect of agricultural commodity prices.
On average, the prices in the years after zone expansions were 40% higher compared to the years before zone expansions. This led to about 6,000 km 2 extra deforestation outside of the conservation zones compared to a situation without the zone expansions. One interpretation of our finding is that the deforestation could have continued into the new protected lands in the absence of the policy. The effects are similar if we include deforestation within the protected areas, which historically had low deforestation rates. Conservation zones take away land from the potential land supply and can thus increase the deforestation pressure on the remaining unprotected land. Our analysis, based on deforestation in non-conserved areas and explicit deforestation pressure, suggests that conservation zones have been less effective in reducing deforestation than existing studies have found. 4 What is the cost of reducing carbon emissions through deforestation? We use data on the initial spatial variation in biomass in combination with deforestation over time to estimate carbon emissions. Comparing these emissions with the average crop production values that could be generated on deforested land, we arrive at carbon prices of between 2.6 and 6.7 USD/tCO 2 . This is based on the unrealistic assumptions that all the carbon held in the cleared forest is emitted and that the mean crop yield per hectare captures the entire value of the additional agricultural activity. Both these assumptions are likely to imply that our calculated carbon prices are too low. Compared to other abatement technologies, our carbon prices do indeed suggest that reducing deforestation is a cheap abatement technology. For comparison, the High-3 Gibbs et al. (2015) find that deforestation for soy dramatically decreased due to the Soy Moratorium, while Nepstad et al. (2014) find only a marginal effect of the Soy Moratorium. Svahn and Brunner (2018) find that the Soy Moratorium reduced deforestation in the Brazilian Amazon biome, but only after it was enforced with satellite monitoring since 2008. 4 Assunção et al. 2015 find that about half of the avoided deforestation in the Brazilian Amazon over the period 2005-2009 was due to conservation policies. Soares-Filho et al. 2010 assign 37% of the reduction in deforestation in the Brazilian Amazon over the period [2004][2005][2006] to expansion of protected areas. Also Nolte et al. 2013 find that protected areas have contributed to reducing deforestation rates. Anderson et al. (2016) find that conservation zones are mostly located in areas where agricultural production is likely to be unprofitable. They find that zones reduce deforestation if the incentives for municipalities to reduce deforestation are high.
Level Commission on Carbon Prices suggested that a global carbon price of USD 40-80/tCO 2 in 2020 and USD 50-100/tCO 2 in 2030 could allow the goals in the Paris climate agreement to be met (Stiglitz et al., 2017). This paper makes two principal contributions to the growing literature on the drivers of deforestation and the effectiveness of policies against deforestation. 5 First, we focus on the effectiveness of policies explicitly accounting for the pressure to deforest, as expressed through international agricultural commodity prices. 6 Our analysis thus tests the robustness of environmental regulation when the pressure on natural resource use is high. A positive price shock resembles a positive shift in the demand curve for agricultural land. The priority list policy and the Soy Moratorium are expected to make the supply curve for agricultural land steeper, i.e. they increase the marginal cost of expanding agricultural land into forested lands (deforestation). A given price increase would then lead to a smaller expansion of agricultural land with the policy in place, compared to a situation without the policy in place. The conservation zones, on the other hand, are expected to shut down parts of the land market. The residual demand for non-protected land then increases, i.e. a given international price increase imposes a higher pressure on the remaining unprotected land. This results in a larger land expansion into unprotected lands with than without the policy in place. Deforestation pressure is discussed in the literature that tests policy effectiveness, e.g. Pfaff et al. (2014) and , but we explicitly bring in demand shocks. Based on our estimates, we graphically demonstrate that the effectiveness of a given policy measure in saving forest, measured in km 2 , depends on the agricultural commodity prices.  Barbier and Burgess (2001), Burgess et al. (2012), Burgess et al. (2017), Chomitz and Thomas (2003), Foster and Rosenzweig (2003), Gibbs et al. (2015) Pfaff (1999, Lopez and Galinato (2005), Rodrigue and Soumonni (2014), Rudel et al. (2005) and Hargrave and Kis-Katos (2013), as well as references therein.
6 There is large empirical literature which has analyzed various impacts of booming commodity prices on commodity-exporting economies, i.e. macroeconomic performance and fluctuations (Deaton et al., 1995;Fernández et al., 2017;Drechsel and Tenreyro, 2018), structural adjustment via Dutch disease mechanisms (Harding and Venables, 2016;Cust et al., ming) and conflict (Dube and Vargas, 2013;Bazzi and Blattman, 2014). 7 Focusing on the interaction between prices and policies also helps with econometric identification, i.e. separating out the effect of the price-policy-interaction from the effect of other factors potentially affecting land demand or land supply. Our specifications allow us to control for a large set of observable and unobservable characteristics, including rich heterogeneity in the effect of prices, and we present evidence that the effect of prices is similar across control and treatment municipalities in absence of the policies. Existing studies have used several approaches to deal with endogenous placement of policies.  use a measure of the tightness of municipal land constraints, which is defined as the share of land that is not legally available to farmers relative to total municipal land, in order to identify the effect of policies across municipalities. Their approach is based on the argument that policies are effective in places where land constraints for agricultural production are tight. Assunção et al. (2017) argue that satellite-based enforcement contributed to reductions in Second, this paper addresses the issue of policy ineffectiveness due to leakage (Aukland et al., 2003;Harstad and Mideksa, 2017). For the Soy Moratorium, we present evidence in support of substitution across crops, as the impact of non-soy prices increases under the moratorium. For conservation zones, we find increasing deforestation pressure due to prices when new areas are put under protection. In contrast, we find that the priority municipality policy is effective in reducing the impact of prices. Within municipality leakage thus reduces the effectiveness of the two policies that zoom in on specific sub-categories of deforestation, whereas the policy that targets deforestation irrespective of its source is effective at the municipality level. While the existing empirical literature has revealed leakage across space, e.g. Pfaff and Robalino (2017) on conservation zones and Gibbs et al. (2015) on the Soy Moratorium, we are not aware that the leakage due to substitution across crops has been documented previously.
The remainder of this paper is organized as follows. Section 2 presents the institutional context. Section 3 discusses the data, the identification strategy and tests of parallel differential pre-trends. Section 4 presents econometric estimates of price effects and how they vary with respect to policy exposure. Section 5 investigates the impact of soy prices versus the prices of other crops under the Soy Moratorium. Section 6 presents robustness checks. Section 7 presents calculations of implicit carbon prices. Section 8 concludes.
2 Background: Key anti-deforestation policies in the Brazilian

Legal Amazon
Our starting point is that agricultural profits are a major driver of deforestation. 8 Since 2004, Brazil has implemented a set of command-and-control policies to avoid the high deforestation rates it experienced in the 1990s and early 2000s, which to a large extent were related to expansion of commercial agriculture. Deforestation on private lands is governed by the Forest Code (FC), which establishes a percentage of rural properties that need to be preserved in the deforestation rates and use cloud cover as an instrument.  use a 2008-change in access to rural credit lines conditional on farmers' environmental compliance in order to show that this policy reduced deforestation rates in municipalities where cattle ranching is a dominant economic activity.
8 But commodity prices may carry not only information about current land use opportunities (forest vs pasture) and manifest through changes in current agricultural profits, but also through expected revenues from future land uses. The latter effect manifests itself through a speculative component of the value of the land. In this paper, we do not differentiate between the effects on deforestation caused by either current or future land opportunities. Blacklisting/priority municipalities policy (PM) was the main component of the second phase of the PPCDAm, launched in 2008. The policy defined a list of 36 municipalities to be prioritized in monitoring and law enforcement due to their high deforestation rates. The priority municipalities were subject to more intense environmental monitoring and enforcement as well as to a number of other administrative measures, such as more stringent conditions applied to the approval of subsidized credit. These measures have increased forest conversion costs and thus reduced incentives to deforest. 9 This group of municipalities accounted for 45% of the deforestation in the Brazilian Amazon in the year before the policy was implemented.
More municipalities were added to the list later. During 2011-2013, eleven municipalities were allowed to leave the list due to a remarkable decline in deforestation. In the data section below we describe in more detail the variation in our sample.
The Soy Moratorium (SM) reflects intensive campaigning by nongovernmental actors and private sector's willingness to adopt sustainable land-use practices. Soy has been Brazil's most profitable crop, with most of it going to exports; 33% in 1996 to 69% in 2004 and to 75% in 2013 (Karstensen et al., 2013;Lawson, 2014). A rapid expansion of soybean plantations on forested lands combined with the strong link to downstream markets in the EU and North America raised international awareness and increased the pressure on soybean producers to reduce deforestation. This led to the announcement of the Soy Moratorium in 2006. Buyers 9 In addition to a more stringent system of monitoring and law enforcement, they also became subject to a series of other measures, not officially established through legislation, such as compromised political reputation of mayors (Abman, 2014), politicians pressuring farmers to comply with environmental legislation. Priority status is determined based on: (a) total deforested area; (b) total deforested area over the past three years; and (c) increase in the deforestation rate in at least three of the past five years. The upper map on the right-hand side of figure 1 shows that these municipalities are mainly located in the southern part of the Amazon region, along the arc of deforestation. who joined the Soy Moratorium banned the purchase of soybeans planted on farmlands cleared after June 2006. The SM was extended to remain in place indefinitely in May 2016. The Soy Moratorium increases the costs of producing soy on newly deforested lands and thus increases the relative attractiveness of alternative uses of deforested lands, which can lead to substitution from soy to other crops.
Conservation zones (CZ) expanded significantly in the Brazilian Legal Amazon in the early 2000s, especially during the first phase of PPCDAm. The areas that we name "conservation zones" in this paper include three types of protected areas: strictly protected areas (SP), sustainable use zones (SU), and indigenous lands (IL). 10 The policy of conservation zones takes away land from the potential land supply, and is thus expected to increase the value of, and the deforestation pressure on, the remaining unprotected areas.
CAR The government has made significant progress towards increasing enforcement of the Forest Code (FC) through mapping properties for environmental registration, first with a number of state-level systems in the Amazon, and more recently with a national "SiCAR" system. 11 The national system was finalized and became operational after 2013, when our sample period ends. However, CAR systems have been used in the zero-deforestation cattle agreements (Gibbs et al., 2016) and the Brazilian Central Bank's (BCB) rural credit policy, mentioned below . Two states, Mato Grosso and Pará, had the most developed state-level property registration systems preceding the SiCAR (INPE, 2015). To make sure that our results are not affected by factors correlated with the property registration, we take into account the area of properties registered in CAR in robustness checks.
Credit In February 2008, the Brazilian Central Bank published Resolution 3545, which conditioned the concession of rural credit for agricultural activities in the Amazon biome upon proof of borrowers' compliance with legal titling requirements and environmental regulation.
Resolution 3545 applied to all rural establishments within the Amazon biome. It was obligatory for all banks and credit cooperatives to implement the terms of the resolution as of July 10 In SP: harvesting of trees or settlements are prohibited completely. In SU zones, extraction of forest resources as well as logging are permitted subject to a sustainable management standard Verissimo et al. (2011). IL are federal territories which are in the permanent possession of indigenous populations, who have exclusive rights to use the natural resources. 11 Sistema Nacional de Cadastro Ambiental Rural, SiCAR 2016. forest or a price index equal to zero. In addition, we drop municipalities with average forest cover below the 1st percentile and above 99th percentile. In our baseline sample, we focus on municipalities located within the forest frontier, the "arc of deforestation," which are to a large extent located along the transition from the Amazon to the Cerrado (tropical savanna) biomes (Levy et al., 2018). Historically, the deforestation in Brazil started in the south east and has swung in the north-western direction over time. The smooth lines in the upper left map of  Note: Maps show, clockwise from the upper left: 1) the forest frontier together with accumulated deforestation over 2002-2013; 2) the municipalities on the priority list; 3) municipalities exposed to the Soy Moratorium as they planted soy in 2005; and 4) the three types of protected lands included in this paper's "conservation zones".
Carbon data We use biomass data from Baccini et al. (2017) and obtain the carbon stock in the year 2000 at the 1 km 2 grid-cell level (C 2000 ). For each grid cell, we calculate the carbon stock in year t as the remaining forest, F t , times the carbon density of the forest in that grid cell in year 2000: C t = F t * C 2000 /F 2000 . Analogously, we calculate the carbon flow as deforestation, DF , times the carbon density in year 2000: DC t = DF t * C 2000 /F 2000 . We recalculate the carbon to CO 2 , i.e. multiply the carbon figures by 44/12. We thus assume, for simplicity, that all the carbon in the cleared forest is turned into omitted CO 2 , which is unrealistically high as, for example, some forest may be used as building materials. The right panel of Figure 2 presents the loss of CO 2 over time in the two samples. To further simplify the cost-benefit analysis, we convert carbon to dollars by value the CO 2 to 50 USD per tonne (2020-prices). This a simple and seemingly not unreasonable estimate for the social cost of carbon in 2020 (see for example Howard and Sylvan (2015)).
When we calculate the implicit price of carbon in section 7, we also add an estimate of sequestration, i.e. the trees could have absorbed carbon continuously if they were kept standing. Hubau et al. (2020) estimate that "intact old-growth tropical forests" in Amazonia sinks about 0.4 ton Carbon per hectare per year, which corresponds to about 1.5 tonnes of CO 2 per hectare forest, or USD 75 at 50 USD/tCO 2 . Our estimates for deforestation and carbon imply CO 2 values between USD 21,000 and USD 43,000 per hectare. The sequestration thus adds only 0.2-0.4% of the carbon stock per hectare per year. 12 Standing forests provide benefits beyond carbon capture and storage, e.g., biodiversity, that we do not pick up with our stylized carbon valuation. If the forest were allowed to grow back instead of the area being turned into nonforest permanently, regrowth of new forest could mean higher absorption of carbon than the previous forest. In our context, this is likely to be rare as we focus on the effect of agricultural commodity prices on deforestation.
Data on production values in agriculture IBGE provides data on annual production value for each crop at the municipality-level. We deflate these values with the deflator used by the World Bank in their Pink Sheet, i.e. the same deflator that is used to deflate our agricultural commodity prices. We recalculate such that the figures are in real 2020-USD. 13 Clearly, there may be other economic benefits related to expanding the agricultural sector that are not captured by crop production values. We also ignore the sales of timber.

Priority municipalities The Brazilian Department of the Environment, Ministério do Meio
Ambiente, MMA, publishes the list of municipalities with a "priority" status, including the date they entered the list. The upper right part of figure 1 shows the listed municipalities. In our sample, a total of 50 municipalities were blacklisted. 33 got on the list in 2008, 8 in 2009, 7 in 12 Clearing a hectare of forest thus corresponds to the removal of a present value of about 750 USD per hectare in terms of lost carbon sequestration at a discount rate of 10%: 75/0.10. We assume that this is the difference between the sequestration of forest and the sequestration of the cleared land. Nauclér and Enkvist (2009) cites research suggesting that "biodiverse forests sequester more carbon than their monoculture equivalents".
13 For the commodity prices in nominal and real values, see Pink Sheet, World Bank. The base year in the deflator is 2010. We use the exchange rate 1BR=USD 0.60, as off 30 dec. 2010 (The Federal Reserve).We arrive at 2020-figures by using the accumulated inflation in the US since 2010, i.e. 18.8% (US CPI)) is also increasingly being planted on recently deforested land in the legal Amazon (Martinelli and Filoso, 2008).
We construct our municipality-specific price index as follows: where P jt is the international price measured in current $US of crop j at time t, normalized to 1 in year 2000. The weights w ij,2002 are calculated based on the size of the planted area of crop j in municipality i in 2002, the initial year in our sample. We use these predetermined weights to avoid that the price index itself is affected by the farmers' behavior during the period we study. The weights sum to one. When we use the soy, non-soy and corn prices separately in the context of the Soy Moratorium, we apply Equation 1 with weights based on 2005, the year before the introduction of the Soy Moratorium. As the weights are then for a subset of crops, they do not sum to one. We provide robustness checks with alternative weights, as described in section 6. Figure 3 presents the price indexes we use.  Note: Upper charts present indexes of the real international agricultural prices, which we combine with municipality weights based on cultivated area in 2002 to construct municipality specific price indexes. Lower left chart shows the average of the general price index across municipalities, with and without soy. Lower right chart presents the mean of the municipality specific price indexes for soy and corn separately.
Controls. We account for: (i) rural credit policy, by including the normalized total value of credit concessions in a given municipality in a given year; (ii) for overall level of stringency of monitoring and law enforcement, by using the log of the annual number of environmental fines applied at the municipality level in the previous year. 15 In addition, we perform a large number of robustness checks in section 6, where we also run robustness with respect to the CAR policy.

Identification strategy
In our empirical strategy we proceed in three steps. First, we estimate the effect of agricultural commodity prices on deforestation. Commodity prices have been used in the literature on conflict (Dube and Vargas, 2013;Bazzi and Blattman, 2014) and in the literature on the Dutch disease (Harding and Venables, 2016;Cust et al., 2019). To identify the direct effect of agriculture commodity prices on deforestation, we estimate various forms of the following equation: where DF it denotes the log of the sum of deforestation in municipality i in year t (August t-1 to August t). P a,it−1 is the log of the municipality-specific price index, with area allocated to the different crops in 2002 as the weights (see section 2). I i and I t refer to municipality and year fixed effects. The coefficient of interest, β 1 , is identified to the extent the error-term it is uncorrelated with P a,it−1 , which is plausible given the pre-determined weights and international prices. Standard errors are clustered at the municipality level.
Second, we estimate how policies aimed at reducing deforestation affect the deforestation's response to international commodity prices. We expand equation 2 with the policy exposure at the municipality level. This amounts to estimating a triple differences model (DDD). Formally, we estimate DDD-models of the following form: The main parameter of interest is β 2 (the triple difference estimate), indicating how the price-effect depends on the presence of the policy. In general, E dum i indicates whether the municipality is ever directly exposed to the policy. For simplicity, we define it as a dummy for all three policies. It takes one if the municipality is ever on the blacklist, the area devoted to soybeans in the year before the Soy Moratorium is larger than zero, 16 or there is an expansion of protected areas in our sample period.
For the blacklisting policy, it takes one if a municipality is on the blacklist in a given year and zero otherwise. For the soy moratorium, the policy treatment variable takes zero for the years before 2006 and then switches to the area devoted to soy production in the year before the moratorium was introduced. For the conservation zones, the policy treatment variable is the area allocated to conservation zones in any given year.
We include the interaction between the price and the ever dummy, E i , allowing for a different price effect across the control and treated municipalities in all years. A full DDD-model requires the price to be interacted with the post dummy. We use instead the more flexible specification of interaction between the price and the year dummies, to allow for a differential price effect across all municipalities over time. 17 We include interactions between the ever dummy and the year dummies, to flexibly allow for different trends between the treatment and control groups. Note that the policy treatment variable E i × A it is not collinear with these time-dummy interactions for the respective reasons: municipalities were put on the blacklist at different times; the area devoted to soy varies across municipalities; and the size and the timing of the conservation zones varies across municipalities. Finally, we include log of lagged forest cover, F it−1 , and its interaction with the price index.
To keep the model tractable, we estimate equation 3 separately for each policy. We present estimates where we include all three policies simultaneously in section 6. 18 We there also discuss threats to identification and show robustness to a host of controls and other policies.

Testing for parallel pre-trends
Our key identifying assumption is that, in absence of the policies, the treated and non-treated municipalities would have had the same difference-in-differences in deforestation with respect to high and low price exposure. This identifying assumption is untestable, but we follow Muralidharan and Prakash (2017) and use the pre-policy data in table 1. As indicated by the first row 16 In Table A.8, we use the log area soy planted in 2005 instead of treatment group dummy for the SM. 17 Note that we sometimes present the direct price effect too, β1, as one of the interactions with the year dummies is dropped. Note that β4 and λ2 are vectors of coefficients.
18 See Appendix Table A.9. in columns 1-3, we can reject parallel trends for the policies in a DD-specification, i.e. when we compare only across the control and treatment group. Bringing in the agriculture commodity prices in columns 4-6, however, we cannot reject common differential trends as seen by the triple interaction term in row 3.
In Tables A.5-A.7, we present pre-trend tests for 12 covariates. The coefficient on the triple interaction term is statistically insignificant in all cases, with the following few exceptions: the size of the area used for agriculture for the priority list policy; agriculture productivity and remaining forest for the conservation zones; and one or more credit measure for all three polices. However, we show in section 6 that our results are robust when we include any of these characteristics as controls.
The pretrend-tests presented above increase our confidence that our key identifying assumption is satisfied. Table 2 presents versions of equation 2, which confirm that higher agriculture commodity prices exert higher pressure on the forest. Column 1 simply includes municipality and year fixed effects in addition to the municipality specific price index. Column 2 adds time trend interaction with the price and column 3 adds time fixed effects interaction, lagged forest cover and interaction between the price and the lagged forest cover. The results show that a one percent increase in the price index increases deforestation by 0.47 percent. As the level of the price index over 2004-2013 was on average 56% higher than in 2003, this estimate implies that the annual deforestation was on average 23% higher than it would have been with the 2003-prices. The higher prices led to about 3.7 km 2 higher annual deforestation per municipality on average, corresponding to a total of about 17,000 km 2 across the 486 municipalities over the 10 years (see Table A.4). 19 The upper left panel of Figure 4 presents the estimated relationships between percentage increases in the price index and percentage increases in deforestation, with the observed price increase of 56% indicated with the vertical dashed line. 20

Agricultural Commodity Prices and Policy Impact
Our main question is whether the priority municipality list (PM), the Soy Moratorium (SM), and conservation zones (CZ) reduce the pressure of higher commodity prices on deforestation. T reatGr × Active × L.P rice). We present the total price effect with and without the policy in the two bottom rows of the table, together with the difference between them and the p-value for the hypothesis test that this difference is equal to zero. Figure 4 shows the total price effects with and without the policies and illustrate the main point of this paper: the effect of the 19 To compute the overall level of deforestation, we multiply the average reduction in deforestation due to the higher prices (∆Y cf ) with the total number of treated municipalities over the period of the policy (N ). 20 In their global study, Busch and Engelmann (2017) find similar price elasticities as we do: "We estimated that every additional US$ 100 ha yr − 1 in potential agricultural revenue increased the rate of deforestation by an average of 0.98% in Latin America, 1.60% in Africa, and 2.42% in Asia, controlling for other factors -a variation of 2.5 across continents. Average potential agricultural revenues were $ 2978 ha yr − 1, $ 2304 ha yr − 1, and $ 3278 ha yr − 1 on each continent respectively, implying a price elasticity of supply of deforestation of 0.29, 0.37, and 0.79 for each continent respectively. Brazil's restrictive policies had the effect of reducing post-2004 deforestation by 47% for a grid-cell with average characteristics, due in part to decoupling potential agricultural revenue from deforestation." 21 The price effect is stronger in municipalities with lower levels of remaining forests, as shown in column 3. Such heterogeneity is not surprising given that our sample covers 486 municipalities and about 3.2 million square km. regulatory policies depends critically on the underlying deforestation pressure.
Comparing the effect of agriculture commodity prices on deforestation with and without the policies, as listed in the bottom rows in Table 3 With the policy in place, the price increase leads instead to a 149% increase in deforestation.
Using the actual observed deforestation for the municipalities in the treatment group over the period 2008-2013, the priority list saved 39 km 2 of forest in every treated municipality on average per year, which sums up to 10,177 km 2 overall (see Table A.4 for the details of these calculations). The upper right chart of Figure 4 illustrates how the policy contributes to avoiding large increases in deforestation when the price growth is high.
The Soy Moratorium does not have a statistically significant effect on how commodity prices affect deforestation, and the sign of the estimated coefficient actually suggests that the Soy  1 and included in the log-form. All area sizes used for the price-weights are measured at the municipality level for the year 2002, the initial year in our sample. Columns include policies as indicated in the column headings. Ever and Active is defined according to the policy type as described in section 3. Columns 1-3 include time trends interacted with the price index as well as with the ever-treated dummy. Columns 4-6 are based on Equation 3 and include interactions between the price and year dummies and interactions between the ever-treated dummy and year dummies. All columns include municipal and year fixed effects and the standard errors shown in parentheses are clustered on the municipality level. The bottom rows give the price effects, with and without the policy for the treated when relevant. The p-values are from an hypothesis test where H0 is that the effect listed above is zero. The marginal effects and the p-values are calculated with the margins package in stata.
Moratorium raised the deforestation pressure. This can also be seen in the lower left chart of Figure 4. We further explore the effects of the Soy Moratorium for the soy price, non-soy prices and the corn price in section 5.
Conservation zones amplify the price effect, which can be seen in the lower right chart of annual deforestation outside of zones by 6.1 km 2 per municipality or a total of 6,039 km 2 (see Table A.4). These results are consistent with zones taking away land from the land supply and Note: The figure illustrates the relative annual deforestation changes (Y ) at different relative price changes (X). The estimates are based on columns 3 in Table 2 and 4-6 in Table 3 and the graphs are based on the formulas shown below each chart. For the three policy-charts, the difference between the two lines is the treatment effect on the treated. Vertical lines indicate the actual average price changes observed for the treated municipalities between the pre-treatment period and the treatment period.
hence they increase the pressure on the remaining land. It is also possible that establishing conservation zones increases rivalry for remaining land and thus increases deforestation as a means of taking land into possession. 22 22 As mentioned in section 2, property rights in the Amazon are not well defined or defended. Thus, deforestation is still seen as a practice to obtain land titles which otherwise could be lost through invasion or expropriation Fearnside (2001). For completeness, we present estimates where the dependent variable is the deforestation within conservation zones only (column 6 in Tables A.10-A.12) and deforestation in the entire municipality (column 7 in Tables A.10-A.12). For deforestation inside zones, we do not find any significant reduction in the price effect. The results based on deforestation in the entire municipality are very similar to the baseline results.

Soy Moratorium and Different Crops
An important finding of this paper is that the Soy Moratorium does not reduce the impact of commodity prices on deforestation. This seems to stand in contrast to the influential study by Gibbs et al. (2015), which found that the Soy Moratorium is effective in reducing deforestation.
The authors studied the extent to which soy has been cultivated on newly deforested land after the Soy Moratorium was introduced. In this section we show that the Soy Moratorium reduced the responsiveness of deforestation to the soy price, but that this was counteracted by an increased responsiveness to the price of other crops.  table repeats column 5 of table 3, but with alternative prices: Column 1 is based on the area of soy planted times the soy price. Column 2 is based on the agricultural price index excluding soy, using the area sizes allocated to each crop as weights (following Equation 1). Column 3 is based on the area of corn planted times the corn price. All area sizes used for the price-weights are measured at the municipality level for the year 2005, the year before the soy moratorium was introduced. All the pricevariables are included in the log-form. All columns include municipal and year fixed effects. Standard errors are clustered at the municipality level. The bottom rows give the price effects, with and without the policy for the treated when relevant. The p-values are from an hypothesis test where H0 is that the effect listed above is zero. The marginal effects and the p-values are calculated with the margins package in stata.
In table 4, we present estimates of our triple difference model, again based on equation 3, for the Soy Moratorium under different commodity price indexes. In column 1 we use a soy price index, in column 2 a price index excluding the soy price and in column 3 a corn price index.
The negative and statistically significant coefficient of the triple interaction term in column 1 suggests that the Soy Moratorium significantly reduced deforestation related to the soy price.
The magnitude means that the policy reduced annual deforestation by 2.3 km 2 per treated municipality and by 2,656 km 2 in total (see Table A.4 for the details).
Column 2, however, indicates that the impact of non-soy prices on deforestation increased significantly in the presence of the Soy Moratorium. The deforested area increased by 5.1 km 2 annually per municipality and 5,847 km 2 in total due to higher prices of other crops. As a result, the net increase in deforestation due to the policy is estimated at 3,191 km 2 (Table A.4).
Corn is a non-soy crop that has experienced remarkable expansion in recent years. While corn was a minor crop in the Brazilian Legal Amazon in 2006, corn production has since then quadrupled and become the second most important crop in the Legal Amazon in terms of export share, after soy (IBGE, 2017). In recent years, soy and corn combined accounted for over 95% of the vegetable exports of the region (SECEX, 2017). Corn has been found to grow under the same climatic and geological conditions as soy, and substitution between soy and corn in the soy producing areas is thus feasible (Jantalia et al., 2007). The Soy Moratorium might therefore have contributed to corn expansions. Our estimates suggest that leakage to corn can account for 20% of the deforestation leakage related to non-soy crops. Specifically, the estimated elasticity of deforestation with respect to the corn price increased by 0.14. This led to 1.0 km 2 higher annual deforestation on average across the treated municipalities and a total of 1,143 km 2 in our sample (Table A.

Robustness Checks
In this section, we present robustness checks for the results presented in table 3 and table 4.
Alternative specifications for baseline models We first present our baseline model for the SM with a continuous treatment variable in Table A.8, and the results stay the same. We then present our baseline model with all three policies included simultaneously, in Table A.9.
We find qualitatively robust and consistent results for PM and CZ. For SM, the soy and corn price results become insignificant. The main results that the SM did not change the effect of the overall price index, but did make deforestation more sensitive to non-soy prices remain robust.
In Tables A.10 Agriculture The municipalities in our sample differ in terms of how developed their agriculture sector already is, which may affect the pressure to deforest further and the implementation of the policies. In columns 2 and 3 in tables A.10-A.15, we address this by controlling for lagged areas allocated to agriculture and for agricultural productivity. In column 4, we control for population. These controls do not affect the conclusions of this study.
Other policies The three policies we focus on in this paper may be correlated with other policy efforts implemented by Brazil, as discussed in section 2. In columns 5-9 in tables A.10-A.15, we control for agricultural credits, given to crop production or cattle production, or the stringency of monitoring and law enforcement measured as the number of fines issued by the environmental police. If anything, our results become stronger with these controls.
Furthermore, the three policies that we consider may in certain municipalities overlap with and complement each other. For example, Abman (2014) points out that international beef and soy companies withdraw from buying these commodities from municipalities with priority status, suggesting a channel through which the policy worked that is similar to the Soy Moratorium.
Seeking to insulate the effect of the different policies, column 10 in tables A.10-A.15 include the two other policies. Our results remain robust to these specifications.
Finally, column 11 in tables A.10-A.15 presents the results where we include as control variables the area registered in CAR at the municipality level interacted with year dummies.
The CAR-variable is based on the CAR-registry as published in 2016 and is time-invariant. The results are very similar to the main results.
Alternative specification In our baseline specification, we include the lagged remaining forest and its interaction with the price as control variables. The purpose is to account for heterogeneity related to the potential for deforestation and earlier development. When we exclude these controls in column 12 in tables A.10-A.15, we obtain similar results as for the baseline.
Sample size Our baseline sample excludes municipalities outside of the forest frontier, or the so-called arch of deforestation (Levy et al., 2018). To deal with the concern that Brazil is large enough to influence the world market price for soy, and hence potentially violating the assumption that the world price is exogenous to events in Brazil, we run robustness checks excluding the municipalities with the largest soy production in the Brazilian Amazon in column 2 in the lower panel of tables A.10-A.15. In our study period, those municipalities were responsible for up to 35% of the total Brazilian Amazonian soy production. In column 3, we exclude instead the 10% municipalities with the highest deforestation rates. Our results are robust to excluding either of these two types of large actors.
The Global Forest Change dataset In our analysis, we use Brazil's National Space Research Institute's (INPE) data on deforestation. The Brazilian government uses these data to monitor deforestation in the Brazilian Amazon. Richards et al. (2017) argue that the decision to use these data as a policing tool has incentivized landowners to find other ways to deforest and avoid compliance with Brazil's official monitoring and enforcement system. They provide evidence of divergence between PRODES and other deforestation indicators after 2008, which implies that INPE's dataset might overestimate the impacts of the policies on deforestation.
In column 6 "Hansen" in Tables A.10-A.15, we therefore use instead the Global Forest Change (GFC) dataset (Hansen et al., 2013). These data cover the entire municipalities. For comparison, we include estimates for deforestation "In zones" only in column 4 and for the "Entire" municipality in column 5. The "Entire"-column is thus comparable to the "Hansen"-column, and they show qualitatively the same effects as our baseline estimates, with two exceptions.
First, the SM is found to increase forest loss related to the general commodity price index when forest loss is measured with the GFC data. Second, and consistent with the first, the SM is found to not reduce the impact of the soy price on forest loss when measured with the GFC data. These two discrepancies are consistent with the observation of Richards et al. (2017) regarding adaptation of different deforestation patterns. However, one caveat is that we do not find similar deviations for PM. Another caveat is that the robustness results for SM with the soy price in general are statistically less robust than our other results. For completeness we note that all coefficients in the column "In zones" across Tables A.10-A.15 are statistically insignificant, which may not be surprising given that the areas within the conservation zones have seen low levels of deforestation.
Exogenous weights in the price index? As weights in the municipality specific price indexes, we use the share of agricultural land devoted to each crop. The weights need to balance two concerns. On the one hand, they need to be relevant and reflect the exposure of each municipality to the international commodity prices. On the other hand, they should be exogenous to unobserved factors determining deforestation and the policies. Our price index thus has similarities with Bartik instruments, which are created by interacting local shares To scrutinize the exogeneity of our price index, we show results for an agricultural price index weighted by potential yields (WPY) in column 9. Out of the 10 crops that we use in our baseline price index, we have the data to do this for 7 crops: Rice, Soybeans, Corn, Sugar cane, Banana, Citrus fruits and Cotton. Potential yields is a measure provided by the FAO GAEZ database, which calculates potential production based on geological and climatic conditions and is available at a pixel level. The data measure PY in kilograms per hectare for a crop in a given location and we constructed the mean PY at the municipality level by statistical zoning in QGIS. To create a sample-wide reference point for crop j, we calculated the mean PY across all municipalities: P Y j = N i=1 P Y j /N . We then calculated the relative PY for crop j in each municipality i: rP Y ij = P Y ij /P Y j . rP Y ij reflects the productivity of the soil in municipality i in producing crop j, relative to the average of the Brazilian Amazon. We use rP Y as the weights in the price index, which for municipality i over all crops j can be expressed as: P P Y,it = j rP Y ij * P jt , where P jt is our standard price from the world bank (set to 1 in 2000). As we use the log of the price index in the regressions, it does not matter that the weights do not necessarily sum to one.
Across all these three alternative weighing schemes, our results remain qualitatively and quantitatively stable (see columns 7-9 in Tables A.10-A.12). Again, the notable exception is the Soy Moratorium, for which the triple interaction takes a negative coefficient in all three cases but is (marginally) statistically significant only with Potential Yield weights.
Spatial correlation In our baseline specification, we cluster the standard errors at the municipality level to deal with serial correlation. In addition, there might be spatial correlation across neighbouring municipalities. 24 We follow Cameron et al. (2011) and use two-way clustered standard errors (on municipality and state-year) in the second most right column in the lower panel of tables A.10-A.15. In the most right column, we include instead state-year fixed effects as control variables. These robustness test do not change the conclusions of this study, although the triple interaction for the Soy Moratorium loses statistical significance for some of the separate price regressions.
Controlling for geographical characteristics. In our setting, characteristics that affect the profitability of agriculture may affect both the pressure to deforest and where the government choose to implement the policies. To test whether our triple-difference estimates pick up the effects of geographic characteristics that may affect the profitability of agriculture, we include interactions between the prices and the following five geographic characteristics: nutrient(1) concentration refers to soil fertility that is particularly important for low input farming; nutrient(2) concentration is particularly relevant for the effectiveness of fertiliser application; oxygen availability in the soil is particularly important for root development; root refers to soil volume limitations of a soil unit, affecting penetration and constraining yield formation; and access provides the estimated travel time to the nearest city with 50 000 or more inhabitants and plausibly also accounts for transportation costs. Tables A.18 and A.19 present   the results for the models in table 3 and table 4, respectively. We conclude that our results are robust and do not reflect variation in these geographic characteristics.

The implicit price of carbon
We have found that higher agricultural commodity prices increase deforestation and hence lead to carbon emissions. If the cleared land is used for crop production, the deforestation will also lead to higher agricultural profits. We now provide a stylized calculation of the implicit price of carbon emissions in our setting by comparing the value of estimated CO 2 emissions with the value of potential agricultural profits.
For agricultural profits we use data on the value of crop production per hectare, measured in 2020-USD. 25 We assume a 15% profit margin and a 10% discount rate, consistent with other 24 Municipalities within a Brazilian state do not only share geographical proximity but also political, legislative and cultural commonalities. 25 We use the mean value of crop production per hectare in a "plot-by-plot calculation". As an example, our studies (Busch and Engelmann, 2017). 26 For carbon loss, we consider the estimated carbon loss for one year. Emission of the carbon stock already stored in forest is a one time emission and is conceptually comparable to the present value of the increase in the profits from crop production. We add an estimate for sequestration, i.e. the carbon the forest could have sinked if it were kept standing, as described in section 3. Table 5 presents our baseline models, equation 2 and equation 3, with carbon loss as the dependent variable. There is a difference between carbon and forest simply because the forest vary in terms of carbon density across municipalities. 27 We find similar effects as for deforestation. In column 1, we estimate that a one percent higher price increases carbon loss by 0.93 percent. In columns 2-4, we find that PM reduces the price effect, CZ amplify the price effect and SM makes no difference on the overall price effect.
The estimated changes in CO 2 translate into a carbon price of 2.6 USD/tCO 2 for the general price increase. 28 Using the mean values of total production value in the treated groups, the corresponding figures are 6.2, 6.2 and 6.7 USD/tCO 2 for PM, SM and CZ, respectively. We think of these implicit carbon prices as a measure of the local social cost of carbon, i.e. what one would need to pay in compensation to make the farmers willing to abstain from an agricultural data suggest a yield of about 3 tonne soy per hectare per year and a price of about 400 USD per tonne. The present values are then 12,000 USD of the revenues and 1,800 USD of the profits. For the overall crop production, the value in our data is 1,473 USD per hectare per year over 2004-2013. With a 15% profit margin and a 10% discount rate, the present value is then 2,210 USD. These figures are in line with other studies. We use the total value of crop production, i.e. the ten crops included in this study. Instead of using the present value of the increase in crop value, we could have used the annual figure as a measure of how much one would need to pay to postpone the deforestation by one year; i.e. the "rental price" of the stock of carbon. 26 The data presented in Zalles et al. (2019) suggest profit margins in the range of less than zero to roughly 40% in Soy Bean production in Mato Grosso over the period 2000-2014. For comparison, McKinsey suggested in their Global Greenhouse Gas Abatement Cost Curve from 2009 a "PV for soy in intensive agriculture at 4% discount rate to be USD 3,000-5,000 per hectare" in South America (Nauclér and Enkvist, 2009). This corresponds to about 1,400 to 2,400 in 2020-USD with a 10% discount rate; we assume they used 2010 prices in the report and use the accumulated inflation of 19% in the US CPI from 2010 to 2020 to convert into 2020-USD. Busch and Engelmann (2017) report potential agricultural revenue of USD 2,978 per hectare per year in Latin America. With a profit margin of 15%, this amounts to a present value of about USD 4,500 per hectare given a 10 % discount rate. Langemeier and Purdy (2019) present examples on soy farming in major soy producing countries over 2013-2017. For soy bean production in Mato Grosso, a state in our sample, they find that the yield is about 3.25 tonne per hectare and gross revenue minus costs pr hectare is about 250 USD per hectare. If so, this would mean a PV of 2,500 USD per hectare with a 10% discount rate. 27 See section 3 for more details on the carbon data. For simplicity, we have measured carbon density in dollars, valued at 50 USD per tonne of CO2 in 2020-USD. 28 We use the estimates for deforestation and carbon loss reported in Tables A.4 and A.21 to arrive at these numbers as follows. If the entire estimated deforested area is used for overall crop production, the present value of the associated profits (per municipality-year) would be: 2, 210 U SD/ha×100 ha/km 2 ×3.7 km 2 = 817, 515 U SD, where 3.7 km 2 is the estimated change in deforestation per municipality-year. The value of the estimated loss in carbon per municipality-year is 16, 000, 000 U SD/50 U SD/tCO2 = 320, 000 tCO2, as we measure the carbon at 50 U SD/tCO2 (in millions). The ratio of the economic value to the carbon loss gives a local carbon price: 817, 515 U SD/320, 000 tCO2 = 2.6 U SD/tCO2. We follow the same procedure for the carbon prices associated with the policies. Note: Dependent variable is log carbon loss in the entire municipality. Column 1 is based on equation 2 and include interactions between the price and year dummies as well as interactions between the policy dummy and year dummies. Columns 2-4 are based on equation 3 and include interactions between the price and year dummies as well as interactions between the policy dummy and year dummies. All columns include municipal and year fixed effects and the standard errors shown in parentheses are clustered on the municipality level. Section 3.1 explains the carbon data.

expansion.
Combining our estimates for deforestation and carbon, we find an average carbon density of about USD 43,000 per hectare of forest cleared due to the general increase in agricultural commodity prices. For the PM, SM and CZ policies, the corresponding figures are USD 22,000, USD 18,000 and USD 24,000, respectively. Thus, the general price increase seems to have induced deforestation in areas with relatively high carbon density compared to the policies.
For given land returns, this contributes to a lower carbon price in the case of the general price increase and a higher carbon price in the policy cases.
Our stylized calculations are informative on the cost of avoided deforestation as an abatement technology. Our figures of 2.6-6.7 USD/tCO 2 are lower than some of the values in the literature. (2017)  Compared to their list, our calculations suggest that reducing deforestation in Brazil may be a cheap alternative.

Busch and Engelmann
Our calculated implicit carbon prices may be on the low side as the plot-by-plot approach may give a too narrow measure of the actual economic benefits of expanding agriculture. There may be important local economies of scale and linkages to other sectors. For example, the development of the agricultural sector is considered to have been a key driver of economic growth in Mato Grosso state (Richards et al., 2015). It could, of course also be that our calculated 30 Kindermann et al. (2008) also find that the cost of reducing emissions through avoided deforestation is lowest in Africa and highest in South East Asia, with Central and South America in between. They find that avoiding deforestation can be cost effective compared to other abatement technologies. 31 Both our economic valuation and carbon loss valuation miss potentially important aspects, as discussed in section 3. minimizing deforestation subject to a monitoring resource constraint. They consider constraints in terms of either the total area that can be monitored or the total number of municipalities that can be on the list. 32 The forgone economic value of protecting a given stock of carbon is another constraints that should be considered by the authorities to achieve economic efficiency.
Our calculations suggest that the three policies we study may have been well targeted.
Given the stylized nature of our calculations, the resulting implicit carbon price should be used with caution. We leave it to future research to evaluate the broader economic effects of deforestation, which will help in establishing the actual abatement costs of reduced deforestation.

Conclusion
Agricultural commodity prices may be high in the coming decades as growth in crop yields may stagnate due to climate change (Iizumi et al., 2017;Wiebe et al., 2015), as the use of land regulation policies may increase (Harstad and Mideksa, 2017), and as the world's population and incomes will increase (FAO, 2017). In this paper, we investigated the effectiveness of three command-and-control policies in protecting tropical forests confronted with higher agricultural commodity prices. We studied the Brazilian Legal Amazon, part of the world's largest tropical rainforest and a key supplier of agriculture commodities such as soy and corn to the world market. Our results showed that protection of specific areas (conservation zones) and targeting of a specific crop (Soy Moratorium) induce leakages within municipalities. Prioritizing entire municipalities in monitoring and law enforcement efforts (blacklisting) is, in contrast, effective in reducing deforestation related to international agricultural commodity prices.
We illustrated the implicit carbon price in our setting, using data on crop production values and carbon loss. Future analysis could investigate the links between deforestation and economic development. This could help local and international policy makers to weight deforestation against other abatement policies. It could also help in designing policies to dampen the negative local economic effects of reduced deforestation or to compensate local stakeholders through transfers.
32 They find that the carbon emissions were at least 8 percent higher than it could have been under their optimal list of municipalities. They also find, however, that selecting municipalities on the list randomly would result in 34 percent higher deforestation.

A Online Appendix: Extra Graphs and Tables
A.1 Descriptive statistics and effect sizes     Note: Table provides treatment effects on the treated, with (i = on) and without (i = of f ) the policy. %∆P = (Ppost/Ppre − 1) * 100 gives the price increase in percent; %∆Y i = (e β i ∆ln(P ) − 1) * 100 gives the increase in Y in percent;Ȳ is the mean of actual DF in the treated municipality-years measured in km 2 , i.e. with policy and price increase. Y i cf is the counterfactual DF, let γ i = %∆Y i /100: Y on cf =Ȳ (1+γ i ) gives the counterfactual Y in the absence of the policy and in absence of the price increase, while Y of f cf = Y on cf (1 + γ i ) gives the counterfactual Y in the presence of the price increase but without the policy; ∆Y cf = Y of f cf −Ȳ gives the difference due to the policy for the actual price increase and %∆Y cf = (Y of f cf /Ȳ − 1) * 100 gives the same in percent. β i estimates in upper panel are from columns 3 and 4-6 in Table 2 and Table 3, respectively. β i estimates in lower panel are from      Note: As columns 4-6 in table 3, but modified by including all policies at the same time. Interactions with time dummies are with the log general price index, as is the interactions with the variables for PM and CZ. For the SM-interactions, we use in column 1 the log general price index, in column2 the log soy price index, in column 3 the log non-soy price index, and in column 4 the log corn price index.       Note: As column 4 in table 3, but modified as indicated in column heading. Columns 1-2 priority municipalities, columns 3-4 soy moratorium, columns 5-6 zones, columns 7-8 soy moratorium with soy-prices only, columns 9-10 soy moratorium with non-soy prices only and columns 11-12 soy moratorium with corn prices only. Note: As column 4 in table 3, but modified as indicated in column heading. Columns 1-2 priority municipalities, columns 3-4 soy moratorium, columns 5-6 zones, columns 7-8 soy moratorium with soy-prices only, columns 9-10 soy moratorium with non-soy prices only and columns 11-12 soy moratorium with corn prices only. Note: As columns 4-6 in table 3, but including price interaction with geographical characteristic as indicated in column headings. See section 6 for a description of these characteristics. Note: As column 5 in table 3, but including price interaction with geographical characteristic as indicated in column headings. See section 6 for a description of these characteristics. Columns 1-5 for soy price only, columns 6-10 for non-soy prices only and columns 11-15 for corn price only. Note: R/ha is the mean value of crop production per hectare. "agriculture" refers to aggregates across the ten crops used in this study. All variables measured at the municipality level. See section 3 for more details on the data. Note: Table provides treatment effects on the treated, with (i = on) and without (i = of f ) the policy. Identical calculations as Table A.4 for the carbon estimates, where Y is carbon loss (measured in USD mn, at 50 USD/tCO2). P V 10 SEQ is the present value of carbon sequestration, if the forest were standing. P V 10 CO2 = P V 10 SEQ + ∆CO2. R/ha is the production value per hectare of the ten crops included in this study. P V 10∆Π is 15% of the present value of R/ha, i.e. profits with a 15% profit margin and a 10% discount rate. SCC local is the local price of carbon: DF (ha) × P V 10∆Π/(∆CO2 * 1000 * 1000/50). β i estimates from