Land Invasions and Contemporary Slavery

Gabriel Cepaluni UNESP-Franca * August 10, 2020

Abstract This paper presents an original panel data set of modern-day slaves from Brazil. We demonstrate that municipalities with more land invasions have fewer numbers of slaves. Based on economic models of labor coercion, we suggest that land invasions offer an outside option in the labor market for poor and unskilled workers. The result is robust to the addition of many relevant controls, a battery of statistical techniques, and estimation with different samples. Additionally, we examine the extent to which municipal characteristics concerning the private sector, public services, and local politics influence the likelihood that municipalities have both land invasions and slavery. Our findings have broad implications for the quality of democratic governance and the working conditions of some of the world’s most vulnerable citizens.

Keywords: Modern-day slaves, land invasion, human rights, development studies, Brazil.

Word count: 8568

*My thanks to Marco Antonio Faganello for his superb research assistance, and Paulo Arvate for his contribution in the early stages of this project. I am grateful to Michael Touchton for their advice and suggestions.

1 1 Introduction

Slavery is an abject institution for today’s moral standards. However, coercive labor relations were fre- quent throughout history. Inevitably, the debate surrounding slave work is part of the history of social thought. Aristotle claims that some people are natural slaves, and it is best for them to serve a master (Aristotle, 1984). John Locke rejects the idea that a person could voluntarily consent to enslavement. However, he argues that enslavement of those who are guilty of capital offenses or captured in war is permissible (Locke, 1947). Karl Marx describes the capitalist-worker relationship as a form of "wage slavery" (Marx, 1971). Recently, Acemoglu and Wolitzky(2011) argues that standard economic models of the labor market assume that transactions in the labor market are "free." However, coercive work transactions were the norm for most human history, and it is still common nowadays. Although traditional slavery — especially African slavery — has been much researched (e.g., Fo- gel and Engerman(1995); Nunn(2008)), only a relatively few quantitative researchers have studied contemporaneous coercive labor relations (Beber and Blattman, 2013; Phillips and Sakamoto, 2012; Hernandez and Rudolph, 2015; Mahmoud and Trebesch, 2010). One reason for this scarcity is that modern slavery is a hidden human rights violation. Researchers or even the State often do not have easy access to some of the remote areas where the modern slaves work (Scott, 2010; McCarthy, 2014; Landman, 2020). In this sense, we take advantage of a rare opportunity to study modern slavery with an extensive original data set. Our data set comprises slaves rescued as a result of inspections from Brazil’s Ministry of Labor and Employment, and a rich set of information at the municipality-level for 1995-2003. We demonstrate that municipalities with more land invasions have fewer slaves than municipalities without invasions. In our most conservative estimates, one land invasion decreases around 0.15 slaves in Brazilian municipalities from 1995 to 2013. Inspired by economic models of labor coercion, we suggest that land invasion offers an outside option in the labor market for poor and unskilled workers (Domar, 1970; Acemoglu and Wolitzky, 2011). Landless movements "compete" with slave-owners, acting in the same marginalized municipalities1, also recruiting unskilled workers. Brazil’s constitu-

1Figure1 shows that the higher incidence of modern-day slavery occurs close to State borders.

2 tion allows the expropriation of land promoted by landless movements that do not comply with the Brazilian labor code that prohibits modern slavery. Consistent with our argument, the members of the most active Brazilian landless movement — the Landless Rural Workers’ Movement, or Movimento dos Trabalhadores Rurais Sem Terra (MST) — have a strong normative stance against slave labor. As a former-modern slave and current member of MST declares: "I was a slave in many farms, and my children, my grandchildren, and great-grandchildren will continue to be slaves. I think we are the ones who have to organize ourselves, get together to stop and say ’no’ to slavery. I believe that the way to overcome world hunger and misery is the distribution of land."2 We find that the number of firms has a positive impact on land invasions. In contrast, the average number of minimum wage salaries decrease them. Concerning public services, both drinking water and sewage increases land invasions. As for local politics, land invasions decrease during electoral years, and the possibility of reelecting the mayor also reduces invasions. Finally, we find that the number of employees increases slavery and drinking water, reducing slavery, while sewage and vaccines increase it. These results imply the private and the public sector, as well as self-interested local politicians, influence the relationship between land invasions and slavery. However, social spending is not reducing land conflicts and the extreme exploitation of workers. Our study builds upon theoretical models of coercive labor relations (Domar, 1970; Acemoglu and Wolitzky, 2011), debating to what extent the predictions of these models work in a real-working setting. We also contribute to the quantitative literature on modern slavery (Beber and Blattman, 2013; Phillips and Sakamoto, 2012), and land conflicts (Hidalgo et al., 2010; Albertus, Brambor and Ceneviva, 2018; Albertus, 2015). As far as we know, we are the first to establish a quantitative link between land invasions and modern-day slavery. Closest to our research finding, Buonanno and Vargas(2019) ex- amines the legacy of traditional slavery on violent crimes in today’s Colombia. In our study, we also explore different local conditions that favor the emergence of modern slavery and land invasions. Fi- nally, our study has implications for the debate on the quality of democracy in developing countries to the extent that human rights and dignity are crucial components for a truly free society (Sen, 2001;

2http://reporterunesp.jor.br/2019/05/14/trabalho-escravo-no-campo/

3 Diamond, Morlino et al., 2005). We organized the paper as follows. We review the literature on labor coercion and land conflicts and make some theoretical predictions. Then, we describe our data and the main variables of this study. After that, we present the methodologies employed in this article and the results of our analysis. We conclude discussing our main findings and their implications for modern slavery studies.

2 Modern Slavery and Land Conflicts

2.1 Labor coercion

The International Labor Organization (ILO) defines forced labor as an involuntary work performed under the menace of a penalty. The employer coerces the person to work through the use of violence or intimidation, or by more subtle means such as debt, retention of identity papers or threats of de- nunciation to immigration authorities (ILO, 1930). In 2016, more than 40 million people in the world were victims of modern slavery, with 24.9 million in forced labor. Of the 24.9 million victims of forced labor, 16 million were in the private economy, another 4.8 million were in forced sexual exploitation, and 4.1 million were forced by State authorities (ILO, 2017). Brazil was the last country in the Western world to officially abolish slavery in 1888. Nevertheless, the country has not extinguished coerced labor completely. Art. 149 of the 1940 Brazil Penal Code asserts that reducing someone to a condition analogous to slavery is a crime. That is, according to the law, it is illegal to subject a person to degrading working conditions or restricting their freedom of movement (Brasil, 1940). The penalty for contemporary slavery established in Art. 149 from the 1940 Penal Code is 2-8 years of imprisonment and a fine. The same penalty applies to employees who restrict the use of any means of transport by the worker and take possession of personal objects to retain them in the workplace. The penalty increases by half if the crime occurs against children or adolescents; or due to the prejudice of race, color, ethnicity, religion, or geographic origin (Brasil, 1940). Although the 1940 Penal Code established the crime of coerced labor, the Brazilian State only started to combat the phenomenon in 1995 after the official recognition of slave labor in its terri-

4 tory. Today, however, the uniqueness of the Brazilian legal and institutional anti-slavery apparatus makes the country a crucial case to study this form of human rights violation. During 1995, the federal government created, through Ordinances No. 549 and 550 and the Presidential Decree No. 1538, the

Special Mobile Inspection Group (Grupo Especial de Fiscalização Móvel — GEFM), within the framework of the Secretariat of Labor Inspection (Secretaria de Fiscalização do Trabalho — SEFIT), and the Executive Group for the Suppression of Forced Labor (GERTRAF). GEFM investigates complaints of exploitation of slave labor in rural areas.3 In 2003, the country created the "Register of employers that kept workers under conditions analogous to slavery," otherwise known as the "Dirty List." The list functioned partly as a "naming and shaming" mechanism. However, it also cuts off flows of government funds to these companies. The National Pact for the Eradication of Slave Labour followed in 2005, inviting firms and employers to commit to the anti-slavery effort. In 2010, the National Pact had more than 130 full signatories representing over 20% of Brazil’s GDP (Phillips and Sakamoto, 2012). Appendix Table B.1 presents a detailed chronology of actions, laws, and public policies against slave labor in Brazil from 1995-2013. The Brazilian law defines slave labor as "reducing someone to a condition analogous to slavery, whether by subjecting them to forced labor or exhaustive working hours or by subjecting them to degrading working conditions or restricting their movement by any means, due to debts to the employer or its representative."4 The definition highlights essential points about the forms that severe exploitation takes place. Often, modern-day slavery is contractual, or the employee has to pay off a debt working under sub-standard labor conditions. Frequently, such relationships of contractual bondage are short-term, lasting the time of a particular contract of work (Breman et al., 2007; Bales, Trodd and Williamson, 2009). Acemoglu and Wolitzky(2011) propose a theory of labor coercion with two characteristics. First, the worker has no wealth, so there is a limited liability constraint. Hence, the principal can punish or

3This study analyzes data produced by the Secretariat of Labor Inspection, combining the information on slaves with other available data, as described in the next section. We obtained this data through a request to the Ministry of Labor and Employment, invoking the Freedom of Information Law (Lei nº 12.527, de 18 de Novembro de 2011). 4Art. 6 of Normative Instruction No. 139, of January 22, 2018, in Art. 149 of Brazil’s Penal Code presents the definition of modern-day slaves.

5 reward the agent according to the her effort level.5 Second, the principal chooses the amount of coercion, which influences the agent’s "outside option." The second point captures the notion that coercion is mainly about forcing workers to accept terms of employment that they would otherwise reject. In this sense, the model suggests that workers who have a lower outside option are coerced more. It also suggests that agents’ outside options affect employers’ coercion and effort choices. Agents who walk away from a coercive relationship, however, may match with another employer. In our study, we argue that social movements promoting land invasions offer "outside options" to modern-day slaves.6 Figure1 suggests that there is a spatial correlation between modern slavery and land invasions. Progressively lighter colors depict the number of slaves and land invasions in the maps. The North and Center-West register fewer slaves and land invasions proportional to their areas, prob- ably due to relatively lower economic activity and population. In contrast, the South, Southeast, and Northeast present a higher incidence of slavery and invasions. We also observe unusual high number of slaves in the State of Pará, and a few municipalities in Bahia, Rio de Janeiro, and Mato Grosso do Sul. White areas are missing data.

2.2 Land Invasions

Brazil ratified its constitution in 1988 after a military dictatorship that lasted from 1964 to 1985. Art. 170 of the constitution asserts that private land has a social function, allowing the federal government to expropriate large portions of land that do not fulfill that function. The notion of social function involves the following criteria: (1) the land must be exploited eco- nomically in a "rational and adequate" manner; (2) the exploitation of available natural resources must preserve the natural environment; (3) the use of land must comply with the Brazilian labor code; (4) the exploitation of land favors the well-being of owners and workers. Relevant for us, the Brazilian law clearly states that the exploitation of land must comply with Brazil’s labor code that prohibits modern- day slavery. The social function of land is a crucial tool for enhancing welfare in such an unequal country as

5See Chwe(1990) for another principal-agent model on labor coercion. 6Consistently with our argument, the landless movement also acts as advocates against modern-day slavery.

6 Figure 1: Geographic distribution of slaves and land invasions

(a) Slaves

(b) Land invasions

Notes: The maps depict the number of slaves (panel a) and land invasions (panel b) in Brazilian municipalities from 1995 to 2013. The legends show the intensity of slaves and land invasions. White solid lines indicate Brazilian States.

7 Brazil. Land occupations are the most relevant public action asserting rights to rural land for members of Brazil’s Landless Rural Workers’ Movement (MST) (Foster and Bonilla, 2011; Ondetti, 2016). MST activists "occupy" the supposedly unproductive land to pressure and call attention to Brazilian authorities. Land redistribution has dramatically increased in response. Land reform allocation in Brazil, as in many other countries, is mostly a demand-driven process. The government reacts and responds to land invasions on private land in more desirable areas with expropriation and subsequent land reform grants. The government reaction to landless movements protects large landowners from broad, top- down land redistribution in countries like Japan, Peru, and Taiwan (Albertus, 2015). Nonetheless, it simultaneously creates incentives for land invasions and other forms of rural conflict (Alston, Libecap and Mueller, 2000). The land problem is not only a characteristic of Brazil. Other developing countries face similar challenges. In Colombia, land titling in conflictive rural areas led to spillover effects in which nearby communities recognized the need to support rebel groups to garner the attention of the government (Albertus and Kaplan, 2013). Similarly, in Russia following the emancipation of serfs in 1861, land- based rural rebellion increased as landlords hijacked the reform implementation process to win favor- able land allotments (Finkel, Gehlbach and Olsen, 2015). Land invasions are generally well-planned occupations of large and frequently underproductive States by land-poor rural workers (Hidalgo et al., 2010). These often sudden invasions usually occurs in areas with high land inequality, given the supply of potentially productive land for potential reform.

3 Theoretical Argument and Predictions

Poor and unskilled workers produce goods in order to survive. These workers have few job opportunities. When they face with two labor options, they choose the one that increases their welfare. Let us say their "first option" is to work as a slave. The working conditions are below what their society agreed upon as minimum labor standards. Imagine now that they have an "outside option." They can participate in a landless movement, and be a land invader. It is legal to invade unproductive land.

8 The new job provides these workers with an opportunity to increase their welfare. They might end up working for themselves in their land or for a collectivist social movement. Hence, we predict that many workers, when faced with these two options, choose the "outside option" over their "first option." As a consequence, holding other factors constant, land invasions decrease slavery.7 Additionally, both land invasions and slavery are embedded in a complex environment that influences workers’ choices and opportunities. Figure2 illustrates some aspects of this complexity. The two rectangles on the bottom part of the figure indicate the effect of land invasions on slavery, as described above. Moreover, we also expect that a stronger private sector, more public services, and more electoral competition and certain political ideologies, as described in the top rectangle, influence land invasions and slavery. Formal jobs and companies might also provide direct or indirect opportunities to some of these poor and unskilled workers.8 Second, as public services should reduce poverty, this should also decrease the workers’ need to work for slave jobs. Finally, electoral incentives and left-wing mayors — arguably, more preoccupied with vulnerable individuals than right-wing politicians (Bob- bio, 1996) — might reduce both land invasions and slavery. Nonetheless, left-wing politicians might be permissive with land invaders and might collaborate to rescue slaves. As a result, these phenomena would be common or visible in their municipalities. In reality, we should expect multiple feedback loops and other omitted intervening factors than we can see in Figure2. 9 We are dealing here with abstract concepts that are difficult to measure and tend to overlap. The complex relationships proposed in this paper are also under-theorized by the still small literature on modern-day slavery. The association between land invasions and slavery has not received much attention in the literature, let alone the setting of this relationship. Hence, we shy away to make strong causal claims, favoring, at this point in the debate, a more elaborate narrative.

7The decrease in slavery can happen because workers choose the "outside option" or because former slave owners im- proved the working conditions of their employees anticipating these "outside option." 8Indirectly, the formal market can also create demands for the informal sector because of its impact on the economy. 9We also explore the influence of social spending on land invasions and slavery in the Appendix. To save space, we excluded the analysis from the main paper because the statistical results were negligible.

9 Figure 2: Determinants of modern-day slavery

4 Data and Measures

4.1 Dependent Variable

We construct an annual panel that comprises 5,494 municipalities from 1995 to 2013. The dependent variable is the number of modern-day slaves in Brazilian municipalities in each year from 1995-2013.10 The minimum number of slaves per municipality is 0, and the maximum is 1,064. The average number of slaves is 0.423. We also created a binomial measure of slavery that we use in some statistical models in this study (see Appendix Table C.7). Table C.2 in the Appendix presents descriptive statistics for all variables used in the main paper.11 The term "slaves" refers to all workers found in conditions analogous to slavery. After the rescu- ing operation from the Ministry of Labor and Employment, the labor auditor removes workers from

10The exact definition of "slavery" was provided by the Ministry of Labor and Employment through a formal request using the Brazilian Freedom of Information Law (Lei nº 12.527, de 18 de novembro de 2011). 11Appendix Figure C.1 plots land invasions and slavery over time. We do not identify any decreasing or increasing clear pattern for those variables.

10 the workplace. Then, she starts a set of procedures to repair the damages, such as social assistance and other measures, so that workers do not return to slavery. The procedures to repair the damages include the termination of the work contract and payment for the end of the contract, payment of unemployment insurance, and job training programs. The term "condition analogous to slavery" is mentioned in Art. 6 of Normative Instruction No. 139, of January 22, 2018, in Art. 149 of Brazil’s Penal Code. For further details, please, return to section2.

4.2 Primary Independent Variable

Land invasions in Brazilian municipalities from 1995 to 2013 is our primary independent variable. In some estimations, we also use an indicator for whether at least one land invasion occurred in a given municipality-year. Land invasions are a crucial variable in this study because the redistribution of land offers new opportunities in the labor market for poor and unskilled workers. We take the data

on land invasions from Dataluta (Girardi, 2014), which is a database on land conflicts hosted by the São Paulo State University. The data set is the most comprehensive and authoritative source of data on land invasions (Albertus, Brambor and Ceneviva, 2018). Most information on land invasions in the

Dataluta comes from the CPT (Comissão Pastoral da Terra), a nongovernmental organization with ties to the National Conference of Bishops of Brazil.12 CPT collects data on land invasions from primary sources such as social movements, trade unions, political parties, government agencies, and churches.

It also gathers data from local, State, and national newspapers and police records. Dataluta documents 4,825 land invasions from 1995 to 2013. The minimum number of invasions is 0, and the maximum is 31 in a given municipality-year. The average number of land invasions is 0.089. In the spirit of Acemoglu, García-Jimeno and Robinson(2015), we create four dummy variables to test spillover effects from one municipality to its neighbors. The variable "neighbors invasions" indicates whether there was an invasion in the adjacent municipality if there were at least one invasions in the municipality of interest. "Neighbor-of-neighbors invasions" measures whether there was at least one invasion on neighbors-of-neighbors municipalities if there was at least one invasion in a

12CPT is also one of the pioneers in denouncing slave labor in Brazil. https://www.cptnacional.org.br/ campanhas-e-articulacoes/campanhas/campanha-de-prevencao-e-combate-ao-trabalho-escravo.

11 given municipality. "Neighbors invasions (inverse)" measures if there was at least one invasion in the municipality but not in the adjacent neighbor. Finally, "neighbors-of-neighbors invasions (inverse)" measures if there was at least one invasion in the municipality but not in the neighbors-of-neighbors municipalities. All the analyses using these variables are in Appendix Figure C.5.

4.3 Control Variables

We choose our controls focusing on variables that the literature on land conflict and modern-day slavery have identified as relevant. The first control variable is land reform, collected from INCRA —

Instituto Nacional de Colonização e Reforma Agrária. Under Brazil’s 1988 constitution, only unproductive land is subject to expropriation. The government responds not only to land invaders but also to the public opinion and the media to start a land reform. These are often areas where land is cheap, and landowners tend to be politically weak. Land reform grants in Brazil typically follow an earlier land invasion (Albertus, Brambor and Ceneviva, 2018). Similarly, existing land reform projects are likely to affect landless agrarian workers’ calculations and increase the risk of land invasions nearby. Land invasions often involve well-organized social movements, such as the MST. Landowners are often members of prominent and wealth groups, such as the National Confederation of Agriculture, or the

UDR (União Democrática Ruralista). These organizations help to inform their members about relevant opportunities for peasants or threats from landowners. Our models include other co-variates whose omission may confound the results. We use the rate of illiteracy in the municipality because many slaves and land invaders are poor and unskilled workers. We collected illiteracy data from 1991, 2000, and 2010 Census, and we employ a linear interpolation with seasonal effects to impute the non-available years. We took the log (+ 1) of Tax Collection because of the skewness in the data. The variable measures the State’s fiscal capacity to extract revenues from its citizens (Weber, 1978; Tilly et al., 1992; Besley and Persson, 2010). We collected data on Tax Collection from the Ministry of Finance and National Treasury (or Ministério da Fazenda/Secretaria do Tesouro Nacional). With this variable, we want to reduce the possibility that municipalities with higher State capacity to raise revenues also have better reporting on slavery. We expect fewer slaves in areas where

12 the State has a higher capacity to collect taxes. We control for land inequality because we anticipate finding more slaves in municipalities with highly unequal land distribution. Land inequality is taken from the detailed agricultural censuses by the Brazilian Institute of Geography and Statistics (IBGE) and measured using a Gini coefficient. We collected murder rates from the Ministry of Health/Datasus. Murder rate is a measure of violent crime that has less severe problems of under-reporting than other crime measures (e.g., rape and robbery), because there is a body that makes the crime harder to hide. Murder is a relevant control because slavery should correlate with other violent crimes. We introduce one control measuring whether or not there is a municipal police guard in the municipality. The local police presence is crucial because we want to control the possibility that there are more slaves in areas with more policing. The percentage of the rural population, taken from IBGE, is measured as the percentage of a municipality’s population that is rural. We use this variable as a control because land invasions and modern-day slavery historically happen more often in rural areas. We also collected yearly population from IBGE. The population is a relevant control in such a diverse and vast country like Brazil. GDP per capita, also taken from IBGE, is an indicator of local development. We took the log (+ 1) of population and GDP per capita because of the skewness in the data.13 Agricultural productivity may affect land invasions or slavery. For that reason, we include different prominent types of crops – cattle, soy, sugar, and coffee — in our regressions. Previous empirical work on slavery and land invasions also included crop types as control variables in their models (Phillips and Sakamoto, 2012; Albertus, Brambor and Ceneviva, 2018).

4.4 Other Variables

We collected four groups of variables to explore further the relationship between land invasions and modern slavery. The four groups of variables are related to the private sector’s workings, public services, local politics, and social spending. For measuring the influence of the private sector, we collected the number of firms, of employees and annual averages of minimum wages in the municipalities from RAIS (Relação Anual de Informações

13We also interpolate the data for the municipal GDP because this information is only available from 1999 to 2013. The data on population is available for the whole period of the analysis.

13 Sociais). With these variables, we can examine outside options for low-skilled and poor workers in the formal labor market. A survey in the State of Maranhão found that at least 28.4% of former slaves households have an income between 1 and 2 minimum wages (in a sample of 2,135) (da Silva, 2018). These wages are not lower than the 1.751 average salaries in our sample. Both Fogel and Engerman (1995) and ILO(2009) also find that coerced laborers often receive income close to that of comparable free laborers. Therefore, we anticipate that a broader market, more job opportunities, and better salaries will have a positive impact on reducing land invasions and slavery. We use three variables for measuring the impact of public services. We take the percentage of households with drinking water and the percentage of houses with plumbing sewage from IBGE. We also include in our models the number of total doses of vaccines given in the municipalities from the Ministry of Health/Datasus. We expect that public services will decrease land invasions and slavery because they should mitigate poverty.

We employ four variables to map municipal politics. All variables are from the TSE (Tribunal Supe- rior Eleitoral). First, we create a variable — Reelection (mayor) — indicating whether the mayor can run in the next election. In Brazil, mayors can only serve for two consecutive terms. Second, we employ an indicator variable defining election years. During our study, there were mayoral elections in 1996, 2000, 2004, 2008, and 2012. Third, we use the number of parties disputing the mayoral election as a measure of electoral competition. The average number of parties is 2.614, with a minimum of only one party and a maximum of 14 parties. Finally, we create a measure of mayors’ party ideology. Mayors’ ideology is a categorical variable with three categories: left, center, and rights. We use the classification of previous work on the ideology of Brazilian parties in the federal legislative branch to define mayors’ ideology (Mainwaring, Meneguello and Power, 2000; Melo, 2004; Power and Zucco Jr, 2009) — see Appendix Table C.3 for further details.14 The expectation is that electoral incentives — the possibility of being reelected, electoral years, and electoral competition — will decrease land invasions and slavery. We also anticipate that left-wing politicians might be more preoccupied with social problems — such as land invasions and slavery — and want to solve or reduce them. However, they may also

14We could not find any work classifying political parties at the municipal level. Therefore, we inferred from the classification from the Brazilian legislative to municipal politics.

14 be more permissive with land invaders and collaborate more to rescue slaves, making these activities more common or visible in their municipalities.

Finally, we also analyze three different measures of social spending in the Appendix: i) spending on social assistance and pensions, ii) education and culture, and iii) health and sanitation. We collected all measures of social spending from the Ministério da Fazenda/Secretaria do Tesouro Nacional. We also logged (+ 1) all measures of social spending to reduce skewness in the data. The expectation is that social spending will decrease land invasions and slavery.

5 Empirical Strategy and Preview of the Results

We first report the results of a series of OLS panel regressions with municipality fixed effects and year dummy fixed effects. Fixed effects models reduce time-invariant and municipality-invariant omitted variable bias. In all models, we cluster the standard errors at the municipal level.15 In this first set of regressions, we analyze the effect of land invasions on modern-day slavery. We present a model with no controls, a model with several controls, and models varying the agricultural controls (cattle, soy, sugar, and coffee dependency). One invasions decrease, on average, 0.15 slaves. We then present the results of a Weighted Unit Fixed Effects Estimator. Imai and Kim(2019) propose a non-parametric framework that connects fixed effects models to matching or weighting methods. The idea is to compare treated (land invasions) and control (no land invasions) groups within the same municipalities. In our application, we weight the covariates, including years, and stratify our data by municipalities.16 All regressions include a model-based robust standard error, allowing for heteroskedasticity. The results of both OLS panel regressions and Weighted Unit Fixed Effects Estimators have the same sign. However, OLS presents the lower bounds of our estimates (see results above), and Weighted Unit Fixed Effects represents the upper bound. With the latter method, one land invasion decrease, on

15Unless we say otherwise, we usually control for year and municipality fixed effects, and cluster the standard errors at municipal level. 16Cases of slaves are the smaller level of observation in our panel. Hence, one municipality might have more than one anti-slavery operation in a given year. Hence, we do not have unique repeated observations across the municipalities and years. As a result, we cannot estimate time effects directly, only adding time as a weighted control variable.

15 average, 0.27 slaves in the municipalities. In the Appendix, we show that the effect is restricted to the municipality of interest, and does not affect neighboring municipalities. Also, the reversal effect of slavery on land invasions is much smaller, if existent, than the effect of invasions on slavery. Together, these results suggest a causal interpretation. Our next step is to use land invasions as our dependent variable. We explore the possibility that some municipal characteristics might influence the likelihood of land invasions. Returning to our OLS panel regressions, we estimate the effect of the private sector, public services, and local politics on modern-day slavery. We find that the number of firms has a positive impact on land invasions. In contrast, the average number of minimum wage salaries decrease land invasions. As for public services, both drinking water and sewage increases land invasions. Finally, land invasions decrease during electoral years, and the possibility of reelecting the mayor also reduces invasions. Afterward, we employ a General Multilevel Structural Equation Modelling (GMSEM) to explore further some municipal characteristics that might affect both land invasions and slavery. Here, land invasions are our mediation variable, slavery is our outcome variable, and municipal characteristics are our independent variables. Specifically, we employ GMSEM to separate the between-municipalities and within-municipalities components of the intervening variable "land invasions," avoiding bias in the indirect effect calculation (Preacher, Zyphur and Zhang, 2010). Throughout the paper, we re- fer to direct, indirect, and total effects to discuss mediation with the predominant language used in the literature (Baron and Kenny, 1986). The direct effect measures the extent to which the outcome variable changes when the independent and mediator variables remain unaltered. The indirect effect constitutes the extent to which the independent variable influences the outcome variable through the mediator. Finally, the total effect is equal to the sum of the direct and indirect effects. With GMSEMs, we confirm the results from the OLS panel models using land invasions as an outcome variable. We also find that the number of employees increases modern slavery. Drinking water and sewages reduce slavery, while vaccines and sewage increase it. Finally, electoral politics does not affect slavery. Using principal component analysis (PCA) indexes, we find that the private sector’s presence increases slavery, and public services both increase land invasions and reduce slavery.

16 The PCA analysis, descriptive statistics, and other robustness checks are in the Appendix. Also, in the Appendix, we explain the equations related to the models employed in the main paper.

6 Results

6.1 Land Invasion and Slavery

Table1 presents OLS Panel regressions. All models include municipality and year dummy fixed effects, and standard errors clustered at municipality-level. In the first column, a bivariate model shows that one land invasion decreases 0.142 slaves in the municipality. The magnitude of the relationship is sizable.17 When controlling for socio-demographic and agricultural variables, the effect remains stable. In columns 2-6, one invasion decreases 0.15 slaves in the municipality. Between the control variables, land inequality is a strong predictor of slavery. As expected, more unequal municipalities have fewer slaves than more equal ones. Finally, other control variables do not reach statistical significance at 0.05 level. Previous research suggests that different agriculture crops, especially cattle businesses, were a determinant of modern slavery in Brazil (Phillips and Sakamoto, 2012), and traditional slavery in general (Engerman, Sokoloff et al., 2000). Against our expectations, the uncertainty of the first set of results from our OLS panel regressions and the negative sign of cattle dependency does not corroborate the results from previous studies. In Appendix Table C.4, we created 10 data sets with multiple imputations by chained equations (MICE) method, and we replicate the same models presented in Table1. The regression coefficients are virtually the same from the analysis in the main paper, which is not surprising because of the size of our data set and its limited amount of missing data. We also replicate Table1 restricting our sample to municipalities that have at least one slave rescued from 1995 to 2013 in Appendix Table C.5. These municipalities are likely more comparable among themselves in their capacity or desire to rescue slaves. The coefficients for land invasions become four times larger. Hence, the analysis presented in the main

17Appendix Table C.2 shows that Brazilian municipalities have, on average, 0.423 slaves and 0.089 land invasions from 1995 to 2003. Appendix Figure C.2 plots OLS and Logit regression lines with the correlation between modern-day slaves and land invasions without any control variables or fixed effects.

17 Table 1: OLS Panel Regressions — Slaves

(1) (2) (3) (4) (5) (6) Dependent Variable: Slaves Invasions (count) -0.142* -0.151** -0.150** -0.151** -0.150** -0.151** (0.057) (0.056) (0.056) (0.056) (0.057) (0.057) Illiteracy 0.157 0.121 0.265 0.081 0.177 (0.567) (0.566) (0.582) (0.555) (0.567) Log (Tax collection) 0.015 0.014 0.013 0.015 0.014 (0.011) (0.011) (0.011) (0.011) (0.011) Reforms 0.122 0.124 0.125 0.124 0.126 (0.110) (0.109) (0.110) (0.110) (0.109) Land (Gini) 1.086** 1.109** 1.142** 1.071** 1.128** (0.417) (0.417) (0.422) (0.415) (0.419) Murders -0.316 -0.315 -0.316 -0.315 -0.315 (0.259) (0.259) (0.259) (0.259) (0.259) Municipal guards 0.059 0.062 0.052 0.071 0.061 (0.093) (0.091) (0.094) (0.091) (0.091) Rural percentage 0.730 0.773 0.801 0.785 0.824 (0.606) (0.610) (0.608) (0.609) (0.611) Log (Population) 0.037 0.051 0.056 0.065 0.068 (0.411) (0.417) (0.411) (0.416) (0.417) Log (PIB per capita) 0.247+ 0.230+ 0.267* 0.243+ 0.247+ (0.135) (0.137) (0.134) (0.137) (0.137) Cattle Dependency -0.115+ -0.115+ (0.062) (0.059) Soy Dependency -0.542 -0.511 (0.545) (0.564) Sugar Dependency 0.405 0.606+ (0.312) (0.333) Coffee Dependency -0.037 -0.201 (0.318) (0.303) Municipalities 5424 5360 5360 5360 5360 5360 N 101870 95323 95323 95323 95323 95323 Notes: All specifications include municipality and year dummy fixed effects (not shown) and cluster standard errors at municipal level (in parenthesis). +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < .001.

document might be a conservative estimate of the true effect of land invasions on slavery. Endogeneity problems plague empirical research, and our study is no exception. For instance, land invasions might reduce the number of slaves because landholding offers an outside option for vulnerable workers (Acemoglu and Wolitzky, 2011). Alternatively, municipalities with more land invasions might attract more attention to the Brazilian political authorities and motivate them to work harder to rescue slaves. Therefore, other research methods are useful to make our estimations more credible.

18 Weighted Unit Fixed Effects regression is a statistical technique equivalent to a within-unit matching estimator used to evaluate the effect of a treatment by comparing the treated and the non-treated units in an observational study (Imai and Kim, 2019). By weighing treated units to make them similar to non-treated units, weighing enables a comparison of outcomes among treated and non-treated units to estimate the effect of the treatment, reducing bias due to confounding. In Figure3, we estimate the same models from column 2 of Table1 using the Weighted Unit Fixed Effects Estimator (Imai and Kim, 2019). All models present robust standard errors, allowing for heteroskedasticity.18 In our estimation, we match the covariates, including the years, and stratify our data by municipalities. Toimplement the estimator, we have to employ a binomial measure of land invasions to match municipalities with and without land invasions. Zero means that a municipality experienced a land invasion in a given year, and one means no invasions. With the Weighted Unit Fixed Effects Es- timator, the effect of invasions on slaves increases in comparison with OLS models: changing from no invasion to one invasion decreases the number of slaves by 0.265. Appendix Figure C.4 uses weights to estimate all models presented in Table1 plus another specification excluding all agricultural variables. The results are remarkably stable. Appendix Figure C.5 estimates possible spillover effects (or "placebo" tests) of invasions in neighboring municipalities on the number of slaves. We create a series of dummy variables to estimate whether there was an invasion in neighbor municipalities or neighbors-of-neighbors municipalities considering that the municipality of interest had at least one invasion. In column 1, neighbor invasions indicate whether there was at least one invasion in the adjacent municipality if there was also an invasion in the municipality of interest. The variable neighbor-of-neighbor invasions in column 2 measures whether there was an invasion on neighbors-of-neighbors municipalities if there was at least one invasion in the municipality of interest. In column 3, the inverse neighbor invasions measures if there was at least one invasion in the municipality but not in the adjacent neighbor. Finally, the inverse of neighbors-of-neighbors invasions in column 4 measures if there was at least one invasion in the municipality but not in the neighbors-of-neighbors municipalities. We do not find any statistically

18These models do not allow for clustering standard errors.

19 significant effect in all these models. These "placebo" tests suggest that invasions only impact slaves in the municipality of interest. In Appendix Figure C.7, we estimated the impact of this variable on land invasions (as continuous and binomial measures) on a binomial measure of slavery. In columns

1 and 2, we do not find a statistically significant relationship at the 0.05 level. The results suggest that the weighting methodology corrects endogeneity problems and that land invasions affect the number of slaves and not the opposite. Two limitations of the approach are that we can only have binomial treatments. The method also uses a different number of observations to construct weights.

6.2 Determinants of Land Invasions and Slavery

We hypothesize that municipal characteristics of the private sector, public services, and local politics influence the likelihood of modern slavery and that this influence can occur either directly or through land invasions.19 To explore these mechanisms, we again employ OLS panel regressions to see to what extent the private sector, public services, and local politics influence land invasions. After that, we employ a generalized multilevel structural equation model (GMSEM) to examine further the impact of the above mention municipal characteristics on both land invasions and slavery (Preacher, Zyphur and Zhang, 2010; Muthén and Asparouhov, 2008). First, we turn to our OLS panel regression estimates in Table2 to explore possible determinants of land invasions. From columns 1-3, we examine the effect of log (n. of firms), log (n. of employees), and minimum wages (average). The results are heterogeneous. While the number of firms increases land invasions (column 1), the number of employees (column 2) and the average minimum wage (column 3) reduces them. However, the effect of log (n. of employees) is not significant at 0.05. The differential results for log (n. of firms), and minimum wages (average) are plausible because we estimate different characteristics of the private sector. More firms might attract more workers to the municipality, but it does not mean that all workers will find a job. Nevertheless, higher averages salaries suggest that many people have a better outside option than expropriating land. Hence, the heterogeneous results do not allow us to make a firm statement about the influence of the labor market on modern-day slavery.

19In the Appendix, we also explore the effect of spending on land invasions and slavery.

20 Figure 3: Weighted Unit Fixed Effects Estimator. Dependent Variable: Slaves. Municipalities: 5360. N (nonzero weigh): 28077. The model include weighted years and municipal fixed effects. Robust standard errors, allowing for heteroskedasticity, are in parentheses. We included the same control variables (not shown) of our baseline model first presented in column 2 of Table1. +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < .001.

21 Therefore, these results might contribute to a more nuanced understanding of the complex impact of the private sector on slavery. From column 4-6, we explore the effect of three public services on land invasions: drinking water, sewage, and log (vaccines). All three variables positively predict land invasions, but the log of vaccines has a statistically insignificant effect. These results are plausible because municipalities with more land invasions should have some infrastructure (drinking water and sewage) to produce agricultural goods. Finally, we estimate the effect of four variables related to the impact of local politics on land invasions. In column 7, we show that the possibility of the mayor’s reelection decreases land invasions. In column 8, we see that land invasions experience a decline during municipal electoral years. Finally, municipalities with more parties disputing mayoral elections (column 9) and with more right-wing ideological mayors (column 10) also have fewer land invasions. Nonetheless, the results from column 9 and 10 are statistically insignificant. In Appendix Table C.6, we create a principal component analysis (PCA) index of (1) the private sector, (2) public services, and (3) spending variables. The public services index is the only statistically significant result at the 0.05 level. It corroborates the analysis in the main paper, which suggests that municipal public services’ provision increases land invasions. Continuing with our exploration of potential mechanisms, we estimate a series of generalized multilevel structural equation models (GMSEMs) to explore possible determinants of land invasions and slavery. All models include standard errors clustered at the municipal level. Most models were estimated using year dummies fixed effects. We do not include year dummies to avoid multicollinearity with three of our political variables: reelection (mayor), election (mayor), and number of parties. All models include the same control variables presented in the baseline model first showed in column 2 of Table1. Table3 employ three variables to explore the influence of the private sector on land invasions and slavery: log (n. of firms), log (n. of firms), and minimum wages (average). In columns 1-3 from Panel A Table3, we find again that land invasions have a direct and stable effect on slavery. Also, in Panel A Table3, only the log (n. of employees) has a direct effect positive and statistically significant effect on slavery. 1% increase in the number of employees increases 0.004 slaves. In Panel B from Table3, we

22 Table 2: Panel OLS regressions — Land invasions

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Dependent Variable: Invasions (count) Log (N. of firms) 0.021** (0.007) Log (N. of employees) -0.004 (0.006) Min Wages (average) -0.005** (0.002) Drinking water 0.045* (0.021) Sewage 0.165***

23 (0.042) Log (Vaccines) 0.002 (0.001) Reelection (mayor) -0.012*** (0.003) Election (mayor) -0.003** (0.001) N. of parties -0.002 (0.003) Mayors’ ideology -0.005 (0.004) Municipalities 5359 5359 5359 5310 5310 5360 5360 5360 5337 5337 N 93211 94340 94340 94033 94033 95323 95323 95323 67570 65750 Notes: Specifications from column 1-6 and 9-10 include municipality and year dummy fixed effects (not shown) and cluster standard errors at municipal level (in parenthesis). Columns 7 and 8 include cluster standard errors at municipal level, but exclude year dummies due to co-linearity. +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < .001. Table 3: GMSEMs — Private sector: Direct and indirect effects

(1) (2) (3) A. Dependent Variable: Slaves Invasions (count) -0.144** -0.143* -0.143* (0.056) (0.056) (0.056) Log (N. of firms) -0.069 (0.046) Log (N. of employees) 0.383*** (0.096) Min Wages (average) 0.001 (0.019) B. Dependent Variable: Invasions (count) Log (N. of firms) 0.025*** (0.004) Log (N. of employees) -0.000 (0.004) Min Wages (average) -0.008*** (0.002) Municipalities 5359 5359 5359 N 93211 94340 94340 Notes: All specifications include municipality and year fixed effects (not shown) and cluster standard errors at municipal level (in parenthesis). We included the same control variables (not shown) of our baseline model first presented in column 2 of Table1. +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < .001.

find a comparable heterogeneous effect to Table2. While the log (n. of firms) has a positive effect on land invasions, the average of the minimum wages decrease invasions. Again, the result is plausible. More firms might attract more workers to the municipality. However, it does not mean that all workers will find a job. Nonetheless, higher averages salaries suggest that many people have a better outside option than invading land in these municipalities. The total effect is the direct effect plus the indirect effect. For instance, the total effect of log (n. of firms) on modern-day slavery is the direct effect (-.069) plus the indirect effect (0.025 × -.069 = -0.002), meaning that the total effect is -0.071.20 Table4 repeats the same exercise with public services measures. Again, columns 1-3 in Panel A

20The result is not statistically significant. We present the rounding numbers with three decimal digits in our tables. Therefore, the calculations made with statistical software might diverge slightly from manual calculations.

24 Table 4: GMSEMs — Public services: Direct and indirect effects

(1) (2) (3) A. Dependent Variable: Slaves Invasions (count) -0.143* -0.145** -0.143** (0.056) (0.056) (0.056) Drinking water -0.615** (0.201) Sewage 1.659*** (0.326) Log (Vaccines) 0.033** (0.010) B. Dependent Variable: Invasions (count) Drinking water 0.066*** (0.017) Sewage 0.087*** (0.025) Log (Vaccines) 0.002 (0.001) Municipalities 5310 5310 5360 N 94033 94033 95323 Notes: All specifications include municipality and year fixed effects (not shown) and cluster standard errors at municipal level (in parenthesis). We included the same control variables (not shown) of our baseline model first presented in column 2 of Table1. +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < 1.

show that invasions have a direct influence on modern slavery. Panel A shows that the three measures of public services have a statistically significant effect on reducing slavery. However, while drinking water reduces the number of slaves, sewage and vaccines increase them. Also interesting, in Panel B, public goods are increasing the likelihood that municipalities have land invasions. The result suggests that it is necessary basic infra-structure to have decent enough land that is worth being invaded. Ap- pendix Table C.7 helps us to have a general idea of the overall effect of the private sector thanks to our principal component analysis (PCA) indexes. Panel A Appendix Table C.7 shows a negative association between the private sector index and slavery, whereas Panel B suggests that public services increase land invasions. Both results are statistically significant. Table5 estimates the influence of local politics on both land invasions and slavery. Here, we see

25 that the mayor’s party ideology (column 4) and the number of parties in the last mayoral elections (column 3) are not relevant predictors of slavery or land invasions. Only the mayor’s eligibility to reelection (column 1) and the year for municipal elections (column 2) influence the likelihood of land invasions. The result is consistent with the assumption that politicians are rational actors that want to win elections and stay in power. Interestingly, these are also the only two variables in our mediation models that have only effect on land invasions but not on slavery. Note that the principle of the social function of the property is part of Brazil’s legal system. The principle implies that the right of private ownership includes an obligation to use the property in ways that contribute to the common good. Hence, landless movements promote land invasions according to a plausible interpretation of the law. On the other hand, slave owners have an incentive to hide their illegal activities from the authorities and the Brazilian public. Thus, politicians seeking and trying to maintain power might want to focus on a more salient and pressing topic (land invasions) than a less salient and pressing one (slavery).

7 Discussion and Conclusions

Adam Smith made two claims about slavery in the context of developing economies. First, Smith argues that slavery was generally inefficient. Second, he observes that, despite its inefficiencies, slavery persists in most of the world (Smith, 1978; Weingast, 2016). The persistence of slavery and forced labor in human history and its long-term effects are also a central concern of modern social scientists (North et al., 2009; Dell, 2010; Nunn, 2008; Nunn and Wantchekon, 2011). Slavery was the most common form of labor arrangements in many ancient civilizations, including Greece, Egypt, Rome, several Islamic, Asian, and pre-Columbian civilizations (Meltzer, 1993; Lovejoy, 2000; Davis, 2006). Slavery was also the basis of agricultural economies in the Caribbean (Curtin and Curtin, 1998; Klein and Vinson III, 2007), in the South of the United States, and in Latin America (Fogel and Engerman, 1995; Lockhart and Schwartz, 1983; Dell, 2010). Although formal slavery has been rare in Europe since the Middle Ages, forced labor was relevant types of "employment contract" until the 19th century (Bloom, 1998; Gingerich and Vogler, 2020). Human trafficking of immigrants

26 Table 5: GMSEMs — Local politics: Direct and indirect effects

(1) (2) (3) (4) A. Dependent Variable: Slaves Invasions (count) -0.132* -0.134* -0.226** -0.239** (0.056) (0.056) (0.085) (0.087) Reelection (mayor) 0.059 (0.037) Election (mayor) -0.003 (0.007) N. of parties -0.018 (0.040) Mayors’ ideology -0.012 (0.062) B. Dependent Variable: Invasions (count) Reelection (mayor) -0.010** (0.003) Election (mayor) -0.003*** (0.001) N. of parties -0.001 (0.003) Mayors’ ideology -0.007+ (0.004) Municipalities 5360 5360 5337 5337 N 95323 95323 67570 65750 Notes: All specifications include municipality and year fixed effects (not shown) and cluster standard errors at municipal level (in parenthesis). We included the same control variables (not shown) of our baseline model first presented in column 2 of Table1. +p < .1, ∗p < .05, ∗∗p < .01, ∗∗∗p < 1.

and refugees, often considered "modern-day slavery," is still recurrent in today’s Europe (Hernandez and Rudolph, 2015; Buonanno and Vargas, 2019). Recently, the United Nations’ International Labor Organization (ILO) estimates that there are over 24.9 million forced laborers worldwide (ILO, 2017). In this paper, we explore a rare opportunity for quantitatively study such a prevalent topic. We estimate that one land invasion decreases around 0.15-0.61 slaves in Brazilian municipalities for the period from 1995 to 2013. Interpreting our results with Acemoglu and Wolitzky(2011)’s formal model of labor coercion, we argue that the landless movements offer an outside option for slaves. Precisely, landless movements "compete" with slave-owners because they both act in municipalities with the same characteristics, recruiting unskilled and impoverished workers. Moreover, landless movements have

27 a normative position against modern-day slavery and other forms of extreme exploitation of labor. There may be under-reporting of modern-day slavery. In some cases, under-reporting could be motivated by a lack of policing, State capacity, or political ideology. First, we could not find any clear temporal pattern related to modern-day slavery in Appendix Figure C.1a. Second, our main results hold employing: (1) a rich set of controls, including measures for policing, State capacity, and political ideology of the mayor; (2) several model specifications; (3) methodologies; and (4) different sample selections. Third, they are robust to estimates based on samples dropping one State or one year at the time — see Appendix Figure C.3.21 Finally, our measure of slavery is taken directly from the Brazilian law and the Brazilian State’s corresponding actions. Therefore, we do not have some measurement bias common in survey research, such as social desirability bias, or a potential mismatch between an abstract category and a real world phenomenon. Nonetheless, despite our best efforts, the reader should be aware of the difficulties of completely overcoming the limitations in the data. Our research faces some of the same challenges of measuring other human rights crimes (Landman, 2020; McCarthy, 2014). Moreover, we examine the extent to what different municipal characteristics influence the likelihood of slavery and land conflicts. First, we find evidence that the number of firms increases land invasions, and average salaries decrease them. We also find that a higher number of employees is asso- ciated with more slaves. Second, we find evidence that public goods increase land invasions because it is necessary to have a basic infrastructure to have productive land that it is worth an invasion. Although we find some heterogeneous results regarding the effectiveness of public services on slavery, globally, the results from Table C.7 suggests that better public services decrease slavery. Finally, the desire to stay in power and win elections is an incentive to local politicians to reduce land invasions, but not to reduce slavery. Slavery is an illegal hidden activity, while land invasions have legal justifications as an instrument for land redistribution. Hence, it is reasonable that politicians seeking power want to focus on a more salient and pressing topic (land invasions) than a less salient and pressing one (slavery).

21Only one regression (when we drop 2009) out of 44 regressions do not reach statistical significance at 0.05 level. Statistically, we should expect that to happen because of random chance. However, even in this case, the point estimate is similar to all other regressions.

28 Finally, Appendix Table C.8 shows that social spending has a negligible effect on land invasions and slavery. The result suggests that local government spending is not clearly or directly ameliorating the lives of Brazil’s most vulnerable individuals. Some researchers associate previous coercive labor relations with different posteriors levels of democratic governance. Paige(1998) argues that the Guatemalan elite was overwhelmingly agrarian, market by debt servitude, serfdom, and other forms of bonded labor. Then, coerced labor institutions inhibited popular mobilization and created a keen interest in authoritarian political structures to control the unfree population. In contrast, in Costa Rica, there was more distribution of land. As a result, Costa Rica’s elite found that land conflicts could be managed by extending the franchise to rural property owners and establishing democratic institutions. Gingerich and Vogler(2020) find that areas hit harder by the Black Death adopted more democratic institutions because the corresponding population reduced the likelihood of repressive labor regimes. Moreover, at least one study found an association between higher levels of democracy and lower cases of modern-day slavery in a cross- country comparison (Landman and Silverman, 2019). In this sense, the study of modern slavery might have long-term consequences for the quality of democracy in countries that still have coercive forms of labor relations (Diamond, Morlino et al., 2005). We predict that the problem of modern-day slavery will not have an easy solution shortly. As the

UN Special Rapporteur on the contemporary forms of slavery declared: "The severe socio-economic effect of the COVID-19 pandemic is likely to increase the scourge of modern-day slavery."22 As national govern- ments and economies suffer because of the COVID-19 pandemic, unfortunately, countries will likely face even more coercive working conditions. More scarce job opportunities for vulnerable people might also promote more land conflicts to compensate for the lack of outside options for unskilled workers. Therefore, the insights provided by our research might illuminate the relationship between labor coercion and land conflicts for the long-term.

22https://www.ohchr.org/SP/NewsEvents/Pages/DisplayNews.aspx?NewsID=25863&LangID=E

29 References

Acemoglu, Daron and Alexander Wolitzky. 2011. “The economics of labor coercion.” Econometrica 79(2):555–600.

Acemoglu, Daron, Camilo García-Jimeno and James A Robinson. 2015. “State capacity and economic

development: A network approach.” American Economic Review 105(8):2364–2409.

Albertus, Michael. 2015. Autocracy and redistribution. Cambridge University Press.

Albertus, Michael and Oliver Kaplan. 2013. “Land reform as a counterinsurgency policy: Evidence

from Colombia.” Journal of Conflict Resolution 57(2):198–231.

Albertus, Michael, Thomas Brambor and Ricardo Ceneviva. 2018. “Land inequality and rural unrest:

theory and evidence from Brazil.” Journal of Conflict Resolution 62(3):557–596.

Alston, Lee J, Gary D Libecap and Bernardo Mueller. 2000. “Land reform policies, the sources of

violent conflict, and implications for deforestation in the Brazilian Amazon.” Journal of environmental economics and management 39(2):162–188.

Aristotle. 1984. Politics. Chicago: The University of Chicago Press.

Bales, Kevin, Zoe Trodd and Alex Kent Williamson. 2009. Modern slavery: The secret world of 27 million people. Oneworld Publications Limited.

Baron, Reuben M and David A Kenny. 1986. “The moderator–mediator variable distinction in social

psychological research: Conceptual, strategic, and statistical considerations.” Journal of personality and social psychology 51(6):1173.

Beber, Bernd and Christopher Blattman. 2013. “The logic of child soldiering and coercion.” Interna- tional Organization 67(1):65–104.

Besley, Timothy and Torsten Persson. 2010. “State capacity, conflict, and development.” Econometrica 78(1):1–34.

30 Bloom, J. 1998. The End of the Old Order in Rural Europe. Princeton: Princeton University Press.

Bobbio, Norberto. 1996. Left and right: The significance of a political distinction. University of Chicago Press.

Brasil. 1940. “Decreto-Lei nº 2.848, de 7 de dezembro de 1940. Código Penal.” Diário Oficial da União .

Breman, Jan et al. 2007. Labour bondage in West India: from past to present. Oxford University Press New Delhi.

Buonanno, Paolo and Juan F Vargas. 2019. “Inequality, crime, and the long run legacy of slavery.”

Journal of Economic Behavior & Organization 159:539–552.

Chwe, Michael Suk-Young. 1990. “Why were workers whipped? Pain in a principal-agent model.” The Economic Journal 100(403):1109–1121.

Curtin, Philip D and Philip DeArmond Curtin. 1998. The rise and fall of the plantation complex: essays in Atlantic history. Cambridge University Press. da Silva, Bráulio Figueiredo Alves. 2018. Mensurando o Trabalho Escravo no Maranhão. Relatório final Organização Internacional do Trabalho (OIT) Criminalidade e Segurança Pública (CRISP/UFMG) Universidade Federal de Minas Gerais (UFMG): .

Davis, David Brion. 2006. Inhuman bondage: The rise and fall of slavery in the new world. Oxford Uni- versity Press.

Dell, Melissa. 2010. “The persistent effects of Peru’s mining mita.” Econometrica 78(6):1863–1903.

Diamond, Larry, Leonardo Morlino et al. 2005. Assessing the quality of democracy. JHU Press.

Domar, Evsey D. 1970. “The causes of slavery and serfdom.” Economic History Review 30(1):18–32.

Engerman, Stanley, Kenneth L Sokoloff et al. 2000. “Institutions, factor endowments, and paths of

development in the new world.” Journal of Economic Perspectives 14(3):217–32.

31 Finkel, Evgeny, Scott Gehlbach and Tricia D Olsen. 2015. “Does reform prevent rebellion? Evidence

from Russia’s emancipation of the serfs.” Comparative Political Studies 48(8):984–1019.

Fogel, Robert William and Stanley L Engerman. 1995. “Engerman. 1974.” Time on the Cross: The Eco- nomics of American Negro Slavery 2.

Foster, Sheila and Daniel Bonilla. 2011. “The social function of property: a comparative law perspec-

tive.” Fordham Law Review 80:101.

Gingerich, Daniel and Jan P Vogler. 2020. “Pandemics and Political Development: The Electoral Legacy

of the Black Death in Germany.” Available at SSRN 3578732 .

Girardi, Eduardo Paulon. 2014. “Dataluta: banco de dados da luta pela terra.” Universidade Estadual Paulista (UNESP), Presidente Prudente, São Paulo.

Hernandez, Diego and Alexandra Rudolph. 2015. “Modern day slavery: What drives human trafficking

in Europe?” European Journal of Political Economy 38:118–139.

Hidalgo, F Daniel, Suresh Naidu, Simeon Nichter and Neal Richardson. 2010. “Economic determinants

of land invasions.” The Review of Economics and Statistics 92(3):505–523.

ILO. 1930. “International Labour Organization — Forced Labour Convention, 1930 (No. 29).” ILO Geneva.

ILO. 2009. “International Labor Organisation - Report of the Director-General: The Cost of Coercion.” ILO Geneva.

ILO. 2017. “International Labour Organization — Global estimates of modern slavery: Forced labour and forced marriage.” ILO Geneva.

Imai, Kosuke and In Song Kim. 2019. “When should we use unit fixed effects regression models for

causal inference with longitudinal data?” American Journal of Political Science 63(2):467–490.

32 Klein, Herbert S and Ben Vinson III. 2007. African slavery in Latin America and the Caribbean. Oxford University Press.

Krull, Jennifer L and David P MacKinnon. 2001. “Multilevel modeling of individual and group level

mediated effects.” Multivariate behavioral research 36(2):249–277.

Landman, Todd. 2020. “Measuring Modern Slavery: Law, Human Rights, and New Forms of Data.”

Human Rights Quarterly 42(2):303–331.

Landman, Todd and Bernard W Silverman. 2019. “Globalization and modern slavery.” Politics and Governance 7(4):275–290.

Locke, John. 1947. Two Treatises of Government: With a Supplement, Patriarcha, by Robert Filmer. Simon and Schuster.

Lockhart, James and Stuart B Schwartz. 1983. Early Latin America: A History of Colonial Spanish America and Brazil. Cambridge University Press.

Lovejoy, Paul E. 2000. Identity in the Shadow of Slavery. Bloomsbury Publishing.

Mahmoud, Toman Omar and Christoph Trebesch. 2010. “The economics of human trafficking and

labour migration: Micro-evidence from Eastern Europe.” Journal of comparative economics 38(2):173– 188.

Mainwaring, Scott, Rachel Meneguello and Timothy Joseph Power. 2000. Partidos conservadores no Brasil contemporâneo: quais são, o que defemdem, quais são suas bases. Paz e Terra.

Marx, Karl. 1971. Theories of Surplus Value: Volume IV of Capital. Vol. 1 Progress Publishers.

McCarthy, Lauren. 2014. “Human Trafficking and the New Slavery.” Annual Review of Law and Social Science 10:221–242.

Melo, Carlos Ranulfo. 2004. Retirando as cadeiras do lugar: Migração partidária na Câmara dos Deputados, 1985-2002. Vol. 23 Editora UFMG.

33 Meltzer, Milton. 1993. Slavery: A world history. Da Capo Press.

Muthén, Bengt and Tihomir Asparouhov. 2008. “Growth mixture modeling: Analysis with non-

Gaussian random effects.” Longitudinal data analysis 143165.

North, Douglass C, John Joseph Wallis, Barry R Weingast et al. 2009. Violence and social orders: A conceptual framework for interpreting recorded human history. Cambridge University Press.

Nunn, Nathan. 2008. “The long-term effects of Africa’s slave trades.” The Quarterly Journal of Economics 123(1):139–176.

Nunn, Nathan and Leonard Wantchekon. 2011. “The slave trade and the origins of mistrust in Africa.”

American Economic Review 101(7):3221–52.

Ondetti, Gabriel. 2016. “The social function of property, land rights and social welfare in Brazil.” Land Use Policy 50:29–37.

Paige, Jeffery M. 1998. Coffee and power: Revolution and the rise of democracy in Central America. Harvard University Press.

Phillips, Nicola and Leonardo Sakamoto. 2012. “Global production networks, chronic poverty and

‘slave labour’in Brazil.” Studies in Comparative International Development 47(3):287–315.

Power, Timothy J and Cesar Zucco Jr. 2009. “Estimating ideology of Brazilian legislative parties, 1990-

2005: a research communication.” Latin American Research Review pp. 218–246.

Preacher, Kristopher J, Michael J Zyphur and Zhen Zhang. 2010. “A general multilevel SEM framework

for assessing multilevel mediation.” Psychological methods 15(3):209.

Scott, James C. 2010. The art of not being governed: An anarchist history of upland Southeast Asia. Nus Press.

Sen, Amartya. 2001. Development as freedom. Oxford Paperbacks.

Smith, Adam. 1978. Lectures on jurisprudence. Vol. 5 VM eBooks.

34 Tilly, Charles et al. 1992. Coercion, capital, and European states, AD 990-1992. Blackwell Oxford.

Weber, Max. 1978. Economy and society: An outline of interpretive sociology. Vol. 1 Univ of California Press.

Weingast, Barry R. 2016. “Persistent inefficiency: Adam Smith’s theory of slavery and its abolition in

Western Europe.” Available at SSRN 2635917 .

35 A Online Appendix – Not for Print Publication

B Chronology — 1995-2013

Table B.1: Chronology of actions, laws and public policies against slave labor in Brazil — 1995-2013

Year Event

1995 Brazil recognizes before the UN the existence of slaves in its territory. Ordinances 549 and 550, of June 14, 1995, and Presidential Decree No. 1538, of June 27, 1995, creates the Spe- cial Mobile Inspection Group (GEFM), the Secretariat of Labor Inspection (SIT), and the Executive Group for the Suppression of Forced Labor (GERTRAF). 1998 The country adopts the "ILO Declaration on Fundamental Principles and Rights at Work." Law 9,777 amends articles 132, 203, and 207 of the Brazilian Penal Code, defining crimes related to slave labor and its legal sanctions. The country decides its first final criminal sen-

tence related to slave labor (Fazenda Alvorada, Água Azul do Norte, Pará). 2002 Resolution No. 306, of November 6, 2002, establishes the Unemployment Insurance for Rescued Workers. After that, rescued slaves can receive three months of unemployment insurance. The Brazilian government also creates a Special Committee in the Council for the Defense of the Human Rights (CDDPH) in the Ministry of Justice, and the National Commit- tee for the Eradication of Slave Labor (CONAETE) in the Labor Public Prosecution Office.

( To be continued)

36 Year Event

2003 Brazil launches the First National Plan for the Eradication of Slave Labor (I PNETE), as part of the National Program for the Eradication of Slave Labor (CONATRAE). The National Plan complies with the provisions of Brazil’s National Human Rights Plan. Brazil signs a friendly settlement regarding the José Pereira case at the Inter-American Court of Human Rights. Ordinance No. 1,234 of the Ministry of Labor and Employment (MTE) creates a "Dirty List on Slave Labor." The list prevents the granting of credits and financing to slave owners. Law No. 10,803 modifies Art. 149 of the Penal Code, defining the following labor conditions analogous to slavery: i) forced labor, ii) exhaustive working hours, iii) degrading working conditions, and iv) debt bondage. 2004 Brazil recognizes before the UN the existence of at least 25 thousand people in slavery per year. During the Unaí massacre, four MTE employees are murdered during an investigation on allegations of slave labor. 2005 The Brazilian government creates the National Pact for the Eradication of Slave Labor. About 400 national and multinational companies are committed to not buying raw mate- rial from suppliers who employ slave labor. 2006 Brazil’s Supreme Federal Court decides that the Federal Justice is responsible for the trials on crimes related to modern slavery. 2008 Conatrae launches the II National Plan for the Eradication of Slave Labor. First expropriation in Brazil of a farm (Cabaceiras, in Marabá / PA) because of slave labor. After five years of experience in the first plan, on April 17, 2008, Brazil launches the II National Plan for the Eradication of Slave Labor. It updates the actions against slave labor. 2009 Brazil creates the National Day to Combat Slave Labor on the date of the Unaí massacre (January 28). The State of Mato Grosso promotes the qualification for workers rescued from slavery.

( To be continued)

37 Year Event

2010 First National Meeting for the Eradication of Slave Labor. 2013 State Law 14,946 cancels the registration of companies involved in slave labor in the State of São Paulo.

Source: da Silva(2018, p. 22-24).

B.1 OLS Panel Regressions

We start our analysis in the main paper with an OLS panel regression with fixed effects:

Yi,t = αi + γt + βTi,t + ΓXi,t + i,t, (1)

Yi,t is our outcome variable, i.e., the number of slaves rescued as a result of inspections by the Ministry

of Labor and Employment in municipality i and year t. αi represents municipality fixed effects, and γt represents year dummy fixed effects. Fixed effects models remove omitted variable bias by measuring changes within groups across time. βTi,t is our key independent variable — land invasions —, and

ΓXi,t is a set of control variables. i,t is error term.

B.2 Weighted Unit Fixed Effects Estimator

The second statistical method employed in the main paper is the Weighted Unit Fixed Effects Estima- tor. Here, we provide a summary of the method, and we illustrate the generality of the proposed framework. The reader can find more details in Imai and Kim(2019)’s work. Following the author’s, equa- tion1 can accommodate diverse weighting estimators through their corresponding matched sets Mi,t.

First, we show how to incorporate year-varying confounders Zi by weighting observations within each municipality based on the values of Zi. For example, the within-unit nearest neighbor matching leads to the following matched set:

38 0 0 0 Mi,t = {(i , t ): i = i, (2) Xi0,t0 = 1 − Xi,t,D(Zi,tZi0,t0 ) = Ji,t}

where D(.,.) is a Mahalanobis distance measure, and

Ji,t = min D(Zi,tZi0,t0 ) (3) (i,t)∈Mi,t

represents the minimum distance between time-varying confounders from observations whose treatment status is opposite — municipalities with land invasions and without land invasions. With this definition of matched sets, we can construct the within-unit nearest neighbor matching estimator using the following equation:

N T 1 X X τˆ = D (Y (1) − Y (0)), (4) PN PT i,t di,t di,t i=1 t=1 Di,t i=1 t=1 where Yi,t(x) is observed when Xi,t = x and is estimated using the average of outcomes among the units of its matched set when Xi,t = 1−x. The argument suggests that this within-unit nearest neighbor matching estimator is consistent for the Average Treatment Effect (ATE) so long as matching on

Zi,t eliminates the confounding bias.

B.3 Generalized Multilevel Structural Equation Models (GMSEMs)

The Generalized Multilevel Structural Equation Model (GMSEM) is the third method presented in the main paper. For more details, see Krull and MacKinnon(2001) for an introduction to multilevel mediation models, and Preacher, Zyphur and Zhang(2010) for a discussion of fitting these models with structural equation modeling. Put it succinctly, GMSEM consists of three matrix equations, simplified here to contain only the essential terms:

39 Yi,j = Λjηi,j

ηi,j = αj + Bjηi,j + ζi,j (5)

ηj = µ + βηj + ζj

The first two equations are, respectively, the measurement and structural models. j indexes municipality clusters, respectively. Yi,j is our outcome variable — modern slavery in municipalities. Λj is the number of random effects (latent variables). ηi,j is a vector of random effects. αj and Bj contain intercepts and path coefficients linking the latent components, and ζi,j is a vector of level-1 residu- als. The primary innovation of the General Multilevel Structural Equation Model (GMSEM) over the

Structural Equation Model (SEM) is the addition of the vector ηj, containing all the random intercepts and slopes from the previous equation. The third equation allows for the level-2 regressions of these random coefficients on one another. This general model contains all previous models for multilevel mediation as constrained individual cases and has the further benefit of providing SEM-style fit statistics in models without random slopes.

40 C Extra analysis: Descriptive Statistics and Robustness Checks

Table C.2: Summary statistics — Panel Data

Variable Obs Mean Std. Dev. Min Max Slaves 101870 .423 8.448 0 1064 Slaves (binomial) 101870 .016 .127 0 1 Invasions (count) 101870 .089 .574 0 31 Invasions (binomial) 101870 .047 .212 0 1 Illiteracy 101181 1.195 .123 .993 1.862 Log (Tax collection) 96153 12.427 2.332 0 23.575 Reforms 101870 .082 .497 0 22 Land (Gini) 101069 .713 .123 .01 .99 Murders 101870 .013 .233 0 22 Municipal guards 101870 .144 .351 0 1 Rural percentage 101180 .393 .226 0 1 Log (Population) 101870 9.376 1.114 6.48 16.285 Log (PIB per capita) 101870 1.862 .757 .01 8.309 Cattle Dependency 101870 1.703 1.196 0 10.208 Soy Dependency 101870 .099 .203 0 1 Sugar Dependency 101870 .107 .232 0 1 Coffee Dependency 101870 .056 .162 0 1 Log (N. of firms) 99505 4.02 1.729 .693 12.467 Log (N. of employees) 100688 2.106 .725 .693 8.538 Min Wages (average) 100688 1.751 1.074 0 30.165 Drinking water 99751 .759 .251 0 1 Sewage 99751 .114 .145 0 .909 Log (Vaccines) 101870 8.957 1.721 0 16.414 Mayors’ ideology 69316 2.227 .721 1 3 Reelection (mayor) 101870 1.663 .808 1 3 Election (mayor) 101870 2001.677 5.604 1992 2012 N. of parties 71110 2.614 1.044 1 14 Neigh invasions 101870 .209 .406 0 1 Neigh-of-neigh invasions 101870 .375 .484 0 1 Neigh invasions (inv) 101870 .016 .126 0 1 Neigh-of-neigh invasions (inv) 101870 .013 .114 0 1 Log (Social and pension) - spending 91350 12.945 2.048 0 22.579 Log (Education and culture) - spending 91350 14.847 1.497 0 22.86 Log (Health and sanitation) - spending 91350 14.42 1.713 0 22.748

41 Figure C.1: Slaves and land invasions over time

(a) Slaves over time

(b) Land invasions over time

Notes: The plots depict the number of slaves (panel a) and land invasions (panel b) in Brazilian municipalities from 1995 to 2013.

42 Table C.3: Mayors’ ideology

Old name Old and new name Ideology DEM DEM/PFL Right NOVO NOVO Right PAN PTB/PAN/PSD(old) Right PC do B PC do B Left PCB PCB Left PCO PCO Left PDT PDT Left PEN PEN Center PFL DEM/PFL Right PGT PR/PL/PRONA/PST/PGT Right PHS PHS Right PL PR/PL/PRONA/PST/PGT Right PMB PMB Center PMDB PMDB Center PMN PMN Center PP PP/PPB Right PPB PP/PPB Right PPL PPL Left PPS PPS Center PR PR/PL/PRONA/PST/PGT Right PRB PRB Right PRN PTC/PRN Right PRONA PR/PL/PRONA/PST/PGT Right PROS PROS Center PRP PRP Right PRTB PRTB Right PSB PSB Left PSC PSC Right PSD PSD Center PSD(old) PTB/PAN/PSD(old) Right PSDB PSDB Center PSDC PSDC Right PSL PSL Right PSOL PSOL Left PST PR/PL/PRONA/PST/PGT Right PSTU PSTU Left PT PT Left PT do B PT do B Right PTB PTB/PAN/PSD(old) Right PTC PTC/PRN Right PTN PTN Right PV PV Center REDE REDE Center SD SD Center Notes: We use the previous work on the ideology of Brazilian parties to construct our classification (Mainwaring, Meneguello and Power, 2000; Melo, 2004; Power and Zucco Jr, 2009).

43 Figure C.2: Regression lines - no controls

44 Table C.4: OLS Panel Regressions — Slaves (Multiple imputation)

(1) (2) (3) (4) (5) (6) Dependent Variable: Slaves Invasions (count) -0.142* -0.135* -0.134* -0.135* -0.134* -0.135* (0.057) (0.055) (0.055) (0.055) (0.055) (0.055) Illiteracy 0.323 0.307 0.430 0.245 0.344 (0.540) (0.548) (0.552) (0.537) (0.546) Log (Tax collection) 0.019 0.018 0.017 0.019 0.017 (0.014) (0.013) (0.013) (0.014) (0.013) Reforms 0.079 0.080 0.080 0.080 0.081 (0.103) (0.103) (0.103) (0.103) (0.103) Land (Gini) 0.993* 1.019* 1.043* 0.972* 1.027* (0.430) (0.428) (0.432) (0.427) (0.429) Murders -0.447+ -0.446+ -0.447+ -0.446+ -0.446+ (0.252) (0.251) (0.251) (0.251) (0.251) Municipal guards 0.086 0.088 0.078 0.097 0.087 (0.093) (0.092) (0.093) (0.091) (0.091) Rural percentage 0.303 0.337 0.352 0.340 0.370 (0.473) (0.469) (0.473) (0.472) (0.471) Log (Population) 0.096 0.107 0.106 0.115 0.117 (0.406) (0.409) (0.405) (0.407) (0.408) Log (PIB per capita) 0.279* 0.262* 0.296* 0.269* 0.274* (0.128) (0.128) (0.128) (0.128) (0.128) Cattle Dependency -0.076 -0.079 (0.066) (0.061) Soy Dependency -0.528 -0.544 (0.539) (0.549) Sugar Dependency 0.477 0.636+ (0.324) (0.326) Coffee Dependency -0.064 -0.227 (0.306) (0.291) Municipalities 5424 5424 5424 5424 5424 5424 N 101870 101870 101870 101870 101870 101870 Notes: All specifications include municipality and year dummy fixed effects (not shown) and cluster standard errors at municipal level (in parenthesis). +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < .001.

45 Table C.5: OLS Panel Regressions — Municipalities with at least one slave (1995-2013)

(1) (2) (3) (4) (5) (6) Dependent Variable: Slaves Invasions (count) -0.603** -0.614** -0.613** -0.634** -0.616** -0.623** (0.202) (0.202) (0.203) (0.203) (0.204) (0.204) Illiteracy -2.533 -1.462 0.817 -1.475 0.170 (5.643) (5.767) (5.812) (5.653) (5.782) Log (Tax collection) 0.055 0.048 0.044 0.047 0.043 (0.083) (0.083) (0.081) (0.082) (0.082) Reforms 0.384 0.412 0.444 0.414 0.442 (0.343) (0.338) (0.336) (0.340) (0.337) Land (Gini) 6.522* 7.077* 6.530* 6.124* 6.658* (2.699) (2.782) (2.826) (2.750) (2.833) Murders -0.494 -0.472 -0.492 -0.483 -0.476 (0.342) (0.343) (0.344) (0.346) (0.346) Municipal guards 0.375 0.587 0.306 0.639 0.557 (0.721) (0.674) (0.712) (0.668) (0.662) Rural percentage -0.364 0.775 1.019 0.489 1.334 (4.175) (4.209) (4.252) (4.161) (4.228) Log (Population) -2.181 -2.015 -2.142 -2.122 -2.056 (1.947) (1.970) (2.005) (1.995) (2.011) Log (PIB per capita) -0.368 -0.416 0.058 -0.302 -0.182 (0.871) (0.896) (0.827) (0.884) (0.867) Cattle Dependency -1.081** -0.807* (0.415) (0.373) Soy Dependency -5.182* -4.728* (2.237) (2.248) Sugar Dependency 9.635* 12.093* (4.476) (4.747) Coffee Dependency -2.135 -2.536 (3.297) (3.119) Municipalities 583 578 578 578 578 578 N 12139 11191 11191 11191 11191 11191 Notes: All specifications include municipality and year dummy fixed effects (not shown) and cluster standard errors at municipal level (in parenthesis). +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < .001.

46 States coefﬁcients 0 -.1 -.2 -.3 PI RJ AL PA AP BA ES PB PE SP TO AC CE PR RS SC MT RN RR AM MA MS RO GO MG

Years coefﬁcients 0 -.1 -.2 -.3 2011 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2012 2013

Figure C.3: Leave-One-Out Checks for the Panel OLS models. Each estimate is based on a sample that omits States and years on the x-axis. The top panel presents the analysis for OLS model dropping one State at the time. Data from Sergipe is not available because the Ministry of Labor and Employment acts independently in this State. The bottom panel leaves out one year at the time. Dots are coefficients; bars are 95% CIs. The dependent variable is the number of slaves. All models include the same control variables (not shown) of our baseline model first presented in column 2 of Table1. All specifications include standard errors cluster at the municipal level.

47 Figure C.4: Weighted Unit Fixed Effects Estimator. Dependent Variable: Slaves. Municipalities: 5360. N (nonzero weigh): 28077. All models include weighted years and municipal fixed effects. Robust standard errors, allowing for heteroskedasticity, are in parentheses. We included the same control variables (not shown) of our baseline model first presented in column 2 of Table1. +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < .001.

48 Figure C.5: Weighted Unit Fixed Effects Estimator. Dependent Variable: Slaves. Municipalities: 5360. N (nonzero weigh): 28077. All models include weighted years and municipal fixed effects. Robust standard errors, allowing for heteroskedasticity, are in parentheses. We included the same control variables (not shown) of our baseline model first presented in column 2 of Table1. +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < .001.

49 Figure C.6: Weighted Unit Fixed Effects Estimator — Slaves

Figure C.7: Weighted Unit Fixed Effects Estimator. Dependent Variable: Invasions (binomial and count). Municipalities: 5360. N (nonzero weigh): 10960. All models include weighted years and municipal fixed effects. Robust standard errors, allowing for heteroskedasticity, are in parentheses. We included the same control variables (not shown) of our baseline model first presented in column 2 of Table1. +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < .001.

50 Table C.6: OLS Panel Regressions — Land Invasions — PCA indexes

(1) (2) (3) Dependent Variable: Land Invasions Private sector index -0.003 (0.004) Public goods index 0.011* (0.005) Spending index -0.003 (0.003) Illiteracy -0.131 -0.051 -0.079 (0.097) (0.091) (0.095) Log (Tax collection) -0.002 -0.003+ -0.001 (0.002) (0.002) (0.002) Reforms 0.047*** 0.049*** 0.043*** (0.012) (0.012) (0.012) Land (Gini) 0.090 0.094 0.085 (0.062) (0.059) (0.062) Murders 0.136 0.136 0.131 (0.095) (0.097) (0.105) Municipal guards -0.013 -0.012 -0.012 (0.016) (0.016) (0.017) Rural percentage -0.134 -0.124 -0.135 (0.088) (0.085) (0.085) Log (Population) -0.003 -0.011 -0.014 (0.020) (0.019) (0.020) Log (PIB per capita) 0.001 -0.001 -0.002 (0.013) (0.012) (0.012) Cattle Dependency 0.003 0.001 0.001 (0.005) (0.005) (0.005) Soy Dependency -0.049 -0.061+ -0.051 (0.032) (0.032) (0.033) Sugar Dependency -0.022 -0.017 -0.015 (0.029) (0.029) (0.028) Coffee Dependency 0.015 0.026 0.027 (0.030) (0.028) (0.032) Municipalities 5359 5310 5353 N 93156 94033 90513 Notes: All specifications include municipality and year dummy fixed effects (not shown) and cluster standard errors at municipal level (in parenthesis). +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < .001.

51 Table C.7: GMSEMs — PCA indexes: Direct and indirect effects

(1) (2) (3) A. Dependent Variable: Slaves Invasions (count) -0.144* -0.143* -0.152* (0.056) (0.056) (0.060) Private sector index 0.283*** (0.071) Public goods index -0.154** (0.050) Spending index -0.025 (0.027) B. Dependent Variable: Invasions (count) Private sector index -0.000 (0.003) Public goods index 0.017*** (0.004) Spending index 0.001 (0.003) Municipalities 5359 5310 5353 N 93156 94033 90513 Notes: All specifications include municipality and year fixed effects (not shown) and cluster standard errors at municipal level (in parenthesis). We included the same control variables (not shown) of our baseline model first presented in column 2 of Table1. +p < .1, ∗p < .05, ∗∗p < .01, ∗∗∗p < 1.

52 Table C.8: GMSEMs — Spending: Direct and indirect effects

(1) (2) (3) A. Dependent Variable: Slaves Invasions (count) -0.152* -0.152* -0.152* (0.060) (0.060) (0.060) Log (Social and pension) - spending -0.012 (0.013) Log (Education and culture) - spending 0.021 (0.020) Log (Health and sanitation) - spending 0.011 (0.015) B. Dependent Variable: Invasions (count) Log (Social and pension) - spending 0.000 (0.001) Log (Education and culture) - spending 0.001 (0.001) Log (Health and sanitation) - spending 0.001 (0.001) Municipalities 5353 5353 5353 N 90513 90513 90513 Notes: All specifications include municipality and year fixed effects (not shown) and cluster standard errors at municipal level (in parenthesis). We included the same control variables (not shown) of our baseline model first presented in column 2 of Table1. +p < .1, ∗p < .05, ∗ ∗ p < .01, ∗ ∗ ∗p < 1.