Content uploaded by Fanny Moffette
Author content
All content in this area was uploaded by Fanny Moffette on May 07, 2021
Content may be subject to copyright.
Global Environmental Change 68 (2021) 102280
Available online 7 May 2021
0959-3780/© 2021 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Cattle ranchers and deforestation in the Brazilian Amazon: Production,
location, and policies
Marin Elisabeth Skidmore
a
,
b
,
*
, Fanny Moffette
a
,
b
, Lisa Rausch
a
, Matthew Christie
a
,
Jacob Munger
a
, Holly K. Gibbs
a
,
c
a
Nelson Institute for Environmental studies, Center for Sustainability and the Global Environment, University of Wisconsin-Madison, United States
b
Department of Agricultural and Applied Economics, University of Wisconsin-Madison, United States
c
Department of Geography, University of Wisconsin-Madison, United States
ARTICLE INFO
Keywords:
Deforestation
Cattle
Supply chain
Amazon
Brazil
Policies
ABSTRACT
Deforestation for cattle production persists in the Brazilian Amazon despite ongoing efforts by the public and
private sectors to combat it. The complexity of the cattle supply chain, which we describe in depth here, creates
challenges for the landmark Zero-Deforestation Cattle Agreements in particular and for enforcement of defor-
estation policies in general. Here, we present a holistic analysis that is increasingly relevant as the number of
policies, initiatives, and markets affecting the region increases. We provide the rst property-level analysis of
which ranchers decided to deforest in the last decade and identify the characteristics that are most related to
deforestation. We rely on newly available animal transit and property boundary data to examine 113,000
properties in the three major cattle-producing states in the Brazilian Amazon. We consider characteristics related
to a property’s role in the supply chain, location, land characteristics, and the policy environment. We nd that
deforestation is most likely to occur on properties that sell fewer cattle and earlier in the supply chain, are
located in remote locations, and have a high percent of remaining forest. Our results can be used to improve
enforcement of existing policies by targeting resources to properties and location where deforestation is more
likely.
1. Introduction
Deforestation of the Brazilian Amazon presents one of the greatest
challenges to global conservation goals. More than 780 thousand square
kilometers of forest has been lost in the last 30 years, causing nearly half
of Brazil’s carbon emissions and the loss of 2,000 species (de Castro
Solar et al., 2015; INPE, 2020; SEEG, 2018). Pasture expansion for cattle
production is the main driver of deforestation and has been linked to
80% of clearing (Global Forest Atlas, 2016). Thanks to the combination
of public policy and supply-chain sustainability initiatives, deforestation
rates declined signicantly between 2004 and 2014 prior to the increase
of recent years (Assunç˜
ao et al., 2015; Nepstad et al., 2014; INPE, 2020).
Yet, despite public and market-based pressure, deforestation for
expanded cattle production continues, and little is known about why
some ranchers continue clearing while others have stopped (Moffette,
2018; Assunç˜
ao et al., 2015). Here, we investigate a multi-dimensional
set of property characteristics to explain deforestation from 2010 to
2018.
Deforestation for cattle production frequently occurs in the forest
frontier, which is fraught with uncertain land rights, violence, and
corruption (Alston et al., 2000; Fetzer and Marden, 2017). Deforestation
has been documented at the hands of sophisticated producers that use
intensive methods of production, smallholders, and those engaged in
land speculation (Godar et al., 2015; Sparovek et al., 2019; Koch et al.,
2019; Miranda et al., 2019). The cattle supply chain includes fattening
and full-cycle farms (often characterized as direct suppliers), as well as
calving ranches and intermediate fattening properties (often character-
ized as indirect suppliers) and connement operations. These operation
types differ in terms of size, forest cover, sales patterns, and location,
among other characteristics, in ways that affect deforestation outcomes
(Andersen, 1996; Godar et al., 2015; Kaczan, 2020; L’Roe et al., 2016).
The cattle industry in the Brazilian Amazon is dominated by large
exporting meatpacking companies that supply both the domestic and
international markets. In 2009, Greenpeace published a report linking
the cattle industry to the deforestation of the Amazon, which caught the
attention of the international community (Greenpeace, 2009). Because
* Corresponding author.
E-mail address: mskidmore@wisc.edu (M.E. Skidmore).
Contents lists available at ScienceDirect
Global Environmental Change
journal homepage: www.elsevier.com/locate/gloenvcha
https://doi.org/10.1016/j.gloenvcha.2021.102280
Received 13 October 2020; Received in revised form 1 February 2021; Accepted 15 April 2021
Global Environmental Change 68 (2021) 102280
2
of the threat to their export markets, major meatpackers signed a rst
agreement with Greenpeace in 2009 to block deforestation from their
supply chains. Simultaneously, Brazilian public prosecutors also began
legal negotiations and signed legally binding “Terms of Adjustment of
Conduct” (TAC) with meatpackers to further pressure the sector. These
interventions together comprise the Zero-Deforestation Cattle Agree-
ments (hereafter CA), and because of the large overlap in their goals and
requirements, we refer to the collectively.
Under the CA, signicant progress has been made on monitoring
direct suppliers for deforestation, but broader impacts on forest con-
servation have lagged behind (Alix-Garcia and Gibbs, 2017; Gibbs et al.,
2015a; Gibbs et al., 2020). For example, a major meatpacking company
reduced its probability of purchasing from properties with recent
deforestation, on average, across the Amazon biome after the CA, but its
actions and adherence were inconsistent across states (Moffette, 2018).
Indeed, deforestation has continued to enter supply chains for all
meatpacking companies. The majority of deforestation in CA supply
chains occurs on unmonitored indirect suppliers that sell to direct sup-
pliers and on additional unmonitored properties that are owned by
direct suppliers (Skidmore, 2020; Klingler et al., 2018; Raj˜
ao et al.,
2020).
In addition to policies that target the cattle sector, other policies aim
at decreasing deforestation, leading to even greater pressure on cattle
ranchers. The Priority Municipalities List, which places additional sur-
veillance on properties in municipalities with high rates of deforesta-
tion, has signicantly slowed deforestation (Assunç˜
ao et al., 2019; Koch
et al., 2019; Cisneros et al., 2015) and indirectly impacted cattle
intensication (Koch et al., 2019; Moffette et al., 2020). Previous work
on protected areas (hereafter PA) found that the direct impacts vary
based on the type of protections (e.g. strictly protected, indigenous
lands, or sustainable use areas) (BenYishay et al., 2017; Anderson et al.,
2016; Pfaff et al., 2015; Amin et al., 2019), although Nolte et al. (2013)
estimate positive impacts across all types. Robalino et al. (2017) provide
some evidence that PA may lead to deforestation leakage when agri-
cultural presence surrounding the PA is high. The Brazilian Forest Code
(hereafter FC), which requires that properties in the Amazon be forested
and that conservation areas be established around riparian areas and
highlands, has not delivered clear results, as most land owners do not
believe it will benet them economically (Azevedo et al., 2017). The Soy
Moratorium, which blocks properties with new clearing for soy from
selling to two major buyers, has achieved avoided deforestation,
although it is mainly efcient as a complement to other policies (Heil-
mayr et al., 2020; Gibbs et al., 2015b). The Brazilian Rural Environ-
mental Registry, which requires landowners to register their property
boundaries with the federal government, and Terra Legal, which is a
program that aims to formalize land rights of farmers in the Amazon,
have been shown to affect deforestation (e.g. Alix-Garcia et al., 2018;
Lipscomb and Prabakaran, 2020), but are beyond the scope of this
paper.
We rely on newly available data to provide the rst property-level
assessment of how a wide variety of factors explain deforestation by
cattle ranchers in the Brazilian Amazon. We consider multiple public
and market-based policies as well as property characteristics that
directly can inuence deforestation or the effectiveness of policies. Pfaff
(1999) and Andersen (1996) took this strategy to describe county-level
deforestation in the 1970s and 1980s, noting that “policy makers must
understand the effects of the full set of potential drivers of deforestation
if they are to respond appropriately to such concerns” (Pfaff, 1999). In
the decades since, researchers have produced precise studies of indi-
vidual policies, increasingly using property-level data. Fine-grained, and
often causally identied, analyses are critical for the body of knowledge.
We complement these analyses by studying deforestation patterns based
on the premise that policies are not implemented in a vacuum and that
non-policy characteristics play an important role in a rancher’s decision
to deforest.
We build an holistic perspective of deforestation patterns in the
cattle supply-chain, a contribution that is increasingly relevant as the
number of policies, initiatives, and markets affecting the region in-
creases. We analyze and discuss the relationship between deforestation
and four broad categories of characteristics. First, we show for the rst
time how a property’s production and role in the cattle supply chain
correlate with the decision to deforest. Ranchers may have access to
more technical assistance or nancial resources based on their produc-
tion volume and whether they sell directly, which could inuence
whether they choose to clear rather than intensify production. Second, a
property’s location affects its access to slaughterhouses, input dealers
and extension services. Properties located closer to the forest frontier
also may receive less oversight from enforcement agencies, have fewer
nearby neighbors that could denounce deforestation activity, and
thereby have a lower risk of being penalized for deforestation. Third,
property land characteristics such as size, forest cover, pasture degra-
dation, and sustainable production capacity may inuence the need for
additional pasture as well as the ability to clear. Finally, the additional
properties owned by the same rancher (hereafter auxiliary properties)
may interact with the need for deforestation on the property that
formally sells cattle (hereafter main properties). Ranchers have more
exibility when selling to CA slaughterhouses if they own multiple
properties, as they can choose which property to sell from and thereby
subject to monitoring. Access to additional properties may also alleviate
the need for clearing on the main property by offering additional area to
expand pasture.
Here, we use three different statistical methods to test all of these
characteristics, as well as the policy context, and discuss which relate to
ongoing deforestation in the cattle supply chain. We contextualize our
statistical results with insights from eld interviews of key supply chain
actors. This analysis is now possible thanks to our database constructed
based on information from millions of cattle transactions linked to more
than one hundred thousand properties. Understanding which properties
continued to deforest can help target efforts to reduce new deforestation.
This information is more critical than ever, as progress is now at risk due
to stagnation of the CA, weakening of the Forest Code, and declining
transparency of public information and general government support of
anti-deforestation efforts (Escobar, 2019b; Escobar, 2019a; Ferrante and
Fearnside, 2019).
2. Materials and methods
2.1. Study area and sample
We study the states of Mato Grosso, Par´
a and Rondˆ
onia, which ac-
count for 80% of deforestation and 80% of the cattle production in the
Legal Amazon since 2000 (IBGE, 2018b; INPE, 2020). We identify
properties that sold cattle in the Amazon Biome portions of these states
using the Guide to Animal Transport (GTA), which tracks movements of
cattle between properties and to slaughterhouses. The Brazilian Ministry
of Agriculture, Food, and Livestock (MAPA) uses the GTA to ensure
animal health standards and track animal vaccinations record, in part to
satisfy international trading partners who want to guarantee the health
of meat or animals (Bowman et al., 2012). Federal law requires that
ranchers register GTA paperwork prior to moving animals between
properties or to slaughter (Law 12.097 and Decree 7.623) and it is
generally considered a reliable record of cattle movement (EU Direc-
torate General, 2011; Klingler et al., 2018; Raj˜
ao et al., 2020).
We use entity matching approaches to link the GTA cattle trans-
actions (ADEPAR´
A, 2019; INDEA, 2019; IDARON, 2018; MAPA, 2018)
with the property boundaries registered in the Rural Environmental
Property Registry (CAR), Terra Legal, and National Institute of Coloni-
zation and Agrarian Reform (INCRA) databases (SICAR, 2019; Minis-
t´
erio do Desenvolvimento, 2015; INCRA, 2017). We analyze all
properties that appeared in the GTA between 2013 and 2018 (2010 and
2018 in Rondˆ
onia) and that could be matched to a property boundary.
This corresponds to 113,008 properties matched to a property registry.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
3
Our sample of properties are linked to 47% of the transactions registered
in the GTA and 55% of the head transported during the study period. It
accounts for 50% of pasture, 29% of deforestation and 23% of total
forest area in 2018 on registered properties in the three states. Properties
in our sample sold more frequently and to slaughterhouses more often
than properties in the GTA that could not be matched to a property
boundary. Our sample properties sold an average of 29 times between
2013 and 2018 with 18% of transactions for slaughter, while unmapped
properties in the GTA sold an average of 16 times with 11% of trans-
action for slaughter. Thus, our sample does not achieve complete
coverage of the supply chain, and it over-represents large direct sup-
pliers. However, our sample does include properties across the full dis-
tribution of property sizes and sales volumes, and is thus well-suited for
our methods, which compare properties with and without deforestation
based on their observable characteristics. Further, to the best of our
knowledge, it is the rst time that a database about cattle-ranchers in the
Amazon with this level of completeness has been created and analyzed.
Additionally, we gather information from a set of 30,000 auxiliary
properties that appear in property map databases but that did not appear
selling cattle in the GTA during our study-period. To be considered
auxiliary properties, these properties must share owner and munici-
pality information with a property in our main sample. Since these
properties do not have the GTA buying or selling records that are
required for many of our variables, auxiliary properties are not part of
our main sample.
Fig. 1 shows a set of ve adjacent properties that are registered to the
same rancher but declared as separate properties in the CAR. In our
sample, the median distance from the main property to the auxiliary
property was 0.4 km and 40% of auxiliary properties were less than 0.01
km from the main property. This suggests that contiguous areas are often
registered under separate properties. Since the CAR system was imple-
mented for the monitoring of deforestation, it is not a system of legal
property title (Republic of Brazil, 2012). Consolidation is common on
the frontier, as sets of small properties are purchased to form larger
properties (Fearnside, 2008). While the 2014 instructions for imple-
mentation of the National CAR System (SICAR) require that consolida-
tions of previously registered properties must be re-registered as a
contiguous area with the CAR (Teixera, 2014), this is often not reected
in the reality, resulting in congurations as we show here.
When their properties are adjacent or nearly adjacent, ranchers may
move animals between the two without registering a GTA. This is due to
differences in the unit of observation between the CAR, which uses
property boundaries, and the GTA, which identies the establishment to
which an animal belongs (Portuguese estabelecimento propriet´
ario) for
purposes of contact tracing (Republic of Brazil, 2009; Republic of Brazil,
2011). A rancher might rightfully not consider movement within their
neighboring properties to be a change in establishment, thereby not
necessitating a GTA. Contiguous properties operated by the same owner
may clearly be considered a singled establishment; indeed, a rancher
may not think overmuch about movements across property boundaries,
as these movements may simply occur as cattle graze. Further, federal
law considers non-contiguous areas as a single agricultural establish-
ment if they are operated by the same producer, are located in the same
municipality, and share inputs (IBGE, 2018a), perhaps leading some
ranchers to not register GTA between non-contiguous properties. Other
ranchers register GTA between their properties; in these cases we can
thoroughly characterize both properties and include them as distinct
properties in our main sample.
If at time of sale the rancher consistently registers the GTA from the
same property, the others, the auxiliary properties, would not appear in
the GTA. Notably, we cannot identify or quantify animal movement
between auxiliary and main properties. Instead, we construct three
variables characterizing our main properties based on the full set of
properties owned by the same rancher. These variables capture how the
ownership of multiple properties (regardless of whether they are used to
produce cattle) inuences deforestation behavior on the main property.
2.2. Data source and description
To measure deforestation, we use the PRODES Amazon deforestation
maps from 2010 to 2018 (INPE, 2020). Our primary outcome of interest
is whether the property had any deforestation, which is equal to one if
the property had at least 6.25 hectares of deforestation in a single year
during that period (6.25 hectares is the minimum mappable unit in
PRODES). We focus on this binary measure as it is best suited to address
rancher’s behavior and the decision to deforest. By testing which
properties chose a zero-deforestation production during the period, it
can be used to target those properties for whom the existing pressures
did not stop deforestation.
As secondary outcomes, we consider total deforestation area from
2010–2018 and the inverse hyperbolic sine transformation of this total.
We exclude patches of less than 6.25 hectares from those outcomes since
they do not correspond to the minimum mappable unit in PRODES.
These continuous outcomes address the fundamentally different ques-
tion of what explains the most deforestation in the supply chain. Because
a large deforestation event is predicated on having a large remaining
forest patch, these results will be strongly inuenced by large properties
with high areas of remaining forest. Thus, it is useful for targeting
properties that account for high deforestation by area, but it may not
highlight the drivers of deforestation that are shared between small and
large properties.
We consider a set of 22 variables that characterize properties in our
main sample. The variables fall into ve categories: production and role
in the supply chain, policy context, property land characteristics, loca-
tion, and auxiliary properties. We compile these variables using the GTA
and other publicly available data sources. Table 1 describes the vari-
ables, their sources, and our hypothesis for the sign of the relationship to
deforestation. We discuss our hypotheses in detail below.
We expect that high-volume properties (those that sell a large
number of cattle) that are well-connected to slaughterhouses will have
less deforestation. Although indirect suppliers are included in the lan-
guage of the CA, they are currently unmonitored. Skidmore (2020)
showed that, among suppliers to the CA, properties were 42% less likely
to deforest if they sold directly to a slaughterhouse at least once. The
cattle supply chain is complex, including calving ranches, full-cycle
farms, indirect suppliers, connement operations, and traders. Many
of these suppliers may sell directly to slaughterhouses a few times
Fig. 1. Example map of a main property and its auxiliary properties.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
4
although their role in the supply chain remains mainly an indirect
supplier. Properties should receive more pressure from the CA, as well as
information about sustainability initiatives, as they interact with
slaughterhouses more. We therefore expect that the percent of head the
property sells directly to slaughterhouses and the number of slaughter-
houses they supply are negatively correlated with deforestation. High-
volume properties are also likely more inuenced by the CA as they
need to maintain market access to sell high volume of cattle; thus, we
hypothesize that cattle outow per year, inow per year, number of
transactions, high volume per transaction, and the number of active
years will be negatively related with deforestation. We consider sixteen
or more head per transaction to be a high average volume as (1) this is
the median value for head per transaction and (2) slaughterhouse trucks
typically hold 16 to 18 head, and a property that can ll a truck in a
single transaction likely has stronger standing at a slaughterhouse. In the
case that neighbors consolidated their cattle to a single truck, each
property should register an independent GTA record. State sanitation
agencies use GTA records to conrm the correct number of animals are
present on a property during their semi-annual vaccination campaigns,
and registering a single GTA per truck in the case of multiple suppliers
would lead to discrepancies for the properties involved. Properties with
high cattle outow per hectare should have less deforestation as they use
land more productively.
We believe that the likelihood of deforestation on ranching proper-
ties will depend on the policies that affect them. CA exposure, or the
proportion of federally-inspected slaughterhouses within the property’s
supply zone that have signed the CA, should be negatively correlated
with deforestation. We follow Alix-Garcia and Gibbs (2017) in dening a
supply zone as a 145 km radius around the property, as this corresponds
to the 75th percentile of the distances traveled by suppliers to JBS, the
world’s largest meatpacking company. Since competition with non-
monitoring slaughterhouses has been shown to undermine the effec-
tiveness of the CA (Moffette, 2018), there could be less deforestation
when local markets are dominated by CA slaughterhouses.
We further expect FC compliance to be negatively correlated with
deforestation. Unlike percent forest cover (a continuous variable that we
present in more details below), FC compliance is a binary measure of
policy compliance. Our statistical model distinguishes between proper-
ties with similarly high levels of forest cover but that have different legal
status because of their location on either side of a legal threshold (80%
forest cover). We expect properties that have this preferred legal status
(FC compliant) to be less likely to deforest in order to maintain that
status.
The presence of the property within a municipality on the Priority
List may also be negatively correlated with deforestation since most
studies that analyzed the impact of this policy, which started in 2008,
Table 1
Description of explanatory variables and their sources.
Category Variable Description Unit Year Source Hyp.
sign
Production and
supply chain
role
Direct supplier Heads sold directly to slaughterhouses as a percent of all
heads sold by the property
Percent of total
heads sold
2013–2018* GTA -
Number of
slaughterhouses
Average number of slaughterhouses they supply Slaughterhouses 2013–2018* GTA –
Outow per year Average volume of sales (cattle outow) per year Heads 2013–2018* GTA –
Inow per year Average volume of purchases (cattle inow) per year Heads 2013–2018* GTA –
High volume per
transaction
Whether average volume per transaction is 16 or greater 0/1 2013–2018* GTA –
Transactions per year Average number of transaction per year Transactions 2013–2018 GTA –
Outow per hectare Average outows per hectare of pasture** per year Heads per year 2013–2018* GTA, Mapbiomas –
Active years Number of years property buys or sells cattle Years 2013–2018* GTA –
Policy context CA exposure Average percent of slaughterhouses within 145 km of the
property that signed the G4 or TAC
Percent 2010–2018 Alix-Garcia and
Gibbs (2017)
–
Forest code
compliance (proxy)
Whether the property reaches the natural vegetation
requirement (% forest =80% in the Amazon biome)
0/1 2009 PRODES +/−
Priority List Percent of years the municipality was on the Ministry of
the Environment’s Priority List
Percent 2010–2018 Minist´
erio do
Meio Ambiente
(2018)
–
Distance to nearest
protected area
Distance to nearest protected area km 2018 MMA, FUNAI +
Property land
characteristics
Property size Property size Hectares 2010–2018 Terra Legal,
INCRA, SICAR,
CAR
+
Forest cover Forest/natural vegetation as a percent of property area Percent 2009 PRODES +
Degraded pasture Average percent of pasture area that is classied as
degraded pasture
Percent 2010, 2012,
2014
TerraClass +
Sustainable capacity Potential capacity Animal unit per
hectare
2017 LAPIG –
Soy/Crop Area of soy production Hectares 2010–2017 Rudorff and Risso +/−
Location Distance to nearest
highway
Distance to nearest highway km 2016 DNIT +
Distance to forest
frontier
Distance to patches greater than 1,000 hectares of
aggregated forest outside of a PA
km 2018 PRODES –
Auxiliary properties Multiple property
owner
Whether any of the owners of a given property owns an
additional property; owners are either listed on the
specic, selected property boundary or on the GTA
transactions
0/1 Details in
Appendix
Terra Legal,
INCRA, SICAR,
CAR
–
Soy/Crop on
auxiliary properties
Whether any auxiliary property owned by the rancher has
soy greater than 50 hectares
0/1 2010–2017 PRODES +/−
Deforestation on
auxiliary properties
Whether an auxiliary property owned by the rancher has
deforestation greater than 6.25 ha in a single year
0/1 2010–2018 PRODES +/−
*
We include years in which the property appears in the GTA with transactions either as a buyer or seller.
**
We include Mapbiomas pasture and mosaic categories.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
5
found that it reduced deforestation (Arima et al., 2014; Assunç˜
ao and
Rocha, 2014; Assunç˜
ao et al., 2019; Cisneros et al., 2015). However, due
to the targeting of the Priority List to municipalities with high rates of
deforestation and the non-causal nature of our analysis, we may observe
more deforestation on properties in Priority Municipalities. Finally, PA
are often located in less developed regions, and there is evidence of
spillover deforestation near PA when agricultural pressure nearby is
high (Robalino et al., 2017). We expect a property’s distance from a PA
will be negatively correlated with deforestation, although we cannot
disentangle potential spillover deforestation due to PA from the overall
higher rates of deforestation in remote areas.
We expect property size and forest cover to be positively correlated
with deforestation. Godar et al. (2015) found that deforestation occurs
in areas dominated by large properties, although they suggested this
trend may be reversing. On the forest frontier, deforestation rates are
high, and large tracts of highly forested land are acquired and cleared.
As an area transitions to a post-frontier status, properties are subdivided,
less forest remains, and deforestation slows (Rodrigues et al., 2009).
Poor quality pasture can motivate clearing; we expect soil degradation
will be positively correlated with deforestation, while the sustainable
capacity to produce cattle, measured by LAPIG as the number of head
per hectare the land can support, will be negatively correlated with
deforestation. Finally, the relationship between cattle production and
soy is complex. Highly productive farms frequently feed cattle with soy
to maximize productivity gains. This soy can be produced on- or off-farm
(Vale et al., 2019) and productivity gains from soy-based feed may lead
to less deforestation (Koch et al., 2019; Moffette and Gibbs, 2021). In-
tegrated crop-livestock systems also improve the productivity of existing
pasture, but ranchers may clear for new pasture as old pasture is
replaced with soy (Pereira et al., 2020). We therefore do not anticipate a
relationship in either direction. Finally, we hypothesize that properties
in remote areas have more deforestation. We specically measure
remote areas as properties that are far from highways and close to the
forest frontier. We dene the forest frontier as patches of forest greater
than 1,000 hectares outside of a PA. This identies regions with large
areas of privately held forest, which is common in new frontier regions
(Rodrigues et al., 2009).
Ranchers with auxiliary properties may expand pasture on those
properties, alleviating need for clearing on the main property. Current
CA monitoring focuses only on the property a rancher lists at the time of
sale, leaving ranchers with multiple properties free to choose which
property they sell from. This allows a system of intra-operational laun-
dering; a rancher can sell cattle through their “clean” property while
continuing to clear their auxiliary properties. Skidmore (2020) found
that unmonitored auxiliary properties are a major source of deforesta-
tion in the CA supply chain. Auxiliary properties may also have been
acquired more recently, and thus be the site of current deforestation.
Following these explanations, we anticipate that multiple property
ownership will be negatively correlated with deforestation on the main
property. Alternatively, in certain cases, auxiliary properties might serve
to compensate for lack of compliance on the main property under the
Brazilian Forest Code (Azevedo et al., 2017). If this is the case, multiple
property ownership would be positively correlated with deforestation,
as auxiliary properties would be used by ranchers with deforestation on
their main property to offset that deforestation. As the relationship be-
tween soy and deforestation for cattle ranchers is complex, we do not
have a clear hypothesis about the relationship between deforestation on
the main properties and production of soy on the auxiliary properties.
Similarly, if the auxiliary properties have been deforested, it is unclear
whether this will decrease or increase the likelihood of deforestation on
the main property.
2.3. Statistical analyses
Here our objective is to provide a broad understanding of the prin-
cipal characteristics that are associated with the decision to deforest.
Our strategy allows us to explore the broader patterns between all var-
iables, including those that are time invariant, and deforestation.
Because of the importance of the relationship of these variables with
deforestation, we choose to transform all variables to a single value and
use a non-time-varying approach. Our choice to average data over
multiple years is explained by three reasons. First, the objective of the
study is not nding the causal effect of any variable on deforestation.
Second, the use of panel data, even in a non-causal set-up, generally
benets from property xed effects; however these xed effects super-
sede the analysis of any variables that are time invariant within the
period. Moreover, a panel analysis would identify effects using small
changes in any explanatory variables that only vary slightly (e.g. pres-
ence on the Priority List, distance from a federal highway). Third, we
don’t have data about the cattle-supply chain commercial activities
before 2013 in Mato Grosso and Par´
a. As such, we averaged all time-
varying variables either from 2010–2018 or from 2013–2018, depend-
ing on the availability of data.
We rst compare properties with and without deforestation using
normalized differences, a scale-free measure of the difference in distri-
butions between samples. Normalized differences have two main ad-
vantages. First, in comparison to a standard t-test, the normalized
difference is not sensitive to the sample size. Second, the result of the
normalized difference presents the difference in means of the two
samples in terms of the standard deviation, which can be easily inter-
preted and compared across variables (Imbens and Wooldridge, 2009).
We test whether properties with and without deforestation were
signicantly different in terms of the 22 characteristics of interest. The
test reports the difference in means in terms of the standard deviation.
We dene signicant differences to be equal to normalized differences
greater than.05.
We further test which of the 22 characteristics are correlated with
deforestation status using regressions. Our main model is a linear
probability model that identies the characteristics that are most related
to the decision to deforest. The advantage of the regressions is to test the
correlation between a single variable and the likelihood of deforestation
while holding all other variables constant, thereby isolating the rela-
tionship between that single variable and deforestation. Specically, we
estimate:
Defori=
α
+β′Prodi+γ′Policyi+δ′Landi+ζ′Loci+
η
′Auxi+∊i.(1)
Here Prodi is the set of variables relating to the property’s production
and role in the cattle supply chain, Policyi the public and market-based
policy context variables, Landi the property land characteristics, Loci
the property location characteristics, and Auxi the set of variables
relating to the auxiliary properties owned by the rancher. All variables
are normalized to have a mean of 0 and standard deviation of 1 to allow
for comparison across variables. The coefcients can be interpreted as
the percent increase in likelihood that the property will have defores-
tation associated with a one standard deviation increase in the explan-
atory variable. Because the variables are normalized, a larger coefcient
(in terms of absolute value) demonstrates a stronger relationship with
deforestation. The constant
α
captures the average deforestation prob-
ability and ∊i is the standard error.
We compare the results from the linear probability model with a
linear LASSO and a logistic regression. To select the reduced set of
variables, LASSO integrates an additional parameter that penalizes large
coefcients with little explanatory power by minimizing or omitting
these coefcients. We select the value of lambda which minimizes the
cross-validation (Ahrens et al., 2018; Athey and Imbens, 2016). The
LASSO model produces estimates of the coefcients without standard
errors because this model’s objective is to select the variables that
enhance prediction accuracy. Omission from the model does not
necessarily mean a variable is uncorrelated with deforestation status,
since omitted variables may be correlated with variables that remained
in the model. Then, to verify that results are robust to relaxing the
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
6
assumption of linearity, we also use a logistic regression. Coefcients
from the logistic regression are presented in terms of marginal effects
and as such, their interpretation is the same as the coefcients of the
linear probability model (discussed above).
Correlation between variables can mask the core relationship be-
tween a variable and deforestation. Because of this, we investigate the
relationship between our 22 variables with a correlation matrix. The
matrix estimates the two-way correlation between every pair of vari-
ables. For example, properties with deforestation may counterintuitively
appear to be more likely to be in compliance with the FC. However, FC
compliance is correlated with the percent forest cover. Thus, the average
values of FC compliance for properties with and without deforestation
may not represent the relationship between compliance and deforesta-
tion without also controlling for forest cover. The linear model and
correlation matrix help us understand these complex relationships and
rene the interpretation of our results.
2.4. Field surveys
We complement our analysis with surveys of key actors in the cattle
supply chain. We conducted eight structured interviews in six locations
in the state of Rondˆ
onia using a convenience sample. Interviews were
conducted with one rancher, one local environmental public servant,
two leaders of local farmers’ associations, two agricultural extension
agents, and two university researchers. Respondents were allowed to
respond freely to a set of structured questions regarding determinants of
deforestation on cattle ranches. Interviews were conducted in October of
2019.
3. Results
We present the results of the normalized differences analysis and
linear regression models by variable category. For each variable, we
report the normalized difference,
α
, and the value of the regression co-
efcient, β. Means and standard deviations for the properties with
deforestation and the properties without deforestation, as well as the
normalized differences are listed in Table 2; we graph coefcients of the
signicant variables in the linear regression models in Fig. 2 and list full
results in SI Table B1. When a result is not statistically signicant, we
report
α
=n.s. or β=n.s. Because we normalize the variables in the
regression, we can interpret larger (absolute) values of β as a stronger
relationship with deforestation. We discuss these coefcients in terms of
the characteristics of properties with deforestation, implicitly in relation
to the base category of properties without deforestation. When appro-
priate, we connect these results with the information obtained from the
correlation matrix (indicating the value of the correlation coefcient,
ρ
)
or the eld survey to provide a more complete interpretation. We
conclude by presenting the results of an analysis of the drivers of
deforestation size.
3.1. Production and supply chain role
Properties with deforestation sold a lower proportion of cattle
directly to slaughterhouses (
α
= − 0.087,β= − 0.011). To provide vi-
sual evidence of this nding, we map the percent of heads sold directly
by suppliers in Fig. 3a,b. Fig. 3a, which maps properties with defores-
tation, contains very few properties that predominantly sell directly
(60–100% of head). Instead, it was most common for these properties to
sell 0–20% of head directly. However, Fig. 3b, which maps properties
without deforestation, shows that relatively more properties without
deforestation specialized in direct sales. Properties with deforestation
also supplied fewer slaughterhouses (
α
= − 0.137,β= − 0.017), and
sold fewer head per transaction (
α
=0.117,β=0.012).
The correlation matrix shows that the variables in this category are
Table 2
Characteristics of suppliers with and without deforestation.
No deforestation Deforestation Norm. diff.
(1) (2) (3) (4) (5)
Production and supply chain role
Direct supplier (% of heads) 15.818 (25.211) 12.879 (22.529) −0.087
Number slaughterhouses 0.343 (0.582) 0.241 (0.464) −0.137
Outow per year (heads) 121.509 (514.510) 125.861 (650.849) 0.005
Inow per year (heads) 112.466 (1140.866) 110.480 (505.025) −0.002
High volume per transaction (0/1) 0.441 (0.496) 0.523 (0.499) 0.117
Transactions per year 5.833 (17.287) 4.904 (15.148) −0.040
Outow per hectare 2.349 (17.877) 1.458 (14.235) −0.039
Active selling years 4.463 (2.129) 3.883 (1.967) −0.200
Policy context
CA exposure 0.582 (0.297) 0.401 (0.371) −0.380
Forest code compliance proxy (0/1) 0.031 (0.173) 0.060 (0.238) 0.100
Priority List 0.260 (0.416) 0.539 (0.469) 0.445
Dist. nearest protected area (km) 33.950 (26.998) 24.346 (21.570) −0.278
Property land characteristics
Property size (ha) 222.711 (987.856) 451.478 (2208.671) 0.095
Forest cover (% property) 2390.386 (2387.445) 4141.183 (2314.058) 0.527
Degraded pasture (% pasture) 10.200 (14.292) 13.827 (15.449) 0.172
Sustainable capacity (au/ha) 3.630 (0.614) 3.454 (0.721) −0.186
Soy/Crop (ha) 4.374 (101.027) 6.112 (102.591) 0.012
Location
Dist. nearest highway (km) 12.439 (17.010) 26.869 (28.251) 0.438
Dist. forest frontier 6.078 (6.918) 2.059 (3.916) −0.506
Auxiliary properties
Multiple property owner (0/1) 0.334 (0.472) 0.332 (0.471) −0.002
Soy/Crop on auxiliary properties (0/1) 0.041 (0.198) 0.048 (0.214) 0.024
Deforestation on auxiliary properties (0/1) 0.159 (0.366) 0.483 (0.500) 0.522
Observations 92983 20025 113008
Note: The rst column presents the mean value for those properties without deforestation and the third column presents the mean value for those properties with
deforestation. The second and fourth columns present the standard deviation in parentheses. The fth column shows the normalized difference.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
7
correlated (Fig. 4); for instance, the proportion of heads sold directly is
highly correlated with the number of slaughterhouses a property sup-
plies (
ρ
=0.63). These correlations show that properties can broadly be
described as active with many connections to slaughterhouses or less
active and often operating earlier in the supply chain. Our results indi-
cate that deforestation occurs on the latter: indirect suppliers selling
earlier in the supply chain, which sell fewer cattle more infrequently.
3.2. Policy context
Our results show that properties with deforestation existed under
different policy conditions than those without deforestation. Properties
with deforestation were less exposed to the CA (
α
= − 0.380,β=
−0.036). Because the location of slaughterhouses is self-selected by the
companies, it may indicate that major meatpackers locate slaughter-
houses in old frontier areas with higher population and larger markets;
these areas typically have low deforestation. Both choice of slaughter-
house location and whether a slaughterhouse signed the CA largely
predate our study; of the 70 slaughterhouses in our sample region and
period, only 6 were opened during the period, and 10 signed the CA
during the period. Instead, these location decisions may be driven by
favorable market and infrastructure conditions. Indeed, CA exposure
was negatively correlated with forest cover (
ρ
= − 0.271) and positively
correlated with distance to the forest frontier (
ρ
=0.212).
Properties with deforestation were more likely to be located in a
Priority Municipality (
α
=0.445,β=0.032). The Priority List explicitly
targeted municipalities with high rates of deforestation. The correlation
matrix shows that properties in Priority Municipalities had less exposure
to the CA, on average (
ρ
= − 0.315).
Although we control for other variables related to location, the non-
random location of CA slaughterhouses means that our results do not
demonstrate that the CA slowed deforestation. However, these results
are useful for targeting and monitoring; the properties that are already
very exposed to the CA are at lower risk of deforesting, and CA
slaughterhouses as well as public enforcement agencies should focus
their monitoring efforts on those properties in areas where the CA are
weak. Similarly, our strategy does not show that the Priority List led to a
decrease in deforestation, as others have shown (Assunç˜
ao et al., 2019;
Koch et al., 2019; Cisneros et al., 2015). In sum, we do not causally
identify the effect of the Priority List as these studies do; the positive
correlation we nd between the Priority List and deforestation is likely
the result of targeting of high-deforestation municipalities for entry on
the Priority List.
The normalized differences show that properties with deforestation
were more likely to be in compliance with the FC during the study period
(
α
=0.100). However, the linear model nds that, after controlling for
other factors, properties with deforestation were actually less likely to be
in compliance with the FC (β= − 0.029). This result highlights the
importance of including all variables together in the model; we nd the
opposite relationship between the FC and deforestation depending on
whether we include other correlated variables. In particular, FC
compliance was highly correlated with percent forest cover (
ρ
=0.47),
the variable that best explains deforestation. In the normalized differ-
ences analysis, we compare properties that were in compliance with the
FC, and thereby had more than 80% forest cover, to properties with any
amount of forest cover. Because continuous forest cover was correlated
with deforestation, we would expect this comparison to yield a positive
relationship between FC compliance and deforestation. In the linear
model, we compare properties on either side of the binary measure of
compliance after accounting for their forest cover. The negative rela-
tionship between deforestation and FC compliance here suggests that,
among properties with high forest cover, being above the threshold of
FC compliance inuences the decision to deforest.
Properties with deforestation were located closer to PA in terms of
normalized differences, but this result is not statistically signicant in
the regression model once we control for other location variables (
α
=
−0.278,β=n.s.). Property distance from a PA was statistically signif-
icantly related to its distance from a highway (
ρ
= − 0.22). Therefore,
the results of the normalized difference primarily highlight a relation-
ship between the remote location of a property and deforestation, rather
than identifying spillover deforestation caused by PA.
3.3. Property land characteristics
Properties with and without deforestation varied greatly in terms of
their land characteristics according to our empirical analysis and eld
interviews. First, properties with deforestation were larger (
α
=0.095,
β=0.017) and had more forest cover as a percent of the total area (
α
=
0.527,β=0.068). Additionally, properties with deforestation had a
higher percentage of degraded pasture (
α
=0.172,β=n.s.) and had a
Fig. 2. Results from our regression models. We graph
all coefcients with p-values smaller than 10% from a
linear probability model that correlates deforestation
with property characteristics. Positive coefcients
mean higher values of the characteristics are associ-
ated with more deforestation, while negative co-
efcients mean higher values of the characteristic are
associated with less deforestation. Coefcients larger
in an absolute value are more correlated with defor-
estation. Coefcients extracted from LASSO and lo-
gistic regression are also presented as robustness.
Results from the linear probability model and the lo-
gistic regression include the 95% condence interval.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
8
Fig. 3. Percent of heads sold directly to a slaughterhouse and location of slaughterhouses.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
9
lower sustainable cattle production capacity (
α
= − 0.186,β=n.s.).
However, neither the percent degraded pasture nor production capacity
were statistically signicant after we controlled for other variables.
We highlight the spatial differences in the percent forest cover in
Fig. 5. Fig. 5a demonstrates that properties that cleared between 2010
and 2018 often had 40% or more forest cover as a percent of the prop-
erty area. Fig. 5b, however, illustrates the predominance of properties
with less than 20% remaining forest cover in the set of properties
without deforestation. Properties without forest were clustered
together, and by comparing the two gures we can see that there are
relatively few properties with deforestation in these regions where most
properties have less than 20% remaining forest cover.
Deforestation occurs on large tracts of remaining forest as it is more
likely to go unnoticed, according to our eld interviews. This corrobo-
rates our ndings that large properties with a high percent of remaining
forest were the most likely to be deforested. We nd that properties with
a high percent of remaining forest were less exposed to the CA, had more
degraded pasture, had a lower sustainable capacity, and were closer to
the forest frontier. This supports the conclusion that deforestation occurs
in remote areas with high remaining forest, as discussed in Section 3.4.
However, because forest cover and size were highly correlated with
deforestation even after controlling for these and other variables, we
conclude that large forested properties are at high risk of deforestation
regardless of where they are located.
3.4. Location
We found both in the empirical analysis and the eld survey that
properties were more likely to have deforestation if they were located in
remote areas. Properties with deforestation were farther from a highway
(
α
=0.438,β=0.064). Similarly, they were closer to the forest frontier
(
α
= − 0.506,β= − 0.034). In all, a property one standard deviation
farther from a highway and one standard deviation closer to the forest
frontier was 10% more likely to clear than a property at the mean
distance.
Fig. 5 also illustrates the difference in location between the proper-
ties with and without deforestation. Properties with deforestation were
often located far from slaughterhouses. Notably, slaughterhouses were
co-located with clusters of properties without remaining forest. These
maps demonstrate the patterns of deforestation and settlement, with
recent deforestation occurring in new frontiers far from
slaughterhouses.
During eld surveys, local farmers, government ofcials, and
extension agents all emphasized the role of remote areas as sources of
deforestation in the cattle supply chain. They reported that deforestation
occurs in remote frontier areas with large areas of remaining forest, with
little occurring in settled “old frontiers.” Remote areas, they explained,
have very low oversight. Therefore, ranchers have a much lower risk of
being punished for clearing. Thus, respondents stated that ranchers in
the frontier have little incentive not to deforest both due to the low risk
of punishment and a lesser inuence of slaughterhouses that monitor for
deforestation close to the forest frontier (
ρ
=0.212).
3.5. Auxiliary properties
Finally, we investigate the deforestation patterns on properties
whose owners hold auxiliary properties. We nd that the main property
was less likely to have deforestation if the owner held auxiliary prop-
erties (
α
=n.s., β= − 0.016). In our sample, properties whose owner
held auxiliary properties had higher outows and more transactions
(
ρ
=0.169 and
ρ
=0.161, respectively). Although neither of these
characteristics are statistically signicant on their own, the relationship
between these variables supports our conclusion that active properties
that sell directly to slaughterhouses often are also less likely to deforest
on the main property.
When we only consider those properties whose owner holds auxiliary
properties, we nd that the main property was more likely to have
deforestation if the auxiliary property also had deforestation (
α
=0.522,
β=0.044). Respondents in the eld also emphasized this relationship.
Rather than limit our understanding of deforestation to the property-
level, interview respondents believed that a key determinant of
whether a property was deforested was its owner, and whether they
followed a “cycle” of deforestation. Amazonian pasture degrades quickly
(Fearnside and Barbosa, 1998), and when it does, a rancher has the
option to clear for additional pasture rather than invest in the existing
pasture. Our results, suggest that some ranchers follow this pattern of
Fig. 4. Correlation coefcients between explanatory variables.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
10
Fig. 5. Percent of remaining forest cover on properties and location of slaughterhouses.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
11
clearing, using land until it degrades, and clearing further; for these
ranchers, we may see deforestation on both their main and auxiliary
properties.
3.6. Implications for continuous deforestation outcomes
Results of models explaining the total area of deforestation on the
property are listed in SI Table B2. In our preferred specication, we
employ an inverse hyperbolic sine (IHS) transformation, which is
identied at zero and reduces the inuence of extreme observations
(Burbidge et al., 1988). The IHS transformation can generally be inter-
preted in the same way as a standard logarithmic dependent variable
and is commonly used in the environmental economics literature (e.g.,
Sims and Alix-Garcia, 2017; Lipscomb and Prabakaran, 2020; Jaya-
chandran et al., 2017; Parker and Thurman, 2018). Coefcients of this
model can be interpreted as the percent changes in deforested area
associated with a one standard deviation change in the explanatory
variable. The coefcients of the IHS transformed model match the co-
efcients of the linear probability model (LPM) in sign and signicance
for all variables with the exception of the area of soy or another crop on
the property, which is statistically signicant in the IHS model but not in
the LPM. This result is in part mechanical as, after controlling for total
area, properties with more area established in soy or other crops had less
area to deforest. As in the LPM, percent forest cover and distance to the
nearest highway had the highest explanatory power. Indeed, the con-
clusions and policy recommendations of our binary models, which
explain likelihood of deforestation, are consistent in this model that
explains deforested area.
We also tested which characteristics explain total (un-transformed)
deforestation on the property. The conclusions of the LPM and IHS
transform were robust, but, as expected, total deforestation was largely
driven by the size of the forested area available to deforest, which is
captured by the property size and percent remaining forest cover.
Property size was the most explanatory variable, and its coefcient was
three times larger than the second largest coefcient (distance to high-
way). While property size was positively and signicantly correlated
with deforestation in the LPM and IHS model, it was the eighth and sixth
largest coefcient in those models, respectively.
The remaining coefcients were largely consistent in sign and sig-
nicance to the LPM and IHS model. In no case did a statistically sig-
nicant characteristic in the LPM or IHS model change sign in this
model. However, we nd that properties with deforestation had statis-
tically signicantly higher sustainable capacities and were statistically
signicantly more likely to have soy or another crop on an auxiliary
property. Because both of these characteristics were correlated with
property size, we posit that these discrepancies were driven by the
relationship between deforestation size and property size. The correla-
tion coefcient between sustainable capacity and property size was
ρ
=
0.042. Main properties with soy on an auxiliary property were 2,120
hectares, on average, while properties without soy on an auxiliary
properties were only 237 hectares, on average.
4. Discussion and conclusions
4.1. Targeting to combat deforestation possible with data science
Deforestation for cattle production is a major threat to the Amazon
biome as well as the global climate. We identify characteristics that were
shared by cattle producing properties with deforestation between 2010
and 2018. We further show that the same set of characteristics explain
the total deforested area on a property. Our results can be used to target
properties at high-risk of any deforestation as well as those at high risk of
large areas of deforestation. This simultaneously supports the goal of
minimizing the number of properties with deforestation, achieving a
“clean” supply chain, as well as minimizing total deforestation.
Until recently, lack of data made this task almost impossible (The
Nature Conservancy, 2021). However, with the availability of the GTA,
large-scale property maps, as well as powerful algorithms of data sci-
ence, this is now possible. Furthermore, new monitoring tools including
near-real-time deforestation alerts freely available on global data plat-
forms (e.g. Global Forest Watch), enhance the capacity of monitoring
actors to act following deforestation alerts and decrease forest loss
(Moffette et al., 2021).
Our investigation of multiple property ownership also highlights the
importance of monitoring and enforcing anti-deforestation policies on
auxiliary properties. We nd that main properties are more likely to
have deforestation when their auxiliary properties have deforestation,
suggesting that properties are at high risk of deforestation if properties
with the same owner have recent deforestation. This policy suggestion is
supported by L’Roe et al. (2016), who showed that ranchers may register
their parcels incompletely or inaccurately with the CAR to take advan-
tage of the current policy system. Monitoring auxiliary properties, which
we demonstrate is possible using publicly available maps and data sci-
ence, would be a marked change from the current strategy of the Cattle
Agreements, which focus monitoring and enforcement on individual
properties.
4.2. Deforestation is more likely outside the scope of the CA
The current system of monitoring of the CA leaves the indirect sup-
plying portions of the supply chain unmonitored (Pereira et al., 2020;
Raj˜
ao et al., 2020), and we nd that deforestation was higher on these
properties. Indeed, deforestation was more likely on properties with low
exposure to the CA and deforestation was three times more likely on
properties that were entirely outside the reach of the policy. We nd that
CA exposure was tightly linked to a property’s location, with low
exposure close to the forest frontier. Further, properties with low
market-access and low CA exposure had a higher likelihood of defor-
estation. Consistent with Moffette (2018), our results demonstrate that,
for meaningful change through the CA, it is not enough for a property to
be in the supply zone of a single CA slaughterhouse; the CA are more
effective when CA slaughterhouses need not compete with non-
monitoring plants.
In all supply chains, whether the slaughterhouse signed the CA or
not, properties were less likely to have deforestation if they sold directly
to slaughterhouses more frequently. These repeated interactions with
slaughterhouses may offer additional pressure not to deforest through
monitoring or may foster communication regarding sustainable ranch-
ing (Skidmore, 2020). Alternatively, direct suppliers may simply be
more organized and their deforestation process occurred before the
time-frame of this analysis.
The fact that deforestation was less likely on properties whose owner
owned multiple properties is also relevant to the CA, as the current
monitoring system only monitors a single property at the time of sale.
Our result may indicate that multiple-property owners strategically
launder cattle through a “clean” property. It may also be that main
properties are older than auxiliary properties, were cleared prior to our
study, and/or are used for sales because it is simpler to gather docu-
ments for a single property (Rausch and Gibbs, 2016).
Expanding the scope of the CA to reach more properties, through
monitoring of indirect suppliers, increasing the number of slaughter-
houses that sign the agreements, and monitoring of auxiliary properties,
could lower deforestation. Leakage of deforestation to unmonitored
areas and laundering of cattle through “clean” properties are established
problems with the current monitoring systems (Alix-Garcia and Gibbs,
2017; Skidmore, 2020; Klingler et al., 2018; Raj˜
ao et al., 2020; Moffette
and Gibbs, 2021). Our analysis does not causally identify the impact of
the CA as the reason deforestation was lower on direct suppliers that
were highly exposed to the CA. Yet whether the CA led direct suppliers
to slow their deforestation or merely pushed deforestation to unmoni-
tored portions of the supply chain, monitoring more properties would be
an important step toward a zero-deforestation cattle supply chain.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
12
4.3. Deforestation for cattle production was more likely in remote areas
Our analyses conrm that a property’s location was highly correlated
with deforestation. Properties with deforestation were three times closer
to the forest frontier and two times farther from the highways.
These ndings support a regional approach to monitoring, as well as
a property-level approach. While remote areas on the forest frontier may
be harder to access, properties in these areas were far more likely to have
deforestation. This may be due to a lack of enforcement because of the
challenge of access or because the lack of nearby slaughterhouses leaves
properties without market pressure to stop clearing. Expansion of the CA
to include more slaughterhouses is therefore unlikely to pressure prop-
erties on the forest frontier as there are few slaughterhouses in these
remote areas.
Spatially targeted monitoring can be a cost- and resource-efcient
way to combat deforestation in these high-deforestation remote areas.
This was the tactic of the Priority Municipalities List, which successfully
focused enforcement of the Brazilian FC in high-deforestation munici-
palities (e.g. see Koch et al., 2019 and Assunç˜
ao et al., 2019). Our results
support the continued targeting of remote areas with high rates of
deforestation for deforestation monitoring and enforcement.
4.4. Properties with deforestation share high-risk characteristics
Certain properties had a higher likelihood of deforestation, even
after controlling for their location and their relationship to the CA. The
amount of remaining forest on the property was the variable in our
model that best explained whether a property had deforestation. Our
results also contribute to the ongoing investigation of the role of large-
and small-holders in deforestation in the Amazon (e.g. see L’Roe et al.,
2016; Lima et al., 2006; Godar et al., 2015); we found that larger
properties were more likely to have deforestation. Both of these char-
acteristics can be used to target properties with high deforestation risk;
large properties with remaining forest, regardless of location could be
easily identied for monitoring as well as support in developing sus-
tainable production practices.
Properties whose owners have cleared other properties could also be
identied as being at high risk of deforestation. While 48 percent of
main properties with deforestation also had deforestation on an auxil-
iary property; only 16 percent of main properties without deforestation
also had deforestation on an auxiliary property. Consequently, while
main properties were less likely to have deforestation overall, there are
differences between the main properties that are part of a holding where
the auxiliary properties do or do not have deforestation. The main
property was more likely to have deforestation if it was owned by a
rancher who deforested its auxiliary properties than if these auxiliary
properties were also free of deforestation. Thus, after deforestation oc-
curs on a property, outreach and monitoring could be targeted to pre-
vent the rancher from clearing their other properties.
In sum, deforestation for cattle production is a major threat to the
Amazon biome as well as the global climate; deforestation has global
ramications due to carbon emissions, regional rainfall patterns, and
biodiversity (de Castro Solar et al., 2015; INPE, 2020; SEEG, 2018).
Efforts to slow deforestation will succeed only if they are designed with
an understanding of the complexity of the context. Our ndings identify
high-risk characteristics that can be used to develop more efcient
monitoring systems that focus on properties and location where defor-
estation is more likely.
Our results can also be used to target outreach to these high-risk
properties. The role of access to nancing in promoting sustainable
development has been documented since the early waves of settlement
in the Amazon (Schneider, 1995). Assistance in the form of technical
education and access to nancing for sustainable practices can support
the development of a sustainable cattle supply chain while minimizing
harm to ranchers’ livelihoods. Since land owners are often reticent to
sacrice income in the short-run to develop sustainable practices
(Schneider, 1995), the decision to invest in those practices, including
new production systems or technologies, requires signicant cost-
benet analyses. Without these analyses, the long-term protability of
the technology is unclear. Further, the complexity of these analyses is
part of the reason ranchers stick to their status quo production methods,
which may include deforestation (Embrapa, 2018). Still, it is possible to
promote sustainable production in the Amazon by pairing farmer
training and education with technology and investment (Ermgassen
et al., 2018). A targeted combination of increased monitoring and pro-
motion of sustainable production techniques has the potential to reduce
deforestation in the Brazilian Amazon, and more research is needed to
understand when and why producers adopt sustainable practices.
CRediT authorship contribution statement
Marin Elisabeth Skidmore: Conceptualization, Formal analysis,
Funding acquisition, Investigation, Methodology, Writing - original
draft, Writing - review & editing. Fanny Moffette: Conceptualization,
Formal analysis, Funding acquisition, Investigation, Methodology,
Writing - original draft, Writing - review & editing. Lisa Rausch:
Conceptualization, Funding acquisition, Writing - review & editing.
Matthew Christie: Data curation. Jacob Munger: Data curation. Holly
Gibbs: Conceptualization, Funding acquisition, Writing - review &
editing.
Declaration of Competing Interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Acknowledgments
We thank participants of the Meridian Institute Supply Chain Sus-
tainability Research and Learning Symposium, including Amintas
Brand˜
ao and Raquel Carvalho, for helpful comments and suggestions.
Ian Schelly prepared the maps. Funding provided by the Meridian
Institute, the Gordon and Betty Moore Foundation, and the Norwegian
Agency for Development Cooperations Department for Civil Society
under the Norwegian Forest and Climate Initiative. HG and LR have an
ongoing consulting relationship with the National Wildlife Federation.
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in the
online version, at https://doi.org/10.1016/j.gloenvcha.2021.102280.
References
ADEPAR´
A, 2019. Consulta E-GTA. http://www.siapec.adepara.pa.gov.br/siapecest/
controletransito/guiatransito/consultapublicagta.wsp.
Ahrens, A., Hansen, C.B., Schaffer, M.E., 2018. LASSOPACK and PDSLASSO: Prediction,
model selection and causal inference with regularized regression. Technical Report
12, Stata Users Group, London Stata Conference 2018.
Alix-Garcia, J., Gibbs, H.K., 2017. Forest conservation effects of Brazil’s zero
deforestation agreements undermined by leakage. Global Environmental Change 47,
201–217.
Alix-Garcia, J., Rausch, L.L., L’Roe, J., Gibbs, H.K., Munger, J., 2018. Avoided
deforestation linked to environmental registration of properties in the Brazilian
Amazon. Conservation Letters 11 (3), 1–8.
Alston, L.J., Libecap, G.D., Mueller, B., 2000. Land reform policies, the sources of violent
conict, and implications for deforestation in the Brazilian Amazon. Journal of
Environmental Economics and Management 39 (2), 162–188.
Amin, A., Choumert-Nkolo, J., Combes, J.-L., Combes Motel, P., K´
er´
e, E., Ongono-
Olinga, J.-G., Schwartz, S., 2019. Neighborhood effects in the Brazilian Amazˆ
onia:
Protected areas and deforestation. Journal of Environmental Economics and
Management 93, 272–288.
Andersen, L.E., 1996. Causes of deforestation in the Brazilian Amazon. The Journal of
Environment & Development 5 (3), 309–328.
Anderson, L.O., De Martino, S., Harding, T., Kuralbayeva, K., Lima, A., 2016. The Effects
of land Use Regulation on Deforestation: Evidence from the Brazilian Amazon.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
13
Technical report, Oxford Centre for the Analysis of Resource Rich Economies.
University of Oxford.
Arima, E.Y., Barreto, P., Araújo, E., Soares-Filho, B., 2014. Public policies can reduce
tropical deforestation: Lessons and challenges from Brazil. Land Use Policy 41,
465–473.
Assunç˜
ao, J., Gandour, C., Rocha, R., 2015. Deforestation slowdown in the Brazilian
Amazon: prices or policies? Environment and Development Economics 1–5.
Assunç˜
ao, J., McMillan, R., Murphy, J., Souza-Rodrigues, E., 2019. Optimal
Environmental Targeting in the Amazon Rainforest. NBER Working Paper 25636.
Assunç˜
ao, J., Rocha, R., 2014. Getting Greener by Going Black: The Priority
Municipalities in Brazil. Climate Policy Initiative, PUC-Rio, (August):18.
Athey, S., Imbens, G., 2016. Recursive partitioning for heterogeneous causal effects.
PNAS 113 (27), 7353–7360.
Azevedo, A.A., Raj˜
ao, R., Costa, M.A., Stabile, M.C.C., Macedo, M.N., dos Reis, T.N.P.,
Alencar, A., Soares-Filho, B.S., Pacheco, R., 2017. Limits of Brazil’s Forest Code as a
means to end illegal deforestation. Proceedings of the National Academy of Sciences,
page 201604768.
BenYishay, A., Heuser, S., Runfola, D., Trichler, R., 2017. Indigenous land rights and
deforestation: Evidence from the Brazilian Amazon. Journal of Environmental
Economics and Management 86, 29–47.
Bowman, Maria S., Soares-FIlho, Britaldo S., Merry, Frank D., Nepstad, Daniel C.,
Rodrigues, Hermann, Almeida, Oriana T., 2012. Persistence of cattle ranching in the
Brazilian Amazon: A spatial analysis of the rationale for beef production. Land Use
Policy 29 (3), 558–568.
Burbidge, J., Magee, L., Robb, A.L., 1988. Alternative transformations to handle extreme
values of the dependent variable. Journal of the American Statistical Association 83
(401), 123–127.
Cisneros, E., Zhou, S.L., B¨
orner, J., 2015. Naming and shaming for conservation:
Evidence from the Brazilian Amazon. PLoS One 10 (9), 1–24.
de Castro Solar, R.R., Barlow, J., Ferreira, J., Berenguer, E., Lees, A.C., Thomson, J.R.,
Louzada, J., Mau´
es, M., Moura, N.G., Oliveira, V.H., Chaul, J.C., Schoereder, J.H.,
Vieira, I.C.G., Mac Nally, R., Gardner, T.A., 2015. How pervasive is biotic
homogenization in human-modied tropical forest landscapes? Ecology Letters 18
(10), 1108–1118.
Embrapa, 2018. Novilho precoce: demandas e caminhos para sua produç˜
ao e
valorizaç˜
ao. Technical report.
Ermgassen, E.K.H.J., Pereira de Alcˆ
antara, M., Balmford, A., Barioni, L., Beduschi Neto,
F., Bettarello, M.M.F., de Brito, G., Carrero, G.C., de A S Florence, E., Garcia, E.,
Trevisan Gonçalves, E., Trajano da Luz, C., Mallman, G.M., Strassburg, B.N.B.,
Valentim, J.F., Latawiec, A., 2018. Results from on-the-ground efforts to promote
sustainable cattle ranching in the Brazilian Amazon. Sustainability 10(1301).
Escobar, H., 2019a. Amazon res clearly linked to deforestation, scientists say. Science
365 (6456), 853.
Escobar, H., 2019b. Brazilian president attacks deforestation data. Science 365 (6452),
419.
EU Directorate General, 2011. Evaluation of animal health controls concerning foot and
mouth disease and certication procedures for fresh bovine meat intended for export
to the European Union. Technical report, European Commission on Health and
Consumers.
Fearnside, P., 2008. The roles and movements of actors in the deforestation of brazilian
amazonia. Ecology and Society 13, 23.
Fearnside, P.M., Barbosa, R.I., 1998. Soil carbon changes from conversion of forest to
pasture in Brazilian Amazonia. Forest Ecology and Management 108 (1–2), 147–166.
Ferrante, L., Fearnside, P.M., 2019. Brazil’s new president and ‘ruralists’ threaten
Amazonia’s environment, traditional peoples and the global climate. Environmental
Conservation 13–15.
Fetzer, T., Marden, S., 2017. Take what you can: property rights, contestability and
conict. The Economic Journal.
Gibbs, H.K., Munger, J., L’Roe, J., Barreto, P., Pereira, R., Christie, M., Amaral, T.,
Walker, N.F., 2015a. Did ranchers and slaughterhouses respond to zero-deforestation
agreements in the Brazilian Amazon? Conservation Letters 9 (1), 32–42.
Gibbs, H.K., Rausch, L., Munger, J., Schelly, I., Morton, D.C., Noojipady, P., Soares-
Filho, B., Barreto, P., Micol, L., Walker, N.F., 2015b. Brazil’s Soy Moratorium.
Science 347 (6220), 377–378.
Gibbs, H.K., Skidmore, M., Rausch, L.L., Christie, M., Munger, J., Brand˜
ao, A., Amaral, T.,
Barreto, P., Hall, S., Walker, N.F., 2020. New Opportunities to reduce pervasive
deforestation under Brazil’s Zero-Deforestation Cattle Agreements.
Global Forest Atlas, 2016. Cattle Ranching in the Amazon Region.
Godar, J., Gardner, T.A., Tizado, E.J., Pacheco, P., 2015. Actor-specic contributions to
the deforestation slowdown in the Brazilian Amazon. Proceedings of the National
Academy of Sciences, 112 (23), 201508418.
Greenpeace, 2009. Slaughtering the Amazon. Technical report.
Heilmayr, R., Rausch, L., Munger, J., Gibbs, H., 2020. The Impacts of the Amazon Soy
Moratorium on Deforestation. In review at Nature Sustainability.
IBGE, 2018a. Censo Agropecu´
ario 2017. Resultados Denitivos.
IBGE, 2018. Pesquisa da Pecu´
aria Municipal. https://www.ibge.gov.br/estatisticas/
economicas/agricultura-e-pecuaria/9107-producao-da-pecuaria-municipal.html?
edicao=25474&t=resultados.
IDARON, 2018. Consulta de GTA. http://www.idaron.ro.gov.br/index.php/consulta-de-
gta/.
Imbens, G.W., Wooldridge, J.M., 2009. Recent developments in the econometrics of
program evaluation. Journal of Economic Literature 47 (1), 5–86.
INCRA, 2017. Certicado de Cadastro do Im´
ovel Rural (CCIR).
INDEA, 2019. Consulta de GTA. https://sca.indea.mt.gov.br/SIA/form.jsp?
sys=SCA&action=openform&formID=467&align=0&mode=-1&goto=-
1<er=&scrolling=no.
INPE, 2020. Acesso aos dados PRODES. http://www.dpi.inpe.br/prodesdigital/dadosn/.
INPE, 2020. PRODES taxas de desmatamento - Amazonia Legal.
Jayachandran, S., Laat, J.D., Lambin, E.F., Stanton, C.Y., Audy, R., Thomas, N.E., 2017.
Cash for carbon: A randomized trial of payments for ecosystem services to reduce
deforestation. Science 357 (6348), 267–273.
Kaczan, D.J., 2020. Can roads contribute to forest transitions? World Development 129.
Klingler, M., Richards, P.D., Ossner, R., 2018. Cattle vaccination records question the
impact of recent zero-deforestation agreements in the Amazon. Regional
Environmental Change 18 (1), 33–46.
Koch, N., zu Ermgassen, E., Wehkamp, J., Oliveira Filho, F., Schwerhoff, G., 2019.
Agricultural productivity and forest conservation: evidence from the Brazilian
Amazon. American Journal of Agricultural Economics, 1–22.
Lima, E., Merry, F., Nepstad, D., Amacher, G., Azevedo-Ramos, C., Lefebvre, P.,
Resque, F., 2006. Searching for sustainability: Forest policies, smallholders, and the
trans-Amazon highway. Environment 48 (1), 26–38.
Lipscomb, M., Prabakaran, N., 2020. Property rights and deforestation: Evidence from
the Terra Legal land reform in the Brazilian Amazon. World Development 129,
104854.
L’Roe, J., Rausch, L., Munger, J., Gibbs, H.K., 2016. Mapping properties to monitor
forests: Landholder response to a large environmental registration program in the
Brazilian Amazon. Land Use Policy 57, 193–203.
MAPA, 2018. Consulta pública de GTA.
Minist´
erio do Desenvolvimento, 2015. Terra Legal. http://terralegal.mda.gov.br/
processo.
Miranda, J., B¨
orner, J., Kalkuhl, M., Soares-Filho, B., 2019. Land speculation and
conservation policy leakage in Brazil. Environmental Research Letters 14 (4),
045006.
Moffette, F., 2018. Forest Conservation and Agriculture: Essays on Land Use Trade-Offs
and Spillovers. PhD thesis, University of Wisconsin-Madison.
Moffette, F., Alix-Garcia, J., Shea, K., Pickens, A.H., 2021. The impact of near-real-time
deforestation alerts across the tropics. Nature Climate Change.
Moffette, F., Gibbs, H., 2021. Agricultural Displacement and Deforestation Leakage in the
Brazilian Legal Amazon. Land Economics 97 (1).
Moffette, F., Skidmore, M., Gibbs, H.K., 2020. Environmental policies that shape
productivity: evidence from cattle ranching in the Amazon. Revise and Resubmit at
Journal of Environmental Economics and Management.
Nepstad, D., Mcgrath, D., Stickler, C., Alencar, A., Azevedo, A., Swette, B., Bezerra, T.,
Digiano, M., Shimada, J., Seroa, R., Armijo, E., Castello, L., Brando, P., Hansen, M.C.,
Mcgrath-horn, M., Carvalho, O., Hess, L., 2014. Slowing Amazon deforestation
through public policy and interventions in beef and soy supply chains. Science 344
(6188), 1118–1123.
Nolte, C., Agrawal, A., Silvius, K.M., Britaldo, S.S.F., 2013. Governance regime and
location inuence avoided deforestation success of protected areas in the Brazilian
Amazon. Proceedings of the National Academy of Sciences of the United States of
America 110 (13), 4956–4961.
Parker, D., Thurman, W., 2018. Tax incentives and the price of conservation. Journal of
the Association of Environmental and Resource Economists 1 (2), 695615.
Pereira, R., Rausch, L.L., Carrara, A., Gibbs, H.K., 2020. Extensive production practices
and incomplete implementation Hinder Brazil’s Zero-Deforestation Cattle
Agreements in Par´
a. Tropical Conservation Science 13.
Pfaff, A., 1999. What drives deforestation in the Brazilian Amazon? Evidence from
satellite and socioeconomic data. Journal of Environmental Economics and
Management 37, 26–43.
Pfaff, A., Robalino, J., Sandoval, C., Herrera, D., 2015. Protected area types, strategies
and impacts in Brazil’s Amazon: Public protected area strategies do not yield a
consistent ranking of protected area types by impact. Philosophical Transactions of
the Royal Society B: Biological Sciences 370 (1681).
Raj˜
ao, R., Soares-lho, B., Nunes, F., B¨
orner, J., Machado, L., Assis, D., Oliviera, A.,
Pinto, L., Ribeiro, V., Rausch, L., Gibbs, H., Figueira, D., 2020. The rotten apples of
Brazil’s agribusiness. Science 369 (6501), 246–248.
Rausch, L.L., Gibbs, H.K., 2016. Property arrangements and soy governance in the
brazilian state of mato grosso: Implications for deforestation-free production. Land 5
(2).
Republic of Brazil, 2009. LEI N 12.097.
Republic of Brazil, 2011. DECRETO N 7.623.
Republic of Brazil, 2012. LEI N 12.651.
Robalino, J., Pfaff, A., Villalobos, L., 2017. Heterogeneous local spillovers from protected
areas in Costa Rica. Journal of the Association of Environmental and Resource
Economists 4 (3), 795–820.
Rodrigues, A.S., Ewers, R.M., Parry, L., Souza, C., Veríssimo, A., Balmford, A., 2009.
Boom-and-bust development patterns across the amazon deforestation frontier.
Science 324 (5933), 1435–1437.
Schneider, R.R., 1995. Government and the Economy on the Amazon Frontier.
Government and the Economy on the Amazon Frontier.
SEEG, 2018. Estimativas de Emiss˜
oes de Gases de Efeito Estufa do Brasil 1970 - 2017.
Technical report.
SICAR, 2019. SFB Sicar. http://www.car.gov.br/#/.
Sims, K.R., Alix-Garcia, J.M., 2017. Parks versus PES: Evaluating direct and incentive-
based land conservation in Mexico. Journal of Environmental Economics and
Management 86, 8–28.
Skidmore, M., 2020. Agriculture, environmental policy, and climate: Essays on cattle
ranching in the Brazilian Amazon. PhD thesis, University of Wisconsin-Madison.
Sparovek, G., Reydon, B.P., Guedes Pinto, L.F., Faria, V., de Freitas, F.L.M., Azevedo-
Ramos, C., Gardner, T., Hamamura, C., Raj˜
ao, R., Cerignoni, F., Siqueira, G.P.,
Carvalho, T., Alencar, A., Ribeiro, V., 2019. Who owns Brazilian lands? Land Use
Policy 87 (March), 104062.
M.E. Skidmore et al.
Global Environmental Change 68 (2021) 102280
14
Teixera, I., 2014. INSTRUǘ
AO NORMATIVA No 2/MMA, DE 06 DE MAIO DE 2014.
Technical report, Minist´
erio do Meio Ambiente.
The Nature Conservancy, 2021. The Forest Code: Using Law to Protect the Amazon.
Vale, P., Gibbs, H., Vale, R., Christie, M., Florence, E., Munger, J., Sabaini, D., 2019. The
expansion of intensive beef farming to the Brazilian Amazon. Global Environmental
Change 57, 101922.
M.E. Skidmore et al.
