LACEA WORKING PAPER SERES. No.

0046

BARREN LIVES: DROUGHT SHOCKS AND AGRICULTURAL VULNERABILITY IN THE

BRAZILIAN SEMI-ARID

Lucas de Almeida Nogueira da Costa

André Albuquerque Sant’Anna
Carlos Eduardo Frickman Young

LATIN AMERICAN AND THE CARIBBEAN ECONOMIC ASSOCIATION

June 2020

The views expressed herein are those of the authors and do not necessarily reflect the views of the
Latin American and the Caribbean Economic Association. Research published in this series may
include views on policy, but LACEA takes no institutional policy positions.
LACEA working papers are circulated for discussion and comment purposes. Citation of such a paper
should account for its provisional character. A revised version may be available directly from the
author.
© 2020 by Lucas de Almeida Nogueira da Costa, André Albuquerque Sant’Anna and Carlos Eduardo
Frickman Young. All rights reserved. Short sections of text, not to exceed two paragraphs, may be
quoted without explicit permission provided that full credit, including © notice, is given to the source.
LACEA WORKING PAPER SERIES No. 0046 June 2020
Barren Lives: drought shocks and agricultural vulnerability in the Brazilian Semi-Arid

Lucas de Almeida Nogueira da Costa

UFRJ
[email protected]
André Albuquerque Sant’Anna
BNDES and UFF
[email protected]
Carlos Eduardo Frickman Young
UFRJ
[email protected]

ABSTRACT
This paper studies the effects of drought shocks in a vulnerable environment – the Brazilian Semi-
Arid. We analyze the impact of drought shocks, measured as deviations from historical averages, on
agricultural outcomes and land-use decisions in a region that suffers recurrently with drought. After
controlling for municipality and year fixed effects, we use weather shocks to exactly identify
outcomes. Our benchmark results show substantial effects on the loss of crop area and on the value
of agricultural output. By investigating distributional effects, we are able to show that crops related
to family farming suffer more from drought shocks. We follow our investigation by testing
heterogeneity effects and show that adequate water provision and maintenance of forest cover help
in reducing the impact of drought shocks. Finally, we show that drought shocks in the previous year
affect deforestation in the following year.

JEL Classification: Q15, Q54.

Keywords: Drought, Climate Change, Agricultural Output; Brazilian Semi-Arid.

ACKNOWLEDGEMENTS AND FINANCIAL DISCLOSURE

We thank José Feres and Romero Rocha for insightful comments. The views expressed here are those
of the authors alone and do not necessarily reflect those of the Brazilian Development Bank.
1 Introduction
It is widely recognized that there is an anthropogenic contribution to the observed changes
in climate. According to statements held by IPCC reports, the agreement among the sci-
entific community has grown stronger regarding the effects of human-based emissions on
the accumulation of carbon dioxide in the atmosphere The effects of climate change are
not only felt with temperature extremes. Besides temperature changes, one expects pre-
cipitation changes, humidity changes, changes in the frequency and intensity of tropical
cyclones, sea-level rise, ocean acidification, effects on droughts and floods and huge im-
pacts on ecosystems, with loss of biodiversity (Hsiang and Kopp, 2018).

As regards droughts, climate change is expected to alter frequency and intensity, since
temperature and precipitation changes affect moisture conditions. Indeed, dry regions are
expected to suffer more with an increase in the frequency of droughts (Collins et al., 2013),
as can already be noticed in the Brazilian Semi-Arid (Brito et al., 2018). As long as these
regions tend to have lower agricultural productivity and need more investments in adap-
tation, the effects of climate change may be especially severe.

This paper analyzes the impact of drought shocks, measured as deviations from historical
averages, on agricultural outcomes and land-use decisions in a region that suffers recur-
rently with drought. The Brazilian Semiarid is the driest and poorest region in the country.
The region has a total area of 1.13 million km2 , covering parts of 9 states in Brazil. To-
tal population living within the Semiarid is 27.5 million, nearly 13% of the country‘s total
population and income per capita is 42% of the Brazilian average (Medeiros et al., 2012;
Silva et al., 2016).1

We investigate how large deviations from historical averages in rainfall averages patterns
affect the area of harvest, production and productivity. To identify causal effects, we use
longitudinal data on Brazilian semi-arid municipalities from 2006 to 2017. After control-
ling for municipality and year fixed effects, we use weather shocks - rainfall idiosyncratic
shocks in our case - to exactly identify outcomes. As argued by Dell et al. (2014), there is
a growing body of the literature that uses weather shocks to exactly identify outcomes,
under the assumption that weather shocks occur randomly in time.2

Our benchmark results show substantial effects on the loss of crop area and on the value
1 As the last demographic census was carried out in 2010 and the new delimitation of the Brazilian

Semi-Arid was in 2017, some demographic data may be out of date. Therefore, in addition to the refer-
ences cited, some updated data can be found at the following link: http://www.sudene.gov.br/images/
arquivos/semiarido/arquivos/Rela%C3%A7%C3%A3o_de_Munic%C3%ADpios_Semi%C3%A1rido.pdf
2 Blanc and Schlenker (2017) provide a discussion on the use of panel models in assessments of climate

impacts on agriculture.

2
of agricultural output: a rainfall deviation of one standard deviation from the historical
average is related to a loss of 3.4% of crop area and 18.4% of output. When we consider non-
linear effects, results are more striking. A year of extreme drought, defined as a dummy
when rainfall deviation is two standard deviations less than the historical average, leads to
a loss of 5.3%-5.9% in crop area and 26%-32% in the value of output.

As investments in adaptation can reduce the impact of those shocks, we focus on the effects
across different crops. Credit and insurance markets imperfections result in underinvest-
ment in adaptation. This is especially important for family farms, which suffer more from
liquidity contraints. In this context, we show that outcomes from crops related to family
farming - beans and corn - are those suffering from drought shocks - as highlighted by
Cirino et al. (2015), whereas business crops, such as sugarcane and coffee, have small ef-
fects. Unfortunately, the harsh effects on family farming cannot be qualified as surprising.
As of 1946, De Castro (1952) had already mapped this pattern in the Brazilian Semiarid.

In order to better understand, mechanisms, we follow our investigation by testing hetero-

geneity effects. We show that adequate water provision and maintenance of forest cover
help in reducing the impact of drought shocks in our measures of agriculture outcome.
Finally, we test whether drought shocks in the previous year affect deforestation in the fol-
lowing year. The intuition behind this test rests on the historical pattern of slash and burn
agricultural in Brazil. We consider these to be the main contributions of this paper, since the
extensive literature on drought shocks in the Brazilian semiarid region has already detailed
this phenomenon - as detailed in Section III -, but slight effort has been made to quantify
the heterogeneous effects of drought shocks on the agricultural output in this region.

To further increase confidence in our results, we conduct a placebo test, by estimating the
effects of previous and forward drought shocks on our main dependent variables. More-
over, it is important to account for spatial and temporal dependence in climatic exposure.
Then, we correct standard errors for spatial dependence using the procedure proposed by
Conley (1999) and results are robust to this correction.

This paper contributes to the literature pioneered by Deschênes and Greenstone (2007),
which uses random fluctuations in weather to assess agricultural impacts of climate change.
Given the greater importance of agriculture and higher levels of poverty, developing coun-
tries are much more vulnerable to these weather shocks on the welfare of its population.
Burgess et al. (2017), for instance, assess the effects of high temperatures in mortality in ru-
ral India. According to the authors, potential mechanisms relate to lower productivity and
wages in seasons with extreme hot days. Taraz (2017) also investigates the effects of climate
change on India‘s farmers. However, the author focuses on adaptation efforts and shows

3
that adaptation only recovers a fraction of lost profits. We contribute to this literature by
providing a specific focus on droughts shocks, instead of temperature, in a developing
country with a significant agricultural production, as it is Brazil.3

We also contribute to the literature that discusses the importance of natural resources to
stabilize effects from drought shocks. Wani et al. (2012) discuss how watershed manage-
ment in dryland tropics increases net returns from crop production, while conserving the
natural resource base. In a paper similar to ours in its conclusions, Noack et al. (2019) re-
late droughts to negative shocks in crop incomes, which are, nevertheless, partly offset by
forest extraction. In addition, the authors show that more biodiversity reduces the effects
of droughts. We show that tree cover attenuates the effects of droughts, as well. Moreover,
our results also imply an increase in the intensity of natural resources use after a year of
drought, which is a mechanism diverse from the one found by Noack et al. (2019).

Our results show how the impacts of droughts in dryland regions can be substantial. Mo-
rover, we are able to show that the adequate provision of public goods, represented by
water supply and native vegetation, can have important effects in reducing the damage
extension of drought shocks.

In the next section, we review and contextualize our object of study, highlighting the his-
torical background of drought in the Brazilian Semi-Arid and reviewing the literature on
climate shocks in this region. In Section III, we describe the database that we set up for this
paper. In Section IV, we explain the empirical strategy used. Section V presents the results
proposed in this paper. Finally, a brief section presents the main conclusions of this study.

2 Background

2.1 Historical Background

The Portuguese occupation of the Brazilian territory in colonial times was concentrated in
the fertile coastal areas of the Atlantic Forest biome. In the Northeast, sugarcane plantations
led to a very profitable trade to local landowners and the Portuguese crown, but extremely
unequal due to their dependence on slave labour, captured mostly from Africa but also
from the indigenous native population. In order to supply beef products to the coastal
population, extensive cattle ranching occupied the interior lands of the Northeast (Prado Jr,
2017; Furtado, 2005).
3 Assunção and Chein (2016) use a mix of Ricardian and production function approaches to simulate the
effects of climate change on agricultural productivity in Brazil. Average effects are expected to decrease yields
by 18%, with a significant variation.

4
Due to low fertility and water scarcity, there was little interest in developing plantations
in the Northeast semiarid. Therefore, cultivation in these interior lands was mainly car-
ried out by subsistence farmers, mixing the remnants of the native indigenous population
with impoverished European descendants and freed or escaped former slaves. Corn and
beans have been the most important crops since these colonial times, even though other
cultivation products were also important, mainly manioc (Prado Jr, 2017; De Castro, 1952).

The rare settlements and villages that catered to merchants, cattle ranchers, and travelers
subsisted on groundwater and subsistence agriculture. Over time, the extensive nature of
the occupation and the long distances to the dynamic centers of the economy have deteri-
orated the average yield of semi-arid production (Furtado, 2005).

The decline of sugarcane plantations since the mid-seventeenth century and demographic
expansion in the coastal areas resulted in a process of migration of the population towards
the semiarid region. However, in spite of occasional booms of demand for ranching prod-
ucts, especially during the “cattle cycle” in the XVIII century, the semi-arid region remained
characterized by high levels of poverty, unemployment and critical dependence on subsis-
tence cultivation (Furtado, 2005; Ab’Sáber, 1999).

The low development of the region if compared to other parts of Brazil has always been
associated to the climate conditions of the drylands, characterized by very low levels of
precipitation and recurrent severe drought events. Under favorable conditions, the Brazil-
ian semi-arid is one of the semi-arid regions of the planet most favorable to occupation, in
terms of water availability and food (De Castro, 1952). However, the frequent repetition of
severe droughts led to extreme events, resulting in food insecurity, poverty and migration
towards other parts of Brazil. But climatic conditions are not the only cause for the social
vulnerability in the region: the concentration of economic and political power in the hands
of the large landowners (locally known as “coronels”), a large portion of landless and job-
less population and the deficit of public policies aiming the access to water resources and
the agrarian issue, maintains a vicious cycle of poverty and social vulnerability, and the
region remains as the poorest part in Brazil (Ab’Sáber, 1999).4

2.2 Recent Background

The Brazilian Semi-Arid region occupies most of the interior land of the Northeast re-
gion. In ecological and biodiversity terms, the Brazilian Semi-Arid is closely related to the
Caatinga, which had its flora adapted to the dry and hot climate that lasts for almost the

4 For an analysis of the political effects of this process, see Leal (2012).

5
whole year. Thus, climatic conditions constituted the only biome that is located exclusively
in Brazilian territory (Ab’Sáber, 1999).

The Brazilian Institute of Geography and Statistics (IBGE) and the Northeast Development
Superintendence (SUDENE) define the Brazilian Semi-Arid region based on specific tech-
nical criteria of very low precipitation (less than 800 mm/year on average) and/or high
daily water deficit of over 60% 5 (Medeiros et al., 2012). Figure 1 shows the 1.13 million
square kilometers area covered by the 1,262 municipalities classified as in the Semi-Arid
region.

Figure 1: Semi-Arid Region and Caatinga Biome

Notes: Own elaboration

The Brazilian Semi-Arid is the largest semiarid territory in the world. It is also the semi-
arid region with the largest human population: 27.5 million people, or around 13% of the
Brazilian population, according to the Brazilian Demographic Census of 2010, most of them
living in rural areas and in municipalities with less than 50,000 inhabitants (Medeiros et al.,
2012).

Most of the agricultural production in the region comes from family farming, either for
subsistence or for commercialization. These farmers have little investment capacity and
5 Evapotranspiration exceeds 60% of precipitation every day of the year

6
 Table 1: Family and Non-Family Crop Patterns in Brazil

Crop Family Non-Family Total Share of Share of

Farming Farming Farming Family Each Crop
Area (ha) Area (ha) Area (ha) Farming on Sum of
Area Crops Area

Semi-Arid
Rice 136,055 33,075 169,130 80.4% 2.5%
Bean 2,436,941 317,758 2,754,699 88.5% 41.5%
Cassava 334,074 48,269 382,343 87.4% 5.8%
Corn 2,329,061 457,872 2,786,933 83.6% 41.9%
Soy 707 456,850 457,557 0.2% 6.9%
Coffee Bean 39,696 53,748 93,444 42.5% 1.4%
Total 5,276,534 1,367,572 6,644,106 79.4% 100.0%
Non-Semi-Arid
Rice 1,028,437 1,190,471 2,218,908 46.3% 6.6%
Bean 952,465 442,162 1,394,627 68.3% 4.1%
Cassava 1,138,825 181,085 1,319,910 86.3% 3.9%
Corn 4,003,675 4,800,206 8,803,881 45.5% 26.1%
Soy 2,697,533 14,499,204 17,196,737 15.7% 50.9%
Wheat 319,515 951,467 1,270,982 25.1% 3.8%
Coffee Bean 726,736 855,193 1,581,929 45.9% 4.7%
Total 10,867,186 22,919,788 33,786,974 32.2% 100.0%
Note: Own elaboration using data from the 2006 IBGE Agricultural Census

low resilience to the increasingly frequent drought events, leading to high social vulner-
ability and major food and economic insecurity during these extreme events (De Castro,
1952; Travassos et al., 2013; Silvia et al., 2013)

Table 1 compares output data, disaggregated in terms of family and non-family farming
and specific crops, for semiarid land and the rest of the country. It is evident that family
farming is far more important in cultivation in the semiarid region than in the rest of the
country, and that corn and beans are responsible for 83% of the family farming area in the
Brazilian Semi-Arid.

The historical vulnerability and poverty of the semi-arid population has caught attention
from policymakers since Brazilian independence. The perceptions on how to proceed,
however, vary widely, usually between two extremes (Campos, 2015):

• Man-made interventions, such as reservoirs and other civil engineering works, can

7
 provide a technically and economically feasible solution to reduce water scarcity;

• Droughts are an inevitable problem caused by climatic conditions, but the social vul-
nerability in the region is a consequence of the inadequate productive and social
structure to this unavoidable natural phenomenon.

The creation of the National Department of Works against Drought (DNOCS in its Por-
tuguese acronym) in 1909, aiming at the construction of cisterns, reservoirs and other hy-
drological infrastructure, has been for a long time associated with the view that proper
engineering would be enough to solve the water problem. The most recent example of this
perception is the construction of the Transposition of São Francisco River Project, initiated
in the mid-2000s and yet to be concluded. This ambitious project aims at the construction
of more than 700 km of channels in order to assure the availability of water, in 2025, to
nearly 12 million inhabitants of cities in the Brazilian Semi-Arid.

However, critics of this view argue that these infrastructure projects have high financial
costs but little efficacy to solve the problems (Cirilo, 2008).6 The creation of the Superinten-
dency for the Development of the Northeast (SUDENE) in 1959, was originally intended to
undertake the structural changes in the productive system in the region, with the support
of many fiscal and credit incentives. Nevertheless, this approach has also failed. Celso Fur-
tado, the man who idealized the creation of SUDENE, recognized the incapacity to surpass
the archaic but well-established political structures that impeded the process of transfor-
mation in the semi-arid, including the unequal distribution of land and other resources
(Furtado, 1989).

As a consequence, despite these initiatives, problems related to droughts and food security
are still very relevant in the region. The El Niño phenomenon, which tends to increase tem-
perature and decrease precipitation in the Brazilian sem-arid region, still generates large
losses in agricultural productivity (Cirino et al., 2015). In turn, the extensive and unsus-
tainable occupation of the semiarid region is further compromising its lands, generating
a phenomenon known as desertification, which decreases its humidity and productivity
(Cirilo, 2008; Travassos et al., 2013; Vieira et al., 2015).7

In 2011 one of the most severe and disastrous droughts in recent decades in the region
began, causing severe social damage and huge agricultural losses (Marengo et al., 2016;

6 Indeed, there is an ongoing problem with the overuse of groundwater for irrigation that is affecting

the supply of water to the São Francisco river. Part of the problem is related to the absence of a fee on
water use. See: https://www1.folha.uol.com.br/mercado/2020/01/agricultura-irrigada-gera-disputa-por-
agua-na-bahia.shtml
7 Droughts are also associated with increased child mortality and various problems in pregnancy

(Rocha and Soares, 2015).

8
Brito et al., 2018; Cunha et al., 2019). As highlighted by Gutiérrez et al. (2014), despite the
public policies implemented in the last decade, it seems that there are very important struc-
tural deficits to mitigate disaster damage, such as this latest drought, but also to prevent
the occurrence of long-term disasters. Therefore, in addition to the historical deficiencies
and vulnerabilities, the increased frequency, duration and severity of droughts due to cli-
mate change – as described by Brito et al. (2018) – will build a new scenario, yet its roots are
already known. Hence, it is essential to understand which factors are capable of mitigating
the damages in the region and guaranteeing food security for the semiarid population in a
context where climate change might bring increasing challenges to the region.

3 Data

3.1 Independent Variables

Our goal is to assess the effects of drought on Brazilian semi-arid agriculture and test
whether municipality-specific structures reflect heterogeneous drought resilience. Thus,
we collected several independent variables that can be divided into two groups: (i) climate
variations, which identify years of drought event; (ii) and structural variables as land use
type and type of access to water, which can capture heterogeneous effects on agriculture.

To identify drought events, we use the database Terrestrial Air Temperature and Terrestrial
Rainfall: 1990-2017 Monthly Monthly Grid Series, Version 5.01 (Matsuura and Willmott, 2018).
This database presents monthly data of georeferenced temperatures and precipitation -
by 0.5x0.5 degree grids - between 1900 and 2017. First, we aggregate the monthly data
into a per year one and, as Rocha and Soares (2015), build a continuous variable of annual
precipitation deviation from the historical average of each grid, according to the following
equation:

mean ( Pg1900−2017 ) − Pgt

Rain f all deviation gt =
sd ( Pig1900−2017 )

Where Rain f all deviation gt is the deviation of annual precipitation from the historical av-
erage for each grid g and year t. Since temperature can be an important control for our
model, we calculate the same metric for temperature. To merge these variables with the
municipalities, we have identified the closest grid to each municipality, using the centroids
for both municipalities and grids. Also, considering the exogeneity of grids, we can correct
problems of spatial correlation of precipitation and temperature variables (Burgess et al.,

9
2018).

One problem with using a continuous variable to identify an extreme event is that drought-
related losses may be not linearly correlated with rainfall (for example, decreasing rainfall
from 80mm to 60mm may not have the same effect as dropping from 50mm to 30mm,
although the difference between the lower and upper values is identical). Therefore, we
also have built two dummy variables to identify drought events, according to the following
principles:

(
if 0 < Rain f all deviationmt < 1, then Drought = 1;
if Rain f all deviationmt ≥ 1, then Extreme Drought = 1.

The first set of structural variables, chosen to identify the heterogeneous resilience of mu-
nicipalities, are the types of access to water in rural households, available in Table 1395 -
2010 Brazilian Demographic Census (IBGE, 2010). From these databases, it was possible to
identify how many rural households per municipality had, in 2010, as main source of wa-
ter the (i) general supply network, (ii) wells or springs within property or village, (iii) wells
or spring outside property or village, (iv) rivers, lakes or streams, or (v) water truck and
rainwater. For practical purposes, we chose to use the percentage of water access type as:

Hma2010
THma2010 =
Hm2010

Where THma2010 is the percentage of rural households in municipality m with type of access
to water a. Hma2010 is the number of rural households in m with access to water a in 2010,
while Hm2010 is the total number of rural households in municipality m in 2010.

It is also an important matter for this paper to study whether conserving vegetation in-
creases agricultural resilience to droughts and to understand the dynamics of deforesta-
tion following drought events. Therefore, we set up an annual forest stock variable from
the MapBiomas platform data, which, through Google Earth Engine, assembles annual
historical series of georeferenced land use data for the entire Brazilian territory. 8 We have
calculated the percentage of forest stock according to the equation:

Fmt
TFmt =
Total Aream

8 We have selected the land use categories 1 to 9 to forest stock, according the Mapbiomas codes. To

measure the forest losses, we considered the transition of categories 1-9 to categories 14-21, 24, 30.

10
Where TFmt is the forest area rate in municipality m in year t, Fmt is the area of forest stock,
and Total Aream is the area occupied by municipality m. Similarly, we set up a ForestLossmt
variable from the same land use database.

3.2 Dependent Variables

To measure the direct and heterogeneous effects of drought on agriculture, we collected
data on agricultural production from Table 5457 - Brazilian Municipal Agricultural Research
(SIDRA/IBGE). The variables used for the purpose of this study were: Planted Area and
Harvested Area, in hectares, Average Productivity, in kilograms per hectare harvested, and
Value of Agricultural Production, in currency units (R$). All these variables are available at
the municipal and year level, by crop type. It is particularly interesting to observe the het-
erogeneity of results by crop type, since crops such as corn and beans are more associated
with family production and have a large weight in the total planted area of the Brazilian
semiarid (IBGE, 2006).

By subtracting the variables Planted Area and Harvested Area, we can calculate a measure
of Lost Area. This new Lost Area variable is convenient for this study, since, to evaluate
drought losses, it is convenient to compare a counterfactual - Planted Area - in relation to
an observed output - Harvested Area. Therefore, the calculation of the Lost Area is given
by the following equation:

CAmtc − H Amtc
TLAmtc =
CAmtc

Where TLAmtc is the percentage of area that was cultivated but not harvested from crop c
in municipality m and year t. We also performed tests for the Value of Agricultural Pro-
duction, in order to give robustness to our tests. However, this variable may suffer from
relative price fluctuations of crop types that do not have relationship with our mesaure of
drought shocks (under the assumption that landholders are price takers). Thus, we also as-
sess the effects on the productivity per hectare harvested, which allows a complementary
analysis, measuring whether there are losses within the harvested areas, in adition to the
loss of the cultivated areas.

3.3 Descriptive Statistics

Our sample is a balanced panel that comprises the 1,262 municipalities within the Brazilian
Semi-Arid. Considering the 12 years of the panel, we have a total of 15,144 observations.

11
Table 2 summarizes our main descriptive statistics. The average municipality has a lost area
of 13.8% yearly. This an extensive area and much higher than the rest of the country, which
loses on average 1.2% of planted area each year. As expected, drought is also important to
the municipalities in our sample: the average rainfall deviation is 0.92. In addition, 45.1%
of the municipalities in an year suffer from droughts and 15.6% from extreme droughts.

Table 2: Summary Statistics: yearly muncipality data 2006-2017, Brazilian Semi-Arid

Variables Mean Std. deviation Min Max Number of observations

Lost Area 13.847 25.045 0 100 15,128

Ln (Output) 7.825 1.872 0 14.394 15,143
Rainfall Deviation 0.120 0.920 -3.435 2.342 15,144
(Dummy of) Drought 0.451 0.497 0 1 15,144
(Dummy of) Extreme Drought 0.156 0.362 0 1 15,144
Forest 0.543 0.250 0.007 .993 15,144
Temperature deviation 1.034 1.121 -1.949 4.939 15,144
Note: Yearly observations by municipality, from 2006 to 2017. Data originally from: (i) IBGE;(ii) the Ter-
restrial Air Temperature and Terrestrial Precipitation: 1900-2017 Gridded Monthly Time Series, Version
5.01; (iii) Mapbiomas.

4 Empirical Strategy
We expect that a drought shock, as defined in the previous section, might have a negative
effect on variables related to agricultural production. For identification, we rely on the
data panel structure, which allows us to control unobservable variables from municipalities
and common annual shocks, including municipality and year fixed effects. Therefore, our
benchmark model to be estimated is:

Yit = β 0 + β 1 Drought Shock it + γXit + αt + λi + ε it (6)

Where Yit is a variable that measures agricultural loss. Throughout the paper, we are going
to use: (i) lost area; (ii) output value and (iii) crop yields, as our main dependent variables.
The coefficient - β 1 - is our coefficient of interest and measures the average treatment ef-
fect in municipalities within the Brazilian Semi-Arid. Xit is a vector of covariates that might
also affect agricultural losses, as temperature deviations from historical averages. The term
αt is a time fixed effect, which captures yearly shocks common to all municipalities, λi is
the municipality fixed effect, which captures effects of unobservable and invariant vari-
ables in time. The model error term is ε it . In alternative specifications, we also allow for

12
municipality-specific time trends.

Our key identifying assumption relies on the hypothesis that drought shocks, defined as
annual rainfall deviations from historical averages (from 1900 to 2017), are uncorrelated
with other determinants of agricultural production, conditional on municipality and year
fixed-effects. In order to establish a causal relationship between drought shocks and agri-
cultural outcomes, we need to take into account the possible omitted variable bias that can
arise from variables that over time are correlated with drought and agricultural results.
As the region is inherently dry, it is plausible that individuals already adapt to local dry-
ness. That is why we scale by standard deviation, which means that we are looking to
intense variations in proportion to the municipalities‘usual variation. Moreover, we in-
clude controls for time-varying factos that could affect agricultural outcomes as well, such
as temperature deviations from historical averages. As it is not possible to control for all
unobservable time-varying factors, we conduct some robustness checks and placebo tests,
which are discussed in the next section.

Moreover, the variable Drought Shock it potentially has spatial correlation problems since it
is originally at the grid level and we calculate the nearest grid to each municipality’s cen-
troid to attribute municipal data. In this case, the attribution of values by municipality is
clustered. Therefore, standard errors must be adjusted, even when the estimate considers
fixed-effects (Abadie et al., 2017). Indeed, it is important to account for spatial and tempo-
ral dependence in climatic exposure (Hsiang, 2016). In order to account for this potential
bias, we make use of two different strategies: (i) robust standard errors are clustered at
the pixel level, since the interpolation necessarily leads to spatial dependence; and (ii) we
apply “Conley” spatial standard errors (Conley, 1999).

5 Results

5.1 Main results

This section presents the main results of this paper. To anticipate part of the discussion,
Figure 2 displays a binned scatterplot generated by regressing rainfall deviation and the
total cropped area lost. Municipality fixed-effects are also included. The figure demonstrats
that, on average, municipalities affected with rainfall deviation higher than average have
higher cropped area lost, on average. Further, as we can check by a visual inspection of
the figure, it appears that there is a non-linear relationship for extreme values of drought
(Rainfall Deviation greater than historical average for more than one standard deviation).
Therefore, this is an important feature to be tested as well in our main results.

13
 Figure 2: Scatter Plot of Residual Rainfall Deviation and Lost Crop Area

Notes: The figure above depicts the relationship between residual Lost Crop Area and residual Rainfall De-
viation, controlling for municipality fixed-effects. Each observation on the plot is averaged over twenty
equal-sized bins.

In Table 3, we report our baseline results of the relationship between drought and agricul-
tural output. In Columns (1) and (2), we report the corelation between rainfall deviation - a
continuous measure of drought - and the total annual crop area that was lost, using Pooled
OLS estimation. From Column (1), a one standard deviation increase in drought in relation
to the historical average is associated with a loss of 6.1% of loss in cropped area, which is
44% of the dependent variables mean. In Column (2), we control for temperature devia-
tion. Results are similar and slightly lower, which is suggestive of no strong correlation
between temperature and rainfall deviation.

In Columns (3) and (4) of Table 3, we include municipality and year fixed effects. The panel
structure of the data allows us to control for unobsorvable time-invariant heterogeneity
that could be correlated with our measure of drought shock. Moreover, the inclusion of
year fixed effects in Column (4) controls for common shocks that might affect every mu-
nicipalities each year. In this context, the inclusion of year fixed effects is important since
it controls for climate shocks that might not be heterogeneous in space and other factors
such as the conditions of farm-credit that can vary on an annual basis. The estimated co-

14
efficient with the addition of municipality fixed effects is even larger - 7.76. However, the
inclusion of year fixed effect reduces the estimated coefficient to 3.38, which is suggestive
of important yearly effects on the loss of cropped area.

As it appears from Figure 2, it seems that there is a non-linear relationship as rainfall devi-
ation achieves higher values. To test this hypothesis, Columns (5) and (6) includes dummy
variables for years with high and extreme drought. We define a variable Drought as a
dummy with value equal to 1 if the variable Rainfall Deviation is between 0 and 1. That is
to say, if, in a given pair municipality x year, rainfall is less than historical average for up
to one standard deviation, the variable Drought equals one. Similarly, we define Extreme
Drought as deviations from historical average above one standard deviation. In Column
(6), we add specific municipality trends, in order to account for specific unobserved factors
varying in time. The results from Columns (5) and (6) are suggestive of a non-linear rela-
tionship, where extreme drought appear to have an important effect in the loss of agricul-
tural area. Compared to years without drought, an year with moderate drought has 3.2%
more of lost area and years with extreme drought have 5.9% more lost area. In addition,
the estimated coefficient for Extreme Drought is 87% larger than the estimated coefficient for
Drought.

Table 3: Pooled OLS and Fixed Effect Models of Drought Shocks Impact on Lost of Crop
Area
(1) (2) (3) (4) (5) (6)
VARIABLES Lost Crop Area Lost Crop Area Lost Crop Area Lost Crop Area Lost Crop Area Lost Crop Area

Rainfall deviation 6.133*** 5.962*** 7.769*** 3.381***

(0.477) (0.580) (0.575) (0.639)
Dummy of Drought 3.843*** 3.205***
(0.881) (0.928)
Dummy of Extreme Drought 5.312*** 5.998***
(1.644) (1.557)
Temperature Deviation 0.273 -0.379 -1.656*** -1.358*** -1.825***
(0.523) (0.493) (0.539) (0.520) (0.527)

Observations 15,128 15,128 15,128 15,128 15,128 15,128

R-squared 0.051 0.051 0.352 0.405 0.404 0.515
Municipality FE N N Y Y Y Y
Year FE N N N Y Y Y
Municipality Trend N N N N N Y
Cluster Grid Grid Grid Grid Grid Grid
Number of municipalities 1262 1262 1262 1262 1262 1262
Number of clusters 359 359 359 359 359 359
Mean of dependent variable 13.85 13.85 13.85 13.85 13.85 13.85

Note: Robust standard errors are clustered by grids. Significance: *** p<0.01, ** p<0.05, * p<0.1.

In Table 3, we analyzed the effects on the loss of cropped area caused by droughts. How-
ever, it is also interesting to understand what happens to agricultural output, which is what
ultimately translates into farm income.

15
In this sense, Table 4 displays results for the effects of drought shocks on the value of agri-
culturual output. Column (1) presents the fixed effects model, adding temperature devia-
tion as a covariate. The estimated coefficient implies a sizable effect on agricultural output:
an year with rainfall less than historical average of one standard deviation implies a reduc-
tion in the value of agricultural output of 16.5%. When we add municipality specific time
trends, as in Column (2), the effect is even larger: 18.4%. Columns (3) and (4) explore possi-
ble non-linear effects in agricultural output. Moderate drought, as our definition, implies a
loss of agriculture output between 17.3% and 20.4%, according with the econometric speci-
fication. Extreme drought implies, as expected, higher losses, ranging from 25.8% to 32.4%.
That is, a municipality that suffers with extreme drought is expected to lose between one
quarter and one third of the value of agriculture output.

Table 4: Fixed Effect Models of Drought Shock Impact on Agricultural Output

(1) (2) (3) (4)

VARIABLES Ln(Output) Ln(Output) Ln(Output) Ln(Output)

Rainfall deviation -0.165*** -0.184***

(0.026) (0.025)
Dummy of Drought -0.204*** -0.173***
(0.036) (0.041)
Dummy of Extreme Drought -0.258*** -0.324***
(0.060) (0.066)

Observations 15,143 15,143 15,143 15,143

R-squared 0.831 0.879 0.830 0.878
Municipality FE Y Y Y Y
Year FE Y Y Y Y
Municipality Trend N Y N Y
Controls Y Y Y Y
Cluster Grid Grid Grid Grid
Number of municipalities 1262 1262 1262 1262
Number of clusters 359 359 359 359
Mean of dependent variable 7.825 7.825 7.825 7.825
Note: Robust standard errors are clustered by grids. Significance: *** p<0.01, ** p<0.05, * p<0.1.

It should be noted that the distributional effects of agricultural losses from droughts can
be quite important, since it is probable that these effects are conditoned on the capacity
to make investments in adaptation. This hypothesis will be investigated in the following
subsections.

16
5.2 Effects on different crops as a proxy for distributive effects
As previously discussed, the Brazilian Semi-Arid has an important share of family farming
in the total share of agricultural activities. The main crops cultivated by families in the
Semi-Arid are corn and beans, whereas business farm especialize in the production of sug-
arcane and, to some extent, coffee. Thus, based on this division of labor, we estimate the
effects of drought shocks on different crops as a way to infer the distributional impacts. Ta-
ble 5 displays the effects, by crop, on lost area and productivity, as measured by the natural
logarithm of each crop specific yield. On Panel A, we present the effects on lost area, by
crops. On columns (1) and (2), we evaluate the effects on crops, which tend to be cultivated
by families - beans and corns. These crops suffer the most when there is a drought: a year
with rainfall one standard deviation below its historical average implies a loss of 5% in
cropped area with beans and 6.3% in cropped area with corn. Columns (3) and (4) measure
the effects on the loss of area on two crops more associated to business farm - sugarcane
and coffee. There is no sizable effect associated to rainfall deviation on lost area for those
crops.

When one evaluates the effects on crop yields, the impact is more widespread. Results
on Panel B show negative effects on yields for the four crops evaluated. However, the
effects are stronger for beans and corns - which lose, respectively, 12% and 21.2% of its
yields - than for sugarcane and coffee, which lose 4.8% and 11.4% of its respective yields.
Therefore, aside from important effects for the agricultural sector, drought shocks have
negative distributive consequences, as well.9

9 Assunção and Chein (2016) discuss the impact of climate change on agricultural productivity and rural

poverty.

17
 Table 5: Effects of Drought Shocks on Different Crops

(1) (2) (3) (4)

Beans Corn Sugarcane Coffee
Panel B: Effects on Lost Area

Rainfall deviation 5.003*** 6.383*** -0.209 1.877

(0.783) (0.889) (0.286) (1.145)

Observations 14,772 14,720 6,363 2,317

R-squared 0.520 0.519 0.393 0.310
Mean of dependent variable 16.21 19.64 1.615 1.659
Panel A: Effects on Yield

Rainfall deviation -0.120*** -0.212*** -0.048*** -0.114***

(0.027) (0.030) (0.015) (0.027)

Observations 14,236 13,718 6,352 2,298

R-squared 0.596 0.619 0.652 0.809
Number of municipalities 1260 1260 644 234
Number of clusters 359 359 270 111
Municipality FE Y Y Y Y
Year FE Y Y Y Y
Municipality Trend Y Y Y Y
Controls Y Y Y Y
Cluster Grid Grid Grid Grid
Note: Robust standard errors are clustered by grids. Significance: *** p<0.01, ** p<0.05, * p<0.1.

5.3 Heterogeneity
The results outlined so forth highlight the need to understand how the characteristics of the
municipalities might affect agricultural output. As previously discussed, drought shocks
might have a heterogeneous impact on agricultural output to the level of investment in
adaptation led by each municipality. Therefore, in this section, we examine heterogeneity
in the treatment effects, since landholders and local governments can decide to invest in
adapation such as water availability and forest cover.10

Table 6 displays heterogeneous effects based on the provision of a fundamental public

good: water. The table is divided in two panels. Panel A displays the effects on the loss of
10 Forestcover can act as a buffer to drought shocks, which protects groundwater or intensifies the hydro-
logical cycle, reducing the high evapotranspiration characteristic of these regions, besides acting as a filter,
improving the quality of both groundwater and surface water (Ellison et al., 2017; Lopes et al., 2019).

18
crop area and Panel B displays the effects on the value of agricultural output. Each column
presents heterogeneous effects according to the municipal level of provision of water from
distinct infrastructure levels. Column (1) presents the effects of rainfall deviation and how
it interacts with a network of rural water supply. In relation to lost area, the estimated co-
efficient of the interaction is negative, albeit not statistically robust. As regards the value of
agricultural output (Panel B), the coefficient of the interaction is positive. Taken together, it
implies that the provision of a network of water supply has an effect of protecting produc-
ers from losing output.

Results from Columns (2) and (3) provide a similar interpretation: having wells to collect
water - independently if being within property - also provide protection against drought
shocks. From Column (2), a back of the envelope calculation implies that a drought with
rainfall deviation equal to one standard deviation below historical average and one-third
of properties having wells within property, lost area would be zero, instead of 5.4% in the
case where zero properties have wells.

Columns (4) and (5) dispay heterogeneous effects of more vulnerable methods of gather-
ing water: collecting it from a river or counting on rainfall to have water. In municipalities
where these methods are predominant, the effects of drought shocks are magnified, with
more lost area and output. These results reinforce the discussion from the previous subsec-
tion, which infers that drought shocks have negative distributive effects. From Table 6, we
can see that more vulnerable municipalities, as measured by the provision of a fundamen-
tal public good - network of water supply - suffer the most with drought.

19
 Table 6: Heterogeneity Effects - Water Supply

(1) (2) (3) (4) (5)

Panel A - Dep. Var: Lost Area

Rainfall deviation 4.198*** 5.386*** 4.915*** 2.647*** 0.397

(0.852) (0.710) (0.771) (0.729) (0.684)
Rainfall deviation x Rural Water Supply -2.706
(1.710)
Rainfall deviation x Well water within property -16.419***
(2.206)
Rainfall deviation x Well water outside property -9.905***
(2.215)
Rainfall deviation x Water supply in river 6.708***
(2.576)
Rainfall deviation x Water supplied by rain 10.790***
(1.940)

Panel B - Dep. Var: Ln(Output)

Rainfall deviation -0.245*** -0.230*** -0.237*** -0.122*** -0.005

(0.034) (0.029) (0.031) (0.028) (0.027)
Rainfall deviation x Rural Water Supply 0.265***
(0.060)
Rainfall deviation x Well water within property 0.537***
(0.091)
Rainfall deviation x Well water outside property 0.466***
(0.095)
Rainfall deviation x Water supply in river -0.386***
(0.110)
Rainfall deviation x Water supplied by rain -0.579***
(0.070)

Observations 15,128 15,128 15,128 15,128 15,128

Municipality FE Y Y Y Y Y
Year FE Y Y Y Y Y
Controls Y Y Y Y Y
Cluster Grid Grid Grid Grid Grid
Number of municipalities 1262 1262 1262 1262 1262
Number of clusters 359 359 359 359 359

Note: Robust standard errors are clustered by grids. Significance: *** p<0.01, ** p<0.05, * p<0.1.

Another possible investment in adaptation is the maintenance of forest cover, since it repre-
sents an important factor in maintaining water supply even in dry environments (Ellison et al.,
2012; Ilstedt et al., 2016).11 In this context, the maintenance of tree cover can be seen as an
investment in adaptation, meanwhile it reduces farming area. Thus, Table 7 presents ev-

11 Sant’Anna (2018) show the importance of forest cover to the protection against extreme rainfall in urban

environments.

20
idence of heterogenous effects on the extension of forest cover at the municipality level.
Columns (1) and (2) display the effects of the interaction between drought and forest cover
on lost area, whereas Columns (3) and (4) present the estimated coefficients for the value
of agricultural output. We present results for the continuous mesure of drought - Rainfall
deviation - and for the two dummies that represent moderate and extreme drought.

Table 7: Heterogeneity Effects - Forest Cover

(1) (2) (3) (4)

VARIABLES Lost Area Lost Area Ln(Output) Ln(Output)

Rainfall deviation 5.303*** -0.161***

(1.227) (0.051)
% of Forest in Municipality Area -42.294*** -39.115*** 1.905*** 2.004***
(9.870) (10.176) (0.478) (0.491)
Rainfall deviation x Forest Area -4.257** -0.003
(1.936) (0.075)
Dummy of Drought 6.780*** -0.110
(1.679) (0.087)
Drought x Forest Area -6.318** -0.164
(2.688) (0.139)
Dummy of Extreme Drought 12.000*** -0.469***
(3.542) (0.146)
Extreme Drought x Forest Area -13.284** 0.380*
(5.817) (0.224)

Observations 15,128 15,128 15,143 15,143

R-squared 0.409 0.409 0.832 0.832
Municipality FE Y Y Y Y
Year FE Y Y Y Y
Controls Y Y Y Y
Cluster Grid Grid Grid Grid
Number of municipalities 1262 1262 1262 1262
Number of clusters 359 359 359 359
Note: Robust standard errors are clustered by grids. Significance: *** p<0.01, ** p<0.05, * p<0.1.

Every specification has the expected results for the variables related to drought, with the
expection of the dummy of moderate drought, on Column (4). More interesting, the pres-
ence of forest cover protects cropped areas from being lost and output losses. Therefore,
forest cover has an effect of acting as a buffer of protection against losses when a drought
shock strikes. A visual inspection of this result can be seen on Figure 3, which shows the

21
margin plots of the interaction based on results from Column (1). It is clear from the Figure
that the extension of forest cover provides an important protection against drought shocks,
even for extreme drought (where rainfall deviation exceeds one, for instance).

Figure 3: Margin Plots of Column (1) of Table 7

Notes: The figure above depicts the predicted lost area based on the results of the interaction between rainfall
deviation and forest cover.

5.4 Placebo and Robusteness tests

As argued before, given the structure of the data, there is spatial dependence in the vari-
ables that measure drought shocks. Therefore, it is necessary to adjust standard errors in
order to overcome spatial correlation problems in the independent variable, as proposed
by Conley (1999) and Hsiang (2010). Table 8 presents results comparable to results from
Column (3), in Table 3, for lost area and from Column (1) - Table 4, for agricultural output.
We work with different distance cutoffs, ranging from 50km to 200km. Columns (1)-(3)
display results for lost area as dependent variable and Columns (4)-(6) present results for
agriculture output.

As expected, the estimated coefficient is measured with increasing uncertainty as distance

increases, however results remain robust to the increase of distance buffers, even for a

22
distance as large as 200 km.12 Thus, when we correct for spatial dependence using Conley
procedures, results seem robust to spatial correlation problems.

Table 8: Effects of Drought Shocks - Conley Correction for Spatial Dependence

(1) (2) (3) (4) (5) (6)

VARIABLES Lost Area Lost Area Lost Area Ln(Output) Ln(Output) Ln(Output)
Buffer 50km Buffer 100km Buffer 200km Buffer 50km Buffer 100km Buffer 200km

Rainfall deviation 3.381*** 3.381*** 3.381*** -0.165*** -0.165*** -0.165***

(0.621) (0.882) (1.134) (0.024) (0.034) (0.045)

Observations 15,128 15,128 15,128 15,143 15,143 15,143

R-squared 0.007 0.007 0.007 0.009 0.009 0.009
Municipality FE Y Y Y Y Y Y
Year FE Y Y Y Y Y Y
Municipality Trend N N N N N N
Controls Y Y Y Y Y Y

Note: Standard errors are corrected for spatial dependence using Conley (1999), with distance cutoffs of
50, 100 and 200 kilometers. Column (1) Significance: *** p<0.01, ** p<0.05, * p<0.1.

An additional test to be performed relates to the timing of treatment. That is to say, one
should expect to find a relationship between drought shocks and agricutltural outcomes
only at the current year when treatment occurs. Therefore, on Table 9, we test whether
rainfall deviation from previous and forward years affect our main dependent variables.
This test works as a placebo where years other than the actual year of a drought should
have no effect on lost area and on agricultural output.

Columns (1) to (3) present results regarding the impact on lost crop area. In Column (1),
we test the effects of rainfall deviation in the previous year. Column (2) presents the results
for forward year effects and Column (3) presents estimates considering previous, current
and forward years. Every specification has municipality and year fixed effects, controls for
temperature deviation and includes a municipality specific trend. Results are reassuring
that our results are not driven by spurious correlation, since there is not any associated
effect of previous and forward drought shocks on lost area.

Columns (4) to (6) reproduce the same structure of estimation using, instead, agricultural
output as the dependent variable. As regards output, there is a positive association be-
tween previous drought shock and output, even when considering current effects (as in
Column (6)). Forward effects are also slight positive in Column (5), but are not robust to
the inclusion of previous and current drough shocks. However, one question remains: why
12 A circle with a radius of 200 km has an area of 125,663 sq km, which is 13% of the Brazilian Semi-Arid

total area.

23
should there be a positive effect on agricultural production one year after a drought shock?
A possible explanation is given in the next subsection.

Table 9: Placebo test: effects of current, previous and forward years

(1) (2) (3) (4) (5) (6)

VARIABLES Lost Crop Area Lost Crop Area Lost Crop Area Ln(Output) Ln(Output) Ln(Output)

Rainfall deviationt−1 -0.696 -0.097 0.082*** 0.036**

(0.553) (0.606) (0.018) (0.017)
Rainfall deviationt 2.960*** -0.117***
(0.679) (0.028)
Rainfall deviationt+1 -0.022 0.442 0.049* 0.034
(0.775) (0.764) (0.029) (0.031)

Observations 15,128 13,866 13,866 15,143 13,881 13,881

R-squared 0.513 0.537 0.539 0.878 0.886 0.887
Municipality FE Y Y Y Y Y Y
Year FE Y Y Y Y Y Y
Municipality Trend Y Y Y Y Y Y
Controls Y Y Y Y Y Y
Cluster Grid Grid Grid Grid Grid Grid
Number of municipalities 1262 1262 1262 1262 1262 1262
Number of clusters 359 359 359 359 359 359

Note: Robust standard errors are clustered by grids. Significance: *** p<0.01, ** p<0.05, * p<0.1.

5.5 Deforestation as a response to drought

Table 9 shows a positive associaton between the deviation of precipitation in the previous
year and current production, with no associaton with the current loss of area. A possible
explanation relies on the fact that forest cover is a stock of biomass that can be converted
into fertile soil, at least in the short term (Silva Neto et al., 2019), through slash and burn
cultivation.13 Therefore, inasmuch rural producers believe a drought shock in the previous
year is random and there will be mean reversion to the historical average in the current
year of cultivation, there is an incentive to clear land, in order to expand production. If this
is so, we should expect a positive relationship between a drought shock in the previous
year and deforestation in the current year.

Table 10 displays results for linear and non-linear effects of drought on the rate of forest
loss. Column (1) displays a positive relationship between previous rainfall deviation and
current deforestation. On Column (2), we provide the estimated coefficient for the current
relationship, which is positive, albeit not statistically significative. When we consider an
estimation with both current and previous drought shocks, as in Column (3), results are

13 Most of the annual crops in Caatinga still use slash and burn system (Menezes et al., 2012). This is a

historical process in Brazil, as described by Dean (1997).

24
robust only to previous rainfall deviation. From Columns (4)-(6), we reproduce the same
exercise, using instead our meausures of moderate and extreme drought, in order to ob-
tain estimates of a non-linear relationship. Indeed, results are more robust and stronger (3
times higher) for the dummy of extreme drought in the previous year. We conclude, thus,
that extreme droughts are conducive to deforestation in the following year, as a way to
recompose soil fertility in the short run.

Table 10: Effects of Drought Shocks on Deforestation

(1) (2) (3) (4) (5) (6)

VARIABLES Forest Loss Forest Loss Forest Loss Forest Loss Forest Loss Forest Loss

Rainfall deviationt−1 0.315* 0.319*

(0.179) (0.183)
Rainfall deviationt 0.041 0.069
(0.242) (0.243)
Dummy of Droughtt−1 0.441 0.460*
(0.271) (0.277)
Dummy of Extreme Droughtt−1 1.480** 1.506***
(0.595) (0.581)
Dummy of Droughtt 0.103 0.190
(0.442) (0.435)
Dummy of Extreme Droughtt 0.012 0.183
(0.612) (0.604)

Observations 15,130 15,130 15,130 15,144 15,144 15,144

R-squared 0.572 0.572 0.572 0.527 0.526 0.527
Municipality FE Y Y Y Y Y Y
Year FE Y Y Y Y Y Y
Control Y Y Y Y Y Y
Cluster Grid Grid Grid Grid Grid Grid
Number of municipalities 1262 1262 1262 1262 1262 1262
Number of clusters 359 359 359 359 359 359
Mean of dependent variable 6.642 6.642 6.642 6.763 6.763 6.763

Note: Robust standard errors are clustered by grids. Significance: *** p<0.01, ** p<0.05, * p<0.1.

6 Conclusion
The Brazilian Semi-Arid is a region prone to suffer droughts. The low development of
the region if compared to other parts of Brazil has always been associated to the climate
conditions of the drylands. However, the frequent repetition of severe droughts led to
extreme events, resulting in food insecurity, poverty and migration towards other parts
of Brazil. But climatic conditions are not the only cause for the social vulnerability in the
region: the concentration of economic and political power and the deficit of public policies
maintains a vicious cycle of poverty and social vulnerability in the region.

25
As climate change is expected to increase the frequency and severity of droughts, we ana-
lyze the impacts of extreme drought shocks on agricultural outputs, namely lost area, value
of agriculture output and yields in the region. Our results show that, even with adaptation,
drought shocks have important impacts, substantially among crops used in family farm-
ing, with impacts on the living conditions of the poorest. We also assess heterogenous
effects according to the provision of water supply and the maintenance of tree cover. These
heterogeneous effects sign for the importance of sound public policies that provide an ade-
quate stream of water and for the maintenance of forest cover, since it acts a buffer against
drought shocks.

Finally, we also show that a common response among municipalities in the Semi-Arid is to
expand the rate of deforestation in the following year after a drought shock. We interpret
this result as an evidence that as rural producers believe a drought shock in the previous
year will present mean reversion to the historical average in the current year of cultivation,
there is an incentive to clear land, in order to expand production. This is so because trees
represent a stock of biomass that can be converted in a fertile soil in the short term through
slash and burn cultivation.

Overall, our results point to important effects of drought shocks in agricultural outputs.
These effects are magnified when adequate infrastructure is lacking. Moreover, crops as-
sociated with family farming suffer significantly more, which we interpret as a sign that
distributive effects are important as well. Therefore, despite climate change might increase
the occurrence of severe droughts, our results show that there is much to be done in terms
of public policies that can reduce the deleterious impacts of drought shocks.

References
Abadie, A., Athey, S., Imbens, G. W., and Wooldridge, J. (2017). When should you adjust
standard errors for clustering? Technical report, National Bureau of Economic Research.

Ab’Sáber, A. N. (1999). Sertões e sertanejos: uma geografia humana sofrida. Estudos avança-
dos, 13(36):7–59.

Assunção, J. and Chein, F. (2016). Climate change and agricultural productivity in brazil:
future perspectives. Environment and Development Economics, 21(5):581–602.

Blanc, E. and Schlenker, W. (2017). The use of panel models in assessments of climate
impacts on agriculture. Review of Environmental Economics and Policy, 11(2):258–279.

Brito, S. S. B., Cunha, A. P. M., Cunningham, C., Alvalá, R. C., Marengo, J. A., and Carvalho,

26
 M. A. (2018). Frequency, duration and severity of drought in the semiarid northeast brazil
region. International Journal of Climatology, 38(2):517–529.

Burgess, R., Costa, F. J., and Olken, B. A. (2018). Wilderness conservation and the reach
of the state: Evidence from national borders in the amazon. National Bureau of Economic
Research.

Burgess, R., Deschenes, O., Donaldson, D., and Greenstone, M. (2017). Weather, climate
change and death in india. University of Chicago.

Campos, J. N. B. (2015). Paradigms and public policies on drought in northeast brazil: A

historical perspective. Environmental management, 55(5):1052–1063.

Cirilo, J. A. (2008). Public water resources policy for the semi-arid region. estudos avançados,
22(63):61–82.

Cirino, P. H., Féres, J. G., Braga, M. J., and Reis, E. (2015). Assessing the impacts of
enso-related weather effects on the brazilian agriculture. Procedia Economics and Finance,
24:146–155.

Collins, M., Knutti, R., Arblaster, J., Dufresne, J.-L., Fichefet, T., Friedlingstein, P., Gao, X.,
Gutowski, W. J., Johns, T., Krinner, G., et al. (2013). Long-term climate change: projec-
tions, commitments and irreversibility. In Climate Change 2013-The Physical Science Basis:
Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel
on Climate Change, pages 1029–1136. Cambridge University Press.

Conley, T. G. (1999). Gmm estimation with cross sectional dependence. Journal of economet-
rics, 92(1):1–45.

Cunha, A. P., Zeri, M., Deusdará Leal, K., Costa, L., Cuartas, L. A., Marengo, J. A.,
Tomasella, J., Vieira, R. M., Barbosa, A. A., Cunningham, C., Garcia, J. V. C., Broedel,
E., Alvalá, R., and Ribeiro-Neto, G. (2019). Extreme drought events over brazil from 2011
to 2019. Atmosphere, 10(11):642.

De Castro, J. (1952). The geography of hunger, volume 74. LWW.

Dean, W. (1997). With broadax and firebrand: the destruction of the Brazilian Atlantic Forest.
Univ of California Press.

Dell, M., Jones, B. F., and Olken, B. A. (2014). What do we learn from the weather? the new
climate–economy literature. Journal of Economic Literature, 52(3):740–798.

Deschênes, O. and Greenstone, M. (2007). The economic impacts of climate change: evi-

27
 dence from agricultural output and random fluctuations in weather. American Economic
Review, 97(1):354–385.

Ellison, D., Morris, C. E., Locatelli, B., Sheil, D., Cohen, J., Murdiyarso, D., Gutierrez, V.,
Van Noordwijk, M., Creed, I. F., Pokorny, J., et al. (2017). Trees, forests and water: Cool
insights for a hot world. Global Environmental Change, 43:51–61.

Ellison, D., N. Futter, M., and Bishop, K. (2012). On the forest cover–water yield debate:
from demand-to supply-side thinking. Global Change Biology, 18(3):806–820.

Furtado, C. (1989). A fantasia desfeita. Paz e Terra, Rio de Janeiro, 2 edition.

Furtado, C. (2005). Formação econômica do Brasil. Companhia Editora Nacional, São Paulo,
32 edition.

Gutiérrez, A. P. A., Engle, N. L., De Nys, E., Molejón, C., and Martins, E. S. (2014). Drought
preparedness in brazil. Weather and Climate Extremes, 3:95–106.

Hsiang, S. (2016). Climate econometrics. Annual Review of Resource Economics, 8:43–75.

Hsiang, S. and Kopp, R. E. (2018). An economist’s guide to climate change science. Journal
of Economic Perspectives, 32(4):3–32.

Hsiang, S. M. (2010). Temperatures and cyclones strongly associated with economic pro-
duction in the caribbean and central america. Proceedings of the National Academy of sci-
ences, 107(35):15367–15372.

Ilstedt, U., Tobella, A. B., Bazié, H., Bayala, J., Verbeeten, E., Nyberg, G., Sanou, J., Bene-
gas, L., Murdiyarso, D., Laudon, H., et al. (2016). Intermediate tree cover can maximize
groundwater recharge in the seasonally dry tropics. Scientific reports, 6:21930.

Leal, V. N. (2012). Coronelismo, enxada e voto: o município e o regime representativo no Brasil.

Editora Companhia das Letras.

Lopes, A. F., Macdonald, J. L., Quinteiro, P., Arroja, L., Carvalho-Santos, C., Cunha-e Sá,
M. A., and Dias, A. C. (2019). Surface vs. groundwater: The effect of forest cover on the
costs of drinking water. Water Resources and Economics, 28:100–123.

Marengo, J. A., Cunha, A. P., and Alves, L. M. (2016). A seca de 2012-15 no semiárido do
nordeste do brasil no contexto histórico. Revista Climanálise, 3:49–54.

Matsuura, K. and Willmott, C. J. (2018). Terrestrial air temperature: 1900-2017 gridded

monthly time series.

28
Medeiros, S. d. S., Pinto, T. F., Hernan Salcedo, I., Cavalcante, A. d. M. B., Perez Marin,
A. M., and Tinôco, L. B. d. M. (2012). Sinopse do censo demográfico para o semiárido brasileiro.
Instituto Nacional de Seminário (INSA).

Menezes, R., Sampaio, E., Giongo, V., and Pérez-Marin, A. (2012). Biogeochemical cycling
in terrestrial ecosystems of the caatinga biome. Brazilian Journal of Biology, 72(3):643–653.

Noack, F., Riekhof, M.-C., and Di Falco, S. (2019). Droughts, biodiversity, and rural incomes
in the tropics. Journal of the Association of Environmental and Resource Economists, 6(4):823–
852.

Prado Jr, C. (2017). História econômica do Brasil. Brasiliense.

Rocha, R. and Soares, R. R. (2015). Water scarcity and birth outcomes in the brazilian semi-
arid. Journal of Development Economics, 112:72–91.

Sant’Anna, A. A. (2018). Not so natural: Unequal effects of public policies on the occurrence
of disasters. Ecological economics, 152:273–281.

Silva, F. P. d., Araújo, J. A., Costa, E. M., and Vieira Filho, J. E. R. (2016). Eficiência produtiva
e pobreza rural no nordeste brasileiro.

Silva Neto, E. C. d., Pereira, M. G., Frade Junior, E. F., Silva, S. B. d., Carvalho Junior, J.
A. d., and Santos, J. C. d. (2019). Temporal evaluation of soil chemical attributes after
slash-and-burn agriculture in the western brazilian amazon. Acta Scientiarum. Agronomy,
41.

Silvia, V. M. A., Patrício, M. C. M., Ribeiro, V. H. A., and Medeiros, R. M. (2013). O desastre
seca no nordeste brasileiro. POLÊM! CA, 12(2):284–293.

Taraz, V. (2017). Adaptation to climate change: Historical evidence from the indian mon-
soon. Environment and Development Economics, 22(5):517–545.

Travassos, I. S., Souza, B. d., and Silva, A. d. (2013). Secas, desertificação e políticas públicas
no semiárido nordestino brasileiro. Okara: Geografia em debate, 7(1):147–164.

Vieira, R. M. d. S. P., Tomasella, J., Alvalá, R., Sestini, M., Affonso, A., Rodriguez, D., Bar-
bosa, A., Cunha, A., Valles, G., Crepani, E., et al. (2015). Identifying areas susceptible to
desertification in the brazilian northeast. Solid Earth, 6(1):347–360.

Wani, S. P., Dixin, Y., Li, Z., Dar, W. D., and Chander, G. (2012). Enhancing agricultural
productivity and rural incomes through sustainable use of natural resources in the semi
arid tropics. Journal of the Science of Food and Agriculture, 92(5):1054–1063.

29