Environmental Regulation, Technology Adoption and Structural
Transformation: Evidence from Brazilian Sugarcane Industry
*
Francisco Costa
„
Francisco Luis Lima Filho
May 26, 2015
[PRELIMINARY AND INCOMPLETE.]
Abstract
We estimate the effects of the adoption of mechanized agriculture led by a new environmental regulation on structural change of local labor markets within a large emerging
country, Brazil. In 2002, the state of São Paulo passed a law outlying the timeline to end
sugarcane pre-harvest burning in the state. The environmental law led to the fast adoption of
mechanized harvest. We investigate if the labor intensity of sugarcane production decreases;
and, if so, if it leads to structural changes in the local labor market. We use satellite data
containing the type of sugarcane harvesting – manual or mechanic harvest – paired with
official labor market data. We find no strong evidence that the mechanization of the field
significantly increased the utilization of formal workers or reduced formal labor intensity in
the sugarcane sector. We find, however, that the new technology led to a reduction in the
share of workers in other agricultural sectors which was compensated by an increase in the
share of workers in the “construction and services” sector and, not significantly, in the manufacturing sector.
Our results also suggests that mechanization improved formal workers’
situation via higher wages in the manufacturing and agriculture sectors, particularly in the
sugarcane sector.
1
Introduction
The main focus of environmental regulation is to correct for externalities and other market
failures.
Very often, in order to comply with new environmental goals, the productive sector
needs to adopt new and cleaner technologies. In particular in agriculture, technological changes
may trigger structural transformation as it changes the factor-content of production in the field
and/or generates demand for services and manufactured goods. Matsuyama (1992) and Bustos
et al. (2013) argue that the positive impacts of agricultural productivity gains on industrialization
occur only in closed economies – e.g.
productivity growth in agriculture can release labor or
generate demand for manufacturing goods. In open economies, however, the authors point out
that the impacts are indeterminate because comparative advantage in agriculture could reduce
industrial growth – e.g.
*
„
the agricultural sector may employ more workers due to the higher
FGV/EPGE, [email protected]
FGV/EPGE, [email protected]
2
productivity, what would harm the manufacturing sector and its benefits from external economies
of scale.
This paper estimates the causal effects of the adoption of mechanized agriculture led by a
new environmental regulation on structural change of local labor markets within a large emerging
country, Brazil.
Our object is an environmental law aimed at ending sugarcane pre-harvest
burning, which promoted the adoption of mechanical sugarcane harvesting. We investigate two
underlying questions: is harvest mechanization a labor saving technical change – i.e., does the
labor intensity of sugarcane production decrease?; and, if so, does harvest mechanization lead to
structural changes in the local labor market?
Sugarcane is one of the main crops in Brazil and world’s largest crop by production quantity
(Walter et al., 2014). In 2012, FAO estimates that sugarcane was cultivated on about 26 million
hectares, in more than 90 countries, with a worldwide harvest of 1.83 billion tons. Sugarcane
can be harvested using two different technologies: the traditional one, by hand which requires
1
pre-harvest field burning; or mechanically, mainly without pre-harvest burning.
Sugarcane
straw burning is responsible for a great amount of pollutant gases in atmosphere (Macedo et al.,
2008) which cause respiratory diseases in the local population (Cançado et al., 2006; Dominici
et al., 2014; Rangel and Vogl, 2015). Many other countries already implemented environmental
regulation to correct for this externality that arises from the production side of the economy
(Greenstone, 2002; Greenstone and Hanna, 2014; Barreca et al., 2014).
In Brazil, the state of São Paulo – which is the richest and largest producer state in the
country – passed a law in 2002 outlying a timeline to end sugarcane pre-harvest burning on
large properties by 2021. In 2007, a cooperation protocol was sealed between São Paulo state
and the Organization of Sugarcane Producers (ORPLANA) in which the phase out process was
accelerated.
The deadline for extinguishing pre-harvest burning was shortened to 2014, and
an agro-environmental certificate for unburnt sugarcane was created.
As a consequence, the
environmental law and the new protocol led to the fast adoption of mechanized harvest. In 2006,
38.2% of the sugarcane harvest area in São Paulo was unburnt, while in 2012 we observe that
73.8% of the harvested area used unburnt mechanical harvesting, despite of the increase in total
harvest area in this period, Figure 1 compares sugarcane harvesting in these two years.
The expansion of mechanical harvest can affect labor demand in the sugarcane sector through
two channels.
First, less workers are needed to harvest the same planted area.
Second, field
mechanization may change the work relations since it may changes the capital-labor ratio used
in production, e.g.
mechanical harvesting uses more formal workers and less temporary, and
mostly informal, workers than hand harvesting. Given the size of this sector in many of these
localities, a technological change of this scale in an important sector could have impacts on the
remaining sectors of the economy.
We use two main datasets to investigate the impact of the adoption of mechanical harvest on local labor market structure: satellite data with sugarcane production by harvest type
(CANASAT/INPE), and the official records with the universe of formal workers in Brazil (RAIS/MTE).
We use the remote sensing data to create an index for the adoption of mechanization in sugarcane
production in different microregions, as defined by the IBGE. In order to characterize the evolu-
1
Although mechanical harvest can be made with or without pre-harvest burning, Nyko et al. (2013) shows
that around 80% of the mechanical harvesting is unburnt.
3
tion of mechanization in sugarcane harvesting, we define a
clean adoption index
as the fraction
of the area (number of pixels) without pre-harvesting burning out of the total area (number of
pixels) with sugarcane production relative to the baseline fraction of mechanical harvest in 2006
(the first year for which we have harvest data). We pair this index with labor market data for
all microregions and municipalities in the state of São Paulo, for 2006 and 2012.
In order to unveil the causal relation between the adoption of mechanized harvest technology
and the evolution of labor market outcomes between 2006 and 2012, we use an instrumental
variable strategy. We use land slope from georeferenced topographical data (TOPODATA) as an
instrument for the adoption of harvest mechanization. The intuition for this instrument is that
it is more costly to use mechanical harvest in steeper plots of land. Our identifying assumption is
that declivity does not affect the evolution of labor market outcome directly, but only indirectly
via the adoption of technologies in the field.
Using the registers of all formally employed workers aggregated at the sector-microregion
(local labor marker) level, we find weak evidence that mechanization led to structural transformation of labor markets. We find that by 2012 the adoption of clean harvesting reduced, not
statistically significantly, the formal labor intensity per plot of land and the employment share
of the sugarcane sector at the microregion level. We find, however, that the new technology led
to a reduction in the share of workers in other agricultural sectors which was compensated by
an increase in the share of workers in the “construction and services” sector and, not statistically
significantly, in the manufacturing sector.
We also find that the adoption of mechanical har-
vesting increased average wages in the agriculture and manufacturing sectors in a economically
meaningful and statistically significant way, particularly in the sugarcane sector. These results
may suggest that mechanization of the field led to an increase in workers’ productivity in a
manner that it reduced formal labor intensity and increased wages in the agriculture sector.
A lot of attention has been drawn to understanding the determinants of technology adoption
in non-developed countries (Esther Duflo and Robinson, 2006; Conley and Udry, 2010; Bandiera
and Rasul, 2002). Foster and Rosenzweig (2010) argues that technology profitability is key for
technology adoption, and that agents decide to use a technology based on the gain in welfare,
not necessarily the pecuniary profits.
Since welfare cannot be measured directly, technologies
which address health externalities – e.g. bed nets, water purifiers, or vaccines – adoption will
depend on how health is valued by the individuals. In our case, environmental regulation aimed
at reducing the level of pollutant gases, may trigger the adoption of improved technology which
may impact labor markets. In this way, this paper contributes to the literature of environmental
regulation consequences on local labor markets (Deschenes, 2010; Kahn and Mansur, 2013), and
more broadly to the literature on the consequences of technology adoption on local development
(Berman et al., 1998; Beaudry et al., 2006), by investigating how sugarcane harvest mechanization
affected workers’ labor market outcomes.
At last, we also contribute to the literature on structural transformation. Matsuyama (1992)
shows that for small open economies facing perfectly elastic demand for agricultural and manufacturing goods the demand and supply channels on labor markets are no longer operative.
His model – which has only one type of production factor, labor – predicts that a Hicks-neutral
increase in agricultural productivity reduces the industrial sector by reallocating labor towards
4
agriculture. Bustos et al. (2013), however, extend Matsuyama (1992) by using two production
factors: land and labor.
In this case, technical change can be factor-biased.
When technical
change in agriculture is strongly labor saving, an increase in agricultural productivity leads to
manufacturing growth as it increases residual labor supply even in open economies. Bustos et al.
(2013) provides direct empirical evidence on the impacts of technical change in agriculture on the
industrial sector by studying the adoption of genetically engineered soybean seeds (a labor saving
technology change), and the adoption of second-harvest maize (a labor demanding technology
change) in Brazil. Their findings corroborate the model predictions. Our findings, although not
all statistically significant, goes in the same direction, the adoption of a labor saving technology
– mechanical harvesting, which weakly reduced the labor-land ratio in production –, weakly
increased the size of the manufacturing and services sector.
The remaining of the paper is organized as follows. Section 2 provides background information
about the sugarcane sector in Brazil. Section 3 describes the data. Section 4 discuss our empirical
strategy, and Section 5 presents the results. Section 6, presents the results using municipality
level data as a robustness check, and section 6 concludes.
2
Background
In this section we briefly discuss the sugarcane industry background in Brazil and in São Paulo
state, the largest producer in the country. Two main aspects are discussed: harvesting and labor
market.
Sugarcane is one of the main crops in Brazil and world’s largest crop by production quantity (Walter et al., 2014).
In 2012, FAO estimates about 26.0 million hectares of cultivated
area, in more than 90 countries, with a worldwide harvest about 1.83 billion tons. Brazil being the top producer responsible for more then a quarter of worldwide production. According
to UNICA (Sugarcane Industry Union), São Paulo state contributes with more then a half of
Brazilian production as in Table 1. Sugarcane is a semi-perennial crop that, in São Paulo State,
reaches its maximum vegetative development in April. Its planting can be done at two moments:
at September–October, when twelve-month sugarcane is planted; or at February–March, when
eighteen-month sugarcane is planted. And the sugarcane harvesting happens between April and
December.
There are two harvesting technologies: the traditional one, by hand with pre-harvest field
burning; and mechanically, without burning. In hand harvesting, field burning is necessary because of ergonomic restrictions, to clean the area from other weeds and to chase away any dangerous animals. However, mechanical harvesting can be made without burning, keeping leaves
and straws intact, what is called called green or clean harvesting.
Leaves and straws can be
used for energy production or as vegetable coverage for conventional or organic agriculture. The
sugarcane straw burning produces negative externalities and is responsible for a great amount of
pollutant gases in atmosphere (Macedo et al., 2008) which cause respiratory diseases in the local
population (Cançado et al., 2006; Dominici et al., 2014; Rangel and Vogl, 2015). Other environmental problems related to sugarcane straw burning were soil and groundwater contamination
(Brasil, 2009).
With the goal of mitigating climate change problems and respiratory diseases, in 2002, the
5
state of São Paulo passed a law outlying the timeline to end sugarcane pre-harvest burning.
There was two calendars with the targets to adopt mechanical harvest according to the slope
of the land: mechanizable areas, properties bigger than 150 hectares and with slope greater or
equal to 12%; and non-mechanizable areas, properties smaller than 150 hectares or with slope
greater than 12% or with soil structures impossible to mechanize. The different calendars are
presented in Table 2.
In 2007, the Cooperation Protocol was sealed between São Paulo state and the Organization
of Sugarcane Producers (ORPLANA). The phase out process was accelerated with the deadline
being shortened to 2014 for properties with slope less or equal 12%, or to 2021 for properties with
slope greater than 12%, and with the creation of an agro-environmental certificate for unburnt
sugarcane. The Protocol also anticipated the reduction of burning practices from 50% to 70%
in 2010 for properties with slope less or equal 12%, and from 10% to 30% in 2010 for properties
with slope greater than 12%. For this protocol, there is no differentiation between mechanizable
or non-mechanizable areas, only in slope level.
2
Novaes et al. (2007) point out three elements about the adoption of mechanized technology.
First, the prohibition of the pre-harvest burning would reduce directly the hand harvesting
productivity. Second, there are natural conditions for machinery introduction, it is easier to do
mechanic harvesting in more flat terrain. Third, the increase in labor productivity and the low
wages in hand harvesting are an obstacle to mechanical harvesting.
Table 3 shows the evolution on mechanical harvest in the Center-South region, where São
Paulo state is located. Since mechanical harvest can be made with or without pre-harvest burning
we make this distinction in the table. We can see that around 80% of the mechanical harvesting
is unburnt. With this table we see a reversal in the type of sugarcane harvesting, in 2006 hand
harvesting comprehended 63.3% while in 2012 this number reduced to 14.9%. Also we see that
in 2006, around 30% of mechanical harvesting used pre-harvesting burning while in 2012 this
number reduced to 15%.
3
The sugarcane ethanol sector has a bad record in labor relations and environmental rules. .
In 2007, there were around 498 thousand formal employees in the sugarcane crop activity, 54%
in the state of São Paulo. In the sugarcane crop, especially in hand harvesting, the most part
are of unskilled and temporary jobs, with different levels according to seasons. Table 4 shows
that aggregating the whole sugarcane industry in 2007, we had a total of 1.26 million formal
workers, 520 thousand in São Paulo, with 270 thousands working directly in sugarcane crop.
91% of the formal workers were under 50 years old, 52% had 4 years of study or less and 7%
were illiterate.
However, in the past few years, with the expansion of mechanical harvesting,
the sugarcane industry started recruiting more skilled workers, substituting the hand worker for
machine operators and other occupations (e.g., engineers and mechanics).
According to some
estimates from the industry (SGPR, 2009), one harvest machine can substitute eighty hand
2
Braunbeck and Magalhães (2010) discuss some issues that might be considered before mechanizing harvest.
One of them is the maximal slope that is able to mechanizing. Theoretically it should be 46% but because of
ground irregularities and other driving problems the critical slope is 12%. It should lead to the end of sugarcane
production in steep areas. However, with advances in the off-road vehicle engineering such as four wheel drive,
harvesters can be used in terrains with slopes greater than 12% around 15%-18%.
3
Reports about criminal recruitment and overexploitation of labor were common, as well as reports showing
precarious accommodation, high level of work place accident, death by exhausting, and child labor (SGPR, 2009)
6
workers.
In 2009, was signed the National Agreement to Improve Sugarcane Working Conditions between UNICA, by the Federal Government and the National Confederation of Agricultural Workers (CONTAG). This Agreement was about improving better labor practices and to promoting
the reintegration of workers unemployed by the advance in mechanized harvesting.
3
Data
In this section we first describe the three main data sources used in this paper and how we treated
the data. We use three main data sources: formal register of employees (the Annual Report of
Social Information - RAIS); spatial data of sugarcane harvesting type (CANASAT-INPE); and
geomorphometric data in Brazil (TOPODATA).
RAIS consists in a national database that processes formal labor market statistics, containing
all information across years from all formal workers in Brazil. Every commercial establishment
must report to the ministry of Employment and Labor (MTE), through RAIS, information about
all employees. The main variables we use from RAIS are: number of workers (permanent and
temporary), total wage (the sum of all monthly wages at microregion and also at municipality)
and average worker wage (the reason of total wage by the number of workers at microregion
and also at municipality).
Using RAIS, we also generate some variables:
employment share
(the fraction of one sector in the total amount of formal workers), and formal labor intensity
(number of workers per harvested area).
Although we have the individual level data, at this
preliminary stage we aggregate all variables at the microregion level to investigate the impact
of the mechanized harvest adoption in the local labor market. We use CNAE 2.0 to separate
the market in four sectors: sugarcane (code 113), agriculture (codes from 1 to 999 and different
from 113), industry (codes from 1000 to 3390), and construction and services (codes from 4100
to 9799).
CANASAT-INPE has two different data base: CANASAT-Planted Area; and CANASATHarvest. CANASAT-Planted Area annually monitors sugarcane planted area in different classes
and regions between 2003/04 and 2013/14.
CANASAT-Harvest monitors sugarcane harvest
of São Paulo state from 2006/07 to 2012/13, distinguishing two types of harvesting: with or
without pre-harvest burning (Rudorff et al., 2010; Adami et al., 2012). CANASAT-Harvest uses
visual interpretation technique of remote sensing images, obtained between April and December
in each crop year, to identify and map the two types of sugarcane harvest practices. Most of
the sugarcane harvesting is performed during the dry season, when it is relatively favourable to
acquire cloud free images. Identification of harvest type is based on the reflectance difference
between green harvested and pre-harvest burned fields (Aguiar et al., 2011). Figure
1 presents
sugarcane harvest differentiating between clean harvesting and pre-harvesting burning.
TOPODATA project offers the Digital Elevation Model (MDE) and its derivations at national
level. MDE is a computational mathematic representation of a spatial phenomenon. The original
data is from the Shuttle Radar Topography Mission (SRTM), which is an international research
effort that obtained digital elevation models on a near-global scale to generate the most complete
high-resolution digital topographic database of Earth.
The resolution of the raw data is one
arcsecond, but this has only been released over United States territory, for Brazil the resolution
7
is 3 arcseconds. The resolution used in this work is thirty arcsecond the same as half arcminute.
We aggregate the remote sensing data at half arcminute level, also. From CANASAT, we have
the distribution of harvesting and planted area across São Paulo state. Using this information
we add up at municipality level and then at microregion level, our variable is the number of
pixels of each harvesting type or the number of pixels with planted area. Planted area is used
to check exogeneity of our instrument variable.
From TOPODATA, we have the distribution
of slope across São Paulo state. Using this distribution we aggregate at municipality level and
calculate some moments and after at microregion level, for this work we use the first and the
second.
São Paulo state has 64 microregions and we observe sugarcane harvesting in 44 of them. We
create a panel of 44 microregions with observations in 2006 and 2012, where we observe at least
one type of sugarcane harvesting in both years.
We also observe sugarcane harvesting in 408
municipalities from a total of 645 of São Paulo state. We then have 408 municipalities with some
sugarcane harvesting in 2006 and 2012.
4
Empirical Strategy
In this section we describe our empirical strategy. We first define a measure used to capture the
evolution of adoption of mechanical harvest. We then present our main regression specification
and discuss potential identification issues. At last, we propose an instrumental variable strategy
to overcome these issues.
In order to measure the evolution of sugarcane clean harvest over time, we first construct
clean harvest index
an
to capture the degree of adoption of mechanical harvesting within a
microregion in one point in time. This index consists in the share of planted area (number of
pixels) with clean harvesting in a microregion
𝑗
𝐶𝑙𝑒𝑎𝑛𝐼𝑛𝑑𝑒𝑥𝑗𝑡 ≡
where
in year
𝑡:
𝐶𝑙𝑒𝑎𝑛𝑗𝑡
𝐻𝑎𝑟𝑣𝑒𝑠𝑡𝑒𝑑𝐴𝑟𝑒𝑎𝑗𝑡
(1)
𝐶𝑙𝑒𝑎𝑛𝑗𝑡 is the area (number of pixels) with mechanical harvesting, and 𝐻𝑎𝑟𝑣𝑒𝑠𝑡𝑒𝑑𝐴𝑟𝑒𝑎𝑗𝑡 is
the total area (number of pixels) with sugarcane harvesting in microregion
𝑗 and year 𝑡.
Note that
this index is bounded between 0 and 1, so microregions in which all production use mechanized
harvest have an index equal to 1, whether microregions using only pre-harvest burning have an
4
index equal to 0.
Our main measure of interest is the
𝐶𝑙𝑒𝑎𝑛𝐴𝑑𝑜𝑝𝑡𝑖𝑜𝑛𝐼𝑛𝑑𝑒𝑥
which is the evolution of clean
adoption between 2006 and 2012:
𝐶𝑙𝑒𝑎𝑛𝐴𝑑𝑜𝑝𝑡𝑖𝑜𝑛𝐼𝑛𝑑𝑒𝑥𝑗 ≡ 𝐶𝑙𝑒𝑎𝑛𝐼𝑛𝑑𝑒𝑥𝑗,2012 − 𝐶𝑙𝑒𝑎𝑛𝐼𝑛𝑑𝑒𝑥𝑗,2006
where
4
𝐶𝑙𝑒𝑎𝑛𝐼𝑛𝑑𝑒𝑥𝑗,2006
and
𝐶𝑙𝑒𝑎𝑛𝐼𝑛𝑑𝑒𝑥𝑗,2012
(2)
were defined in 1.
Note that microregions with 𝐶𝑙𝑒𝑎𝑛𝑗𝑡 equal to 1 not necessarily have no pre-harvest burning microregions,
because we can have pre-harvest burning and still use mechanical harvesting, as discussed in background section.
8
4.1
Main Regression
Our variables of interest are labor market outcomes, e.g.
labor intensity in the sugarcane in-
dustry, employment share, number of permanent employees, and average wage.
the evolution of these variables relative to 2006 (∆𝑌𝑖𝑗
them consistent with our clean adoption
5
index .
≡ 𝑌𝑖𝑗,2012 − 𝑌𝑖𝑗,2006 ),
We evaluate
in order to make
Variables are indexed by industry sector (𝑖) and
microregion (𝑗 ). Our main regression equation is:
∆𝑌𝑖𝑗 = 𝛽𝐶𝑙𝑒𝑎𝑛𝐴𝑑𝑜𝑝𝑡𝑖𝑜𝑛𝐼𝑛𝑑𝑒𝑥𝑗 + 𝛾𝑋𝑗 + 𝜈𝑖𝑗
where
𝑋𝑗
(3)
is a vector of controls, including total harvested area in 2006 and the share of mechanical
harvest in 2006 (𝐶𝑙𝑒𝑎𝑛𝐼𝑛𝑑𝑒𝑥𝑗,2006 ). The coefficient of interest is
𝛽
which capture the effect of
the adoption of mechanical harvesting on the evolution of labor market outcome
∆𝑌𝑖𝑗 .
We use
heteroskedasticity-consistent standard error estimators. Tables 5 and 6 report summary statistics
of our explanatory variables variables and dependent variables, respectively.
Reduced form estimates of the equation above may have identification issues if there are
any endogeneity between labor market development and the adoption of new technology in the
field.For example, consider a shock in labor market which increases wages at a level that producers
opt to pay the fixed costs of acquiring a new technology.Alternatively, consider, for example, the
case of a positive shock to investments in the local economy, such as government subsidies, which
affects directly both technology adoption and labor market outcomes. In these cases, ordinary
least square estimates would be biased and one would not unveil the causal relation we are
interested at first. Next, we propose a strategy to overcome this issue.
4.2
Instrumental Variable
We use instrumental variable strategy to overcome potential endogeneity between labor market
development and adoption of mechanical harvesting. We use the land slope as an instrument
for the adoption of mechanical harvesting in the sugarcane industry. The intuition for using this
instrument is that it is more costly to mechanize steep areas, so producers with land in steep
terrains would resist more to mechanize their harvest. This is a natural candidate, because even
the original law passed in 2002 acknowledged that the slope of the terrain could impose technical
problems to mechanization. Figure 2 plots the total adoption (number of pixels) of mechanical
harvesting relative to 2006 against the slope of the terrain, as presented in the georeferenced
data. We can see a quadratic relation between the adoption of clean harvesting and the slope of
the terrain. The relation of our Clean Adoption Index and the land slope can be seen in Figure
3.
Our estimates using land slope as an instrument for the adoption of mechanical harvesting
will capture the causal relation between technology adoption and labor market outcomes if: (i)
slope is correlated to the
𝜈𝑖𝑗
𝐶𝑙𝑒𝑎𝑛𝐴𝑑𝑜𝑝𝑡𝑖𝑜𝑛𝐼𝑛𝑑𝑒𝑥;
and (ii) slope is uncorrelated to the error term
in equation 3. In words, slope must be correlated with the adoption of mechanical harvesting
in the sugarcane industry, and slope must have no direct influence on the evolution of labor
5
In Appendix A we also analyze the evolution of each year for sugarcane sector, so we calculate the difference
for years from 2007 to 2012 and compare.
9
market outcomes, except via the mechanical harvesting adopted in the sugarcane sector.
To shed light on the exogeneity assumption (ii), we analyze a panel data of sugarcane planted
area for the whole São Paulo state from 2003 to 2013. Figure 4 plots expansion of planted areas
since 2003 across different slope levels. We can see that the growth of planted areas, which is
the planted area in year
𝑡
divided by planted area in 2003, is relatively homogeneous across
slope levels at least until 2008. We can observe an increase in total planted area which is not
concentrated in terrains with certain slopes.
After 2008, however, we se a greater growth in
planted areas in land with moderate slopes, between 3 and 8 degrees.
Our intuition for this
evidence is that if producers used to choose more flat terrains to begin with, we would observe
a negative relation between expansion of sugarcane production and land slope since before the
Cooperation Protocol was signed, but it seems not to be the case.
Our first stage regression equation is:
𝐶𝑙𝑒𝑎𝑛𝐴𝑑𝑜𝑝𝑡𝑖𝑜𝑛𝐼𝑛𝑑𝑒𝑥𝑗 =Π1 𝑆ℎ𝑎𝑟𝑒𝑆𝑙𝑜𝑝𝑒𝑗0−4 + Π2 𝑆ℎ𝑎𝑟𝑒𝑆𝑙𝑜𝑝𝑒8−12
+ Π3 𝑆ℎ𝑎𝑟𝑒𝑆𝑙𝑜𝑝𝑒12−16
𝑗
𝑗
+ Π4 𝑆ℎ𝑎𝑟𝑒𝑆𝑙𝑜𝑝𝑒16−20
+ Π5 𝑆ℎ𝑎𝑟𝑒𝑆𝑙𝑜𝑝𝑒20
𝑗 + 𝜆𝑋𝑗 + 𝜀𝑗
𝑗
where
𝑋
is a vector of controls, and
microregion
𝑗
𝑆ℎ𝑎𝑟𝑒𝑆𝑙𝑜𝑝𝑒𝑖−𝑖
′
with slope in interval [𝑖,𝑖 ) and
′
(4)
is the share of sugarcane planted area in
𝑆ℎ𝑎𝑟𝑒𝑆𝑙𝑜𝑝𝑒20 is the share of planted area with slope
greater than 20. Note that the omitted category is the share of land with sugarcane planted in
6
slopes between 4 and 8 degrees.
One potential problem that may arise in our regression is the weak instrument problem. This
would be the case if land slope is a poor predictor of the Clean Adoption Index. If this is the
case, our fitted Clean Adoption Index will have little variation. We test for weak instruments
analyzing Kleibergen-Paap F statistic and the Montiel-Pflueger clustered weak instrument test.
Kleibergen-Paap statistic tests whether the instruments jointly explain enough variation in the
multiple endogenous regressors to conduct meaningful hypothesis tests of causal effects (Kleibergen and Paap, 2006). Weak instruments can bias point estimates and lead to substantial test size
distortions. The null hypothesis for Montiel-Pflueger weak instrument test is that the estimator
presents weak instrument bias (Olea and Pflueger, 2013; Pflueger and Wang, 2014).
Table 7 presents the first stage results. The signs of the coefficients are consistent with the
relation between Clean Adoption Index and land slope observed in Figures 2 and 3. As we can
see in the table, our KP F statistic is 14.20, above standard levels. The Montiel-Pflueger weak
instrument test present different values and significance levels for each outcome variable.
We
reject the weak IV bias hypothesis at 20% level for all different outcomes. In sum, these tests
makes us believe that our instrument is sufficiently strong.
5
Results
In this section, we present the results of the impacts of the adoption of clean harvesting in
Brazilian sugarcane industry on local labor market outcomes.
6
Since standard OLS estimates
Notice that slope does not change between years, so we assume that producers don’t directly change terrain,
like terrace for example. Even if this is the case, this would represent part of the differential cost of adopting
mechanical harvesting.
10
are potentially biased do to endogeneity and omitted variables, and focus our discussion on the
estimates from our 2SLS regressions.
Table 8 presents the main results of equation 3, dividing by sectors with a total of four
panels. We report the coefficient of
Clean Adoption Index, robust standard errors and the mean
of each dependent variable in 2006. The unit of observation is a microregion, and all regressions
controls for total sugarcane harvest area in 2006 and the share of mechanical harvesting in 2006.
Coefficients should be interpret as follows.
If a microregion which initial production is fully
based on pre-harvest burning technique fully converts to mechanical harvesting, mechanization
will change the outcome in question by
100𝛽%.
To put in perspective, the mean (median)
microregion had clean adoption index equal to 0.189 (0.167).
We can look at columns 1 and 2, which presents the results for sugarcane industry, to compare
the OLS and 2SLS estimates. We can observe that the OLS and 2SLS coefficients have similar
size, but the point estimate of the instrumented coefficient is smaller in magnitude than the OLS
one. We observe a similar pattern in all four panels, what suggests that OLS biases the estimates
towards zero.
Our first question is if harvest mechanization is a labor saving technical change. Technical
reports claim that one harvest machine may substitute around eighty hand workers, but productivity gains can lead to an increase in the harvested area what would increase the demand
for workers. Looking at Table 8 Panel A, column (12), we find that mechanical harvesting not
significantly reduced the labor intensity of sugarcane production – i.e. there are less workers per
planted area. We also find in column (2) that it reduced, not significantly, the share of formal
workers in the sugarcane sector. The magnitude of both point estimates suggest that labor intensity and the employment share of sugarcane in the average microregion reduced by around
20% relative to baseline due to the adoption of the new technology.
Although we do not have a measure of labor intensity, we can look at employment share
across sectors in column (2). We see that our 2SLS estimates suggest that mechanization led
to a significant reduction in employment share in the agriculture sector (Panel B), which was
more than compensated by a significant increase in the employment share in the “construction
& services” sectors (Panel D) and a non-significant increase in the manufacturing sector (Panel
C).
Although we cannot pinpoint that the adoption of mechanical harvesting in the sugarcane
industry statistically reduced the labor intensity in this sector, we find suggestive evidence that
it actually reduced the employment share in the agriculture sector overall and increase the share
of workers in the other private sector of the economy. One can interpret this as weak evidence
of structural transformation of the local labor market, shifting the composition of sectors away
from the field into more “urban” sectors.
We next, look at the average wage of workers in these sectors, as shown in column 6. For
all sectors the coefficient is positive, what suggests that harvest mechanization increases workers
wage relative to 2006.
Coefficients are significant for the sugarcane sector, at 1% level, and
for agricultural and manufacturing sectors at 5% level.
For sugarcane sector, the adoption of
mechanical harvest increased average formal wage by 7.26% (6.41%) in the average (median)
local labor market. The impact in the manufacturing and agriculture sectors were a half and
11
third, respectively, of the wage increase of workers in the sugarcane sector.
With these results together we can answer our two main questions. First, we find weak evidence that sugarcane mechanization is a labor saving technical change and led to a structural
change in local labor markets, measured by the sector composition of the workforce. The productivity gains from adopting mechanical harvesting in large scale seem to have impacted workers
positively via higher wages in most sectors.
6
Robustness Check
In this section we replicate all our results disaggregating the local labour market at the municipality level as a robustness check. We keep in our sample only municipalities where we observed
any type of sugarcane harvest in the period, leaving us with 408 municipalities.
7
We still cluster
the standard errors at the microregion level to account for correlation within local labor markets.
Table 9 presents the estimates of the first stage regression using municipality-level data. We estimate a KP F statistic equal to 14.93, and the Montiel-Pflueger clustered weak instrument test
reject the null hypothesis at 10% level for all outcomes.
Our results estimated using municipality level data are shown in Table 10. The coefficients
of employment share and wages are similar to the ones found previously, however we do not find
any statistically significant result on our 2SLS estimations. When we make the analysis at the
municipality level, we lose a lot of information and add noise to the estimates because we cannot
capture local work migration, commuting.
For example, larger formal firms tend to be based
on the main local city and all labour market data would be concentrated in these local urban
centres.
7
Conclusion
We study the effects of environmental regulation and technology adoption on structural transformation in local labor markets.
For this purpose, we analyze the adoption of mechanical
harvesting in sugarcane plantations led by a new environmental regulation aimed at extinguishing pre-harversting burning in São Paulo, Brazil. We estimate the impacts of large scale field
mechanization on local labor markets. Using satellite and administrative data paired with an
instrumental variables strategy, we fail to find clear evidence that the adoption of clean harvesting affected formal employment share and formal labor intensity in the sugarcane sector.
We find evidence that the adoption of mechanical harvesting reduced the labor share in other
agricultural sectors, which was compensated by an increase in the employment share in services
and manufacturing sectors. This is consistent with a labor saving technological shock as in Bustos et al. (2013). We also find that the harvest mechanization increased average formal workers
wage, in particular in the sugarcane sector. We interpret our results as
preliminary evidence that
mechanization leads to less formal workers in agricultural sector, releasing manpower to other
sectors such as construction and services and manufacturing.
7
Three municipalities were dropped out because we observed harvesting just some years, what could be a error
because of aggregation level we worked in our spatial data.
12
Bibliography
Adami, M., M. P. Mello, D. A. Aguiar, B. F. T. Rudorff, and A. F. d. Souza (2012).
A web
platform development to perform thematic accuracy assessment of sugarcane mapping in southcentral brazil.
Remote sensing 4 (10), 3201–3214.
Aguiar, D. A., B. F. T. Rudorff, W. F. Silva, M. Adami, and M. P. Mello (2011). Remote sensing
images in support of environmental protocol: Monitoring the sugarcane harvest in são paulo
state, brazil.
Remote Sensing 3 (12), 2682–2703.
Bandiera, O. and I. Rasul (2002, June). Social Networks and Technology Adoption in Northern
Mozambique. STICERD - Development Economics Papers - From 2008 this series has been
superseded by Economic Organisation and Public Policy Discussion Papers 35, Suntory and
Toyota International Centres for Economics and Related Disciplines, LSE.
Barreca, A., K. Clay, and J. Tarr (2014, February). Coal, Smoke, and Death: Bituminous Coal
and American Home Heating. NBER Working Papers 19881, National Bureau of Economic
Research, Inc.
Beaudry, P., M. Doms, and E. Lewis (2006). Endogenous skill bias in technology adoption: citylevel evidence from the IT revolution. Working Paper Series 2006-24, Federal Reserve Bank
of San Francisco.
Berman, E., J. Bound, and S. Machin (1998, November). Implications Of Skill-Biased Technological Change: International Evidence.
The Quarterly Journal of Economics 113 (4), 1245–1279.
Braunbeck, O. and P. Magalhães (2010).
açúcar.
Avaliação tecnológica da mecanização da cana-de-
CORTEZ, LAB Bioetanol de cana-de-açúcar 1, 451–475.
Bustos, P., B. Caprettini, and J. Ponticelli (2013, August). Agricultural Productivity and Structural Transformation. Evidence from Brazil. Working Papers 736, Barcelona Graduate School
of Economics.
Cançado, J. E., P. H. Saldiva, L. A. Pereira, L. B. Lara, P. Artaxo, L. A. Martinelli, M. A.
Arbex, A. Zanobetti, and A. L. Braga (2006). The impact of sugar cane-burning emissions on
the respiratory system of children and the elderly.
Environmental health perspectives , 725–729.
Conley, T. G. and C. R. Udry (2010, March). Learning about a New Technology: Pineapple in
Ghana.
American Economic Review 100 (1), 35–69.
Deschenes, O. (2010, June). Climate Policy and Labor Markets. NBER Working Papers 16111,
National Bureau of Economic Research, Inc.
Dominici, F., M. Greenstone, and C. R. Sunstein (2014). Particulate matter matters.
(New York, NY) 344 (6181), 257.
Science
Esther Duflo, M. K. and J. Robinson (2006, April). Understanding technology adoption: Fertilizer in western kenya evidence from field experiments. Unpublished working paper.
13
Foster, A. D. and M. R. Rosenzweig (2010, January). Microeconomics of Technology Adoption.
Working Papers 984, Economic Growth Center, Yale University.
Greenstone, M. (2002, December).
The Impacts of Environmental Regulations on Industrial
Activity: Evidence from the 1970 and 1977 Clean Air Act Amendments and the Census of
Manufactures.
Journal of Political Economy 110 (6), 1175–1219.
Greenstone, M. and R. Hanna (2014, October).
Pollution, and Infant Mortality in India.
Environmental Regulations, Air and Water
American Economic Review 104 (10), 3038–72.
Kahn, M. E. and E. T. Mansur (2013). Do local energy prices and regulation affect the geographic
concentration of employment?
Journal of Public Economics 101 (C), 105–114.
Kleibergen, F. and R. Paap (2006, July). Generalized reduced rank tests using the singular value
decomposition.
Journal of Econometrics 133 (1), 97–126.
Macedo, I. C., J. E. Seabra, and J. E. Silva (2008). Green house gases emissions in the production
and use of ethanol from sugarcane in brazil: The 2005/2006 averages and a prediction for 2020.
Biomass and bioenergy 32 (7), 582–595.
Matsuyama, K. (1992, December). Agricultural productivity, comparative advantage, and economic growth.
Journal of Economic Theory 58 (2), 317–334.
Relatório
de estudo desenvolvido para a pesquisa Juventude e integração sul-americana: caracterização
de situações-tipo e organizações juvenis, realizada pelo Ibase e pelo Instituo Pólis .
Novaes, J. R. P. et al. (2007). Jovens migrantes canavieiros: entre a enxada e o facão.
Nyko, D., M. S. Valente, A. Y. Milanez, A. K. R. Tanaka, and A. V. P. Rodrigues (2013). A
evolução das tecnologias agrícolas do setor sucroenergético:
estrutural?
BNDES Setorial, n. 37, mar. 2013, p 399-442 .
estagnação passageira ou crise
Olea, J. L. M. and C. Pflueger (2013, July). A Robust Test for Weak Instruments.
Business & Economic Statistics 31 (3), 358–369.
Pflueger, C. and S. Wang (2014). A robust test for weak instruments in stata.
(forthcoming).
Journal of
Stata Journal
Rangel, M. A. and T. Vogl (2015). The dirty side of clean fuel: Sustainable development, preharvest sugarcane burning and infant health in brazil. mimeo.
Rudorff, B. F. T., D. A. Aguiar, W. F. Silva, L. M. Sugawara, M. Adami, and M. A. Moreira
(2010). Studies on the rapid expansion of sugarcane for ethanol production in são paulo state
(brazil) using landsat data.
SGPR (2009).
Remote sensing 2 (4), 1057–1076.
Compromisso nacional para aperfeiçoar as condições de trabalho na cana de
açucar. Technical report, Secretaria Geral da Presidência da Republica.
Walter, A., M. V. Galdos, F. V. Scarpare, M. R. L. V. Leal, J. E. A. Seabra, M. P. da Cunha,
M. C. A. Picoli, and C. Ortol (2014, 01). Brazilian sugarcane ethanol: developments so far
14
and challenges for the future.
70–92.
Wiley Interdisciplinary Reviews: Energy and Environment 3 (1),
15
(a) Harvest in 2006
(b) Harvest in 2012
Figure 1: Sugarcane Harvest
These figures presents sugarcane harvest. Figure 1a presents sugarcane harvest in 2006 and figure 1b presents
sugarcane harvest in 2012. In both figures green means clean harvesting and red means pre-harvest burning. We
see São Paulo state divided by microregions.
Figure 2: Clean Harvesting vs Slope first moment
The figure presents clean harvesting in year 𝑡 divided by clean harvesting in 2006 per value of slope first moment.
To do this we round slope first moment to the first decimal place, so we have more observations at one slope first
moment point. We drop slope first moment values greater than 15% because of very few observations.
16
Figure 3: Clean Adoption Index vs Slope first moment
This figure presents clean adoption index in year 𝑡 per value of slope first moment, locally weighted scatterplot
smoothing (lowess). To do this we round slope first moment to the first decimal place and then calculate clean
adoption index at each rounded value and each year 𝑡. We drop slope first moment values greater than 15%
because of very few observations.
Figure 4: Planted Area Expansion vs Slope first moment
This figure presents presents planted area in year 𝑡 divided by clean harvesting in 2003 per value of slope first
moment. To do this we round slope first moment to the first decimal place, so we have more observations at one
slope first moment value. We drop slope first moment values greater than 15% because of very few observations.
17
Table 1: Sugarcane Production (thousands
ton)
Year
Brazil
São Paulo
%
2006
427,723
264,339
62
2007
495,723
296,243
60
2008
569,216
346,293
61
2009
602,193
361,261
60
2010
620,409
359,503
58
2011
559,215
304,230
54
2012
588,478
329,923
56
This table presents sugarcane production in thousands of ton from Brazil and São Paulo state.
Source: UNICADATA
Table 2: Timeline
Year
Mechanizable Areas
Non-mechanizable Areas
2002
20% reduction
2006
30% reduction
2011
50% reduction
10% reduction
2016
80% reduction
20% reduction
2021
100% reduction
30% reduction
2026
50% reduction
2031
100% reduction
This table resumes the reduction timeline to end sugarcane pre-harvest
burning.
Table 3: Evolution of Mechanical Harvesting (%)
Center-South
Mechanical Harvesting
2006
2007
2008
2009
2010
2011
2012
36,7
42,8
53,4
60,1
72,8
79,2
85,1
Green
25,1
29,9
38,2
42,9
52,5
66,3
73,8
Burning
11,6
12,9
15,2
17,1
20,3
13,0
10,8
Hand Harvesting
63,3
57,2
46,6
39,9
27,2
20,8
14,9
This table presents mechanical harvesting evolution in percentages for the Brazilian Center-South region.
Source: Nyko et al. (2013)
18
Table 4: Employees in the Sugarcane Industry (2007)
Region
North/ Northeast
Sugarcane crop
Sugar Production
Ethanol Production
Total
83,843
16.85%
252,250
44.09%
40,348
21.14%
Center-South
413,827
83.15%
319,897
55.91%
150,546
78.86%
376,441
884,270
São Paulo
268,282
53.91%
199,512
34.87%
51,824
27.15%
519,618
BRAZIL
497,670
100.0%
512,147
100.0%
190,894
100.0%
1,260,711
This table presents formal workers in sugarcane industry in Brazil at 2007, differentiating between workers of sugarcane crop,
sugar production, and ethanol production.
Source: SGPR (2009)
Table 5: Descriptive Statistics: Explanatory Variables
0−4
Mean
Median
0.082
0.046
(0.088)
4−8
0.607
0.620
(0.189)
8 − 12
0.258
0.229
(0.169)
12 − 16
0.035
0.014
(0.046)
16 − 20
0.010
0.002
(0.024)
> 20
0.007
0.007
(0.020)
Harvest Area 2006
951.932
Clean 2006
314.136
685
(896.834)
210
(314.615)
Clean Adoption Index
0.360
0.371
(0.164)
Note: The table reports descriptive statistics of the independent variables used in regressions. Clean Adoption Index is calculated for 2012. Standard deviation reported in brackets.
Table 6: Descriptive Statistics: Dependent Variables
Employment
Total
Average
Total
Permanent
Labor
Share
Wage
Worker Wage
Employment
Employment
Intensity
Level
Dif
Level
Dif
Level
Dif
Level
Dif
Level
Dif
Level
Dif
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
Mean
0.047
-0.033
5615.730
1092.078
1.389
0.665
3589.318
-2008.795
3588.273
-2003.545
2.813
-3.131
Std. Deviation
(0.049)
(0.046)
(7136.924)
(3750.399)
(0.275)
(0.223)
(4086.091)
(2960.208)
(4085.989)
(2957.870)
(2.455)
(3.038)
Median
0.029
-0.013
2836.434
221.550
1.355
0.640
2134.000
-1235.500
2134.000
-1235.500
3.022
-2.651
Mean
0.169
-0.033
10471.930
5039.132
1.129
0.531
9207.136
117.114
9206.864
137.159
Std. Deviation
(0.080)
(0.045)
(9957.010)
(5659.632)
(0.136)
(0.083)
(8575.012)
(2411.181)
(8574.821)
(2402.434)
Median
0.153
-0.024
7880.709
3755.775
1.099
0.515
7386.500
204.000
7386.500
213.000
Panel A: Sugarcane
Panel B: Agriculture
Panel C: Manufacturing
Mean
0.247
0.021
74225.620
37566.880
1.591
0.603
40485.840
9389.295
40469.250
9381.864
Std. Deviation
(0.084)
(0.045)
(118945.500)
(58390.760)
(0.370)
(0.136)
(49620.050)
(12356.260)
(49613.720)
(12358.740)
Median
0.252
0.014
32263.930
16621.300
1.576
0.599
21761.500
3370.500
21759.500
3372.500
Panel D: Construction & Services
Mean
0.537
0.044
155859.300
89267.910
1.421
0.533
100946.00
34169.480
98689.410
34431.180
Std. Deviation
(0.077)
(0.043)
(290509.300)
(167053.600)
(0.174)
(0.121)
(161846.100)
(58665.600)
(155331.600)
(59167.680)
Median
0.518
0.050
69816.86
40681.26
1.398
0.509
50725.000
16070.500
50504.500
16234.500
This table reports descriptive statistics of dependent variables used in regressions. Variables in level are considered observations at microregion level in 2012. Variables in difference are
calculated at microregion level taking the difference between 2012 and 2006.
19
20
Table 7: First Stage Regression
CleanAdoptionIndex
0−4
0.249
8 − 12
-0.691***
12 − 16
1.539**
(0.222)
(0.183)
(0.709)
16 − 20
0.510
(2.889)
> 20
-2.809
(2.736)
Kleibergen-Paap Wald rk F statistic
14.20
This table reports the estimates of the first-stage regression from equation 4. The
dependent variable is the difference between Clean Adoption Index in 2012 and
in 2006. The unit of observation is a microregion, we have 44 observations. Robust standard errors are reported in parentheses. Significance level: ***p<0.01,
**p<0.05, *p<0.10.
Table 8: Main Results from OLS and Instrumental Variables Regressions
Panel A: Sugarcane
Clean Adoption Index
Change in Em-
Change in Total
Change in Average
Change in Total
Change in Perma-
ployment Share
Wage
Worker Wage
Employment
nent Employment
Clean Adoption Index
2SLS
OLS
2SLS
OLS
2SLS
OLS
2SLS
OLS
2SLS
OLS
2SLS
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
-0.025
-0.089
1.071
0.786
0.333***
0.384***
0.770
0.097
0.770
(0.050)
(0.069)
0.080
(0.120)
(0.129)
-0.100*
(0.050)
Average Dependent Variable 2006
Panel C: Manufacturing
Clean Adoption Index
(0.071)
0.201
0.037
0.120
(0.099)
0.226
Panel D: Construction & Services
0.088
(0.056)
Average Dependent Variable 2006
-0.155**
(0.064)
Average Dependent Variable 2006
Clean Adoption Index
Intensity
OLS
Average Dependent Variable 2006
Panel B: Agriculture
Change in Labor
0.124**
(0.051)
0.493
(1.798)
(1.983)
4523.652
-0.223
0.723
0.038
0.058
0.112**
(0.300)
(0.414)
5432.793
(0.045)
(0.048)
0.301
0.598
0.097
-2.787
-7.529
(1.754)
(1.925)
5598.114
(1.754)
(1.925)
5591.818
(3.782)
(5.116)
5.944
-0.371
-0.375
-0.209
-0.205
(0.311)
(0.459)
9090.023
(0.311)
(0.458)
9069.705
-0.005
-0.002
0.477
0.111*
0.183**
(0.311)
(0.378)
36658.740
(0.060)
(0.082)
(0.260)
(0.302)
31096.550
(0.260)
(0.303)
31087.390
-0.072
0.081
-0.045
0.041
-0.009
-0.020
(0.084)
(0.133)
66591.360
(0.058)
(0.042)
(0.106)
(0.145)
66776.500
0.988
0.889
-0.139
-0.040
-0.137
-0.078
(0.104)
(0.143)
64258.230
This table reports the estimates of the impact of 𝐶𝑙𝑒𝑎𝑛𝐴𝑑𝑜𝑝𝑡𝑖𝑜𝑛𝐼𝑛𝑑𝑒𝑥𝑗𝑡 on labor market outcomes, 𝛽 , from equation 3. Columns indicated with 2SLS were estimated using land slope as instrument
for 𝐶𝑙𝑒𝑎𝑛𝐴𝑑𝑜𝑝𝑡𝑖𝑜𝑛𝐼𝑛𝑑𝑒𝑥𝑗𝑡 as in equation 4. The dependent variables are the difference between the variable indicated in the columns in 2012 and in 2006. Employment Share and Labor Intensity are
fraction, and all other variables are in logs. All regressions include controls for the total harvest area in 2006 and the share of clean harvesting in 2006. The unit of observation is a microregion, we have
44 observations. Robust standard errors are reported in parentheses. Significance level: ***p<0.01, **p<0.05, *p<0.10.
21
22
Table 9: First Stage using municipality level data (Robustness)
CleanAdoptionIndex
0−4
0.115
8 − 12
-0.500***
12 − 16
0.765***
(0.129)
(0.075)
(0.260)
16 − 20
-0.406
(0.543)
> 20
-1.056***
(0.406)
Kleibergen-Paap Wald rk F statistic
14.93
This table reports the estimates of the first-stage regression from equation 4. The
dependent variable is the difference between Clean Adoption Index in 2012 and in
2006. The unit of observation is a municipality, we have 408 observations. Standard
errors clustered at the microregion level are reported in parentheses, 44 clusters.
Significance level: ***p<0.01, **p<0.05, *p<0.10.
Table 10: Estimates using municipality level data (Robustness)
Change in Em-
Change in Total
Change in Average
Change in Total
Change in Perma-
ployment Share
Wage
Worker Wage
Employment
nent Employment
Change in Labor
Intensity
OLS
2SLS
OLS
2SLS
OLS
2SLS
OLS
2SLS
OLS
2SLS
OLS
2SLS
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
Clean Adoption Index
0.016
-0.051
-0.111
0.012
0.179
-0.026
-0.203
-0.028
-0.198
2.256
0.870
(0.420)
(0.802)
603.039
(3.054)
Average Dependent Variable 2006
(0.031) (0.079)
0.093
(8.449)
4.986
Panel A: Sugarcane
Panel B: Agriculture
Clean Adoption Index
Average Dependent Variable 2006
Panel C: Manufacturing
-0.043*
-0.053
(0.021) (0.049)
0.189
Clean Adoption Index
0.020
Average Dependent Variable 2006
(0.024) (0.068)
0.229
Panel D: Construction & Services
0.095
Clean Adoption Index
0.006
Average Dependent Variable 2006
(0.039) (0.064)
0.489
0.010
(0.425)
((0.798)
487.845
-0.231**
0.006
((0.074)
(0.124)
0.594
0.071
0.008
0.049
(0.112)
(0.302)
585.890
(0.013)
(0.037)
0.607
0.771
0.014
0.024
(0.286)
(0.484)
3953.393
0.198
(0.048)
(0.088)
0.817
0.013
0.123
-0.010
0.060
(0.072)
(0.157)
7181.421
(0.013)
(0.037)
0.848
(0.420)
(0.803)
603.718
-0.250**
0.017
(0.111)
(0.329)
980.297
0.178
0.694
(0.292)
(0.465)
3353.549
0.030
0.035
(0.066)
(0.156)
7201.387
-0.252**
0.019
(0.110)
(0.329)
978.105
0.179
0.695
(0.292)
(0.465)
3352.561
0.046
0.032
(0.065)
(0.149)
6929.809
This table reports the estimates of the impact of 𝐶𝑙𝑒𝑎𝑛𝐴𝑑𝑜𝑝𝑡𝑖𝑜𝑛𝐼𝑛𝑑𝑒𝑥𝑗𝑡 on labor market outcomes, 𝛽 , from equation 3. Columns indicated with 2SLS were estimated using land slope as instrument
for 𝐶𝑙𝑒𝑎𝑛𝐴𝑑𝑜𝑝𝑡𝑖𝑜𝑛𝐼𝑛𝑑𝑒𝑥𝑗𝑡 as in equation 4. The dependent variables are the difference between the variable indicated in the columns in 2012 and in 2006. Employment Share and Labor Intensity are
fraction, and all other variables are in logs. All regressions include controls for the total harvest area in 2006 and the share of clean harvesting in 2006. The unit of observation is a municipality, we have
408 observations. Standard errors clustered at the microregion level are reported in parentheses, 44 clusters. Significance level: ***p<0.01, **p<0.05, *p<0.10.
23
24
Appendix A
In this appendix we estimate regression 3 for each year between 2007 and 2012 for sugarcane
sector only. We can see that the effects of mechanized harvest adoption for the latest two years
of the sample, 2011 and 2012, are quite the same, and may be different when compared to the
early years of technology adoption.
Table 11: Microregion Level: Sugarcane Sector Second Stage by Year
2012
2011
2010
2009
2008
2007
Change in Em-
Change in Total
Change in Average
Change in Total
Change in Perma-
Change in Labor
ployment Share
Wage
Worker Wage
Employment
nent Employment
Intensity
(1)
(2)
(3)
(4)
(5)
(6)
-0.089
0.786
0.384***
0.097
0.097
-7.529
(0.069)
(1.983)
(0.129)
(1.925)
(1.925)
(5.116)
-0.029
1.756
0.342**
1.137
1.138
-6.124
(0.044)
(1.588)
(0.135)
(1.513)
(1.513)
(4.340)
-0.108
0.191
0.231
-0.506
-0.507
-8.515
(0.070)
(1.511)
(0.218)
(1.550)
(1.552)
(6.812)
-0.126
0.231
0.393
-1.293
-1.303
-10.907
(0.123)
(1.994)
(0.304)
(2.430)
(2.433)
(11.169)
-0.096
-0.444
-0.080
-0.729
-0.736
-3.336
(0.078)
(0.989)
(0.171)
(1.015)
(1.015)
(5.582)
-0.074
-1.660*
-0.165
-1.864**
-1.869**
-1.089
(0.049)
(0.909)
(0.177)
(0.945)
(0.945)
(3.985)
Note: The table reports 2SLS results of equation 3 in the text, for sugarcane sector only and for each year from 2007 to 2012. The unit of observation is microregion. Robust
standard errors are reported in parentheses. We have 44 microregions each year. All outcomes are in difference with 2006 value. Employment Share and Labor Intensity are
fraction, and all other variables are in log. All regressions are controlled for total harvest in 2006 and clean harvesting in 2006. Significance level: ***p<0.01, **p<0.05, *p<0.10.
25