Impact of the Luz Para Todos program : an empirical paper analyzing the effect of electrification on child labor

(1)

MS

C

E

CONOMICS

T

RACK

:

D

EVELOPMENT

E

CONOMICS

M

ASTER

T

HESIS

Impact of the Luz Para Todos Program:

An Empirical Paper Analyzing the Effect of Electrification on

Child Labor

by

T

HUY-VY

N

GUYEN

11808535

August 12, 2018

Supervisor:

(2)

Statement of Originality

This document is written by Student Thuy-vy Nguyen who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

Abstract

This thesis uses data from the Pesquisa Nacional por Amostra de Domicílios (PNAD, “Brazilian National Household Sample Survey”) to analyze the impact of the Luz Para Todos program (LPT, “Light for All”) on child labor through its impacts on adult labor. The LPT program is a national Brazilian federal government program that launched in November 2003 with the initiative of ending electric exclusion across the country. The program was more specifically targeted to reach rural and peri-urban areas. The goal was to increase access to electricity through grid expansion, distributed generating systems with isolated grids and stand-alone systems. Two methodologies are utilized in this paper to estimate the effect: difference-in-differences and fixed effects at two geographical units: state and municipality. I find that an increase in the average access to electricity in a state has a positive effect on adult income while having a negative significant effect on the percentage of working children in the state.

(4)

Table of Contents 1. Introduction ... 5 1.1 Literature Review ... 5 1.1.1 Socio-economic Outcomes ... 5 1.1.2 Employment outcomes ... 6 1.2 Thesis Outline ... 8

2. Background and Context: Luz Para Todos (LPT) Program ... 9

3. Data ... 10

3.1 Pesquisa Nacional por Amostra de Domicílios (PNAD) ... 10

3.1.1 State Selection ... 11

3.1.2 Municipality Selection ... 12

3.1.3 Child Labor Measure ... 14

4. Methodology ... 15

4.1 Difference-in-Differences ... 15

4.2 Two-way Fixed Effects ... 21

5. Results ... 22

5.1 Adult Labor Outcomes ... 22

5.2 Child Labors ... 29

5.3 Robustness Check ... 32

5.3.1 Programa de Aceleração do Crescimento (PAC) ... 33

5.3.2 Results ... 33

6. Discussion ... 38

7. Conclusion ... 39

8. Appendix ... 40

8.1 Municipality Coding ... 40

8.2 Estimated Propensity Scores ... 41

8.3 School Attendance ... 42

(5)

1. Introduction

As one in five people worldwide lack access to modern electricity, it is imperative to work towards solutions to achieve universal access. Many households still rely on alternative forms of electricity such as firewood, oil and oil products. In general, the areas that suffer the most from lack of electricity access are the rural, more isolated regions. In 2000, the Brazilian Institute of Geography and Statistics (IBGE) reported roughly 13 million people did not have access to electricity. Because electricity is undoubtedly a necessity for socio-economic development, the Brazilian government has invested greatly in efforts to achieve universal electricity access. Luz Para Todos (LPT, “Light for All”), Brazil’s universal electricity program, launched in 2003 has reached over 15 million individuals (da Silveira Bezerra, 2017).

1.1 Literature Review

There is considerable literature on the estimation of the effect of rural electrification programs in developing countries. A variety of socio-economic indicators as well as employment outcomes have been studied. I will briefly discuss the socio-economic outcomes that have been studied in relation to electrification programs, but will mainly focus on the studies on employment outcomes which are most relevant to my research.

1.1.1 Socio-economic Outcomes

Socio-economic outcomes that have been studied include energy poverty, the Municipal Human Development Index (MHDI) and fertility. Pereira et al. (2010) evaluate the impact of the LPT program on energy poverty. The authors used self-formulated questionnaires and surveyed a universally representative sample during two different time periods: April and May 2003, before the launch of the program, and November and December 2004, a year after the launch of the program. Using these surveys, they create and compare a sample of “electrified” households, that covered households that would be provided access due to the program, to a control sample of those dwellings that would not be serviced during this period. Using data from Eletrobras, an electric power holding company, and Cepel, an electric energy research center, the authors generated a set of poverty indicators: poverty line, poverty gap, gap quadratic, gini coefficient,

(6)

Lorenz curve and Sen index and apply a difference-on-differences approach. The authors find that rural electrification significantly reduced the energy poverty level and as a result, improved energy equity. This paper utilizes data that adds to the precision of the analysis their through questionnaires because it includes information on which dwellings would be affected or not, rather than instrumenting or assuming about program take up like many other studies.

Similarly, Borges et. al. (2017) use a fixed effects model to test the effect the LPT program has on the three dimensions of the Municipal Human Development Index (MHDI): income, education and health. The authors use Brazilian population data, data from the Atlas of Human Development in Brazil and a constructed panel at the municipality level. The authors decided to use a fixed effects specification rather than a random effects model based off of the results of a Hausman test. The authors find that electrification has a positive effect on all dimensions of the MHDI, with the education component having the strongest effect. Their results show that electricity access is a major requirement to improve quality of life, however increasing access was more effective in higher HDI regions.

1.1.2 Employment outcomes

There are many other studies that focus on the effect of rural electrification on employment. Dasso and Fernandez (2015) study the effects of a rural electrification program in Peru on individual-level labor market outcomes. The program, Plan Nacional de Electrification Rural (PER), is comprised of many projects. The authors study the outcomes: labor force participation, employment status, hours of work, earnings, hourly wages, number of jobs, and whether an individual is a wage-earner, working in the agriculture sector, or self-employed. Program placement was not random; districts were chosen based off of initial electrification rates, poverty rates, subsidies and costs per connection, and use of renewable energy. Using repeated cross-section data from an annual representative household survey that includes data on labor market outcomes and individual characteristics, the authors use a difference-in-differences approach for the years 2006 to 2012. The authors also employ a fixed effects method for the years 2007 to 2010 using a unique household panel data set from the National Office of Statistics. Dasso and Fernandez include district, time and individual fixed effects. They use data

(7)

on projects that concluded in the 2007 to 2010 time period. They find that males are less likely to have a second job and work more hours, whereas females are more likely to have a job and have higher hours of work. The authors also find that electrification promotes formulation on microenterprises and suggest that electricity increases available time to work, and that economies of scale increase the payoff of focusing on one occupation.

Because the authors have included individual fixed effects, they are not able to include time-invariant individual characteristics in their model, which are included in their difference-in-differences model. While individual fixed effects account for some heterogeneity among individuals, characteristics such as sex and gender are significantly influential on employment outcomes, so this may impact estimates. Additionally, in the difference-in-differences equation, the indicator variable for treatment is defined as 1 if a program was implemented in a district in year t, and 0 otherwise, whereas in the fixed effects model, the indicator variable is defined as 1 if an individual lives in a district where at least one project had been concluded by year t. The different definitions for the treatment indicator variable leads me to believe that these estimates are not truly comparable.

In another paper that studies employment outcomes, Akpandjar and Kitchens (2017) evaluate the National Electrification Program (NEP) in Ghana. More specifically, the authors focus on the impact of residential electrification on employment and the household composition. Many of the outcome variables estimated in this paper are similar to those in Dasso and Fernandez (2015). They include: whether the head of household runs a small business, is employed in the agricultural sector, unemployment status, occupational score, whether they are a wage-earner or self-employed whether wood fuel is used in the home, whether the home has an electric stove, the number of children under 5 years old in the household, and child employment. They do so by using repeated cross-sectional census data. Because the program initially targeted district capitals, individuals were located using their reported district-level code and chosen based on their reported district code. The authors employ a fixed effects specification controlling for time and district fixed effects. With respect to household composition, the authors find that families with electricity had reduced fertility and invested more in their children’s human capital. Additionally, they find that access to electricity leads to decreases in agricultural employment,

(8)

increases in wage-earning occupations rather than self-employment, decreases in the fraction of working children, and higher incidences of (nonagricultural) small business formation. The authors find that access to electricity causes a shift in the labor market job composition – individual occupational scores increase by 2.3 points which is equivalent to an individual moving from being a self-employed farmer to a semi-skilled machine operator or retail sales employee. Akpandjar and Kitchens (2017) add to the credibility of their analysis by calculating sensitivity ratios1_{to test whether their specification suffers from great omitted variable bias. The}

sensitivity ratio tests the severity of selection on unobservable characteristics. Following a method from Altonji et al. (2005) and Bellows and Miguel (2009), they compare their main regression results to a regression that includes a fully interacted set of covariates. Because they find that the sensitivity ratio on their outcome variables exceeds 1, their model does not suffer from omitted variable bias.

1.2 Thesis Outline

Other than electrification, another social issue that the Brazilian government places emphasis on is child labor. Although there are laws in place to combat child labor, children in Brazil still engage in some of the worst forms of child labor: production of coffee and commercial sexual exploitation. In Brazil, the minimum working age is sixteen for most work and eighteen for hazardous work. According to an analysis done by the United States Department of Labor, the 2015 Brazilian National Household Survey (PNAD) found that there were more than 400 thousand child laborers in Brazil, although there has been a downward trend. Their analysis uses children aged 5 to 13. Undoubtedly, involvement in work as a child can have lasting effects on the health and education development of children. Children who are economically active are likely to work in hazardous conditions and are less likely to attend school full-time, or even part-time (DOL, 2017). In hopes of overcoming this development challenge, the Brazilian government also funds and participates in many programs that aim to eliminate or prevent child labor.

1_{The sensitivity ratio is given by:}𝛼"#

(𝛼"%− 𝛼"#)

( where 𝛼"% is the coefficient on the outcome variable in the main regression

(without full interacted covariates) and 𝛼"#is the coefficient on the outcome variable in the regression including the interacted

(9)

The purpose of this thesis is to unify the two social issues in Brazil: electricity access and child labor. I evaluate the causal relationship between the LPT program and child labor. Although the universal electrification program was not formulated specifically to fight child labor, it is a social program that could potentially have lasting effects on the livelihoods of Brazilians, thus impacting their economic activity. Access to electricity could potentially release people from hard work, increase productive working hours and provide more opportunities for nonfarm self-employment. Electrification is expected to provide a channel through which jobs are created and incomes increase resulting in welfare improvements so that kinds do not have to work.

This relationship will be assessed in two stages. First, considering adult labor – do adults earn more income as a result of this program, and do they work more hours? Secondly, analyzing changes in the percentage of children working within a state to determine if there is a substitution effect occurring between adult labor and child labor. Two assumptions are crucial for this analysis: the luxury axiom and substitution axiom. The luxury axiom, in this context, states that a family will send a child to work only if the household requires an additional source of income for survival. The substitution axiom states that, from a firm’s point of view, adult labor and child labor are substitutes (Basu and Van, 1998). I theorize that the electrification program will cause adults to become more productive – electrification will free additional hours in the day from chores and other time-consuming activities, thus hours of work will increase and therefore income will increase. Because much of Brazil’s child labor is fueled by poverty, under the assumptions of the luxury and substitution axioms, I hypothesize there will be a negative effect on child labor.

This paper is structured as follows: Section 2 provides background and context. Section 3 describes the data. Section 4 introduces the methodologies used for this analysis. Section 5 presents results. Section 6 includes a discussion and Section 7 concludes.

2. Background and Context: Luz Para Todos (LPT) Program

In Brazil’s effort to achieve universal electrification, the LPT program was launched by the federal government in November 2003. LPT places an obligation on electricity distribution

(10)

companies to provide electricity to those without power within their concession area. The distribution companies must provide electricity to every citizen that requests service. While the LPT program is nationwide program, the government placed priority on certain municipalities based on social welfare parameters such as their electricity access coverage and human development index in 2000 – there was no definitive roll-out scheme for the LPT program (da Silveira Bezerra, 2017).

As a first step of the program, all electricity companies were asked to develop an energy access strategy within their coverage area – a Universalization Plan. Providers were not restricted in their method to connect people. They could opt to extend their network, create decentralized generation systems with isolated networks or use individual generation systems. LPT agents were used in the process of implementing the program to identify new electricity demand by working close with local communities. Unlike many social programs, the resources to develop the new infrastructure come from federal and state government funds and distribution companies’ own finances – beneficiaries do not contribute financially (IEA, 2017).

The initial goal for this social policy was to reach 12 million unconnected people by 2008. Because the program goal was not met within the initial program time, 2003 to 2008, the implementation of additional phases of the program was necessary. The program was extended a total of four times. By May 2016, over 15.6 million people had been reached, however, the project has been extended to 2018 to reach those still unserved, mainly in the Amazon region (da Silveira Bezerra, 2017).

3. Data

3.1 Pesquisa Nacional por Amostra de Domicílios (PNAD)

The microdata used in this analysis is the PNAD (Brazilian National Household Sample Survey) for the years: 2001 to 2008 and 2011 to 2015. These are the years currently published on the IBGE website. The PNAD is an ongoing annual national survey conducted at the household and individual level by the IBGE. The PNAD provides data on several characteristics of the population including: household composition, education, labor, income, education, health,

(11)

professional training and food security. Each year, the PNAD surveys over 100 thousand households and well over 400 thousand individuals. The formulation of this survey focuses on the ability to evaluate policies oriented to socio-economic development and the improvement of living conditions in Brazil (Datazoom, 2011). I merged the household-level data and individual-level data for the 13 available years of the PNAD to do this analysis.

The dependent variables of interest in this analysis are: adult income, average weekly adult working hours from all jobs, and child labor. I restrict the survey to persons aged 5 to 80. Although the retirement age is 65, people often work at least part-time after retirement. According to the International Labor Organization, child laborers are those working between the ages of 5 and 17 (ILO). However, because the legal working age in Brazil is 16, I define children to be those aged 5 to 15 and adults to be those aged 16 and up.

3.1.1 State Selection

Although the PNAD is a nationally representative survey with data for all 27 states of Brazil, only some are included in this analysis. Since there is no official data on the states that took part in the LPT program, it is necessary to identify and select states that were served by the program based on the variation and trends in electricity access. For this analysis, I selected states that had an initial electrification rate below 85 percent with similar electrification trends over the 2001 to 2015 period and states that were already at full electricity access at baseline. I use 85 percent because this was a criteria outlined in program to establish priority to certain municipalities. This analysis will focus specifically on the selected states: Rio de Janeiro, Sao Paolo, Santa Catarina, Distrito Federal, Tocantins, Maranhao and Piaui. For the purposes of the traditional difference-in-differences analysis, these states are aggregated into control and treatment groups. The selection process is described below.

Electricity access is determined by what households reported as their “electricity used for lighting.” Households were categorized as having access under the assumption that if they responded “electrical,” they have access, and no access otherwise. Other responses included not applicable, oil/kerosene/canister gas and other. An average access to electricity variable was

(12)

then calculated per state. Rio de Janeiro, Sao Paolo, Santa Catarina and Distrito Federal were chosen as the ‘control’ group as they were the states with already high (almost 100 percent) electricity access in the pre-program period, 2001 and 2002. Tocantins, Maranhao and Piaui were aggregated to create the ‘treatment’ group as they started at a much lower initial average access, roughly 80 percent. Figure 1 depicts the differences in electricity access levels in the treatment and control group for the full sample.

Figure 1: State level: Average Access to Electricity (%)

3.1.2 Municipality Selection

In order to increase the precision of my analysis, it is possible to analyze the causal relationship at a smaller geographical unit. Analyzing at the municipality level rather than the state level would increase variation between groups, but decrease variation within groups and could affect the point estimates. Because program priority was placed on municipalities’ characteristics, estimates at this level are more relevant. There are some downsides to doing the analysis at this level, mostly pertaining to data. I do not have information on the names of the municipalities, only the state they are in. However, I am able to calculate their “municipality codes” and number the unique values to see which municipalities are reported on in each year2_{. Without}

information on methodology about changes in municipalities and coding, it is only possible to

2_{See Appendix for formulation of unique municipalities.}

(13)

compare the years 2001 through 2008 or 2011 through 20153_{. For this analysis, it makes sense to}

focus on the former time period, restricting the sample size to 8 years of data. Within the seven states selected for this analysis, there are 224 municipalities reported on in 2001 to 2008. Municipalities will be sorted into treatment and control groups based on the average access to electricity at baseline, 2001. Those municipalities with 80 percent or lower will be grouped into the treatment group – others will be grouped into the control group. Figure 2 depicts the differences in electricity access levels in the treatment and control group for 2001 to 2008.

Figure 2: Municipality Level: Average Access to Electricity (%)

Similar to Figure 1, Figure 2 shows that there is an upward trend in electricity access in the treatment group. In the municipality grouping, initial access in the treatment group starts significantly below the state grouping. This makes sense because, in some cases, municipalities may have had an initial electrification rate well below 80 percent, but when aggregated at the state level, these much lower values are averaged out by municipalities in that state that had a high or full access level. There is more within-state variation, which is shown through the differences in access levels between municipalities, than there is between states. Therefore, the analysis at the municipality level should hone more accurate results. For the purpose of this thesis, I include results for both geographical units.

3_{The same municipalities are reported on for this time period, whereas in the 2011 to 2015 time period, the municipality coding}

(14)

3.1.3 Child Labor Measure

The outcome variables to assess adult labor are derived directly from existing survey variables, with few data transformations. However, to measure child labor, I generated a new variable, percentage of working children. The percentage of working children is calculated based on people aged 5 to 15 who reported working hours in the PNAD survey. The variable is estimated first by categorizing each child in the survey as a child laborer or not based on their survey response – equals 0 if they do not report hours and 1 if they do report hours. The percentage of working children is calculated based on that categorization at the state and municipality-level and then aggregated to calculate figures for the treatment and control groups. Figures 3 and 4 depict the difference in Child Labor trends between the treatment and control groups for both samples.

Figure 3: State Level: Percentage of Working Children

Figure 4: Municipality Level: Percentage of Working Children

(15)

4. Methodology

This section introduces the two methodological approaches used to investigate the impact of LPT on both adult labor and child labor. The methodology in Galiani et. al. (2005) on water privatization in Argentina is used as a basis for the structural modeling of this analysis. This analysis will be done at both the state-level as well as the municipality-level, however, due to data restrictions, the analyses will be done on different samples which are defined below. Because employment composition is likely correlated within local labor markets, I consider clustering standard errors at the state and municipality-level. Because the state geographical unit is so large, it is not likely to make a significant difference; therefore I use robust standard errors in the state-level analysis. However, at the municipality-level, the unit is much narrower and more likely to influence results, so they are clustered into 224 clusters for each of the reported municipalities in that sample.

First, I use repeated cross-sections from the PNAD to adopt a difference-in-differences approach based on state and municipality selection into treatment and control. Secondly, I use a two-way fixed effects specification to assess the marginal effect of increasing electricity access by using the average access to electricity in a state as a continuous treatment variable. I also perform this analysis at the municipality level with the restricted sample. These specifications capture the aggregate time trend as well as state/municipality-level differences. The fixed effects model is more suitable for the data I have and provides more insightful findings about the marginal effect than the difference-in-differences approach. Overall, tikkuhe marginal effect would be more useful for policy recommendations and evaluation.

4.1 Difference-in-Differences

To calculate an appropriate counterfactual to evaluate the causal effect of the LPT program on child labor, I use a difference-in-differences approach. Difference-in-differences is a common statistical technique used in program evaluation. It allows you to compare treatment groups with control groups pre- and post-intervention. For both the state and municipality-level analysis, the pre-treatment period is 2001 and 2002, the two available datasets prior to the implementation of the LPT program. Since the program was launched in November 2003, the

(16)

post-intervention period refers to the years 2003 to 2008 and 2011 to 2015 for the state-level analysis and only 2003 to 2008 for the municipality-level. The difference-in-differences approach removes biases in post-intervention period comparisons between the treatment and control group that could be the result of systematic differences between those groups. Treatment, in this case, is simplified to whether the household is in a treatment state/municipality, as specified in section 3.1.1 and 3.1.2, or not, and whether the time period is pre- or post-program.

The key identifying assumption of a difference-in-differences analysis is that the treatment and control group have parallel trends in absence of the treatment. Notice the pre-treatment trends in Figure 3 and 4 for the main outcome variable, Percentage of Working Children, above. I acknowledge that the trends in percentage of working children for the treatment and control groups are not perfectly parallel. Because this analysis is based on observational data rather than experimental data and not perfectly randomized, there is likely to be treatment selection bias which makes it difficult to compare groups. In an effort to account for the imperfect parallel trend, I tested a propensity score matching approach.

The key underlying assumption for propensity score matching is that, if selection bias is eliminated by matching on observable characteristics, matching on propensity scores, a function of observable characteristics, would also eliminate any present selection bias. Table 1 and 2 present the baseline statistics for adults and children by control and treatment groups at the state-level. Table 3 and 4 present the baseline statistics for adults and children by treatment and control groups as defined at the municipality-level. As shown in the tables, there are significant differences between the control and treatment groups in most of the baseline characteristics for both geographical units. All outcome variables are significantly different at baseline.

(17)

Table 1: Baseline Descriptive Statistics, State-level: Adult Population

Variable Control Treatment Difference

Avg. Access to Electricity (%) 99.614 81.433 18.181*** (0.000) (0.004) (0.002) Gender 0.527 0.507 0.020***

(0.002) (0.005) (0.005) Years of Education 7.506 4.874 2.632***

(0.017) (0.039) (0.043) Age Started Working 14.729 12.390 2.339***

(0.025) (0.054) (0.061)

Urban 0.938 0.685 0.253***

(0.001) (0.004) (0.003) Mother in Household 0.364 0.368 -0.003

(0.002) (0.005) (0.006) Number of Sleeping Rooms 2.085 2.330 -0.245***

(0.003) (0.009) (0.009) Ln(income) 7.076 6.115 0.961***

(0.005) (0.013) (0.013) Avg. Weekly Hours Worked 43.573 40.456 3.117***

(0.069) (0.190) (0.177)

This table includes the results of t-tests on the equality of means between the control and treatment groups. Means are calculated based on the adult population (persons aged 16 to 80) of the 2001 PNAD survey. Robust standard errors are in parentheses, Two-tailed test. * p < 0.1, ** p < 0.05, *** p < 0.01

(18)

Table 2: Baseline Descriptive Statistics State-level: Child Population

Avg. Access to Electricity (%) 99.499 77.363 22.136*** (0.001) (0.006) (0.177) Gender 0.494 0.496 -0.002 (0.002) (0.005) (0.005) Urban 0.924 0.629 0.294*** (0.002) (0.007) (0.005) Mother in Household 0.935 0.864 0.071*** (0.002) (0.005) (0.004) Number of Sleeping Rooms 2.169 2.466 -0.297***

(0.006) (0.013) (0.013) Working Children (%) 3.851 15.335 -11.484***

(0.017) (0.036) (0.037) Attends 4+ Hours of School 0.611 0.185 0.426***

(0.004) (0.006) (0.008) This table includes the results of t-tests on the equality of means between the control and treatment groups. Means are calculated based on the child population (persons aged 5 to 15) of the 2001 PNAD survey. Robust standard errors are in parentheses, Two-tailed test. * p < 0.1, ** p < 0.05, *** p < 0.01

Table 3: Baseline Descriptive Statistics, Municipality-level: Adult Population

Avg. Access to Electricity (%) 97.614 80.152 17.462*** (0.022) (0.010) (0.091)

Gender 0.526 0.479 0.041***

(0.002) (0.008) (0.008) Years of Education 7.319 3.428 3.890***

(0.016) (0.056) (0.069) Age Started Working 14.544 11.348 3.196***

(0.024) (0.074) (0.092)

Urban 0.927 0.424 0.503***

(0.001) (0.008) (0.004) Mother in Household 0.365 0.354 0.011

(0.002) (0.009) (0.009) Number of Sleeping Rooms 2.112 2.289 -0.177***

(0.003) (0.015) (0.014) Ln(income) 6.998 5.786 1.212***

(0.005) (0.013) (0.022) Avg. Weekly Hours Worked 43.363 38.648 4.715***

(0.067) (0.312) (0.268) This table includes the results of t-tests on the equality of means between the control and treatment groups. Means are calculated based on the adult population (persons aged 16 to 80) of the 2001 PNAD survey. Robust standard errors are in parentheses, Two-tailed test. * p < 0.1, ** p < 0.05, *** p < 0.01

(19)

Table 4: Baseline Descriptive Statistics Municipality -level: Child Population

Avg. Access to Electricity

(%) 99.499 77.363 16.714*** (0.001) (0.006) (0.151) Gender 0.496 0.486 0.010 (0.004) (0.011) (0.012) Urban 0.906 0.390 0.294*** (0.002) (0.011) (0.005) Mother in Household 0.923 0.885 0.038*** (0.002) (0.007) (0.006) Number of Sleeping Rooms 2.208 2.482 -0.274***

(0.006) (0.020) (0.019) Working Children (%) 5.502 14.753 -9.251***

(0.033) (0.055) (0.107) Attends 4+ Hours of School 0.563 0.067 0.496***

(0.004) (0.006) (0.012) This table includes the results of t-tests on the equality of means between the control and treatment groups. Means are calculated based on the child population (persons aged 5 to 15) of the 2001 PNAD survey. Robust standard errors are in parentheses, Two-tailed test. * p < 0.1, ** p < 0.05, *** p < 0.01

Instead of using sensitivity ratios as in Akpandjar and Kitchens (2017) to detect selection bias, I estimate equation (1):

Pr (𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 = 1) = 𝛽₆+ 𝛿𝑿_:;<= _{+ 𝜀}

:;< (1)

I use the following time-invariant individual and household characteristics to estimate propensity scores: gender, race, age started working, and years of education as well as urban/rural status, number of sleeping rooms in the house, and whether the mother is living in the house, given by 𝛿𝑿:;<= . This equation estimates the probability of being in the treatment

group based on the vector of controls. Using these variables, I was unable to create a proper control group with enough common support, therefore propensity score matching is not used in the remaining parts of this analysis and for the purpose of this thesis, I assume a parallel trend. See Appendix Figure (5) and (6) for the plot of the propensity score distributions for both samples.

(20)

To perform the difference-in-differences analysis, I estimate the following regression model, Equation (2), for the adult employment outcome variables at both geographical units:

𝑌_:;= 𝛽₆+ 𝛽_@𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡_:+ 𝛽_A𝑇𝑖𝑚𝑒_;+ 𝛽_C(𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 ∙ 𝑇𝑖𝑚𝑒)_:;+ 𝜂𝑿_:;<= _{+ 𝜀}

:; (2)

where 𝑌:; refers to the outcome variables: log of adult income and average weekly working hours

of adults for individual 𝑖 at time 𝑡. 𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡: equals 1 if the individual is in a treatment state

or municipality and 0 if not, 𝑇𝑖𝑚𝑒; equals 1 if the observation is in the post-treatment period

(2003 to 2015 at the state-level and 2003 to 2008 at the municipality level) and 0 if it is in the pre-treatment period (2001 and 2002), 𝑿:;<= is the vector of control variables for adults and 𝜀:; is

an error term. In this model, the coefficient, 𝛽C, on the interaction term is the

difference-in-differences estimator of the causal effect.

For the regression of the treatment on the percentage of working children in a state, I estimate Equation (3) at both geographical units:

𝐶_;< = 𝛽₆+ 𝛽_@𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡_:+ 𝛽_A𝑇𝑖𝑚𝑒_;+ 𝛽_C(𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡 ∙ 𝑇𝑖𝑚𝑒)_;:+ 𝜑𝑿_:;<H + 𝜀_;<₍₃₎ where 𝐶;< refers to the percentage of working children in a given state/municipality. The

aggregated state-level value is given to every individual in that state. 𝑇𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡: equals 1 if the

individual is in a treatment state or municipality and 0 if not, 𝑇𝑖𝑚𝑒; equals 1 if the observation

is in the post-treatment period and 0 if it is in the pre-treatment period (same as defined above), 𝑿:;<H 4 is a vector of control variables for children and 𝜀;: is an error term. Years of

education and age started working were omitted from the regressions pertaining to child labor because these two variables are likely to be influenced by the employment status of a child. Similarly, the coefficient, 𝛽C, on the interaction term is the difference-in-differences estimator of

the causal effect. Note the pre-treatment periods are the same for both levels of analysis, 2001 and 2002. However, post-treatment differs in that for the municipality-level analysis, 2011 to 2015 are left out.

4_{The vector includes: gender, race, household’s urban/rural status, number of sleeping rooms in the house, and whether the mother}

(21)

4.2 Two-way Fixed Effects

To deepen the analysis, I test another method. I employ a two-way fixed effects specification. This model controls for systematic differences between states/municipalities, time trends as well as presents the marginal effect of increasing access to electricity. With fixed effects estimation, you can either demean each variable across time, or add dummy variables to each group. The best suitable option for my data, because I do not have proper panel data, is to create dummy variables for each state/municipality and year. Using a fixed effects approach greatly reduces the chance of omitted variable bias in the regression results. Formally, I do so by estimating Equation (4) and (5):

𝑊_:;= 𝛽₆+ 𝛼_<+ 𝜆_;+ 𝜎𝐴𝑐𝑐𝑒𝑠𝑠_<;+ 𝜌𝑿_:;<= _{+ 𝜀}

:;< (4)

where 𝑊:; refers to the outcome variables: adult income and average weekly working hours of

adults for individual 𝑖 at time 𝑡. 𝜆; is a vector of year dummies that control for temporal

variation in the outcome variables. 𝛼<, a vector of state dummies is included to control for

systematic differences between states/municipalities, 𝑿:;<= is a vector of control variables and 𝜀:;

is an error term. Although the LPT program is a national program, many areas were affected more than others because priority was placed on areas with the lowest access and lowest human development index. This results in different program intensities in each state. Rather than evaluating the LPT program as a strict binary treatment, I use percent access to electricity as a proxy for the LPT program intensity. This variable is given by 𝐴𝑐𝑐𝑒𝑠𝑠<; – the average access to

electricity at time 𝑡, in state/municipality 𝑠. This average access to electricity is given to every individual in that given state at time 𝑡.

Equation (5) is estimated to assess the relationship between child labor and access to electricity in a two-way fixed effects specification:

𝐺;< = 𝛽6+ 𝛼<+ 𝜆; + 𝜔𝐴𝑐𝑐𝑒𝑠𝑠<;+ 𝛾𝑿:;<H + 𝜀:;< (5)

where 𝐺;< refers to the outcome variable, percentage of working children, at time 𝑡, in

state/municipality 𝑠. This percentage is given to every individual in that state. 𝜆; is a vector of

(22)

error term. Again, the variable 𝐴𝑐𝑐𝑒𝑠𝑠<; represents for the average access to electricity at time 𝑡,

in state/municipality 𝑠.

5. Results

5.1 Adult Labor Outcomes

I estimate equation (2) and (4) for the two adult labor outcome variables, log income and average weekly hours worked, using only the adult subpopulation. Table 5 shows the regression results of the impact of increased electricity access on adult income for both the difference-in-differences and fixed effects models at the state grouping level. Column (1) and (2) report results for the difference-in-differences analysis. Without including any controls, the results indicate that being in a treatment state significantly increases income by 17.2 percent. When controlling for race, gender, years of education, age when started working, household urban status, whether the mother lives in the household, and the number of sleeping rooms, the estimate drops to an increase of 9.7 percent. Columns (3) and (4) display the results for the fixed effects specification. Column (3) is inclusive of state and year dummies, but exclusive of additional controls. The results show that a 1 percentage point increase in access to electricity increases income by 1.4 percent. With controls, the estimated impact of a 1 percentage point increase in access to electricity falls to a 0.9 percent increase in income. Scaling up the marginal estimate from the fixed effects model, the average effect is 26.0 percent and 16.7 percent, without and with controls respectively5_{. Compared to the 17.2 percent and 9.7 percent, the fixed}

effects estimates seem large, but is likely more accurate of a treatment effect because it accounts for time trends as well as systematic differences between the states. Additionally, because there is a large number of observations in the study, my results are significant at the 1 percent level. Table 6 shows results for the same regression as in Table 5 but at the municipality-level. The estimates at the municipality level are lower than at the state-level for both of the specifications – more significant in the difference-in-differences analysis. The results show that solely being in a treatment municipality, without controlling for individual and household characteristics,

5_{The difference in average electricity access between 2001 (baseline) and 2015 is 18.6 percent. By multiplying this difference with}

(23)

increases income by 4.5 percent. With controls, the results show a decrease of 5.6 percent. Neither of these results are significant. The results for the fixed effects specification are much closer – without controls, a 1 percentage point increase in access to electricity increases income by 1.3 percent. With controls, the estimated impact of a 1 percentage point increase in access to electricity falls to a 0.7 percent increase in income. Scaling up the marginal estimate from the fixed effects model, the average effect is 24.2 percent and 13.0 percent, without and with controls respectively.

(24)

Table 5: State-level: Impact of Electrification on Adult Income

Diff-in-Diff Fixed Effects

(1) (2) (3) (4) Treatment (=1) -0.999*** -0.594*** (0.010) (0.009) Time (=1) 0.627*** 0.505*** (0.004) (0.003) Treatment x Time 0.172*** 0.097*** (0.011) (0.010)

Avg. Access to Electricity (%) 0.014*** 0.009*** (0.001) (0.000) Race White 0.037** 0.121*** (0.027) (0.025) Black -0.081** -0.075** (0.027) (0.025) Asian 0.267*** 0.345*** (0.031) (0.029) Multiracial -0.077*** -0.061** (0.027) (0.025) Gender (=1) -0.458*** -0.450*** (0.002) (0.002) Years of Education 0.118*** 0.105*** (0.000) (0.000) Age Started Working 0.012*** 0.009***

(0.000) (0.000) Urban (=1) 0.233*** 0.235*** (0.005) (0.004) Mother in Household (=1) -0.526*** -0.493*** (0.003) (0.002) # of Sleeping Rooms 0.081*** 0.085*** (0.002) (0.001)

Controls No Yes No Yes

State Dummies No No Yes Yes

Year Dummies No No Yes Yes

N 600,332 447,548 600,332 447,548 Using data from the 2001 – 2008, 2011 – 2015 PNAD survey: Columns (1) and (2) show results for a traditional Diff-in-Diff where ‘Treatment’ refers to whether an individual is in a treatment state or not. Columns (3) and (4) show estimates for a fixed effects specification using a continuous treatment variable. Robust standard errors in parenthesis.

(25)

Table 6: Municipality-level: Impact of Electrification on Adult Income

(1) (2) (3) (4) Treatment (=1) -1.285*** -0.631*** (0.120) (0.074) Time (=1) 0.010 -0.056*** (0.015) (0.011) Treatment x Time 0.045 -0.028 (0.046) (0.038)

Avg. Access to Electricity (%) 0.013*** 0.007*** (0.002) (0.002) Race White 0.107** 0.175*** (0.050) (0.044) Black -0.105** -0.029 (0.048) (0.042) Asian 0.349*** 0.380*** (0.112) (0.092) Multiracial -0.133*** -0.009 (0.046) (0.038) Gender (=1) -0.475*** -0.470*** (0.017) (0.014) Years of Education 0.115*** 0.106*** (0.006) (0.006) Age Started Working 0.008*** 0.004***

(0.001) (0.001) Urban (=1) 0.245*** 0.125*** (0.035) (0.018) Mother in Household (=1) -0.526*** -0.529*** (0.003) (0.019) # of Sleeping Rooms 0.083*** 0.099*** (0.012) (0.004)

Municipality Dummies No No Yes Yes

N 371,766 277,032 600,332 277,032

Using data from the 2001 – 2008 PNAD survey: Columns (1) and (2) show results for a traditional Diff-in-Diff where ‘Treatment’ refers to whether an individual is in a treatment state or not. Columns (3) and (4) show estimates for a fixed effects specification using a continuous treatment variable. Clustered standard errors in parenthesis – 224 clusters for municipalities. * p < 0.1, ** p < 0.05, *** p < 0.01

(26)

Table 7 shows the results of the impact of increased electricity access on average weekly adult working hours for both the difference-in-differences and fixed effects models at the state-level. Column (1), not including controls, indicates that being in a treatment state significantly decreases average weekly working hours by 1.5 hours. When controlling for race, gender, years of education, age started working, household urban status, whether the mother lives in the household, and the number of sleeping rooms, the estimate remains unchanged – results are shown in Column (2). Columns (3) and (4) display the results for the fixed effects specification. Without controls, I find that a 1 percentage point increase in the state’s average electricity access rate decreases average weekly working hours by 0.112 hours. When the same controls are added, the estimated impact of a 1 percentage point increase in access to electricity is a 0.115 hour decrease in working hours. Results are significant at the 1 percent level. When scaling of the marginal results, I find that without controls, being treated, on average, decreases work hours by 2.1 hours and with controls, a decrease of 2.1 hours6_{. This estimate is not far from the}

1.5 hours found in the difference-in-differences analysis.

Table 8 shows the results of the impact of increased electricity access on average weekly adult working hours for both the difference-in-differences and fixed effects models at the municipality-level. Column (1) indicates that being in a treatment state significantly decreases average weekly working hours by 1.4 hours. When controlling for individual and household characteristics, the estimate decreases minimally, essentially unchanged. Without controls in the fixed effects specification, I find that a 1 percentage point increase in the municipality’s average electricity access rate decreases average weekly working hours by 0.162 hours. When the same controls are added, the estimated impact of a 1 percentage point increase in access to electricity is a 0.151 hour decrease in working hours. Results are significant at the 1 percent level. When scaling of the marginal results, I find that without controls, being treated, on average, decreases work hours by 3.0 hours and with controls, a decrease of 2.8 hours7_.

(27)

Table 7: State-level: Impact of Electrification on Average Adult Working Hours

(1) (2) (3) (4) Treatment (=1) -3.309*** -2.279*** (0.143) (0.159) Time (=1) -1.440*** -1.229*** (0.052) (0.057) Treatment x Time -1.506 -1.477*** (0.154) (0.167)

Avg. Access to Electricity (%) -0.112*** -0.115*** (0.007) (0.008) Race White -0.034 -0.337 (0.464) (0.464) Black -0.678 -0.502 (0.469) (0.468) Asian 0.843* 0.422 (0.513) (0.513) Multiracial -0.606 -0.513 (0.465) (0.465) Gender (=1) -6.160*** -6.190*** (0.039) (0.039) Years of Education -0.006*** 0.026*** (0.005) (0.005) Age Started Working -0.138*** -0.112***

(0.006) (0.006) Urban (=1) 2.913*** 2.845*** (0.076) (0.077) Mother in Household (=1) -1.782*** -1.910*** (0.041) (0.041) # of Sleeping Rooms -0.036 -0.043* (0.023) (0.024)

N 657,232 485,255 657,232 485,255

Using data from the 2001 – 2008, 2011 – 2015 PNAD survey: Columns (1) and (2) show results for a traditional Diff-in-Diff where ‘Treatment’ refers to whether an individual is in a treatment state or not. Columns (3) and (4) show estimates for a fixed effects specification using a continuous treatment variable. Robust standard errors in parenthesis.

(28)

Table 8: Municipality-level: Impact of Electrification on Average Adult Working Hours

(1) (2) (3) (4) Treatment (=1) -5.379*** -4.098*** (0.629) (0.742) Time (=1) -0.932 -0.739*** (0.103) (0.118) Treatment x Time -1.449** -1.384** (0.588) (0.579)

Avg. Access to Electricity (%) -0.162*** -0.151*** (0.039) (0.041) Race White -0.050 -0.273 (0.833) (0.769) Black -0.589 -0.460 (0.798) (0.757) Asian 0.939 0.687 (0.747) (0.721) Multiracial -0.816*** -0.396 (0.807) (0.757) Gender (=1) -6.913*** -6.884*** (0.495) (0.490) Years of Education -0.005 -0.013 (0.028) (0.022) Age Started Working -0.126*** -0.119***

(0.016) (0.013) Urban (=1) 2.317*** 1.841*** (0.594) (0.469) Mother in Household (=1) -1.902*** -1.990*** (0.075) (0.093) # of Sleeping Rooms -0.121** -0.026 (0.059) (0.042)

N 407,052 300,471 407,052 300,471

Using data from the 2001 – 2008 PNAD survey: Columns (1) and (2) show results for a traditional Diff-in-Diff where ‘Treatment’ refers to whether an individual is in a treatment state or not. Columns (3) and (4) show estimates for a fixed effects specification using a continuous treatment variable. Clustered standard errors in parenthesis – 224 clusters for municipalities.

(29)

While the results on adult income are in line with my expectations (an increase), it is not achieved through the same theory of change I hypothesized. Adults do not show be earning a higher income through an increased number of working hours. This can be explained by numerous economic reasons. One possible economic explanation is that there are less people working in the agricultural sector where jobs are generally low-paying. Using the PNAD data from 2002 and 2015, I find that there was a decrease in agricultural jobs held from 12.2 percent to 8.7 percent, with a shift into higher-wage jobs. This makes sense because, as Brazil becomes more developed, there has been a shift towards more service-oriented jobs. Although there has been a decrease in the share of agricultural employment, the sector is still a crucial part of the labor market. Therefore, another explanation of the increase in income and decrease in working hours is that there may have been an increase in productivity in the agricultural sector itself – for example, many agricultural producers may have been able to switch from manual to grid-powered irrigation would allow them to spend fewer hours on irrigation yet facilitate a higher crop yield. This productivity increase could be arising in other small industries as well as they switch from manual labor to using electric tools and machinery. Lastly, as seen in many other papers covering rural electrification, it is possible that the program has stimulated the creation of small businesses, possibly causing an increase in income while decreasing working hours.

5.2 Child Labors

Child labor is measured as the percentage of working children in state or municipality. I estimate equation (3) and (5) to assess the impact of increased electricity access on child labor. Table 9 shows the results for both the difference-in-differences and fixed effects models. Column (1), not including controls, indicates that being in a treatment state, significantly decreases the prevalence of working children in a state by 3.36 percent. When controlling for race, gender, household urban status, whether the mother lives in the household, and the number of sleeping rooms, the estimate increases slightly to a 3.39 percent decrease in child labor – results are shown in Column (2). Columns (3) and (4) display the results for the fixed effects specification. Without controls, the results show that a 1 percentage point increase in access to electricity decreases child labor in a state by 0.33 percent. When the same controls are added, the estimated impact of a 1 percentage point increase in access to electricity results in the same

(30)

percentage decrease, 0.33 percent. Scaling up these marginal results, I get 6.08 percent decrease8_.

These results are significant at the 1 percent significance level. These results are in line with my expectations based on the luxury and substitute axioms. Since adults are earning a higher income, they can afford to remove their kids from the labor force.

In the municipality-level analysis, I find a smaller effect than in the state-level analysis. Table 10 shows results for the difference-in-differences and fixed effects specifications. Column (1), without controlling for individual and household characteristics, shows that being in a treatment municipality decreases the percentage of working children by 1.2 percent. With controls, being in a treatment municipality decreases child labor by 1.0 percent. As for the fixed effects specification, I find that, with and without controls, a 1 percentage point increase in access to electricity in a municipality decreases child labor by 0.25 percent. The average effect of electrification on child labor is a decrease of 4.6 percent.

As a result of the decrease in child labor in a state, I find that children are more likely to attend school (nearly) full-time. That is, children are more likely to attend school for four or more hours a day, where a normal full school day in Brazil is roughly five hours. This finding is in line with the luxury axiom. Adults can now afford to take their children out of the workforce, thus are more likely to send them to school. See appendix for these results.

8_{The difference in access to electricity for the child subpopulation from 2001 to 2015 is 18.65 percent. To scale up the results,}

(31)

Table 9: State-level: Impact of Electrification on Child Labor

(1) (2) (3) (4) Treatment (=1) 10.770*** 10.807*** (0.026) (0.027) Time (=1) -1.307*** -1.256*** (0.013) (0.003) Treatment x Time -3.357*** -3.388*** (0.031) (0.031)

Avg. Access to Electricity (%) -0.326*** -0.326*** (0.001) (0.010) Race White 0.472*** 0.021 (0.125) (0.051) Black -0.331** 0.062 (0.136) (0.052) Asian -0.011 0.063* (0.134) (0.056) Multiracial -0.198 0.014 (0.125) (0.051) Gender (=1) -0.007 0.004 (0.009) (0.004) Urban (=1) -0.678*** -0.018*** (0.019) (0.005) Mother in Household (=1) 0.111*** -0.015** (0.017) (0.007) # of Sleeping Rooms 0.183*** 0.010*** (0.006) (0.003)

N 260,050 255,670 260,050 255,670

Using data from the 2001 – 2008, 2011 – 2015 PNAD survey: Columns (1) and (2) show results for a traditional Diff-in-Diff where ‘Treatment’ refers to whether an individual is in a treatment state or not. Columns (3) and (4) show estimates for a fixed effects specification using a continuous treatment variable. Robust standard errors in parenthesis.

(32)

Table 10: Municipality-level: Impact of Electrification on Child Labor

(1) (2) (3) (4) Treatment (=1) 8.972*** 7.156*** (0.701) (0.927) Time (=1) -0.991*** -1.040*** (0.102) (0.113) Treatment x Time -1.160*** -1.019*** (0.213) (0.217)

Avg. Access to Electricity (%) -0.254*** -0.254*** (0.035) (0.035) Race White -0.070 0.010 (0.649) (0.082) Black -0.317 0.035 (0.612) (0.086) Asian -0.118*** 0.037 (0.756) (0.089) Multiracial 1.057** 0.007 (0.424) (0.079) Gender (=1) 0.039 0.003 (0.029) (0.003) Urban (=1) -2.280*** -0.030** (0.501) (0.015) Mother in Household (=1) -0.921*** -0.009 (0.147) (0.011) # of Sleeping Rooms 0.630*** 0.008 (0.190) (0.005)

N 172,475 169,554 172,475 169,554

Using data from the 2001 – 2008 PNAD survey: Columns (1) and (2) show results for a traditional Diff-in-Diff where ‘Treatment’ refers to whether an individual is in a treatment state or not. Columns (3) and (4) show estimates for a fixed effects specification using a continuous treatment variable. Clustered standard errors in parenthesis – 224 clusters for municipalities. * p < 0.1, ** p < 0.05, *** p < 0.01

5.3 Robustness Check

I evaluate whether the identifying assumption of no omitted time-varying and region-specific effects being correlated with the program is violated. I used a method similar to that used in

(33)

Duflo (2001) where, as a robustness check, she tests whether the allocation of other governmental programs was correlated with the INPRES school construction program. More specifically, I test whether the Programa de Aceleração do Crescimento (PAC, Growth Acceleration Program) is correlated with the LPT program.

5.3.1 Programa de Aceleração do Crescimento (PAC)

Programa de Aceleração do Crescimento (PAC), known as the Growth Acceleration Program by English translation, was a major strategic infrastructure investment program launched in January 2007 by the Brazilian federal government. PAC policies are geared towards encouraging higher private investments in Brazil. PAC projects are aimed at accelerating economic growth by increasing employment, reducing inequality and sustaining macroeconomic foundations (OECD, 2011). Initially, the more disadvantaged communities were the target of PAC projects; more specifically, those living in favelas9_{. Because some of the target communities may overlap}

with those of the LPT, it is appropriate to assess whether the estimates presented in Section 4 are upwardly biased through this program. There have been two phases of this program: PAC from 2007 to 2010 and PAC-2 from 2011 to 2014. I present specifications that control for the presence of the PAC program. For the purpose of this analysis, it is only appropriate to assess the impact of the first phase of PAC because funding for LPT was included in the PAC-2 budget. Due to missing PNAD data for 2009 and 2010, I will create a dummy variable for PAC, but also include 2011 as an “active” year of the program to get a fuller, more representative result.

5.3.2 Results

Table 11 shows results for the impact of electrification on adult income. Additionally, these regressions control for the PAC program. In the difference-in-differences specifications, because our coefficient decreases slightly, I find that the estimates in the difference-in-differences approach may be upwardly biased by omitted programs. However, when assessing the fixed effects specification, the results show a minor increase in the specification without controls. With

(34)

controls, the estimate is unchanged therefore, the results are robust and do not show signs of omitted program bias.

Table 12 shows results for the impact of electrification on average adult weekly working hours. Additionally, these regressions control for the PAC program. In the difference-in-differences specifications, the coefficient decreases slightly without controls and increases slightly with controls. Estimates are unchanged in the fixed effects specification; therefore, the estimates are not upwardly biased by omitted programs.

Table 13 shows results for the impact of electrification on child labor including a control for the PAC program. Similar to the regression for adult income, the coefficients on the difference-in-differences specifications decreases slightly, so estimates may be upwardly biased by omitted programs. However, the fixed effects specification is robust with an unchanged coefficient in both models.

Table 11: Robustness Check: Adult Income

Diff-in-Diff Fixed Effects (1) (2) (3) (4) Treatment (=1) -1.003*** -0.605*** (0.010) (0.004) Time (=1) 0.143*** 0.030*** (0.004) (0.003) Treatment · Time 0.158*** 0.085*** (0.011) (0.010)

Avg. Access to Electricity (%) 0.015*** 0.009*** (0.001) (0.000) Growth Acceleration Program (=1) -0.061*** -0.041*** -0.061*** -0.142***

(0.003) (0.003) (0.006) (0.005)

N 600,332 447,548 600,332 447,548

Using data from 2001 – 2008, 2011 -2015 PNAD survey: Columns (1) and (2) show results for a traditional diff-in-diff where “Treatment” refers to whether an individual is on a treatment state or not. Columns (3) and (4) show estimates for a fixed effects specification using an intensity of treatment measure; model is inclusive of time and state dummies. Robust standard errors are in parentheses, * p < 0.1, ** p < 0.05, *** p < 0.01

(35)

Table 12: Robustness Check: Average Adult Weekly Working Hours

Diff-in-Diff Fixed Effects (1) (2) (3) (4) Treatment (=1) -3.309*** -2.277*** (0.143) (0.159) Time (=1) -1.518*** -1.316*** (0.053) (0.057) Treatment · Time -1.500*** -1.472*** (0.154) (0.167)

Avg. Access to Electricity (%) -0.112*** -0.115*** (0.007) (0.008) Growth Acceleration Program (=1) 0.408***

(0.047) 0.447*** (0.051) 2.215*** (0.084) 2.279*** (0.093)

N 657,232 485,255 657,232 485,255

Table 13: Robustness Check: Child Labor

Diff-in-Diff Fixed Effects (1) (2) (3) (4) Treatment (=1) 10.770*** 10.808*** (0.026) (0.027) Time (=1) -1.402*** -1.351*** (0.013) (0.012) Treatment · Time -3.349*** -3.379*** (0.031) (0.031)

Avg. Access to Electricity (%) -0.326*** -0.326*** (0.010) (0.001) Growth Acceleration Program (=1) 0.496*** 0.490*** 1.827*** 2.471***

(0.010) (0.010) (0.013) (0.013)

N 260,050 255,670 260,050 255,670

(36)

Table 14 shows results for the impact of electrification on adult income at the municipality-level. As a robustness check, these regressions control for the PAC program. In both the difference-in-differences and fixed effects specifications the estimates are unchanged from the estimates found in Table 6 with the inclusion of the PAC program. This finding suggests that estimates are not upwardly biased by omitted programs.

Table 15 shows results for the impact of electrification on average adult weekly working hours. Following the same exercise, the difference-in-differences and fixed effects specifications’ coefficients are essentially unchanged from the results in Table 8 meaning that there are no signs of omitted program bias.

Again, comparing the results from Table 10 on child labor to the results in Table 16 shows that there is no sign that the model suffers from omitted program bias. Similar to the checks executed on adult labor and adult working hours, when controlling for the PAC program, the effect of electrification remains unchanged. Overall, the results at the municipality-level show to be more robust than at the state-level.

Table 14: Robustness Check: Adult Income

Diff-in-Diff Fixed Effects (1) (2) (3) (4) Treatment (=1) -1.285*** -0.632*** (0.120) (0.074) Time (=1) -0.035*** -0.083*** (0.013) (0.010) Treatment · Time 0.046 -0.028 (0.046) (0.039)

Avg. Access to Electricity (%) 0.013*** 0.007*** (0.002) (0.002) Growth Acceleration Program (=1) 0.133*** 0.082*** 0.061*** -0.021

(0.008) (0.007) (0.017) (0.013)

N 371,766 277,032 371,766 277,032 Using data from the 2001 – 2008 PNAD survey: Columns (1) and (2) show results for a traditional Diff-in-Diff where ‘Treatment’ refers to whether an individual is in a treatment state or not. Columns (3) and (4) show estimates for a fixed effects specification using a continuous treatment variable. Clustered standard errors in parenthesis – 224 clusters for municipalities. * p < 0.1, ** p < 0.05, *** p < 0.01

(37)

Table 15: Robustness Check: Average Adult Weekly Working Hours

Diff-in-Diff Fixed Effects (1) (2) (3) (4) Treatment (=1) -5.380*** -4.094*** (0.629) (0.742) Time (=1) -0.747*** -0.588*** (0.098) (0.108) Treatment · Time -1.457** -1.388** (0.589) (0.495)

Avg. Access to Electricity (%) -0.162*** -0.151*** (0.039) (0.041) Growth Acceleration Program (=1) -0.541*** -0.452*** -1.355*** -1.049***

(0.106) (0.103) (0.182) (0.193)

Table 16: Robustness Check: Child Labor

Diff-in-Diff Fixed Effects (1) (2) (3) (4) Treatment (=1) 8.972*** 7.153*** (0.701) (0.926) Time (=1) -0.768*** -0.809*** (0.072) (0.079) Treatment · Time -1.165*** -1.023*** (0.213) (0.216)

Avg. Access to Electricity (%) -0.254*** -0.254*** (0.035) (0.001) Growth Acceleration Program (=1) -0.700*** -0.727*** -0.897 *** -0.893***

(0.144) (0.151) (0.102) (0.102)

(38)

6. Discussion

In this section I will include a few discussion points as well as comments about this thesis. Outside of the possibly violated parallel trend assumption in the difference-in-differences model discussed in Section 4.1, there are some other concerns that may be an obstacle to causal inference.

Firstly, the matter of comparability of the state-level analysis and municipality-level analysis is of concern as the variation in access to electricity gains between the two time periods is significant. In the state-level analysis, the change in access to electricity from baseline to the end of the data is roughly 18 percent whereas in the municipality level, it is about 10 to 11 percent. There is clearly a tradeoff between using the two different geographical units. One offers more data and larger variation, whereas the other uses less data, but more precise estimates.

Secondly, while the results show to be generally robust and indicate that the LPT program has a significant effect on adult employment and child labor, it can be difficult to attribute all these findings to the program. Firstly, there is no publicly available data on the program itself, therefore these figures are based solely on average access to electricity rates, a proxy for the program. This could lead to issues with the causal interpretation of these results. Another issue that makes it difficult to attribute all success to this program is its interaction with the Bolsa Familia Program, a conditional cash transfer program aimed at reducing inequality and poverty as well as breaking the generational aspect of poverty. Since there are cash transfers taking place due to Bolsa Familia, the program is likely to influence child labor as it serves as a supplement to household income. However, because this program and the LPT program began at the same time, it is not possible to disentangle the effects mainly in the difference-in-differences approach. Additionally, this thesis does not include information on migration trends and how they may affect the results. Many of those living without access to electricity may move to a different state because of this. To get a more accurate causal effect of the program, it would be interesting to see if migration flows influence the results.