Causes of and Strategies out of poverty in Brazil. Econometric analysis of the 2012 Brazilian National Sample Survey (PNAD) and continuation of David Lam and Deborah Levison's analysis from their 1992 paper: Age, Experience, and Schooling: Decomposing Earnings Inequality in the United States and Brazil, and Lorene Yap's 1976 paper: Internal Migration and Economic Development in Brazil.
Q3 2024 Earnings Conference Call and Webcast Slides
Brazil - Strategies out of Poverty
1. STRATEGIES OUT OF POVERTY
Microeconomic Data Management
DECEMBER 14, 2015
COMPILED BY: BRIAN LYNCH AND MICHAEL THOMOPOULOS
2. -1-
Table of Contents
Introduction… (2)
Background Information … (3-10)
Changes in Distribution of Schooling… (11-32)
Migration… (33-41)
Child Labor… (42-59)
Resource Pooling… (60-65)
Conclusion… (66-67)
Appendix… (69-85)
Works Cited… (86)
3. -2-
1.1: Introduction
The purpose of this study is to understand strategies adopted by Brazilians in order to
escape poverty. To fully understand these strategies, we begin by looking at the historical
development of Brazil’s political-economy and how it has promoted the current economic state.
We then look at the changing education system in Brazil and its impact on income inequality.
From here, we look at strategies Brazilians use to escape poverty.
We break down these strategies into two decision points. The first being short term
poverty alleviation, which is done by parents entering their children in to the labor force in order
to see current income gains. We also see families moving in together in order to minimize living
costs. We classify these as short term income alleviation strategies, however they come at a long
term cost of lower income for future generations. The second decision would be long term
poverty alleviation through parents enrolling their children in school. This tradeoff is the
decision parents have to make between short term poverty alleviation and future alleviation for
their children.
First we look at migration and the returns migrants receive from making the decision to
migrate. We then look at the role children play in poverty alleviation and the decisions families
make in order to escape poverty.
1.2: Data
The reason for the choice of Brazil as the focus of this study is due to its large population
of 200 million people. Brazil also demonstrates a high level of income inequality with a GINI
coefficient of 52.67 (World Bank). There is also an ample amount of data available for use
through the Pesquisa Nacional por Amostra de Domicílios or PNAD. The PNAD is a survey,
which ascertains the general household characteristics of the population in respect to educational
attainment, labor, income, and housing. The survey also procures data concerning migration,
fertility, marriage, and health and food safety to help influence government policy. The dataset
includes 362,451 observations, which allowed us to parse the data while maintaining large,
significant sample sizes. We cleaned the data, removing survey respondents with miscoded data
points or entries with missing responses.
5. -4-
2.1: Urbanization and Population Growth in Brazil, 1960-2012
Brazil has experienced significant population growth and urbanization in the last 50
years. Table 2.1 shows a decrease in rural population (as a %age of total population) from 54%
to 15% and a corresponding increase in urban population from 46% to 85% in 1960 and 2012
respectively. Brazil also saw rapid population growth from 72.8 million in 1960 to 198.7 million
in 2012. In the graph below, Brazil has seen a steady decline in population growth rate since
1960. There is a steady decrease from 3% in 1960 to just under 1% in 2012.
Table 2.1
Rural/Urban Population Comparison 1960, 2012
^^ 1960 2012
Rural Population (% of total) 53.86 15.1
Urban^
Population (% of total) 46.14 84.9
Urban >1M (% of total) 21.06 39.66
Total Population 72.7759M 198.656M
^Urban defined by Brazilian government and reported to World Bank (exact definition unknown). YAP(1975) reported 30% in areas >20,000 in
1960
^^All statistics compiled from the World Bank (http://data.worldbank.org/indicator/SP.URB.TOTL.IN.ZS/countries,
http://data.worldbank.org/indicator/EN.URB.MCTY.TL.ZS/countries)
Data from World Bank (http://data.worldbank.org/indicator/SP.POP.GROW/countries/1W?display=default), graph generated by Google on
10/5/15
6. -5-
Brazil’s economy has grown rapidly since 1960, especially in the 1960s and 1970s. The
average annual growth rate was 6.2% in the 1960s and 8.5% in the 1970s (Devlin, 1995).
However, Brazil’s GDP growth has consistently outpaced Brazil’s GDP per capita growth
illustrating Brazil’s struggle to keep up with its growing population.
The 4.73% decline in GDP per capita for the 1980s can be explained by significant
contractions in the Brazilian economy in the years 1981-3 as well as contractions in 1988, and
1991-3. This period is known as the “Lost Decade” due to the Latin American debt crisis during
this time. The majority of Latin American governments (most notably Mexico, Argentina, and
Brazil) acquired large loans from foreign banks in order to finance their unprecedented growth.
The economic recessions faced by several major economies in the 1970s and 1980s coupled with
rising oil prices slowed economic growth in Brazil. In addition, the US and Europe both
increased interest rates making debt repayment for Brazil increasingly difficult. Eventually, some
banks called for immediate loan repayments forcing Brazil (as well as other governments) to
enact widespread austerity measures further deepening the recession. (Devlin, 1995)
Table 2.2
Brazil’s GDP and GDP per capita Growth since 1960^
Year Range %Change in GDP %Change in GDP/capita
1961-70 61.92 32.87
1971-80 85.09 59.69
1981-90 20.89 -4.73
1991-00 26.26 10.63
2001-2012 42.42 29.09
Total (1960-2012) 236.58 127.55
^Data compiled from the World Bank (http://data.worldbank.org/indicator/NY.GDP.MKTP.KD.ZG/countries/1W?display=default,
http://data.worldbank.org/indicator/NY.GDP.PCAP.KD.ZG?display=default)
Brazil has the 8th largest GNP, but one of the highest levels of income inequality in the
world. However, from 2003-2013 inequality was greatly reduced. Since 2013, 26 million people
have been lifted out of poverty, causing the Gini coefficient to fall by 6% (The World Bank,
2004). Income of the bottom 40% grew nearly twice as much as the income of the total
population. However gender inequality is still prevalent, where average female income is 29%
less than average male income. When controlling for age, education, and hours worked, the
difference rises to 34%. (The World Bank, 2004)
To further understand the regional, gender, and racial inequalities discussed in this paper,
it is important to understand the history of slavery in Brazil and the significant role it has played
in Brazil’s development.
2.2: Slavery in Brazil
Slaves were first brought to Brazil in the 1500s, and Brazil quickly became the largest
importer of African slaves in the world (Setti, 2015). When slavery was finally abolished in
7. -6-
1888, almost 5 million slaves had been imported from Africa (Setti, 2015). Slaves provided an
essential labor force for the growing sugar industry of colonial Brazil, Brazil’s chief export at the
time. The Northeast was home to some of the largest sugar plantations in the world, and
subsequently a large share of Brazilian slaves were sent there. In the years leading up to 1888,
many slaves escaped the sugar plantations and formed their own colonies in Northeast Brazil.
These colonies grew substantially after 1888 as freed slaves migrated to these already established
colonies. To this day the Northeast is still largely dominated by those of African descent.
Unfortunately, the economies of these colonies never grew substantially as the Northeast still
lags behind the rest of the country in economic growth.
2.3: Migration/Urbanization Policies
Migration and urbanization has long played a significant role in the development of
Brazil and has been a main topic of political debate and reform in the last 50 years. This started
in 1964 when a military regime took control of the country and attempted to stimulate rural
economies in order to decrease migration to larger cities (Martine and McGranahan, 2010).
These attempts failed and after the fall of the military regime in 1986, a different approach was
needed.
In 1986 a democratic regime took power and voters signaled a desire for better urban
planning. Overall, the most successful urbanization policy was the Statue of the City (Estatuto da
Cidade) passed in 1988, but not implemented until 2001. The Statue of the City regulates the
master plans of all municipalities with more than 20,000 urban inhabitants. In addition, the
Statue now allows squatting under certain conditions and the local government can now tax
uninhabited property or property that is not being used in a socially beneficial manner (Martine
and McGranaham, 2010). These policies have created additional opportunity for migration to
urban areas while “organizing” it in a socially efficient way. This signaled a change in Brazilian
politics as urban policy became a significant tool in fighting social inequality. The Statue of the
City has generated much debate over social inequality and pushed the political process of
urbanization in that direction. It is important to understand the political pressures of migration
and urbanization in Brazil when assessing the analysis of later sections.
2.4: New Brazilian States
Migration and Urbanization policies have played a large part in the development of
Brazil. Brazil has experienced significant growth, particularly in the Northern region. The
population and economy of Brazil has been drifting from the crowded coast inwards since 1900.
This shift was fueled in part by the relocation of the capital from Rio de Janeiro to Brasília in
1960 (Wade, 2011). In addition, six new states were created since 1960 (Acre, Rondônia, Mato
Grosso do Sul, Amapá, Roraima, and Tocantins) and one has been dissolved (Guanabara). Five
out of the six new states are located in the North, with Mato Grosso do Sul in the center.
The territory of Acre became important due to its rubber trees in the early 1900s but its
economic importance was short-lived. During the Second World War the rubber industry was
revived, attracting migrants from the crowded East coast and poor Northeast regions. In 1962
Acre won its statehood after complaining of official neglect (Minahan, 2007). After Rondônia
became a state in 1981, the Brazilian government launched “Polonordeste,” an initiative to
distribute land in Rondônia to poor settlers (Wade, 2011). As a result, the population of
8. -7-
Rondônia exploded from 116,620 in 1970 to 1,130,874 in 1991 (Wade, 2011). The last
restructuring act of Brazil was the 1988 Constitution which granted the Northern territories of
Amapá and Roraima statehood and split the state of Tocantins from Goiás.
Table 2.3
New Brazilian States
Year New States
1962 Acre gains statehood
1975 Guanabara merges with Rio de Janeiro
1977 Mato Grosso do Sul splits from Mato Grosso
1981 Rondônia gains statehood
1988 Amapá and Roraima gain statehood and Tocantins splits from Goiás
Figure 2.1
Map of the New Brazilian States
http://www.brazil-help.com/brazilian_states.htm
The reclassification of territories and changes to state lines can be attributed in part to
population growth. While the population of Brazil has nearly tripled in the last 60 years, the
populations of all new states have grown much faster. As shown in Table 5, populations of the
youngest states have at least quadrupled in size. Most notably, Roraima and Rondônia have
experienced a staggering 1528% and 2207% growth, respectively.
9. -8-
Table 2.4
Population by Major region from 1960-2010 from the Demographic Census^
Federative Units
and Household
Situation
Population from the Demographic Census
9/1/1960
(1)
9/1/1970
(1)
9/1/1980
(1)
9/1/1991
(2)
9/1/2000
(2)
8/1/2010
(2)
Brazil 70,992,343 94,508,583 121,150,573 146,917,459 169,590,693 190,755,799
Rondônia 70,783 116,620 503,125 1,130,874 1,377,792 1,562,409
Acre 160,208 218,006 306,893 417,165 557,226 733,559
Roraima 29,489 41,638 82,018 215,950 324,152 450,478
Amapá 68,889 116,480 180,078 288,690 475,843 669,526
Tocantins 328,486 537,563 738,688 920,116 1,155,913 1,138,445
Rio de Janeiro 6,709,891 9,110,324 11,489,797 12,783,761 14,367,083 15,989,929
Mato Grosso do
Sul
579,652 1,010,731 1,401,151 1,778,741 2,074,877 2,449,024
Mato Grosso 330,610 612,887 1,169,812 2,022,524 2,502,260 3,035,122
^Data from the Instituto Brasileiro de Geografia e Estatística 2010 census
Table 2.5
Population Growth 1990-2010^
State % Population Growth
Brazil 269%
Rondônia 2,207%
Acre 458%
Roraima 1,528%
Amapá 972%
Tocantins 421%
Mato Grosso do Sul 422%
^Data calculated from the Instituto Brasileiro de Geografia e Estatística 2010 census
2.5: Education Policies
In addition to infrastructure developments, Brazil’s education systems and policies have
improved drastically since the 1960s. Even during the military regime two major educational
improvements took place. The first was the MOBRAL, Movimento Brasileiro de Alfabetizacao.
10. -9-
This was a large movement to eradicate adult literacy from the country. It was mainly
experimental, however its success led to similar future illiteracy-combatting programs. The
second major event was Law 5,692 passed in 1971. This law completely changed the structure of
the entire education system. It led to a sharp increase in average education. Progression to
secondary school increased 40% to 85% in two years. (Schwartzman, 2003)
Data from the World Bank: (http://data.worldbank.org/indicator/SE.SEC.ENRR/countries/1W?display=graph)
In the 1990s after the military regime, Brazil created the National Program of Literacy
and Citizenship to further combat illiteracy. It reduced illiteracy in Brazil by 70%. A new model
of elementary school was also put into place in the mid-90s called CIACs, centro integrado de
Educacao Popular (Integrated Center for General Education). CIACs supported children from
poor families with education and food. In 1997, mandatory exams were implemented to assess
primary education quality, called Povaos (Schwartzman, 2003). The program was then extended
to high schools and called ENEN (Exame Nacional do Ensino Medio). Finally the program was
expanded to elementary schools and called SAEB (Sistema de Avaliacao da Educacao Basica).
Since the successful policies in the 90s, Brazil has given educational policy much greater
attention.
11. -10-
A three-pronged approach has been adopted:
1. Education of finance equalization
2. Conditional cash transfers
3. Education results measurement
Taking what we know about the current and historical political and economic situation ln
Brazil, we now look at the income effect of educational expansion and attainment. We find that
the returns to education have changed over time with respect to race, gender, and geography. We
also find that there is a gap in educational quality based on geographic location'
12. -11-
Section 3
Changes in Distribution of Schooling and its Impact on
Income Inequality and Disparity in Brazil
Joel Ferguson
Porter Reim
Frederic Chasin
13. -12-
3.1: Introduction
It is well established that education is one of the most important determinants of income
and that the distribution of education has some impact on income inequality. The question of
how much of an impact the distribution of schooling has on income distribution is an important
question that has been studied before. Lam and Levinson (1992) found that changes in the
distribution of education were working to equalize incomes in Brazil despite the increase in
income inequality that had occurred over the time period studied. This conclusion was reached
by analyzing the changes in mean education, variance of schooling, and returns to education
among cohorts.
We expand upon the work done by Lam and Levinson by looking at how the schooling
inequality between different groups of people has affected income inequality, attempting to
quantify the impact of education on income inequality, and predicting future trends of income
inequality in Brazil. In section 3.2 we replicate Lam and Levinson’s study using more recent
survey data to see if it is still true that changes in the distribution of education have continued to
be an equalizing force for income. We find that current trends are in line with those found in
1992. In fact, we find that there has been significantly more convergence in the distribution of
schooling and steep declines in returns to education for working age males since the original
study was published.
Lam and Levinson’s original only looked at the distribution of education and income for
working-aged men. In section 3.3 we extend the methodology used by Lam and Levinson to get
a sense of how the distribution of schooling between groups of people has changed over time. In
particular, we look at how differences in educational attainment and returns to education for
Afro-Brazilians and rural Brazilians as well as labor force participation of women have changed
over time and affected income inequality. We find a general fall in disparity between these
groups.
In section 3.4 we try to understand how the distribution of schooling affects income
inequality by simulating different distributions of education while keeping all other
characteristics the same. Despite significantly lower returns to education than in the past,
inequality in schooling still has a very large effect on the distribution of income.
In section 3.5 we forecast how education inequality will continue to change in Brazil and
provide some concluding remarks.
3.2: Lam and Levinson replication
Lam and Levinson (1992) split income-earning men born between 1925 and 1963 into
three-year birth cohorts in order to approximate secular changes in mean education and variance
of education as well as returns to schooling. By using the same survey we can easily compare our
results with those found by Lam and Levinson. Using the same age group as Lam and Levinson,
22-60, there are four cohorts that perfectly overlap with the 1985 data. This allows us to directly
compare those four cohorts between 1985 and 2012. For the purpose of this section, we use the
exact same subsample of the survey, 22-60 year old men with positive earnings.
By splitting up the population into three-year cohorts it is possible to estimate change
over time with a single cross-section of data. We utilize the power of this novel methodology
later on in our analysis of changes in income between different groups. Thanks to the large size
14. -13-
of the PNAD data sets, the cohort samples still have adequately large sizes with the smallest
having more than 4000 observations.
Educational attainment and distribution of schooling
Table 3.1
Educational attainment for three-year birth cohorts, Brazilian males, 2012 and 1985
PNAD
Age
Group
(1)
Birth
cohort,
1985
(2)
Birth
cohort,
2012
(3)
Sample
size,
1985
(4)
Sample
size,
2012
(5)
Mean,
1985
(6)
Mean,
2012
(7)
Variance,
1985
(8)
Variance,
2012
(9)
Coeff. of
variation,
1985
(10)
Coeff of
variation,
2012
(11)
22-24
1961-
1963
1988-
1990
13,937 8,730 5.98 9.52 16.08 11.36 0.67 0.32
25-27
1958-
1960
1985-
1987
13,024 8,482 5.93 9.58 17.80 13.25 0.71 0.34
28-30
1955-
1957
1982-
1984
11,734 8,662 5.89 9.26 19.33 15.52 0.75 0.38
31-33
1952-
1954
1979-
1981
10,622 8,378 5.77 8.80 20.66 17.47 0.79 0.42
34-36
1949-
1951
1976-
1978
9,643 7,674 5.24 8.31 21.00 19.10 0.87 0.47
37-39
1946-
1948
1973-
1975
8,386 7,283 4.95 7.87 20.84 19.27 0.92 0.49
40-42
1943-
1945
1970-
1972
7,634 7,160 4.43 7.61 19.12 19.54 0.99 0.51
43-45
1940-
1942
1967-
1969
7,123 6,590 4.08 7.93 17.79 20.98 1.03 0.51
46-48
1937-
1939
1964-
1966
6,109 6,511 3.92 7.38 16.75 22.37 1.04 0.56
49-51
1934-
1936
1961-
1963
5,588 6,096 3.78 7.03 16.52 21.72 1.08 0.57
52-54
1931-
1933
1958-
1960
4,942 5,526 3.58 6.83 15.70 21.44 1.11 0.69
55-57
1928-
1930
1955-
1957
4,590 4,820 3.32 6.46 14.84 21.90 1.16 0.62
58-60
1925-
1927
1952-
1954
4,099 4,246 3.05 6.07 14.03 22.94 1.23 0.67
22-60
1925-
1961
1952-
1990
107,431 90,158 4.98 8.28 19.11 19.18 0.88 0.53
Table 3.1 shows descriptive statistics on the educational attainment of males by age
group for both the PNAD data sets. Columns (6) and (7) show mean educational attainment in
1985 and 2012 respectively. Columns (8) and (9) show the variance in schooling which can be
considered a measure of schooling inequality. Columns (10) and (11) present the coefficient of
variation, a measure of inequality independent of mean.
In line with what Lam and Levinson found, mean educational attainment has continued to
rise since 1985 with the average 22-24 year old male getting 3.5 more years of education in 2012
than in 1985. From the1925-1927 cohort to the 1988-1990 cohort mean education attainment has
more than tripled. Notably, the increase in mean educational attainment has been relatively
steady over the entire observed period as seen in fig. 3.1. One notable exception to this steady
increase can be seen in the 1970-1972 cohort. This may have to do with the fact that they grew
15. -14-
up during Brazil’s “lost decade,” when economic growth stagnated and pessimism was high,
perhaps leading parents to pull their children out of school.
Variance in level of schooling and the coefficient of variation have also, for the most
part, been decreasing. The continued fall in the coefficient of variation is particularly important,
showing that schooling inequality has continued decreasing between 1985 and 2012. This and
the rise in mean educational attainment suggest that there has been significant convergence in
educational attainment since 1985.
Interestingly, the four cohorts that overlap between the 1985 and 2012 data sets (1952-54,
1955-57, 1958-60, and 1961-63) have all experienced increases in schooling variance. However,
as shown in columns (10) and (11) these four cohorts have also all experienced decreases in the
coefficient of variation of schooling, suggesting that the distribution of education among them
has become more equal since 1985.
Figure 3.1
Mean educational attainment by birth cohort, Brazilian Males: PNAD 1985 and PNAD
2012
Another comparison of note is that the mean level of schooling has increased for all of
the overlapping cohorts. The inverse relationship between education and mortality suggests that
the changes we find by using cross-sectioned data to estimate time trends have a downward bias.
However, as can be seen in fig. 3.1, the greatest increase in educational attainment came to the
cohort that was the youngest in 1985, rather than the oldest. Therefore, it is safe to say that this
increase is due at least in part to continued educational attainment between 1985 and 2012. This
16. -15-
makes intuitive sense as Lam and Levinson noted that the youngest cohorts might not have
completely finished schooling (202).
Overall, these trends show a convergence in educational attainment. With a rising mean
schooling level and coefficient of variation, it appears that people with lower levels of education
have been catching up since 1985. To get a better sense of this phenomenon, we look at
frequency of completion by year of schooling for three different cohorts, the oldest, middle, and
youngest, shown in Fig. 3.2. It is easy to see that educational attainment has been converging to
eleven years, or the equivalent of secondary school completion in the Brazilian schooling
system. Other smaller but still notable peaks occur at the end of the first half of primary
schooling and the end of middle school. Unfortunately, we did not have access to Lam and
Levinson’s complete PNAD 1985 data set, but fig. 3.2 can also be compared to fig. 2 in their
paper (205). The most striking difference between the figures is how the largest peak in
educational attainment has shifted over time from no education to completion of high school.
The other peaks are now much lower in relative size than they were in 1985, showing a more
equal distribution of schooling.
In order to visualize the decline in inequality in schooling we present the education
Lorenz curves for the youngest cohort, the cohort with the highest variance in educational
attainment, and the oldest cohort in fig. 3.3. What we see is that distribution of schooling has
unambiguously and substantially improved over the observed period. The Gini coefficients for
the 1952-1954, 1964-1966, and 1988-1990 cohorts are .45, .35, and .18 respectively, showing
just how great this improvement has been.
Figure 3.2
Frequency distribution, years completed schooling: 1952-54, 1964-66, 1988-90 birth
cohorts, Brazilian males, PNAD 2012
17. -16-
Figure 3.3
Lorenz curves, years of completed schooling: 1952-54, 1964-66, 1988-90 birth cohorts,
Brazilian males, 2012
Returns to education and income inequality
While the convergence of educational attainment is important for ensuring that education
acts as an equalizing force on income inequality, we must also look at returns to education. We
follow Lam and Levinson (1992) and estimate these returns using the same simple Mincer OLS
regression
ln 𝑌𝑌𝑖𝑖𝑖𝑖 = 𝛼𝛼𝑐𝑐 + 𝛽𝛽𝑐𝑐 𝑆𝑆𝑖𝑖𝑖𝑖 (Equation 3.1)
Where Y is income, S is an educational attainment, and subscripts i and c index individual ‘i’ in
cohort ‘c’. Because the cohorts have so little variation in age, the coefficient found for schooling
can be seen as indicative of mean returns to schooling. As shown in table 3.2, returns to
education have steadily decreased over time. This, coupled with the fact that the distribution of
schooling has become more equal suggests that education has continued to equalize incomes in
Brazil.
We calculate the impact of variance in education on variance of the log of income in
column (10) of table 3.2 and find a nearly unambiguous downward trend. Additionally, column
(5) shows that the variance in log income has been generally declining over the observed period,
18. -17-
so not only has education’s contribution to income inequality been decreasing, but income
inequality within cohorts has also been decreasing. These trends are also shown in fig. 3.4. This
is in contrast to what was found by Lam and Levinson using the 1985 data, which was that
variance in income peaked for the 37-39 year old age group and declined for younger groups
(211). It is important to note that due to currency reforms, this table cannot be directly compared
to the results found by Lam and Levinson.
Figure 3.4.
Earnings inequality, three-year birth cohorts: Total variance of log earnings, explained
variance, and residual variance, males with positive earnings, PNAD 2012
In general, we find that trends have continued since 1985. The distribution of schooling
has continued to become more equal with educational attainment converging to completion of
high school. In contrast with the 1985 results, variance in the log of income has steadily declined
between cohorts. This along with the decline in schooling inequality and reduction of returns to
education shows that changes in the distribution of education for the observed cohorts have been
equalizing.
As shown in column (5), mean returns to education have also fallen. This is noteworthy
because holding all else equal, declines in returns to education lead to a more equal distribution
of income. Not only is this true, but as shown by Lam and Levinson the effect a change in the
mean return to education has on variance of income is the square of the magnitude of the change
19. -18-
(1992, 209). This means that returns to education have an especially significant role in
determining income inequality.
The effect of the decline in returns to education as well as the convergence in mean
education found in table 3.1 can be seen in column (10), which presents the portion of variance
of income explained by education, as well as in Fig. 3.4. The clear downward trend in the
explained portion of income inequality shows that changes in the distribution of education and
returns to education have been equalizing.
Table 3.2
Monthly income by birth cohort: Descriptive statistics and cohort-specific earnings
equations, Brazilian males with positive earnings, 2012 PNAD
Age
group
(1)
Birth cohort
(2)
Sample
size
(3)
Mean log
earnings
(4)
Variance
log earnings
(5)
β
(6)
Std. err.
(7)
R2
(8)
V(u)
(9)
β2
V(S)
(10)
22-24 1988-1990 6,881 6.77 0.375 0.066 0.0021 0.126 0.325 0.049
25-27 1985-1987 7,229 6.91 0.482 0.091 0.0020 0.217 0.372 0.110
28-30 1982-1984 7,703 6.99 0.573 0.100 0.0019 0.267 0.418 0.155
31-33 1979-1981 7,610 7.07 0.634 0.104 0.0019 0.288 0.455 0.189
34-36 1976-1978 6,976 7.09 0.681 0.105 0.0019 0.303 0.470 0.211
37-39 1973-1975 6,717 7.08 0.724 0.107 0.0020 0.294 0.504 0.221
40-42 1970-1972 6,607 7.11 0.721 0.110 0.0020 0.321 0.484 0.236
43-45 1967-1969 6,123 7.13 0.797 0.100 0.0022 0.260 0.588 0.210
46-48 1964-1966 6,004 7.14 0.828 0.106 0.0021 0.298 0.577 0.251
49-51 1961-1963 5,624 7.16 0.872 0.120 0.0022 0.351 0.560 0.313
52-54 1958-1960 5,065 7.15 0.908 0.121 0.0024 0.339 0.594 0.314
55-57 1955-1957 4,387 7.14 0.968 0.130 0.0025 0.379 0.598 0.370
58-60 1952-1954 3,869 7.16 0.978 0.120 0.0027 0.334 0.648 0.330
22-60 1952-1990 80,795 7.07 0.719 0.096 0.0005 0.246 0.542 .177
Additionally, the unexplained portion of income inequality, shown in column (9) has also
been decreasing, leading the decreases in income variance seen in column (5). However, as seen
in Fig. 3.4, the unexplained portion of inequality has grown in terms of its share of total income
variance. So while education still plays an important role in determining income inequality, that
role is diminishing.
While the analysis in this section shows how changes in income affect overall income
inequality, it does not show anything about how income disparity between groups has changed.
Lam and Levinson focused on the picture of Brazil as a nation, ignoring differences between
groups of people in order to gain a broad understanding of how income inequality was changing.
We now extend the Lam and Levinson (1992) methodology to analyze the differences in
education and income between different groups.
3.3: Changes of differences in educational attainment and income inequality
between groups
One interesting way we can use Lam and Levinson’s methodology to learn more about
how changes in schooling inequality have affected inequality in general is to look at how the
distribution of education between groups of people has changed over time and how that has
affected the incomes of these groups. We look at Afro-Brazilians, women, and people living in
rural areas.
20. -19-
Afro-Brazilians
Figure 3.5 compares how mean educational attainment has changed for Afro-Brazilians
and White Brazilians. While Afro-Brazilians have lower mean schooling for every observed
cohort, we can see that the gap between them has been slowly closing. The difference in average
years completed of schooling between Afro and White Brazilians has decreased from 2.36 for
the 1952-54 cohort to 1.47 for the 1988-90 cohort. While this graph shows great advancement in
mean educational attainment for both groups, it also shows the extent to which Afro-Brazilians
continue to lag behind whites. The mean schooling of Afro-Brazilians in the youngest cohort is
only slightly above that of whites in the 1964-66 cohort, suggesting that Afro-Brazilians get the
same education as White Brazilians did about 24 years ago.
Figure 3.5
Mean educational attainment by birth cohort: Afro-Brazilians and white Brazilians, males,
PNAD 2012
This disparity in educational attainment may be partially due to the fact that Afro-
Brazilians experience lower returns to education than whites for every cohort observed. Lower
returns to education make the investment in education less attractive for Afro-Brazilians, which
can lead to lower mean educational attainment. Fig. 3.6 compares returns to education for White
21. -20-
and Afro-Brazilians using the same simple Mincer regression as in the previous section. As the
variable being modeled is log of income, returns to schooling are defined as the %age increase in
income that is the result of one additional year of schooling. It is clear from Fig. 3.6 that Afro-
Brazilians experience systematically lower returns to education.
Take the example of a 28-31 year old male. The expected incomes for an Afro-Brazilian
and a white Brazilian with no education in this age group are the same. However, if they have
completed high school, the white Brazilian can expect to earn around 300 more reais a month
than his Afro-Brazilian counterpart. This is equivalent to nearly a third of a standard deviation
difference in log of income.
This is problematic for ensuring that distribution of education is changing in a way that is
conducive to decreasing income inequality. As it stands, Afro-Brazilians not only earn
significantly less per year of education than white counterparts, but also get less education,
possibly due to these lower returns. Both effects contribute to income disparity between the two
groups.
Also problematic is the fact the difference in returns to education appears to be
increasing. As the youngest cohorts are generally still in school, considering their returns to
education is problematic. However, the difference in log of income between White and Afro-
Brazilians has decreased1
, suggesting that reduction of the educational attainment gap has
overpowered changes in returns to education.
Figure 3.6
Returns to education by birth cohort: Afro-Brazilians and white Brazilians, males, PNAD
2012
1 Appendix A3.1
22. -21-
Women
Women were left out of Lam and Levinson’s original study due to differences in their
labor markets that would have obscured results. While these differences still exist, it is important
to understand how these differences affect income disparity between men and women. For this
reason we analyze differences educational attainment and remuneration for men and women.
As seen in figure 3.7, women tend to have completed more schooling than men and the
difference has been growing over the years. Women also generally have higher returns to
education than men2
, so it makes sense that women get more education than men as they can
expect more for their investment. In general both these effects should be equalizing or even lead
to women having higher mean incomes, but this is not the case.
Figure 3.7
Mean educational attainment by birth cohort: Men and women, income earners, PNAD
2012
There are two difficulties in comparing women with men. Firstly, women have not
experienced decreases in returns to education. On the contrary, younger women tend to have
higher returns to education than older women. The average %age increase in income for one
additional year of education has risen from 9.9% to 13.8%. This may be due to the fact that
young women must be significantly better remunerated to entice them to enter the labor market.
2 Appendix A3.2
23. -22-
Secondly, women have a much lower predicted income if they are uneducated. Taking the
example of a 28-31 year old again, an uneducated man can expect to earn almost 300 reais more
per month than an uneducated woman. This is again about a third of a standard deviation in log
of income for males of that age group. Not only that, but a woman in this age group would need
on average about seven years of schooling to earn the same income as man with no education.
To better understand how income disparity between men and women has changed
between the observed cohorts we graph the expected log of income for a man and woman with
the mean education for each cohort in fig. 3.8. Except for some progress in the youngest cohorts,
the gap in earnings has remained relatively constant with an average difference of .61. So while
women continue to get more education than men and experience higher returns to education the
income disparity between men and women has not changed much.
Figure 3.8
Predicted log of income using mean educational attainment by birth cohort: Men and
women, income earners, PNAD 2012
Rural
Rural Brazilians also face disparities in educational attainment and income as
illustrated by Fig. 3.9. There has been slow but steady progress with the difference in mean
educational attainment shrinking from 3.73 years for the oldest cohort to 2.60 years for the
24. -23-
youngest. This could be due to a number of things, among them the educational reform that took
place in 1971, which raised the mandatory amount of schooling. Since mean educational
attainment in rural areas was very low before the reform, if the new schooling requirements were
well enforce they most likely had a significant impact on mean education in rural areas.
Individuals living in rural areas also experience lower returns to education and similar to Afro-
Brazilians, the difference in returns to education between rural and urban Brazilians has
generally been increasing, peaking for 25-27 and 37-39 year olds3
.
Figure 3.9
Mean educational attainment and log income by birth cohort: Rural and urban, Brazilian
males, 2012
This makes the rural case particularly interesting. For the two oldest cohorts, returns to
education in rural areas are higher than in urban areas despite the mean educational attainment in
rural areas for these cohorts being less than half that of in urban areas. This may be due to an
extra large premium put on relatively scarce education in rural areas for these cohorts.
Additionally, despite the fact that the gap in returns to education has grown, with returns to
education in rural areas now substantially lower in rural than in urban areas, the gap in mean
educational attainment has steadily shrunk. This may be due to differences in education quality.
As will be shown in Section 4, rural areas, particularly those in the North East of Brazil, have
lower than average education quality, which could lead to lower returns to education.
3 See Appendix A3.3
25. -24-
It is important to note that the difference in mean log earnings between urban and rural
individuals has been generally decreasing. This suggests that despite the changes in returns to
education, the ground being made up in rural areas in terms of educational attainment has been
working to lower disparity between urban and rural Brazil.
Summary
In general, we have seen that differences in education and earnings profiles of Afro-
Brazilians, women, and rural Brazilians have led to disparities in income for between groups. For
Afro-Brazilians, lower returns to education and lower mean educational attainment contribute to
income disparity between them and whites, but there has been some progress for the observed
cohorts. The gap between White and Afro-Brazilian mean educational attainment has decreased
from 2.36 years to 1.47 and the log earnings gap has decreased from 0.24 to 0.11 over all
observed cohorts. Women’s earnings profiles have kept mean income below that of men despite
having higher rates of return to education and on .6 more years of schooling on average. For
those living in rural areas, the education gap has been closing while the difference in returns to
education has been generally increasing, in the end reducing disparity between rural and urban
individuals. In the next section, we look at how much these differences in education and earnings
profiles affect overall income inequality.
3.4: Income inequality simulations
To determine the impact of education on income disparities, we estimated inequality
models incorporating three distinct educational adjustments, and then compared each to the
existing system. We tested the effects on income of the rural-urban, afro-white, and the female-
male education gaps. Additionally, we simulated income inequality if every male had the mean
years of education and compared it to income inequality under the existing education
distribution.
Method
We predicted current income inequality by first calculating determinants of monthly
income with:
𝑙𝑙𝑙𝑙 𝑌𝑌𝑖𝑖 =∝ +𝛽𝛽1 𝐸𝐸𝑖𝑖 + 𝛽𝛽2 𝐴𝐴𝑖𝑖 + 𝛿𝛿𝑽𝑽�𝒊𝒊 (Equation 3.2)
where for individual i, E represents years of education, A represents age, and V represents
gender, race, rural status, and migration status. We held V fixed, applied educational
corrections, and predicted new income distributions with:
𝑌𝑌𝑖𝑖 = 𝑒𝑒^(∝ +𝛽𝛽1 𝐸𝐸𝑖𝑖 + 𝛽𝛽2 𝐴𝐴𝑖𝑖 + 𝛿𝛿𝑽𝑽�𝒊𝒊) (Equation 3.3)
Lorenz curves were drawn to compare hypothetical inequalities with current inequality, and Gini
coefficients were calculated to quantify disparities. The actual GINI coefficient associated with
existing inequality was calculated to be .33.
Rural
26. -25-
We found that people living in rural communities had an average of 3.24 fewer years of
education than urban Brazilians. We amended education by 3.24 years for all rural-located
individuals. This adjustment had a low impact on earnings, as shown in figure 3.9, improving
the GINI coefficient from .33 to .31. Of the tests done, this modification produced the least
impact per additional years of education. This small change for such an improvement in
education indicates that rural communities experience significantly lower returns to education
than urban counterparts, and benefit less from additional education than Afro-Brazilians
(discussed below).
Figure 3.9
Lorenz curve of real earnings vs earnings without rural-urban educational gap
Race
To determine economic inequality not due to the education race gap, we applied an
additional 1.93 years of schooling to each Afro-Brazilian. This amount represents the difference
in years between the mean Afro-Brazilian education and the mean White-Brazilian education.
This was done to more closely align average Afro-Brazilian cohort education levels with those of
the same age White-Brazilian cohorts. Due to this reassignment, income inequality improved
slightly, and GINI improved from .33 to .31 (figure 3.10). While the Gini improvement is equal
to that of the rural education simulation improvement, the lesser educational curve indicates that
Afro-Brazilians respond more positively to educational corrections than rural Brazilians. When
27. -26-
compared to White-Brazilians, Afro-Brazilian earnings are impacted by education and
educational inequality more than rural Brazilians are when compared to urban Brazilians.
Figure 3.10
Lorenz curve of real earnings vs earnings without education-race gap
Gender
Investigating our findings that women have an average of one year more of education
than men, are experiencing rising returns to education, and still earn significantly less; we
predicted the real income distribution for men and women and compared it to a simulated
income distribution if women were subjected to the same pay scheme as men. To calculate the
real income distribution, we first regressed income against years of education and gender using
the equation:
𝑙𝑙𝑙𝑙 𝑌𝑌𝑖𝑖 =∝ +𝛽𝛽1 𝐸𝐸𝑖𝑖𝑖𝑖 + 𝛽𝛽2 𝐺𝐺𝑖𝑖𝑖𝑖 (Equation 3.4)
where, for individual i in cohort c, E represents years of education and G is a dummy variable
that represents gender. The outputted coefficients 𝛽𝛽1 and 𝛽𝛽2 were then used to estimate average
income per years of education and gender with the equation:
𝑌𝑌𝚤𝚤
� = 𝑒𝑒^(∝ +𝛽𝛽1 𝐸𝐸𝑖𝑖𝑖𝑖 + 𝛽𝛽2 𝐺𝐺𝑖𝑖𝑖𝑖) (Equation 3.5)
28. -27-
Both of these equations were used in conjunction with the entire data set; working men and
working women, ages 22-60.
We then calculated corrected earnings using similar equations, with the following
modification:
𝑙𝑙𝑙𝑙 𝑌𝑌𝑖𝑖 =∝ +𝛽𝛽1 𝐸𝐸𝑖𝑖𝑖𝑖 (Equation 3.6)
In this equation (3.6), gender was removed as an independent variable, and only data for
men was processed in the regression. The outputted 𝛽𝛽1 was then used as a coefficient for income
prediction in equation:
𝑌𝑌𝚤𝚤
� = 𝑒𝑒^(∝ +𝛽𝛽1 𝐸𝐸𝑖𝑖𝑖𝑖) (Equation 3.7)
into which we inputted data for both men and women, incorporating the true number of years of
education for each individual. This was done to discern impact on earnings of non-educational
gender-linked factors. As above, Lorenz curves were draw for both curves (Figure 3.11).
Following is method, the actual Gini was calculated to be .30 and the Gini coefficient
after equalization was .27. A larger improvement in inequality than that of the race or rural
simulations, this rise indicates that women are subjected to greater inequality due to non-
educational factors than Afro-Brazilians are due to educational-factors. Figures 3.7 and 3.8,
respectively indicating the greater education levels of women and significantly lower wages
support these findings.
Figure 3.11
Lorenz Curve of overall real earnings vs if men and women were paid at same rate
29. -28-
Overall Equality
Finally, to test income disparity due to educational distribution, all working men were
assigned the mean level of education, 8.28 years. As opposed to the previous trials which only
raised educational levels, this modification both lowered and raised educational attainment
depending on where the individual fell. Understandably, this had a large effect on improving
income inequality within the country (figure 3.11). Improving the GINI coefficient from .33 to
.23, this test shows that a significant portion of income disparities within Brazil is due to overall
educational inequality and not strictly due to racial, geographical, or biological inequality. While
there is no single categorically defined cause of income inequality within Brazil, there is still
great economic inequality due to overall schooling inequality.
Figure 3.12
Lorenz curve of real earnings vs earnings if all had mean years of education
3.5: Inequality predictions and conclusions
Predictions
To get an idea of how the distribution of education will continue to affect income
inequality we predict mean educational attainment for the birth cohorts that form the 22-60 age
group in 2021 and use these figures to compare future inequality to current inequality.
30. -29-
To estimate mean education in 2021 for each cohort in the observed age group, we look
at the percentage of each age group still in school. In order to predict mean education in three
years, we use the formula:
𝑃𝑃𝑎𝑎 =
𝑛𝑛𝑎𝑎 𝐸𝐸𝑎𝑎+3𝑛𝑛𝑎𝑎 𝑅𝑅𝑎𝑎+1+𝑛𝑛𝑎𝑎(𝑅𝑅𝑎𝑎−𝑅𝑅𝑎𝑎+1)
𝑛𝑛𝑎𝑎
(Equation 3.8)
or equivalently:
𝑃𝑃𝑎𝑎 = 𝐸𝐸𝑎𝑎 + 𝑅𝑅𝑎𝑎 + 2𝑅𝑅𝑎𝑎+1 (Equation 3.9)
where:
P is predicted mean educational attainment in three years
E is mean educational attainment
R is the %age of individuals in school
n is the number of observations
a is an age group index
Equations (3.8) estimates the mean schooling for a cohort in three years by first finding
the total years of schooling for an age group. Then the number of students who will still be in
school in three years is estimated using the % of the population of the next oldest cohort still in
school. This number is multiplied by three and added to the total years of schooling. Assuming
that all other students get one year of education, this number is also added to the total years of
schooling to generate an estimated number of total years of schooling in three years. Finally, this
new total number of years of schooling is divided by the number of observations in the cohort to
generate a predicted mean level of education in three years. This prediction method provides a
lower bound because as Fig 3.2 shows, more Brazilians are staying in school longer, suggesting
the %age of students in each age group is rising rather than remaining constant.
We use this estimation method three times to predict mean educational attainment for the
observed age group in 2015, 2018, and 2021, presented in table 3.4. As expected, we see mean
educational attainment increase for each age group in each three-year period. One interesting
prediction is that the best-educated age group becomes 28-30 in 2021 rather than 25-27 as seen
in 2012. This is feasible because as people continue to attain higher and higher levels of
education, they take longer to do so. Thus, it will most likely take longer to reach a terminal level
of mean education for a cohort. We saw evidence of this in the increase in mean schooling for
the four cohorts born in 1952-1963 between 1985 and 2012.
31. -30-
Table 3.4
Educational attainment, Brazilian males, PNAD 2012
Age group
(1)
Mean, 2012
(2)
Predicted mean,
2015
(3)
Predicted mean,
2018
(4)
Predicted mean,
2021
(5)
Increase in
mean from
2012 to 2012
(6)
22-24 9.52 9.87 9.86 9.98 0.46
25-27 9.58 9.93 10.27 10.27 0.69
28-30 9.26 9.85 10.20 10.54 1.28
31-33 8.80 9.44 10.03 10.38 1.58
34-36 8.31 8.94 9.58 10.17 1.86
37-39 7.87 8.42 9.05 9.70 1.83
40-42 7.61 7.96 8.51 9.14 1.53
43-45 7.93 7.69 8.04 8.59 0.66
46-48 7.38 7.99 7.75 8.10 0.72
49-51 7.03 7.43 8.04 7.80 0.77
52-54 6.83 7.07 7.47 8.08 1.25
55-57 6.46 6.86 7.11 7.51 1.05
58-60 6.07 6.49 6.89 7.13 1.06
22-60 8.28 8.30 8.68 9.19 0.91
In order to estimate how these changes will affect income inequality, we graph predicted
Lorenz curves for the 22-60 age group in 2012, 2015, 2018, and 2021 using the methodology
used in the previous section. However, for these predictions individuals are assigned their
cohort’s (predicted) mean educational attainment. In order to account for decreasing returns to
education, we reduce the return to education by .002, the average change between cohorts in
2012, every three-year period. Because the Lorenz curves for the three future cohorts are so
close, only those for 2012 and 2021 are shown in fig. 3.10.
Figure 3.13
Estimated Income Lorenz Curves, 2012 and 2015
32. -31-
We find that the estimated Gini coefficient using this prediction method went from .34 in
2012 to .22 in 2021. While it is not likely that actual income inequality in Brazil will experience
so extreme a decline by 2021, this result is still valuable. We see from this prediction that even if
the %age of the population in school for each age group does not change, higher mean education
of younger cohorts and expected decreases in returns to education will continue to equalize both
the distribution of schooling and income.
Conclusions
By replicating the methodology used in Lam and Levinson (1992) and expanding on their
work, we found that education and schooling inequality continue to be extremely important
factors of income and income inequality.
Mean educational attainment of Brazilian males has continued to rise since 1985. The
youngest adult cohort in 2012 has on average more than 3.5 more years of education than the
youngest cohort in 1985. Returns to education and the share of income inequality attributable to
schooling inequality have also continued to fall, showing that changes in the distribution of
income have been equalizing. This is shown by the fact that the proportion of income variance
explained by education has fallen continuously. However, inequality not due to changes in
education or returns to education persists and has grown significantly as a proportion of total
income inequality.
Although mean educational attainment is increasing, we observe significant differences
in educational attainment and earnings profiles between groups. We find that although the
difference in mean education between white and Afro-Brazilians has largely remained stable, it
is beginning to decrease amongst the youngest cohorts. This coupled with a shrinking gap in
returns to income suggests that disparity between white and Afro-Brazilians is decreasing. For
women, we find that different earnings profiles have kept disparity between men and women
relatively stable despite women continuing to receive more education and experience higher
returns to education than men. For rural areas, we find that the gap in educational attainment has
been steadily decreasing despite little obvious progress in terms of returns to education.
However, the effect of increased education has overpowered any changes in returns to education
and urban-rural income disparity has decreased.
In order to estimate how much of an impact schooling inequality has on income inequality
we run simulations using different hypothetical distributions of education. We find that
distributing education equally across races explains about one third of income inequality in
Brazil. Although this is a very large proportion of existing income inequality, there still remains
a very significant portion of inequality that is unexplained by the distribution of education. The
effects of eliminating the gaps in educational attainment between white and Afro-Brazilians and
between rural and urban Brazilians also have significant impacts on the predicted Gini
coefficient, showing that disparity between groups as a result of unequal distribution of
education has a significant effect on overall income inequality. Eliminating differences in the
earnings profiles of men and women also has a large effect on income inequality, showing how
different returns to education between groups also have a significant effect on overall income
inequality
Looking to the future of inequality in Brazil, we estimate a lower bound for the education
of the observed age group in the year 2021 and find that mean educational attainment continues
to increase and converge among all cohorts. This coupled with decreasing returns to education
suggest that education will continue to have an equalizing effect on income inequality in Brazil.
34. -33-
4.1: Introduction
This section examines the effectiveness of migration within Brazil as a long-term strategy
out of poverty. We present Yap’s findings from 1960 and offer our analysis to determine income
returns to migration in 1960. We then replicate Yap using data from 2012 to see whether
migration remains an effective means of escaping poverty half a century later. We conclude by
identifying similar trends in income returns to rural-urban migration and migrant assimilation
into the urban workforce between the two periods, suggesting similar workforce dynamics in
urban Brazil.
4.2: Yap (1976) Methodology
In the middle of the last century, policy makers considered rural-urban migration a
catalyst in the expansion of urban poverty and the dilution of the traditional, low-productivity
labor market in metropolitan areas. However, Yap’s empirical analysis of the economic returns
to migration told a different story. Yap (1976) used the 1960 population census in Brazil to
examine the income gains associated with migration as well as the assimilation of migrants into
the urban labor market. She found both significant income benefits and rapid assimilation of
migrants.
The author estimated OLS earnings equations to examine the returns to migration and
migrant assimilation into urban areas. Her sample included 56,000 people currently living in
three major regions in Brazil – Northeast, East, and South. Appendix 4.2 offers an explanation as
to why the North was excluded from Yap’s analysis and our results for returns to migration in
the region. Other variables that may affect income, sex, age, race and completed education, are
accounted for in her analysis. These variables are held constant when interpreting the effect of
migrant status on income.
Yap also compared incomes of migrants living in urban areas to their foregone earnings,
estimated as the current mean income of non-migrants in rural areas. Differences in age,
educational attainment, sex and race are taken into consideration in order to isolate income
differences attributable to migration status from differences in human capital and other
characteristics that may affect pay. The functions are of the following form:
𝐿𝐿𝐿𝐿𝑌𝑌𝑖𝑖 = 𝛼𝛼𝑜𝑜 + ∑𝑏𝑏𝑖𝑖 𝐸𝐸𝑖𝑖𝑖𝑖 + ∑𝑐𝑐𝑖𝑖 𝐴𝐴𝑖𝑖𝑖𝑖 + 𝑑𝑑1 𝑆𝑆𝑗𝑗 + 𝑑𝑑2 𝑅𝑅𝑗𝑗 + ∑𝑒𝑒𝑖𝑖 𝑀𝑀𝑖𝑖𝑖𝑖 (Equation 4.1)
Where:
Yj = average monthly income of individual j,
Eij = education level i of individual j [no formal education (omitted), some primary education,
primary education completed, more than a primary education],
Aij = age group i of individual j [10-19, 20-29 (omitted), 30-39, 40-49, 50 and over],
Sj = sex of individual j [male (omitted), female]
Rj = race of individual j [white (omitted), nonwhite]
Mij = migration status I of individual j [rural non-migrant (omitted), rural-rural migrant, recent
rural-urban migrant (0-4 years), less recent rural-urban migrant (5+ years)]
35. -34-
Two additional variables were also used in subsequent regressions: employment status [wage
earners (omitted), self-employed, employer] and sector of employment [traditional (omitted),
modern].
4.3: Description of Yap’s Results
Yap found significant income gains associated with migration in all three regions. As
shown in Table A 4.1, rural-urban migrants in the South earned about 38% more than non-
migrants, on average, holding all else constant, which is similar to 40% in the Northeast, but
significantly higher than Eastern states, whose recent immigrants only experienced an average of
20% improvement in income, holding all other variables constant. Less recent rural-urban
migrants gained even higher income returns across all region, ranging from 43% in the South,
45% in the Northeast and almost 60% in the East, on average, holding all else constant.
The returns for rural-rural migrants, however, were not as substantial. In fact, rural-rural
migrants in the South and East experienced negative returns to migration compared to non-
migrants of, on average, holding all other variables constant, 13% and 1% respectively. While
the Northeast did not have negative income returns, the effect of being a rural-rural migrant was
only a 4% increase in income compared to non-migrants. Overall, Yap found that, in 1960, it was
financially beneficial to migrate from a rural area to a non-rural area. This is likely due to more
job opportunities educated or non-educated workers in cities. In rural areas, job opportunities are
most likely limited to agriculture.
4.4: Replication Methodology
We replicate Yap’s article using data from the 2012 PNAD. The dependent variable is the
natural log of income, so that the results may be interpreted in ages. Migrant status is divided
into the same groupings as Yap’s article: recent rural-urban migrants (0 to four years since
migration), less recent rural-urban migrant (greater than five years since migration) and rural-
rural migrant. The PNAD survey does not specify from where individuals migrated, so we
assume that migrants originated from a rural area. The group “non-migrant” is excluded so that
we can compare the incomes of various migrant types to non-migrants living in the same region.
The replication also includes the same control variables: sex, age, race and completed education
(one to three year, primary completed or more than primary completed).
4.5: Replication Results: Income Returns to Migration in Brazil, 2012
The income returns to migration have significantly changed since 1960 across all regions.
In the South, recent rural-urban migrants experience a 15% increase in income, the East and
Northeast have a 30% increase compared to non-migrants, on average, holding all other variables
constant. Table A 4.2.9 in the Appendix shows that recent rural-urban migrants in Northern
states have about a 13% greater income than non-migrants, holding all else constant. Income
returns for less-recent migrants is highest in the Northeast, 28% then drops to 21% in the
Northeast and finally only 8% in the South, on average, holding all else constant. The North
actually experiences negative income returns of 20%, on average, holding all other variables
constant. However, income returns for rural-rural migrants are negative for the South, Northeast
and East compared to non-migrants. The North is the only region where it is beneficial to be a
36. -35-
rural-rural migrant, who experiences an average of 8% increase in income, holding all else
constant.
4.6: The Control Group: Returns to White Males, 20-29, with No Formal Education
Tables 4.2a and 4.2b show the log of predicted monthly income for white males, 20-29
years old, with no formal education, separated by migration status. No significant conclusions
come from comparing the regional returns to migration due to many changes in Brazil’s currency
since 1960. The results are separated into two tables, table 4.2a showing Yap’s results and table
4.2b showing the results from 2012.
The 2012 differ from those found in Yap in several ways depending on the region
observed. In the South in 1960, migrating correlated with an increase in monthly income of
about 1.719 R$, and correlated with an increase of about 1.337 R$ per month. While average
monthly predicted income was relatively similar for both recent and non-recent migrants in these
regions, they differed significantly in the East, where recent migrants experienced an increase in
monthly income of about .764 R$, and non-recent migrants experienced an increase in monthly
income of 1.216 R$. These correspond to about a 45-50% increase in the monthly-predicted
income in the South, around 35% increase in the Northeast, and 20% increase for recent migrants
and 32% increase for non-recent migrants in the East.
However, the percent change in monthly-predicted income based on being a migrant in
2012 was significantly smaller. In the South, recent migrants experienced a 15% increase and an
11% increase for non-recent migrants versus rural non-migrants, 21% for recent and 19%
increase for non-recent migrants in the Northeast, and a 9% increase for non-recent migrants and,
interestingly, a 19.2% decrease for recent migrants in the East. It is unclear why there is
suddenly a decrease in expected monthly income in the East.
These numbers were found by using the same equations used by Yap as well as 𝛼𝛼𝑜𝑜 by
region from Table 4.1. To obtain our regression estimates of returns to migration for rural non-
migrants, recent rural-urban migrants, and less recent rural urban migrants, we use:
a: =
𝑒𝑒α0
10
b: =
𝑒𝑒(α0+Rr)
10
c: =
𝑒𝑒(α0+Rl)
10
(Equation 4.2)
Where:
Rr
is returns to recent migrants (1-5 years) and
Rl
is returns to less recent migrants (5+ years)
37. -36-
4.7: Urban Labor Force Assimilation
Table 4.3 shows the effects of the assimilation of rural-urban migrants into the labor
force. For this regression, we again make the assumption that all migrants came from rural areas.
This table shows the log of the total monthly income per capita regressed against when they
migrated (non-migrants), sex (male), age (20-29), race (white), educational attainment (no
education), which sector they worked in (Traditional Sector), and their employment status
(employed). We include the sector in which they work post-migration to see if they are
assimilated into the formal or informal workforce. To do this, we looked to see if migrants paid
income taxes, indicating work in the modern sector.
Interestingly, and perhaps expectedly, there is larger return to working in the modern
sector compared to the traditional sector. The magnitudes of the increase in returns are 24.5%,
42%, and a 46% increase in income from working in the modern sector in 2012 compared to
those in the traditional sector in the South Northeast and East respectively, versus a 27.3%
increase in income in 1960 in the South (the only region for which the results for Yap were
statistically significant.) Also, returns for employers have increased dramatically in the south by
about 60%, decreased very slightly but still retain a huge magnitude in the Northeast from 123 %
Table 4.1
Changes in Returns to Migration in Brazil
Independent
Variable
Yap South
(1960)
(1)
PNAD
2012
South
(2)
Change
in
South
(3)
Yap
Northeast
(1960)
(4)
PNAD
2012
Northeast
(5)
Change
in
Northeast
(6)
Yap
East
(1960)
(7)
PNAD
2012
East
(8)
Change
in East
(9)
Recent (0-4)
rural-urban
migrant
(Compared to
Rural non-
migrants)
.379**
(2.59)
0.115**
(7.05)
-0.264
0.412
(2.11)
0.303**
(11.46)
-0.109
0.179
(0.99)
0.299**
(15.34)
0.120
Recent (0-4)
rural-urban
migrant
(Compared to
Urban non-
migrants)
0.011
(0.09)
-0.043**
(-2.00)
-0.054
0.245
(1.3)
0.184**
(7.87)
-0.061
0.100
(0.81)
0.111**
(5.47)
-0.011
Less recent (5+)
rural-urban
migrant
(Compared to
Rural non-
migrants)
.428**
(3.77)
0.083**
(9.54)
-0.345
0.454
(2.11)
0.279**
(20.96)
-0.175
0.585**
(4.95)
0.212**
(20.37)
-0.373
Less recent (5+)
rural-urban
migrant
(Compared to
Urban non-
migrants)
0.08
(-1.13)
-0.029**
(-3.55)
-0.109
0.289**
(3.16)
-0.195
(-18.18)
-0.484
0.296**
(4.18)
0.121**
(13.87)
-0.175
38. -37-
to 117% increased returns, and have become statistically significant in the East, at 88.1%
increased returns for being an employer. These results imply that many of the changes in Brazil
between 1960 and 2012 have significantly benefitted the bottom line of employers.
4.8: YAP Comparison
The coefficients associated with both recent and less recent rural-urban migrants when
compared to rural non-migrants have generally decreased compared to Yap’s figures, suggesting
that there are now lower income returns to migration. There is 26.4% decrease in returns to
migrating in the South between 1960 and 2012, an 11% decrease to migrating in the Northeast
between 1960 and 2012 and a 12 % decrease in migrating in the East between 1960 and 2012 for
recent rural-urban migrants when compared to rural non-migrants, on average, holding all else
constant. When compared to urban non-migrants the coefficients for both recent and less-recent
migrants have decreased in all regions, but the data from 1960 is not statistically significant.
Recent rural-urban migrants earn 4 % less than urban non-migrants in the South. However, rural-
urban migrants earn 18.5 % more in the Northeast, and 11 % more in the East when compared to
urban non-migrants.
It is not certain why urban non-migrants earn more than migrants in the South but less
than migrants in the Northeast and East, but one possible explanation is that migrants have
characteristics that cause them to migrate. These characteristics may be positively correlated
with higher earnings, like grit, hard-work or determination. These characteristics are difficult to
capture, but may express themselves through higher levels of returns. Income returns to age have
been relatively stable except for a notable increase for ages 50 and above in almost every case.
For example, in the South (where the effect appears to be weakest) those 50 or older have
expected incomes on average 38.2 % to 56.5 % higher than 20 to 29 year olds (holding all else
equal). This could be interpreted as an increase in returns to experience, as shown in Table 4.5.
Table 4.2
Gender and Race Effects in Brazil
Independent Variable
PNAD 2012
South
(1)
Yap South
(1960)
(2)
PNAD
2012
Northeast
(3)
Yap Northeast
(1960)
(4)
PNAD 2012
East
(5)
Yap East
(1960)
(6)
Rural-born Female
(compared to Male)
-0.444**
(-55.38)
-.572**
(-6.17)
-0.450**
(-38.96)
-0.337
(-4.33)
-0.494**
(52.55)
-0.551**
(6.91)
Urban Female (compared to
Male)
-0.366**
(-51.85)
-0.549**
(-11.82)
-0.416**
(-44.47)
-0.588**
(-10.46)
-0.464**
(-61.81)
-0.503**
(-11.07)
Rural-born Nonwhite
(compared to white)
-0.186**
(-21.61)
-.214**
(-2.16)
-0.175**
(-13.82)
-0.135
(-2.69)
-0.291**
(29.58)
-0.076
(-1.32)
Urban Nonwhite (compared
to white)
-0.275**
(-36.37)
-0.175**
(-3.25)
-0.008
(-0.82)
-0.118**
(-2.21)
-0.324**
(42.29)
-0.286**
(6.81)
39. -38-
The effects of being female and non-white have both remained negative. Being rural-born
female compared to a male results in 44 % less income in the south, 45 % less income in the
northeast, and 49.4 % less income in the East in 2012. However, these magnitudes have slightly
decreased since 1960; although the Northeast is statistically insignificant, being a rural-born
woman in the South results in a 13 % smaller decrease in income in 2012 than it did in 1960, and
a 6 % smaller decrease in 2012 versus 1960 for a woman in the East. Females living in urban
areas experience an even greater decline in gender-related pay differences. Compared to 1960, a
woman in 2012 earns 18 % less in the South, 17 % less in the Northeast, and 4 % less in the East.
Table 4.3
Returns to Age in Brazil
Independent
Variable
PNAD 2012
South
(1)
Yap South
(1960)
(2)
PNAD 2012
Northeast
(3)
Yap Northeast
(1960)
(4)
PNAD 2012
East
(5)
Yap East
(1960)
(6)
50 and Over,
Rural-born
(compared to
20-29 year old)
0.382**
(28.92)
.444**
(4.06)
0.794**
(43.55)
0.114
(1.49)
0.696**
(44.92)
0.392**
(4.32)
50 and Over,
Urban
(compared to
20-29 year old)
0.568**
(53.48)
0.193**
(-3.25)
0.778**
(58.08)
0.656**
(7.75)
0.662**
(60.91)
0.369**
(5.80)
Another important change is the decrease of income returns as a result of being non-
white. Where they are statistically significant in the South and East, returns have increased by 10
% and 4 % respectively from 1960 to 2012 among the urban non-white. For the urban non-white,
there are also significant negative returns associated with race: people receive almost 19 % less
income in the South, 17.5 % less income in the Northeast, and 29 % less income in the East if
they are non-white. We cannot observe the net effect of these changes, since 1960 values are
statistically insignificant in the East and Northeast, but we can be certain of a 3 % decrease in
income returns in the South.
Income returns to age have remained constant except for a notable increase for ages 50
and above across all regions. For example, in the South (where the effect appears to be weakest)
those 50 or older have expected incomes on average 38.2 % to 56.5 % higher than 20 to 29 year
olds (holding all else equal). In the Northeast, where effects appear strongest, those 50 years or
older have 77.8 to 79.4 % higher incomes when compared to 20-29 year olds. This could be
interpreted as an increase in returns to experience over time, or possibly an increase in the return
to experience based on education.
An interesting finding is that returns to education above the primary level have decreased
substantially, as illustrated by Table 4.4. The urban Northeast is an example of where the returns
to having more than primary education have dropped from over 100 % to 54.8 % from 1960 to
2012. This decline is most likely due to the education reforms mandating that children stay in
school longer. Since most people now have higher than a primary education, the income returns
to that education have diminished. However, education is still clearly a strong, positive force in
increasing income. By attaining more education, people are unequivocally better off compared to
those with less education. When compared to those with no education, having more than primary
education increased a rural-born migrant’s earnings by almost 60 %. Those increases also hold
40. -39-
true in the East, where earnings increase by almost 73 %, and in the Northeast, where earnings
increase by 58 %. It is important to note that this data captures all people with more than primary
education, so there is some significant effect from those with very high levels of education, but
overall it is clear that education has very positive effects on income.
Table 4.4
Changes in Returns to Education in Brazil
Independent
Variable
Yap
South
(1960)
(1)
PNAD
2012
South
(2)
Change
in
South
since
1960
(3)
Yap
Northeast
(1960)
(4)
PNAD
2012
Northeast
(5)
Change
in
Northeast
since
1960
(6)
Yap
East
(1960)
(7)
PNAD
2012
East
(8)
Change
in East
since
1960
(9)
More than
Primary,
Rural-born
(Compared to
no education)
.931**
(3.78)
0.592**
(28.92)
-0.339
1.148
(3.83)
0.580**
(36.06)
-0.568
1.043**
(4.54)
0.727**
(48.93)
-0.316
More than
Primary,
Urban
(Compared to
no education)
0.754**
(9.86)
0.577**
(32.25)
-0.177
1.285**
(14.66)
0.548**
(34.02)
-0.737
1.077**
(16.20)
0.588**
(41.28)
-0.489
The returns to working in a modern sector have remained about the same in the South,
decreasing by about 3 % (from around 27 % to 24 % increase in wages.) However, income
effects of working in the modern sector have become statistically significant Northeast and East
and relatively large, at 42 % and 46 % respectively (Table 4.5). These changes are important for
two reasons. First, they show evidence that the modern sector has moved into both the Northeast
and the East, and is having significantly positive effects on employee income. Secondly, for the
South, they show that gains to income based on working in the modern sector have persisted, and
may imply some ways to increase income through moving those in the traditional sector in to the
private sector. The returns to being an employer have decreased in all regions but still remain
relatively high with a 77 % increase in income in the South where this estimate is lowest.
Finally, self-employment still has a wide difference in effects between regions, between a 9 %
decrease and a 5 % increase, but the effect has declined in all regions compared to 1960 (Table
4.5).
Table 4.5
Returns to Self-Employed and Employers in Brazil
Employment
Status
PNAD 2012
South
(1)
Yap South
(1960)
(2)
PNAD 2012
Northeast
(3)
Yap Northeast
(1960)
(4)
PNAD 2012
East
(5)
Yap East (1960)
(6)
Self-employed
0.052**
(4.58)
0.137**
(2.57)
-0.085**
(-6.71)
-0.015
(-0.26)
0.024**
(2.11)
-0.130**
(2.56)
Employer
0.765**
(39.02)
0.112**
(8.02)
1.169**
(37.70)
1.226**
(5.41)
0.881**
(34.24)
0.232
(1.76)
41. -40-
4.9: Conclusions
Migration still remains an effective strategy out of poverty in some regions, but it appears
that migration is not as lucrative as it was in 1960. Brazil may have even attained the maximum
benefit of migration, as shown by decreasing returns to migration in every region: around 26 %
less for recent migrants and 35 % less for non-recent migrants in the 2012 South compared to the
1960 South, around 11 % and 18 % for recent and non-recent migrants respectively in the 2012
Northeast compared to those in the 1960 Northeast, and 37 % less returns to migration for those
in the 2012 East than those in the 1960 East, on average, holding all else constant (changes in the
incomes of recent migrants in the 1960 Census were not statistically significant). Furthermore, a
20% income decrease for non-recent migrants in the North compared to non-migrants is
observed in our new analysis using 2012 data. Due to the preceding effects of both decreasing
returns to migration as time goes on (and even negative returns) it will be interesting to observe
if Brazil’s population continues to move away from the countryside or if the tide of migration
will stop or perhaps even reverse itself. It is also important to note that differential educational
quality by region may also play some part in returns to migrating. A migrant from the Northeast
may be paid less than a migrant from the South simply because they suffer poorer quality
education. This phenomenon will be explored further in this report.
However, if the returns to work in modern sectors continue be significantly higher than
the return to work in traditional sectors, at 24.5 %, 42 %, and 46 % higher for the South,
Northeast, and East respectively using 2012 data, it may not make sense for large groups of
people to move back to rural areas, and instead they may try to enter modern labor markets. It
will be interesting to see how much Brazil’s migration patterns shift over the next forty years.
42. -41-
Section 5:
Analyzing the Trade Off Between Labor and School
By Alexandra Bryant, Rachel Butler, Natalie Melville, Alexander Montiel, and Thaddeus
Pinakiewicz
43. -42-
5.1: Introduction
In section 4, the difference in school quality due to geographical differences is apparent.
This leads to a new choice parents may take, choosing to put their students in school and where
to put them in school. To look at this analysis this next section will look at whether parents are
choosing to put their children in school or in work. As discussed in section three, education does
have returns. Even though these returns are decreasing, they are still positive. Therefore, the
choice between school and labor is a choice between a long-term investment in your child and a
short term gains from the child working.
To start off this analysis, section 5.2 will describe the characteristics of the students in
school and in the labor workforce. This will show which people are in school, in work, or a mix
of the two by age, race, and geographical factors. To further understand these relations, section
5.3 will analyze how these various characteristics affect the decision between entering the labor
force and attending school. This will show how various factors influence these decisions
including personal characteristics, geographical characteristics, home characteristics, and income
factors.
Once the choices have been analyzed, the next step is to look at the outcome of these
choices. This will be done by analyzing how far behind students are in school based on the
typical grade level for a child’s grade. The effect of work on causing students to be behind in
school will be further addressed to find the opportunity cost of choosing to put a child in the
labor force.
The following analysis will be based upon 4 possible decisions of child labor and
schooling.
Figure 5.1
Child Schooling and Labor Choice Set
Figure 5.1 shows the four possible outcomes.
1. In school and working
2. In school and not working
3. Not in school and working
4. Not in school and not working
44. -43-
Table 5.1
Proportion of Children Ages 10 to 16 in
School and Working
Attends School (%)
Yes No
Works (%)
Yes 8.1 1.1
No 88.0 2.8
Figure 5.1 shows the %age of children in each choice set. The vast majority of children
attend school (96.1 %). Of those who attend school, 8.2 % also work. Since our research is
primarily concerned with the trade-off between schooling and work, we exclude the children
who neither work nor attend school from our analysis.
5.2: Descriptive Statistics
The data used in this paper is a subset of the data from the 2012 PNAD. We looked at
children ages 5 to 16 and the choices their families made about sending their children to school
and having them work. The survey includes data for 68,439 children in this age range.
There was a slight majority of male respondents, 51.48% to 48.52% female. The racial
breakdown of respondents was 37.68% White, 0.28% Asian, 61.59% Afro-Brazilian (Black or
“brown”), and 0.44% Indigenous. The majority of respondents lived in urban environments with
only 17.69% living in rural environments. The children lived mostly in the more densely
populated Northeast (30.87%) and Southeast (25.76%), followed by the Northwest (19.06%),
South (13.87%) and finally, Central Brazil (10.45%).
Families ranged in size from 2 to 16 members with an average of 4.47 members. These
families had from 1 to 11 children age 16 or below with an average of 2.25 children age 16 and
below. They had an average monthly household income of $138.87 per capita. 15.65% of
children live below the poverty line and 8.01% of children live below the destitute line, which
we defined as making $38.74 per capita or less and $26.25 per capita or less, respectively.
24.89% of households only had one parent, while 87.35% of children live with their mother.
The majority of children age 5 to 16, 95.89%, attend school and this rises to 96.11%
when looking at children age 10 to 16. Enrollment peaks at age 11 with 99.26% of children
attending.
For the purpose of this section, we limit the scope of our analysis to children between 10
and 16 years of age. First, approximately 96% of the entire child labor force (2.2 million) falls
within this age range. Figure 5.2 shows that a very small proportion children below the age of 10
are in the labor force. At any given age between 5 and 9, less than 1% of children have a job.
This proportion begins to grow around the age of 10 and then sharply increases beyond the age
of 14, when children are no longer required to attend school by law. At age 16, children become
independent economic agents and can legally enter the workforce full-time, without restriction.
45. -44-
Therefore, we exclude observations beyond the age of 16 because parents are no longer the
primary decision maker in the schooling-labor tradeoff.
Figure 5.1
Child Labor Force Participation by Age
This age sample is consistent with historical trends, which show a changing child labor
demographic. Figure 5.2 shows the mean age of entry into the labor force over time. The
average age of entry into the labor force for the 1947-1949 birth cohort was approximately 13.
This age has increased steadily over time until the most recent cohort, born between 1992 and
1994, with a mean age of 16.
Figure 5.2
Mean Age of Labor Force Entry Over Time
46. -45-
9.19% of children 10 to 16 have worked in the past week, and of them 88.12% also
attended school. Children who reported working in the past week worked from 1 to 98 hours per
week, with an average of 24.45 hours.
Upon looking at demographics of those who attended school and those who worked, a
common trend appeared. Those who attend school at higher rates are less likely to work,
although by varying degrees.
Girls were only slightly more likely attend school, but boys reported working at a rate of
12.03% versus 6.17% of girls. Indigenous respondents reported the lowest enrollment %age and
Asian respondents reported the highest, although both groups have less than 175 respondents.
Indigenous children also reported working at much higher rates than other children. 30.43% of
indigenous children work while every other race reported from 7.03 to 9.68% of children
working. While rural children reported only slightly lower enrollment, they reported working at
over three times higher rates than their urban counterparts. Region had little impact on
enrollment %age, but it did affect the %age of children working. 11.04% of children in the
Northwest reported working, the highest of any region, while only 6.94% of children in the
Southeast reported working, the lowest of any region. There was a positive a positive correlation
between the number of children under age 16 and below and employment. The employment rate
increased from 8.97% with only one child 16 and under to 70% with 11 children age 16 and
under. Schooling level of the head of household and attendance also appeared to have a positive
correlation, and schooling level of the head of household and rate of employment of the children
appeared to have a negative correlation.
5.3: Factors that Influence Decisions
Having discussed the characteristics of the Brazilian family, we now consider how these
characteristics affect the parent’s decision to send their child to school or to work. Under the
human capital model of education, education increases future income while working only
increases current income. Therefore, the decision to work or go to school is a trade-off between
current and future income.
To show the effects of the school-work decision we utilize a logistic regression. A
logistic regression analyzes the dependent binary variable in the equation where: p() is the
probability of a variable being equal to 1, x is a vector composed of our regression inputs, α is
the intercept, and β is the regression coefficient, and ɛ is the error term. A logistic regression can
be represented as follows:
𝑙𝑙𝑙𝑙 𝑙𝑙𝑙𝑙𝑙𝑙�𝑝𝑝(𝑥𝑥)� = log �
𝑝𝑝(𝑥𝑥)
1−𝑝𝑝(𝑥𝑥)
� = 𝛼𝛼 + 𝛽𝛽𝑥𝑥 𝑋𝑋𝑥𝑥 + 𝜀𝜀 (Equation 5.1)
Where 𝑋𝑋𝑥𝑥 is a vector composed of variables for:
Personal characteristics
Geographical characteristics
Home characteristics
Income factors
We can interpret the βx coefficients of the regression via their odds ratio point estimates.
There is an estimate for every independent variable in the regression and the value ranges from
*See appendix A5.2 for list of all variables used
in regression
47. -46-
zero to infinity. An odds ratio point estimate with a value greater than one indicates that an event
is more likely to occur, whereas a value less than one suggests that it is less likely to occur. The
% concordance is determined by taking every data point from the dependent variable, and
calculating the % of the cases where the model correctly estimates the variable.
48. -47-
4 Values in parenthesis are the p-value results associated with each result.
5 This predicted wage was found by analyzing the working students and finding how different factors such as race, geographic locations,
education, and age affect their wage. The results of this analysis can be found in appendix table A.5.2. This creates an equation that can be used
on each child who works or does not work. This information will then be used to see how it affects the choice to go into the labor force. By
analyzing this information, it will be clearer if a higher wage causes a student to be more likely to work due to the increased benefit the child can
have from working
Table 5.2
Logistic Model Assessing Likelihood of School-Work Choices
Dependent Variable4
Attending School
Attending School and
Working
Not Attending School
and Working
Age 2.6** 0.95** 13**
(<.0001) (<.0001) (<.0001)
Age2 0.95** 1.0** 0.93**
(<.0001) (<.0001) (<.0001)
Female 1.1** 0.43** 0.27**
(<.0001) (<.0001) (<.0001)
Rural 1.34** 2.9** 2.1**
(<.0001) (<.0001) (<.0001)
Birth Order 0.93** 1.1** 1.3**
(0.508) (<.0001) (<.0001)
Number of Children in Family 1.0** 1.1** 1.0**
(<.0001) (<.0001) (0.0002)
Head of Household's Years of Schooling
1.1** 0.96** 0.96**
(<.0001) (<.0001) (<.0001)
Head of Household's Spouse's Years of
Schooling
1.1** 0.98** 1.0**
(<.0001) (<.0001) (<.0001)
Afro-Brazilian 0.81** 0.95** 1.1**
(<.0001) (<.0001) (<.0001)
Asian 0.55** 0.52** <0.001
(<.0001) (<.0001) (0.592)
Indigenous 0.90** 2.1** 0.35**
(<.0001) (<.0001) (<.0001)
Migrant 0.89** 1.0** 0.96**
(<.0001) (<.0001) (<.0001)
North Region 0.69** 1.2** 1.0**
(<.0001) (<.0001) (0.0009)
North-East Region 0.72** 0.83** 0.82**
(<.0001) (<.0001) (<.0001)
South Region 0.76** 1.7** 1.2**
(<.0001) (<.0001) (<.0001)
Central Region 0.77** 1.6** 1.0**
(<.0001) (<.0001) (0.0043)
Federal District 1.3** 4.4** 0.77**
(<.0001) (<.0001) (<.0001)
Predicted Income5 1.0** 1.0** 1.0**
(<.0001) (<.0001) (<.0001)
Poverty Status 0.76** 1.2** 0.95**
(<.0001) (<.0001) (<.0001)
% Concordance 81% 81% 75%
# of Observations 22,457,290 21,586,850 870,440
49. -48-
An interesting result is the insignificant effect of one’s predicted income on determining
child labor. Traditional labor theory would suggest an upward sloping labor supply curve. As
wages increase, the incentive to work increases and labor supply increases. However, traditional
theory does not apply to the market for child labor. In all three logistic models, the point
estimate for a child’s predicted wage was approximately 1.0, indicating no effect. This means
that a child’s wage does not function as an incentive to work, as theory would suggest. A
number of trends in child labor can help to explain this result. First, a large proportion of
children work unpaid jobs, which include jobs in the informal sector, family businesses, and
domestic work at home. Secondly, the insignificant effect of predicted wage on childhood labor
can be explained by framing child labor as something undesirable-an inferior good. The strong,
positive relationship between poverty incidence and child labor suggests that child labor is a
necessity for low-income families to alleviate short term poverty.
One of the most significant determinants of child labor and child schooling is family
income. This is unsurprising, due to the opportunity cost associated with sending a child to
school. The family must take into account the child’s foregone earnings or the value of domestic
labor done at home. This opportunity cost is relatively greater for low income families. To
analyze the impact of on a family’s decision to enroll a child in school or put him in the labor
force, we compare children belonging to poor and non-poor households. Utilizing data on
Brazil’s poverty lines from the World Bank, we define poor to be all individuals living in
households with a monthly per capita income of less than 150 reals. This equates to
approximately $40 per month. Our analysis shows a significant relationship between family
income and both child labor and schooling. First, children from poor households are significantly
less likely to attend school and more likely to be employed. These results support the trade-off
between labor and schooling. Table 5.3 shows that children from poor households are 1.31
times less likely to attend school than children who are not poor and 1.15 times more likely to
work while in school. Although, poor children are less likely to not attend school and work than
the children who are not poor. This has significant implications for school performance, which
we will further discuss later in our analysis.
Gender also functions as an important determinant of child labor and schooling. Table
5.3 shows that female children are less than half as likely to work as their male counterparts.
This can be attributed to two factors: 1) Wage differentials between male and female children
and 2) societal attitudes surrounding gender roles. Gender norms are a socially constructed set of
“rules” that define appropriate behavior and roles of men and women within a society. For
example, in Western culture, the traditional role of women surrounds motherhood and domestic
responsibilities within the home. These unequal attitudes surrounding gender originated
thousands of years ago but persist to this day, manifested in relatively low female labor force
participation rates (Alesina et. al). Societal gender roles clearly have an impact on child labor
force participation. Furthermore, because girls are less likely to have a job, they do not
experience the same trade-off between schooling and work as boys experience. Consequently,
female children are more likely to be enrolled and school and on average have higher educational
attainment relative to male children.
Living in a rural area, rather than an urbanized one, has significant effects of the school-
work choice. Those children that are living in rural areas are more than twice as likely to work
than those living in urban areas. This is mostly due to the nature of work in a rural area compared
to an urban area. Those children in urban areas are restricted in the extent that they can work,
they must get a job at a business or hawk their own wares. Those children in rural areas on the
50. -49-
other hand are not as restricted in their ability to work, as many more of them are able to and do
work on family farms and for their own consumption. This is reflected in the distribution of the
job types between rural children and urban children. Those children from rural areas make up
82% of children who are producing goods for their own consumption and 68% of children who
are unpaid (presumably working on a family farm), while only 18% of children are classified as
living in rural areas. Furthermore, children from rural areas are far more active, children from
rural areas are more like to work, more likely to go to school and more likely to attend school
while working.
5.4: Student Efficiency
In regards to the schooling system of Brazil an important factor to consider is the “school
for age” of Brazilian schools. School-for-age refers to the level of education that children have
attained in relation to the level of education that they ought to have attained at their age. In other
words, we examine how far ahead or behind Brazilian students are in their schooling.
In our study we created a simple definition for school-for-age, as shown in equation 5.5.
This equation will yield a number (normally between 1 and -4) that tells us how many years of
schooling each student has completed in relation to how many years of schooling he or she
should have completed. A positive number means that the student is ahead of where they should
be, a negative number means he or she is behind, and if school-for-age is zero the student has
completed exactly the amount of schooling he or she should have at that age.
𝜀𝜀 = 𝛼𝛼𝐹𝐹 − 𝛼𝛼𝐸𝐸 (Equation 5.2)
𝜀𝜀: School for Age
𝛼𝛼𝐹𝐹: Years of schooling student has finished
𝛼𝛼𝐸𝐸: Age of student – 6
We first calculated overall school-for-age for all children in Brazil, as shown by Figure
5.2. We chose to start at age six because schooling is mandatory at age six. While it ceases to be
mandatory after age fourteen, we did not stop until age sixteen because that is when secondary
schooling ends, and students choose whether to work or continue on to higher education. As the
graph shows, average efficiency is initially positive for children of age six, then immediately
becomes increasingly negative. By age sixteen, average school-for-age is down to almost
negative two and a half.
Afterwards, we calculated school-for-age for various sub populations, comparing the
averages by age for gender, race, geographic region, and employment status. Figure 5.3 shows
the average school-for-age of boys and girls. School-for-age for girls was more efficient on
average than for boys. The two genders begin to split at age 8, and the gap becomes increasingly
larger at every additional year old. The largest gap is at the end at age 16, with girls being ahead
of boys by almost a full point.
The results of the race calculations were what we had expected, with white leading black
and pardo by about half a point. Black and pardo were close together for all ages. Indigenous,
51. -50-
however, trails by quite a large margin. The furthest Indigenous fell behind was at age 15,
trailing behind black and pardo by almost two full points. These results are shown in Figure 5.4.
Figure 5.5 shows the average efficiency by geographic region. The regions are arranged
from highest returns to education to lowest returns to education, with South being the highest and
Northeast being the lowest. South, Central West, and Southeast are all roughly the same.
However the Northeast, unsurprisingly, fell behind the rest, ending about half a point lower than
the other three by age sixteen.
Finally we compared working children to non-working children, shown by Figure 5.6.
We only looked at ages thirteen to sixteen because the %age of working children was extremely
low before age thirteen. As expected working children are behind those who are solely attending
school, although not by as much as we had expected (ending at less than half a point at age
sixteen).
Figure 5.3
Overall School-For-Age
54. -53-
While comparing school-for-age gives a good picture of where students are compared to
where they should be for different characteristics, the relation is clearly a multivariate one. To
capture these effects, we ran two OLS regressions, the first being for general characteristics and
the second including household environment characteristics.
The restricted regression is shown in equation 5.6. The left side of Table 5.6 shows the
results of this regression. All coefficients were statistically significant to 1%. Employment had a
negative coefficient, meaning that a student was likely to have a lower school-for-age if they
were working. All other coefficients were positive, meaning they were positively correlated with
school-for-age. The strongest effect was that of working, while the weakest was the region
variable.
𝑌𝑌� = 𝛼𝛼 + 𝛽𝛽1 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹 + 𝛽𝛽2 𝑊𝑊𝑊𝑊𝑊𝑊𝑊𝑊 𝑊𝑊 + 𝛽𝛽3 𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 + 𝛽𝛽4 𝑊𝑊𝑊𝑊𝑊𝑊𝑊𝑊 (Equation 5.3)
Where:
𝛼𝛼: Intercept
FEMALE: 0 if Male, 1 if Female
WHITE: 0 if Nonwhite, 1 if White
REGION: Variable indicating returns to education of student’s region, higher values for
higher returns to education
WORK: 0 if Not Working, 1 if Working
In our expanded OLS regression we included several household environment variables,
as detailed in equation 5.7. The results from this regression are in the right half of Table 5.6. The
addition of the new variables did not dramatically change the estimates on any of the other
coefficients, with the exception of the effect of being white being reduced by about 50%. The
employment coefficient is again negative, along with the coefficients on age, age squared,
number of kids in family younger than 16, and sex of head household. These variables are
negatively correlated with school-for-age, while all other variables are positively correlated with
school-for-age. The strongest effect was that of age, while the weakest effect was that of
geographic region.
𝑌𝑌�� = 𝛼𝛼 + 𝛽𝛽𝛽𝛽𝛽𝛽𝛽𝛽 + 𝛽𝛽̂2 𝐴𝐴𝐺𝐺𝐺𝐺2
+ 𝛽𝛽3 𝐴𝐴𝐴𝐴𝐴𝐴3
+ 𝛽𝛽4 𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝐸𝐸16 + 𝛽𝛽5 𝐵𝐵𝐵𝐵𝐵𝐵𝐵𝐵ℎ𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂𝑂 +
𝛽𝛽6 𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝑑𝑑𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 + 𝛽𝛽7 𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝑒𝑒𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 + 𝛽𝛽8 𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝐻𝑑𝑑𝑆𝑆𝑆𝑆𝑆𝑆 + 𝛽𝛽9 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆
𝛼𝛼: Intercept (Equation 5.4)
AGE: Basic age variable
AGE2: AGE * AGE
AGE3: AGE * AGE * AGE
NumKidsLE_16: Number of children in family younger than 16
BirthOrder: Where child is in family’s birth order
HHHead_Skul: Schooling level of household head
HHSpouse_Skul: Schooling level of spouse of head of household
HHHead_Sex: 1 if female head of household, 0 if male
SingleParent: 0 if single parent, 1 if not