Frontier Planters, Immigrants, and the Abolition of Slavery in Brazil

Home , Empire of Brazil, Quilombo, Quilombola, Slavery in Brazil

Frontier planters, immigrants, and the abolition of slavery in Brazil

François Seyler∗ Arthur Silve† August 13, 2020

Abstract A protracted legislative battle culminated in the abolition of slavery in Brazil in 1888. In this paper, we unbundle the material interests of individual electoral districts in the persistence of the coercion system. We build a new data set of roll-call votes on 1884-1888 emancipation bills in the legislature, and connect it to local features of the districts. In line with previous works, we find a robust positive association between local levels of slavery and the likelihood that the representative votes against emancipation. We extend this literature by considering two sources of conflict within the enfranchised elite, and thus revisit the influential Nieboer-Domar hypothesis. We find higher support for emancipation laws where slaves could more easily escape, and also where immigrants provided an alternative source of labor for landowners. A two-pronged instrumental variables strategy that leverages historical and topographic determinants of the location of slaves and maroons across space as well as heteroskedasticity with respect to the regressors supports a causal interpretation of our main results.

Key words: labor coercion, Brazil, slavery, immigration, institutional change, intra-elite conﬂict. JEL codes: J15, J47, N36, N46, O54.

∗Université Laval, Department of Economics. Email: [email protected] †Université Laval, Department of Economics. Email: [email protected].

1 Art. 1.o: É declarada extinta desde a data desta lei a escravidão no Brasil. Art. 2.o: Revogam-se as disposições em contrário.

Lei Áurea, May 13, 1888

1 Introduction

The Lei Áurea that abolished slavery in Brazil in 1888 also put an end to legal slavery in the Western Hemisphere. It was the culmination of a protracted process that had started before the foundation of the Empire of Brazil in 1822. By 1807, the United Kingdom and the United States had abolished the Atlantic slave trade, and started pushing other countries to do the same. The young empire accepted to ban the trade in 1831, but it took 20 years (and forceful action by the British crown) to get Brazil to effectively act against it (de Alencastro, 1979). Another 20 years later, the 1871 Lei do Ventre Livre (that liberated children born of enslaved mothers) temporarily placated abolitionist sentiments stirred by the American Civil War (Conrad, 1972). The question of abolition was only brought back to the forefront of the legislative agenda in the 1880s, and it took the Câmara dos Deputados (the lower house of the Brazilian General Assembly) one decade and a succession of laws (notably the 1885 Lei dos Sexagenários that emancipated slaves over 60 years old) to finally abolish slavery. With the benefit of hindsight, it is maybe easy to believe that slavery was doomed to disappear eventually1 under a combination of international and domestic pres- sures. This teleological view of history overlooks a unique combination of obstacles to abolition in nineteenth-century Brazil. The roots of slavery are deep in the Amer- icas. After the sixteenth-century repression of the Tupi-Guaraní populations along the Brazilian coastline (Schwartz, 1978), the success of large-scale sugar production and the massive arrival of African slaves marked the birth of a deeply coercive system that permeated through the New World. The Brazilian economy was entirely 1At least, legal slavery. In the Brazilian Amazon, tens of thousands of workers (mostly internal migrants from rural areas) continue to live under coercive labor arrangements. The exact number of coerced workers is not known, but according to Repórter Brasil (2015), nearly 50,000 workers were freed between 1995 and 2015. Illegal slavery is very much a modern problem, and not only in Brazil: from the kafala system in Qatar to debt-bondage in India, examples abound globally. The International Labour Organization and the Walk Free Foundation (2017) estimate that in 2016, 40 million people were victims of modern slavery, 71% of whom women and girls.

2 driven by agricultural exports and heavily dependent on captive labor. According to the latest estimates of the Slave Voyages online database (Eltis, 2007), 5,532,120 Africans arrived in Brazil between 1550 and 1866. In comparison, ‘only’ 472,382 slaves disembarked in mainland North America, nearly twelve times fewer. This economic dependence was reflected in the oligarchical political system of the Empire. Among many such statements, we can quote Conrad (1972, p. 16): “[M]uch of the real power in the provinces was in the hands of the slaveholding landlord class,” or Viotti da Costa (1989, p. 179): “[P]oliticians often represented in the Chamber, the Senate, or the Council of State the interests of plantation owners and merchants to whom they were tied by links of patronage and clientele.” Maybe the real puzzle is not why Brazil abolished slavery so late, but how it came to abolish it at all. Abolition was the result of a protracted legislative battle between members of the elite, who did not all profit from slavery equally. As the center of gravity of the economy moved from north to south, and from sugarcane to coffee,2 so did a massive number of slaves.3 It is thus maybe unsurprising that the northern elites were ready earlier for a transition to free labor, and that the elite as a whole was divided on the issue of slavery. Conrad (1972, p. 67) quotes the northern representative Araújo Lima who in 1854 established a parallel with the United States: “Be certain that you will have opposite interests, provinces with slaves, provinces without slaves (...). You will have the kind of struggles and antagonisms (...) which have placed the American Union in such imminent danger.” Even within the elite of the Centro- Sul, the coffee-growing South-Central region, frontier planters and old latifundiários started adopting somewhat antagonistic stances after the 1871 law. Frontier planters struggled to attract an adequate supply of labor for abundant land. Meanwhile, landowners in older settlement regions were quickly exhausting their land (Reis and Reis, 1988).4 The maps in figure 1 illustrate the rapid expansion of the agricultural ‘frontier’ in the province of São Paulo.

2In the second half of the century, coffee exports and export prices respectively increased by 341% and 91%, while exports of sugar continued growing by a mere 33% and prices actually decreased by 11% (Viotti da Costa, 1989). Hence, whereas sugar represented 49% of the country’s exports and coffee 19% in 1822, by 1913 sugar and cotton together accounted for less than 3% of Brazil’s exports, while coffee had gone up to 60% (Leff, 1991). 3Slaves represented 23% of the northeastern population in 1823 and less than 10% in 1872, a 30% nominal decline. In 1874, more than 50% of the country’s slave population was located in the Centro-Sul (Stein, 1957). 4The slave population decreased everywhere between 1873 and 1882, except in Rio de Janeiro’s coffee municípios (+4.85%), in the Paulista part of the Paraíba Valley (+10.54%), and in the new coffee region of São Paulo (+45.51%). In 1882, the total slave population of Rio’s coffee region comprised 156,009 individuals, 37,649 in the part of the Paraíba Valley belonging to São Paulo, and 38,242 in the new coffee region to the northwest of São Paulo (Conrad, 1972, p. 294-95).

3 Figure 1: Geographic distribution of coﬀee production in the province of São Paulo. Geo- referenced and digitized from Milliet (1941).

This article establishes the importance of intra-elite divisions as a driver of institutional persistence and change. Our three main results correspond to three sources of variation in local elites’ support for slavery. i) Elites differed in the vested interests they had in the slaves that they or their patrons detained. We show that the local prevalence of slavery was robustly associated with a legislator’s likelihood to vote against emancipation bills. To gather an abolitionist coalition, legislators from low prevalence municipalities needed to find allies. We investigate two sources of heterogeneity within high prevalence municipalities: the cost of the enforcement of slavery, and the presence of alternative sources of manual labor. ii) The enforcement of the coercive institution depended on how difficult it was for slaves to escape their condition. We show that ‘distance to freedom’ is a predictor of high-prevalence legislators’ likelihood to vote against emancipation bills. iii) There may have been local alternatives to slave labor, more in some areas than in others. We find that the local

4 presence of immigrant communities facilitated the support of emancipation bills by high-prevalence legislators. We draw from a wealth of archival records, census surveys, geo-referenced spatial sources and historical maps. We retrieve roll-call votes on each emancipation-related bill in the Empire’s three last legislatures, ie. between 1882 and 1888,5 from the annals of the Câmara dos Deputados. We match every single legislator recorded in these votes to the electoral district in which they were elected, and we construct a database of relevant descriptors of each municipality in each electoral district. To establish our ﬁrst result, we capture the vested interests of local elites using the local prevalence of slavery, as measured by the share of slaves within a district’s population according to Brazil’s 1872 demographic census (see ﬁgure 2). We address the simultaneous determination of voting decisions and of the number of slaves with a two-pronged instrumental variables strategy. First, our approach leverages two historical determinants of the location of slaves in Brazil: the distance to the caminhos do ouro, the trade routes that were built in the eighteenth century to transport gold, cattle and slaves, and the local repression (and enslavement) of natives in the sixteenth century. Second, we supplement our external instrument using internally generated heteroskedasticity-based instrumental variables. To establish our second result, we capture the local cost of coercion by a collection of measures of the distance to the nearest quilombos, communities of maroons that were able to escape and hide in the hinterland. Unlike in the United States, escaping slavery did not ‘simply’ mean heading north and trying to reach Canada (Allen, 2015). With data from Fundação Palmares (2020) and INCRA (2020), we construct a collection of measures of the distance to the nearest quilombos. Here as well, we mitigate concerns of endogeneity with the same two-pronged instrumental variables strategy, now leveraging topographical features favorable to the establishment of runaway communities in addition to heteroskedasticity-based instruments. To establish our third result, we capture the presence of núcleos coloniais, state- sponsored colonial settlements, with data from Fundação Getúlio Vargas (2016). Unlike the 30 million European immigrants that settled in the United States between 1850 and 1915 (Tabellini, 2020), immigrants in Brazil were overwhelmingly debt- bonded laborers, credit constrained and tied to the land (Rocha et al., 2017; de Souza, 2019). In the absence of a convincing external instrument, we mitigate concerns of endogeneity with a heteroskedasticity-based instrumental variables strategy. The remainder of the paper proceeds as follows. Section 2 reviews the related

5These three legislatures follow a major electoral reform in 1881 that redeﬁned voting rights and the electoral map (Lei Saraiva, Jan 9, 1881), and end with a military coup on Nov 15, 1889, which established the ﬁrst Brazilian Republic.

5 Figure 2: Enslaved population by municipalities in 1872 (Brazil, 1874). Municipality boundaries in 1872 come from IBGE (2010). literature. Section 3 provides additional elements of historical context. Section 4 describes our data sources. Section 5 presents our empirical strategy. Section 6 lays out the results. Section 7 discusses robustness. Section 8 concludes.

2 A selective review of the literature

With this paper, we make a new contribution to the old debate that surrounded the so-called Nieboer-Domar hypothesis. Domar (1970) revived a conjecture ﬁrst formulated by Nieboer (1900), by which we may never simultaneously ﬁnd in an agrarian economy “free land, free peasants, and non-working landowners.” If land is abundant and labor is scarce, individuals can only derive rents from owning land if its acquisition is restricted for workers and/or if the peasants’ freedom is restricted.

6 Brenner (1976) noted that historians had argued that the scarcity of labor after the Black Death was responsible for the progressive decline of coercive feudal arrangements in fourteenth- to sixteenth-century Western Europe, but also for the rise of the ‘second serfdom’ in late fifteenth-century Eastern Europe, in apparent contradic- tion with one another (see Aston and Philpin, 1985, for a summary of the Brenner Debate). More recently, Acemoglu and Wolitzky (2011) proposed that a shift in the land-to-labor ratio has two antagonistic effects on the prevalence of coercion: a labor demand effect, in line with Domar (1970), and an outside option effect, in line with neo-Malthusian theories of feudal decline (Postan, 1937; Le Roy Ladurie, 1969; North and Thomas, 1973). Recent empirical contributions include Fenske (2012) in southwestern Nigeria, Green (2014) in the Cape colony, and Klein and Ogilvie (2017) in Bohemia. Perhaps closest to the spirit of our work, Ashraf et al. (2018) examine the influence of capital-skill complementarity on labor emancipation in Prussia. In the second half of the nineteenth century, half of the province of São Paulo was still branded as terreno desconhecido, uncharted land, on historical maps: with plenty of free land and an open agricultural frontier, Brazil allows us to provide new evidence consistent with the existence of the outside option effect theorized by Acemoglu and Wolitzky (2011). We also contribute to the debate on the causes of the decline of the institution of slavery. Williams (1944) was first to argue that slavery had started to decline because it was simply no longer productive (his predecessors had promoted the role of humanistic sentiments). Although Fogel and Engerman (1974) defended that slavery was still productive in the antebellum South when it was abolished, other scholars have argued that its inefficiency condemned slavery to eventually disappear (see Sutch, 2018, for a related discussion). Still, this begs the question as to why the abolitionist movement managed to assemble a winning coalition at a time where slavery was still profitable to many members of the elite. Engerman (1973) suggested that once a coercion system is established, strong incentives for its perpetuation arise to avoid capital loss for the slave-holders. In this paper, we argue that even if the elite collectively profits from the institution of slavery, it may face a coordination problem that eventually leads to abolition. Brazil had little international competition at the time. The country became the world’s largest coffee exporter in 1831 (Nützenadel and Trentmann, 2008; Klein, 2010), and it supplied 80% of the world’s coffee well into the 1920s. Even in the slavery-intensive coffee-growing regions, some planters felt that slavery profited their domestic competitors comparatively more than would ‘free’ labor. Finally, our work is connected to a growing literature that examines the relationship between elite behavior and coercive institutions (and long-run development).

7 Naidu and Yuchtman (2013) examined how labor demand shocks influenced working conditions under coercive legislation in industrial Britain. Bobonis and Mor- row (2014) investigated the influence of labor coercion on human capital accumulation and show that, consistently with a reduction of plantation workers’ outside options, positive shocks to coffee prices decreased public expenditure on education in nineteenth-century Puerto Rico. Dippel et al. (2020) documented how export- driven demand shocks may incentivize elites to invest in coercive institutions, and Carvalho and Dippel (2020) studied how elite identity relates to political accountability in Caribbean plantation islands. Another strand of the literature documents the persistent influence of labor coercion institutions and their extinction on development (Nunn, 2008; Nunn and Wantchekon, 2011; Dell, 2010; Acemoglu et al., 2012; Acharya et al., 2016; Bertocchi and Dimico, 2014; Markevich and Zhuravskaya, 2018). In Brazil, Summerhill (2010) found little association between colonial institutions, including slavery, and contemporary outcomes, but Naritomi et al. (2012) found that differences in sub-national colonial institutions do matter for development. More recently, Papadia (2019) investigated the influence of slavery on fiscal capacity in Brazil’s main coffee provinces, and Fujiwara et al. (2019) showed that slavery had a persistent influence on contemporary income inequality. In this quickly growing literature, we document the mechanisms by which institutions persist and change, using what is arguably one of the most important episodes of institutional change in modern history.

3 Abolition laws in Brazil6

Emperor Pedro II charged the liberal senator Sousa Dantas to constitute a new cabinet in 1884 and to move towards emancipation. The bill he presented, known as the Dantas project, gave rise to an unprecedented rallying of pro-slavery interest groups (Ridings, 1994) and a parliamentary crisis of unrivaled proportions (Conrad, 1972; Viotti da Costa, 1989). Sousa Dantas was ousted and replaced by a cabinet more amicable to slave-holders’ interests. The new cabinet proposed the Saraiva- Cotegipe bill, described as a “lame and odious caricature of the Dantas project” by leading abolitionist Joaquim Nabuco’s daughter in a biography of her father (Nabuco, 1950, p. 61). The bill faced opposition from some liberals disappointed that the law did not go far enough, from some conservatives (mostly from Minas Gerais and Rio

6For additional references and a broader discussion of the political economy of Brazil in the nineteenth century, also see Stein (1957); Leﬀ (1972, 1991); Curtin (1972); de Alencastro (1979); Reis and Reis (1988); Bethell (1989, 2002).

8 de Janeiro) opposed to any change in the existing institution, and from the hardcore slave-holders from the Paraíba valley, but it was adopted with 81% of the votes. The law, known as the Lei dos Sexagenários, was disappointing to many, but it marked the point where abolition started gaining supporters in the Centro-Sul (see figure 3). Scholars disagree about the fundamental causes of São Paulo’s gradual conversion to abolitionism. Morse (1958), Graham (1968) and Dean (2012) proposed that new planters were progressive and keen on turning to immigrant workers as a substitute to coerced ones. However, Conrad (1972) observes that the growth of the slave population was largest in newly cultivated areas of São Paulo (see also Lowrie, 1938). Be that as it may, and although Paulista representatives were still cautious in 1885, merely willing to accept the timid provisions of the Saraiva-Cotegipe bill, their conversion played a fundamental role in the abolition of slavery. In the late 1880s, emancipation movements radicalized, with frequent rebellions, scenes of violence and flights from plantation (Conrad, 1972; Reis and Reis, 1988). Military aid was sent to (reluctantly) help persecute runaways (Toplin, 1969). At the same time, efforts to attract European immigrants to work on the plantations started paying off. 6,500 immigrants entered São Paulo in 1885, 32,000 in 1887, and 90,000 in 1888 Conrad (1972). It is under these circumstances that many Paulista planters converted to abolitionism (Luna, 1976), and liberated around 100,000 slaves in the first months of 1888. By early 1888, slavery was almost extinguished in the province (and coffee production was continuing almost unperturbed). Other provinces followed in the steps of São Paulo, leaving only planters in Rio de Janeiro and a few recalcitrant latifundiários from São Paulo and Minas Gerais to defend the coercion system. The legislative session that opened in May 1888 had but one priority: bring a definitive solution to the question of emancipation. A bill proclaiming the immediate abolition of slavery in two short articles was voted in the Chamber on May 9. As illustrated in figure 4, Rio de Janeiro was by spring 1888 the very last bastion of slavery in the Empire. Of the nine legislators that voted against the bill, eight were representatives from the province’s electoral districts. At the time of abolition, Rio had remained unaffected by immigration and its planters were threatened by bankruptcy, with little more wealth than that represented by their slaves (Conrad, 1972; Viotti da Costa, 1989). The bill was sanctioned by over 90% of representatives and passed by the Senate a few days later. It was soon approved by the Princess Regent as the Lei Áurea.

9 Figure 3: Vote by district on the 1884 Dantas Project and the 1885 Lei dos Sexagenários.

10 Figure 4: Vote by district on the 1888 Lei Áurea.

4 Data

We use archival records, census surveys, historical maps, and geo-referenced data sources to conduct our empirical analysis. In order to extract information from historical sources, we use a combination of geo-referencing, optical character recognition, text mining, and manual coding when document quality leaves no other option. Ex- cept for votes, we aggregated most of our variables from the level of the municipality to the district (642 municipalities in 1872 to match with the 122 districts after the 1881 electoral reform). Table 1 provides summary statistics of our main variables. Appendix A.4 elaborates on the construction of our variables. Political variables. Our main dependent variable captures the vote of legislators in the Câmara dos Deputados. We identified 13 pieces of legislation in the annals of the chamber related to the emancipation of slaves from the onset of the eighteenth legislature (1882) to the end of the twentieth legislature (1889). Together, they constitute the universe of relevant votes, starting with the first no-confidence vote against the Dantas Cabinet in 1884 and ending with the vote for the Lei Áurea

11 in 1888. We explored the archival records of parliamentary debates and collected the vote (or absence) of all legislators on each instance of roll-call vote during that period. Jobim and Porto (1996) report the post-1881 district-level electoral division, which we geo-reference using IBGE (2010). Nogueira and Firmo’s (1973) encyclopedia of parliamentarians allows us to match each legislator to a unique district, and helps us control for individual legislators’ characteristics, such as their party affiliation. Demographic variables. One of our main explanatory variables is the share of slaves in a district’s total population in 1872, which we compute from the rather extraordinarily detailed first nation-wide demographic census in the country’s history (Brazil, 1874). In the same census, we use two variables as controls: i) the share of the non-captive colored population (defined as the sum of blacks, brown-skinned and mixed-race), in case they helped the cause of enslaved colored, and ii) the literacy rate, in case abolitionism was driven by better educated citizens. These variables are aggregated from the municipality to the district level. In section 6, we also control for the share of capitalists and factory owners to proxy for industrialization. Quilombos. Quilombos were communities founded by maroons as early as the sixteenth century (most of them before the nineteenth), that offered refuge to other runaway slaves (Anderson, 1996). We posit that the proximity of one or several quilombolas communities in the relative vicinity of a plantation facilitated evasion for the slaves and increased the costs of coercion for the slave-owners. To capture the disaggregated costs of coercion, we use the data from Fundação Palmares (2020) on all certified quilombos, as well as shapefiles from INCRA (2020). We build several proxies of the ‘distance to freedom,’ ie. the number of quilombos and the total area quilombola in the district.7 We also consider the distance from each municipality’s head town to the closest municipality with at least n quilombos, which we average at the level of the district.8 Immigration variables. The Atlas Histórico do Brasil (Fundação Getúlio Vargas, 2016) describes immigration in Brazil in a succession of waves: 1748-1800 (Azorean immigration only), 1800-1870, 1870-1930 and 1930-1970. State-sponsored colonial settlements became the preferred alternative to address the shortage of la-

7The distribution of quilombos across districts is skewed. This explains the pattern of maximum average distances as n increases (the first district of the region of Amazonas is responsible for the maximum values of the distance for all n ≥ 2, and the second for n = 1). The largest quilombo in the Fundação Palmares (2020) data set is Tambor, in the municipality of Manaus, first district of Amazonas, measuring an impressive 7197 km2. The second one, Kalunga, in the municipality of Cavalcante, second district of Goyaz, is significantly smaller, still measuring 2618 km2. 8Exploiting sub-district variation offers increased precision compared to computing distance measures from each district’s centroid.

12 bor from the 1870s onward (Martins, 1973; Rocha et al., 2017). For want of a more precise measure of the population of immigrants in each district, we use these data to measure the number of núcleos coloniais. Geographical variables. We use IIASA/FAO (2012) to build measures of rainfall and land-suitability for sugarcane and coﬀee at the district level, in combination with 1872 municipality boundaries from IBGE (2010). We compute each district’s area, population density in 1872, latitude and longitude of each district’s centroid, as well as measures of remoteness (notably the distance from each district’s centroid to the coast). We also use Nunn and Puga’s (2012) data on terrain ruggedness (TRI) at the 30”×30” grid level to construct a measure of each district’s TRI (using Riley et al., 1999) as an instrument for the location of quilombos. In secondary speciﬁ- cations, we also make use of the terrain slope index (TSI) at the district-level from IIASA/FAO (2012). Miscellaneous. We geo-reference several maps from the Atlas Histórico do Brasil (Fundação Getúlio Vargas, 2016) in order to calculate a number of distances: i) the closest supply line to eighteenth-century mining areas, averaged across municipalities, and the distance from each district’s centroid to: ii) the closest eighteenth- century diamond mine, iii) the closest abolitionist journal, iv) the closest pro- abolition club, and v) the closest clube da lavoura, pro-slavery interest groups. Fi- nally, we compute the share of each district’s area that corresponds to zones of sixteenth and seventeenth-and-eighteenth centuries Indian enslavement.

5 Empirical approach

To establish the importance of intra-elite divisions as a driver of institutional persistence and change, and of material interests in the divisions of the elite, we consider three sources of variation of local elites’ support for slavery. Our ﬁrst hypothesis focuses on the vested interests of certain regions in slavery: we expect that support for slavery goes hand in hand with the local prevalence of slavery in each municipality/district. For our second hypothesis, we consider local variations in the cost of coercion: we expect that support for slavery in high-prevalence municipalities is stronger if it is harder for slaves to escape. Finally, our third hypothesis examines the availability of alternative sources of manual labor: we expect that support for slavery in high-prevalence municipalities is stronger if there are fewer immigrant workers available locally.

13 Table 1: Summary statistics

Statistic N* Mean St. Dev. Min Max Political variables (by district × vote) Abolitionist vote 1,586 0.549 0.447 0 1 Absence on roll call day 1,586 0.190 0.393 0 1 Liberal 1,586 0.466 0.499 0 1 Demographic variables (by district) Share of slaves 122 0.147 0.091 0.013 0.534 Share of free colored 122 0.556 0.160 0.128 0.888 Share of literates 122 0.154 0.061 0.073 0.378 Share of capitalists/owners 122 0.003 0.004 0.000 0.025 Coercion variables (by district) Total number of quilombos 122 27.55 49.26 0 416 Av. dist. to closest 1-quilombo municipality 122 21.93 29.00 0.0 196.6 Av. dist. to closest 2-quilombos municipality 122 48.96 85.81 0.0 737.1 Av. dist. to closest 3-quilombos municipality 122 67.67 92.42 0.0 737.1 Av. dist. to closest 4-quilombos municipality 122 76.82 92.48 0.0 737.1 Av. dist. to closest 5-quilombos municipality 122 86.58 92.94 0.0 737.1 Av. dist. to closest 6-quilombos municipality 122 111.1 111.2 0.0 861.4 Av. dist. to closest 7-quilombos municipality 122 117.8 112.8 0.0 861.4 Av. dist. to closest 8-quilombos municipality 122 118.8 112.7 0.0 861.4 Av. dist. to closest 9-quilombos municipality 122 138.3 121.5 0.0 861.4 Av. dist. to closest 10-quilombos municipality 122 161.0 130.2 0.0 861.4 Total area quilombola 122 622.91 3,337.47 0.01 29,712.50 Immigration variables (by district) Colonial settlements during 1748-1800 122 0.11 0.59 0 4 Colonial settlements during 1800-1870 122 0.16 0.66 0 4 Colonial settlements during 1870-1930 122 0.94 3.64 0 36 Colonial settlements during 1930-1970 122 0.76 3.47 0 28 Geographical variables (by district) Coﬀee suitability index 122 26.35 12.24 0.00 50.48 Sugarcane suitability index 122 24.86 9.70 5.16 59.85 Terrain slope index 122 72.00 15.03 37.71 95.19 Terrain ruggedness index 122 0.43 0.34 0.04 2.31 Rainfall 122 1,389.9 366.5 733.7 2,506.3 Latitude 122 −43.02 5.95 −65.87 −35.04 Longitude 122 −14.54 7.83 −31.48 −2.44 Distance to the coast 122 242.62 311.95 12.08 1,825.87 Population density 122 19.53 41.41 0.02 232.30 Miscellaneous variables (by district) Av. distance to nearest gold supply road 122 229.19 276.15 2.07 1,379.93 Distance to nearest diamond mine 122 657.85 369.80 64.74 1,594.41 Share of 16th cent. Indian enslavement area 122 0.09 0.23 0 1 Share of 17-18th cent. Indian enslavement area 122 0.40 0.45 0 1 Distance to nearest abolitionist journal 122 468.71 450.21 16.73 2,980.43 Distance to nearest abolitionist club 122 278.15 229.33 10.78 1,165.07 Distance to nearest Clube da lavoura 122 396.27 395.89 6.51 2,405.81

* 122 districts × 13 laws = 1586 obs. Dist. in km, surf. in km2, dens. in km−2.

14 5.1 Empirical specifications Our baseline specifications rely on a linear probability model and within estimators with a combination of instrumental variables strategies to establish a plausibly causal interpretation of our results. Hypothesis 1. The likelihood that a legislator votes in favor of abolition bills decreases with the local prevalence of slavery. To test hypothesis 1, we use the following baseline specification

1884−1888 1872 0 Pijv = Sij β + xijvγ + δv + ζj + εijv, (1) where subscript i denotes the district, j the province in which i is located, and v the 1884−1888 vote. Pijv is the vote of the representative of district i on emancipation bill v, that we take equal to 1 when the representative votes in favor the bill, against the 1872 status quo. Sij measures the share of slaves in the population of district i. The vector xijv controls for the party of the representative of district i at the time of vote v, and an array of geographic and demographic descriptors of district i (ethnicity and literacy rates, population density, crop suitability, rainfall, distance to the coast, latitude and longitude). Coefficients δv and ζj are vote and province fixed effects (FE). Hypothesis 1 corresponds to β < 0. Hypothesis 2. The likelihood that a representative from a high-prevalence area votes in favor of abolition increases when it is easier for slaves to escape. To test hypothesis 2, we use the following specification

1884−1888 1872 1872 0 Pijv = Sij β + Oijφ + Sij × Oijψ + xijvγ + δv + ζj + εijv (2) where Oij captures the proximity of freedom for slaves in district i of province j and, correspondingly, the cost of coercion for slave-holders. Our measure of the distance to freedom for escapees relies on the location of quilombos. All other notations are the same as for hypothesis 1. The marginal effect of the proximity of freedom 1884−1888 is ∂Pijv /∂Oij = φ + Sijψ. Our focus is on the intra-elite conflict: what we predict is that slave-holders who face higher costs of coercion have a competitive disadvantage relative to other slave-holders. Consequently, hypothesis 2 corresponds to ψ > 0. Hypothesis 3. The likelihood that a representative from a high-prevalence area votes in favor of abolition increases when there are local alternatives to slave labor. To test hypothesis 3, we use the following specification

15 1884−1888 1872 X k k X k 1872 k 0 Pijv = Sij β + Nijλ + Nij × Sij µ + xijvγ + δv + ζj + εijv, (3) k k

k where Nij captures the number of núcleos coloniais located in the district i of province j during period k, with k ∈ {1748-1800, 1800-1870, 1870-1930, 1930-1970}. All other notations are the same as above. The marginal effect of the substitution with immi- 1884−1888 1870−1930 1870−1930 1872 1870−1930 grant labor is ∂Pij /∂Nij = λ + Sij µ . Again, our focus is on the intra-elite conflict: what we predict is that slave-holders who can rely more on immigrant labor, or expect to rely more soon, are more willing to abolish the institution of slavery. Hypothesis 3 corresponds to µ1870−1930 > 0. Taken together, hypotheses 1, 2 and 3 paint a nuanced picture that lends credit to certain aspects of both the Nieboer-Domar hypothesis and of neo-Malthusian theories of feudal decline. Hypothesis 1 is in line with a labor demand effect, and hypotheses 2 and 3 suggest the existence of an outside option effect. As a country with an open agricultural frontier and large stretches of unchartered territory, nineteenth- century Brazil is an ideal case-study to verify, on the one hand, that landowners needed slavery to sustain their status and, on the other hand, that intra-elite conflict was instrumental to ending slavery. This is not to say that slave-holders were not conscious that slavery was condemned to disappear eventually. Still, empirical val- idation of the three hypotheses would suggest that among slave-holders, those who were facing higher costs of coercion and those who could count on alternative forms of labor found it in their best interest to hasten the transition to free labor.

5.2 A discussion of identification Identification is complicated for each of our three main specifications by the non- random allocation of slaves, quilombos, and núcleos coloniais across municipalities (and possibly by measurement errors, a common issue with nineteenth-century census data). The coefficient β is a biased estimate of the effect of vested interests on voting patterns if, because districts are abolitionist, they have fewer slaves (reverse causality), or if omitted variables (for instance deeply rooted norms) determine abolitionism and the local prevalence of slavery simultaneously. The coefficients φ and ψ are biased estimates of the effect of evasion on voting patterns if quilombos tend to be located in abolition-friendly districts or, conversely, if the lower prevalence of slavery in abolitionist districts meant that fewer quilombos needed to be established. Possibly, abolitionism and the location of quilombos are also simultaneously determined. Moreover, the arrival of migrants, captured with the establishment of

16 state-sponsored colonies, is hardly separable from voting patterns on emancipation laws, so that λ1870−1930 and µ1870−1930 may also be biased estimates of the effect of immigrants on voting patterns. It is apparent from the annals of the Câmara dos Deputados that questions of immigration and abolition were closely linked. From the 1880s onward, the planters themselves exerted much effort to attract immigrant workers (Conrad, 1972). Moreover, only a fraction of immigrants settled in núcleos coloniais, and we cannot control for a selection effect in our data. To address this, our baseline specifications exploit within-province variation, which means that variations in norms, culture and other unobserved variables across provinces are captured in our province FE. Additionally, we capture variation across bills with our vote FE. Consequently, the specifics of each bill in our data are (almost) irrelevant.9 Hence, a causal interpretation of our coefficients of interests is equivalent to assuming that the allocation of slaves, quilombos and state-sponsored colonies within provinces and bills is as good as random, conditionally on a wide array of political, demographic and geographical controls. There are two remaining issues for identification. For hypothesis 1, there may be a concern that other within-province district-specific determinants of slavery in 1872 also affect voting patterns in 1884-1888. Similarly for hypothesis 2, within-province district-specific determinants of the location of quilombos (and of their survival to- day) may affect voting patterns. For hypothesis 3, it it clear that núcleos coloniais and voting patterns on emancipation laws are tightly linked. For the three hypotheses, we are also concerned with reverse causality (abolitionist landowners having fewer slaves, allowing more quilombos or making quilombos redundant, núcleos coloniais distributed as rewards or punishment for voting patterns among districts, etc.). For hypotheses 1 and 2, we address these issue using a two-pronged instrumental variables strategy. We leverage long-predating historical determinants of the location of slaves across Brazil and topographic determinants of the location of quilombos to build instrumental variables that plausibly respect the exclusion restriction. We then supplement these instruments with heteroskedasticity-based instruments built as simple functions of the model’s data, following the procedure of Lewbel (2012). These generated instruments are in general not required (external instrument suf- fice), but they help with within-province identification and significantly improve our estimates’ precision, in addition to allowing us to run overidentification tests whenever external instruments would leave our models just-identified. In the absence of

9The specifics of each bill still matter if legislators decide to oppose an emancipation bill that they judge insufficient. In that case, we are misinterpreting the voting decision of the most active abolitionists, and our estimation of β should be interpreted as a conservative estimate of the effect of prevalence on voting patterns.

17 a convincing external instrument of the location of núcleos coloniais, we only use heteroskedasticity-based instruments to identify equation 3. More speciﬁcally, we run the following ﬁrst-stage regressions:

1872 0 S S S S Sijv = xijvσ + δv + ζj + ξijv 1872 O 0 O O O O Oijv = Sij β + xijvσ + δv + ζj + ξijv (4) 1870−1930 1872 N 0 N N N N Nijv = Sij β + xijvσ + δv + ζj + ξij . 0 ˆ The estimated residuals are then used to create instruments as (xijv −xijv) ξijv, where xijv is the mean-centered vector xijv. Identification requires that the error terms be heteroskedastic, which we verify using Breusch-Pagan tests. Out of this large array of possible instruments, we select those that are empirically the most relevant. We use geographical coordinates and potential yield to build the corresponding internal instruments for the share of slaves in 1872; we use geographical coordinates, potential yield and/or rainfall to build the corresponding instruments for the location of quilombos; and rainfall with population density for the location of núcleos coloniais (as well as, in some specifications, literacy rates and distance to the coast). Our external instrument for the share of slaves in equation 1 interacts the average distance from municipalities’ head towns to the nearest eighteenth-century caminho do ouro and an indicator variable for the enslavement of Indians in the sixteenth century. The former captures the intensive margin of the distribution of slaves across districts, while the latter captures its extensive margin.10 First, in the eighteenth century, major trade routes known as caminhos do ouro were established between the current province of Minas Gerais (literally, ‘General Mines’) and coastal areas throughout the country (Zemella, 1951). They were made necessary by the discovery of large alluvial deposits towards the end of the seventeenth century (soon followed by diamonds), and they also helped ensure the continuous supply of mining areas in slaves (from the north) and cattle (from the south). They were built by slaves, making the distance to these roads an interesting instrument of the prevalence of slavery into the nineteenth century (Klein and Luna, 2009). We provide an unconditional plot of this relationship in figure A.5 in the Appendix A.1. Second, when sixteenth-century Portuguese settlers were laying the foundations of the plantation system, they first started experimenting with an enslaved Indian labor force. At its peak in the 1560s, Indian slavery counted tens of thousands of individuals. Because of a combination of widespread epidemics, continuous conflict with free Indian peoples, and increasing discomfort from the Crown with Indian enslavement after the Valladolid debate, the Portuguese turned away from Indian to African slave labor (Klein and Luna, 2009).

10Map A.12 in Appendix A.3 oﬀers a visual representation of this instrument.

18 To be valid, this instrument should not be correlated with the decision to vote in favor or against abolition in the 1880s, except through its influence on the local prevalence of slavery. First, whereas the coffee planter elite’s interests played a fundamental role in perpetuating the coercion system in the second half of the nineteenth century, they are plausibly orthogonal to both Indian repression in the sixteenth century and mining interests in the eighteenth century. The emergence of the mining interests had large consequences on the configuration of economic activity, even leading to the relocation of the colony’s capital from Salvador to Rio de Janeiro in 1763. However, the economy of the region was declining in the early nineteenth century, only to be revived in its second half by the cultivation of coffee and, to a lesser extent, sugar. Second, we might be concerned with Indian repression and/or the location of the caminhos do ouro having other impacts on the location of economic activity in the 1880s, besides the prevalence of slavery. This is less of an issue within provinces, and we control for agricultural interests with land suitability controls for the two main export crops, and for other aspects of economic activity with additional geographical covariates (population, density, average rainfall, distance to the coast, and geographical coordinates). With the Indian repression dummy, we filter the risk that the distance to the gold paths predicts economic activity rather than the prevalence of slavery. In table 7 in the Appendix A.1, we also flexibly control for the distance to eighteenth-century diamond mines. Finally, we may worry that a ‘preference for slavery’ or a ‘culture of abolitionism’ predated the enslavement of Indians and the establishment of the caminhos do ouro, and persisted into the nineteenth century. This seems implausible: it was northern districts that held the highest number of slaves during the sugar boom, well into the nineteenth century. Unlike in the United States, slavery was never defended in Brazil on the grounds that it was a positive institution (Klein and Luna, 2009). Even the slaveholding elite appeared to defend slavery only as a necessary evil. Our main external instrument for the location of quilombos in equation 2 is the average terrain ruggedness. There is little doubt that the decision of where to establish quilombos was driven by considerations of security and remoteness. Nunn and Puga (2012) showed that in Africa, terrain ruggedness discouraged slave trades and facilitated escape. In Brazil, Klein and Luna (2009, p. 195) point out that a per- manent escape depended on “the existence of dense forests or inaccessible mountains within a short distance from their homes.” Again, we provide an unconditional plot of the relationship between the location of quilombos and the TRI in figure A.5 in the Appendix A.1. To be valid, this instrument should not be correlated with the decision to vote in favor or against abolition in the 1880s, except through its influence on the location

19 of quilombos. A concern is that the TRI tends to be negatively associated with economic outcomes at the country level, because it increases transport costs and makes trading more difficult (Nunn and Puga, 2012). This is however unlikely to be an issue within provinces and at a highly disaggregated level. In addition, determinants of economic activity are captured by our control variables (in particular ethnicity and literacy shares, crop suitability measures, population density, average rainfall, distance to the coast, coordinates and the share of slaves). 1870−1930 Our efforts to offset the endogeneity of Nijv in equation 3 are limited to heteroskedasticity-based instruments, and we take the resulting evidence as sugges- tive, rather than conclusive. However, because we are interested in heterogeneity in the voting behavior of legislators representing high prevalence districts, this is not necessarily an issue. Indeed, exerting efforts to attract immigrant workers was one of the strategies adopted by a number of planters and their representatives to address labor shortage and the possible end of the coercion system. In fact, we would be keen on estimating the strength of these political pull factors if there was a credible way to isolate them. Hence, regardless of the direction of causality, we are interested in the sign and magnitude of the association between 1870-1930 sponsored settlements and abolition voting decisions. Our preferred specifications also include others waves of immigration to verify that immigrant labor was a substitute to slave labor only during the relevant period.

6 Main results

Before turning to the results, we brieﬂy discuss the issue of auto-correlation. Through- out this section, we use two-way clustered standard errors. Clustering at the level of the district allows all the votes of a district’s representative to be correlated non- parametrically. Clustering at the level of the vote allows arbitrary dependence between districts for each bill.

6.1 Anti-abolition voting and local preponderance of slavery In table 2, we report the results for hypothesis 1. Columns 1 to 5 present the result of our baseline OLS specification with province and vote FE. In column 1, we report the ‘raw’ influence of the prevalence of slavery on our binary abolitionist voting outcome. In column 2 to 4, we consider how this association is affected by the introduction of political, demographic, and geographical controls. In column 5, we take into account the whole set of controls. Column 6 presents the result of our 2SLS specification, using both external and heteroskedasticity-based instruments.

20 Table 2: H1 – Prevalence of slavery and voting decisions

1(Abolition vote) OLS 2SLS (1) (2) (3) (4) (5) (6) Share of slaves −1.113∗∗∗ −0.702∗∗∗ −0.903∗∗ −1.079∗∗∗ −0.704∗∗∗ −0.724∗∗ (0.341) (0.209) (0.357) (0.373) (0.245) (0.304)

Controls None Pol. Dem. Geo. All All Province & vote FEs Yes Yes Yes Yes Yes Yes Observations 1,586 1,586 1,586 1,586 1,586 1,586 R2 0.196 0.323 0.201 0.203 0.327 0.327 1st stage F (standard iid) 159.7*** 1st stage F (HAC excl. inst.) 5.72*** Breusch-Pagan p-value < 2.2e-16 Wu-Hausman p-value 0.96 Sargan-Hansen p-value 0.57 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

In all specifications, a higher prevalence of slavery in a district is robustly associated with a higher probability that the district’s representative votes against abolitionist bills. This relationship is robust to the introduction of all controls in the OLS specification. Its magnitude decreases when we introduce political controls. In our sample, 46% of members of the Câmara dos Deputados are Conservatives, almost 47% are Liberals (the party of abolition), and around 2% are Republicans (the remainder’s affiliation – about 5% – is unknown, sometimes because legislators simultaneously occupy Ministerial positions and are thus unaffiliated). Since liberal representatives tended to be elected in districts with a lower prevalence of slavery (and conservatives in high-prevalence districts), it is unsurprising for part of the effect observed in column 1 to be captured by political affiliation. The magnitude of the relationship also seems to decrease when we introduce demographic controls, which supports our conjecture that coerced workers are better represented in districts with a higher share of free colored population and with a higher literacy rate. Per- haps more surprisingly, the magnitude of the relationship does not decrease when we consider geographic controls (log of the population density, land suitability indices, average annual rainfall, coordinates, and distance to the coast). Their effect is actually not significant, suggesting that inasmuch as geography plays a role in the voting patterns in 1884-1888, this role is captured (within provinces) by the prevalence of slavery in 1872. We interpret columns 4 and 5 as giving two alternative estimates of our main effect. The political affiliation of a representative is possibly endogenously determined

21 by both the share of slaves and the vote for or against abolition in this regression (as are, to a lesser degree, demographic variables). We want to take column 4’s estimate if we believe demographic and political variables to be colliders, and column 5 if they are confounders. According to column 5, which includes the whole set of controls and FE, a 1 percentage point increase in the share of slaves in the 1872 population is associated with a .70 percentage point increase in the probability to vote against an emancipation bill. The share of slaves ranges from 0.0% in the first district of the Amazonas province to 53.4% in the tenth district of Rio de Janeiro, with a standard deviation of 9.1%: these results imply that a standard deviation increase in the share of slaves increases a representative’s probability to vote against emancipation by 6.4 percentage points. According to column 4, a standard deviation increase in the share of slaves increases the probability by 9.8 percentage points. We keep column 5 as our preferred, most conservative specification. This relationship is also robust to instrumenting the share of slaves in 1872 by our external instrument (distance to nearest caminho do ouro × Indian repression) supplemented by heteroskedasticity-based instruments.11 The external instrument alone becomes weak once province FE are added, as reported in table 6 in Appendix A.1.12 However, in combination with internal instruments, the first stage regression shows a comfortably large association between the prevalence of slavery and our set of instruments (a heteroskedasticity- and autocorrelation-consistent [HAC] joint significance test on excluded instruments yields a p-value below 0.0001).13 The strength of generated instruments depends on the degree of heteroskedasticity with respect to first stage regressors, which we verify using a Breusch-Pagan test. This confirms the relevance of using heteroskedasticity-based instruments to supplement our external instrument. The (non-significant) value of the Sargan-Hansen test statistic implies that over-identifying restrictions cannot be rejected (i.e. the joint validity of our instruments cannot be rejected). 2SLS estimates are somewhat larger than their OLS analog in column 5. This is consistent with a small attenuation bias in our OLS estimates, as could for instance be the case if the share of slaves was mismeasured in the 1872 census. Our 2SLS results imply that a standard deviation increase in the share of slaves increases a representative’s probability to vote against emancipation by 6.6 percentage points.

11The Durbin-Wu-Hausman test, in combination with the Sargan-Hansen test, suggests that we cannot reject the hypothesis that the prevalence of slavery is exogenous, making the OLS specification consistent, and more efficient than the 2SLS. 12Table 6 also suggests that alternative definitions of the external instrument yield more precise estimates of comparable magnitude. We choose to present the instrument with the easiest interpretation as a baseline, despite the loss in precision. 13Table 5 in Appendix A.1 provides additional details

22 6.2 Anti-abolition voting and the cost of coercion In table 3, we report the results for hypothesis 2. Columns 1 to 4 present the result of our baseline OLS specification with province and vote FE, with four different measures of the ‘distance to freedom.’ In column 1, our preferred specification, we capture how easy it is for slaves to escape by the area of all quilombos in the district. In columns 2, 3, and 4, we capture how hard it is for slaves to escape by the (log of the) average distance the nearest, nearest four, and nearest seven quilombos respectively. Columns 5 to 8 present the result of the 2SLS specification with the same four measures, using both the external and the heteroskedasticity-based instruments. In all specifications, we include province and vote FE and the whole set of controls. In most specifications, distance to freedom is robustly associated with a higher probability that the district’s representative votes against abolition bills (in districts with a high prevalence of slavery). In all four OLS specifications, the relationship has the right sign, although it is not statistically significant when we take the distance to the nearest quilombo as a measure of the distance to freedom. In the other three specifications, however, the relationship is significant even with the whole set of controls and both province and vote FE. According to column 1, a 1 km2 (247 acres) increase in the area of a district’s quilombos is associated with a −.000019 + .00059 × 50% = .028 percentage point increase in the probability to vote in favor of an emancipation bill in a district with a 50% prevalence of slavery, and a .0013 percentage point increase in the probability to vote against an emancipation bill in a municipality with a 1% prevalence. The total area of quilombos in a district ranges from .006 km2 (1.5 acres) to 29,713 km2 (7.3 million acres). As the distribution is skewed (see figure A.6 in Appendix A.2), instead of the standard deviation, we consider the interquartile range [IQR] (131.5 km2) as the relevant unit of comparison (in table 13 in Appendix A.2, we verify that the regression is not driven by outlier districts). An increase in the area of quilombos by the IQR in a district with a 50% prevalence increases a representative’s probability to vote in favor of abolition by 3.6 percentage points. This relationship is also robust to instrumenting the location of quilombos with our external instrument (terrain ruggedness) supplemented by heteroskedasticity- based instruments, as well as to instrumenting the location of quilombos with our extended external instrument (terrain ruggedness and slope). The external instruments alone (barely) provide enough association to identify the model with all FE, as we report in table 7 in Appendix A.1, but it does suggest that the OLS regression is not consistent (the Wu-Hausman test statistic is significant). As before, the combination with internal instruments yields a reassuringly large association (the p-value for our HAC joint significance test on excluded instruments is below .0005).

23 Table 3: H2 – Outside options and voting decisions

1(Abolition vote) OLS 2SLS (1) (2) (3) (4) (5) (6) (7) (8) Share of slaves −0.757∗∗∗ −0.528∗∗ −0.392 −0.098 −0.810∗∗∗ −0.221 −0.101 0.306 (0.255) (0.242) (0.280) (0.379) (0.258) (0.242) (0.357) (0.391)

Quilombo area −1.9e-05 −3.0e-05 (2.5e-05) (2.7e-05) Share of slaves × Quilombo 5.9e-04 ∗∗ 1.1e-03∗∗∗ area (2.8e-04) (4.2e-04)

Ln av. dist. closest 1 0.0048 0.0036 quilombo-muni. (0.0066) (0.0093) Sh. of slaves × Ln av. dist. −0.059 −0.128∗∗ closest 1 quilombo-muni. (0.038) (0.050)

Ln av. dist. closest 4 0.0071 0.015∗ quilombos-muni. (0.0064) (0.008) Sh. of slaves × Ln av. dist. −0.074∗ −0.146∗∗∗ closest 4 quilombos-muni. (0.040) (0.055)

Ln av. dist. closest 7 0.034∗∗∗ 0.052∗∗∗ quilombos-muni. (0.013) (0.015) Sh. of slaves × Ln av. dist. −0.155∗∗ −0.260∗∗∗ closest 7 quilombos-muni. (0.067) (0.073)

Controls & FEs All All All All All All All All Observations 1,586 1,586 1,586 1,586 1,586 1,586 1,586 1,586 R2 0.328 0.330 0.329 0.330 0.327 0.318 0.328 0.329 1st stage F (standard iid) 441.3*** 37.96*** 95.75*** 186.7*** 1st stage F (HAC excl. inst.) 72.89*** 3.65*** 26.11*** 42.41*** Breusch-Pagan p-value < 2.2e-16 < 2.2e-16 < 2.2e-16 < 2.2e-16 Wu-Hausman p-value 0.21 0.11 0.003 0.02 Sargan-Hansen p-value 0.22 0.23 0.28 0.65 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 For multiple endogenous variables, we always report the most conservative diagnostic tests’ statistics.

Again, we verify the presence of heteroskedasticity with respect to the regressors in the first-stage regression, confirming the relevance of using heteroskedasticity-based instruments. The (non-significant) value of the Sargan-Hansen test statistic indicates that over-identifying restrictions cannot be rejected. Although the null hypothesis of the Wu-Hausman test fails to be rejected in our preferred specification (indicating that we cannot conclude that the OLS estimator is not consistent), the variation in our estimated effect between the OLS and the 2SLS specifications with and without internal instruments invites to a cautious interpretation of its magnitude. According to column 5, our preferred specification, an increase in the area of quilombos by the

24 IQR in a district with 50% prevalence of slavery increases a representative’s probability to vote in favor of abolition by 7.1 percentage points. The higher magnitude of the eﬀect in the 2SLS estimation (with the four measures of distance to freedom) suggests that the OLS estimates might be downward biased by a reverse causation eﬀect: more abolitionist districts may have had, as a result of their culture or institutions, less need for quilombos (even controlling for the prevalence of slavery), or more abolitionist legislators may have self-selected into districts with larger/closer quilombos, possibly more open to abolition.

6.3 Anti-abolition voting and immigration In table 4, we report the results for hypothesis 3. Columns 1 to 3 present the result of our baseline OLS specifications, including all controls and vote FE. Column 2 differs from column 1 simply by the exclusion of province FE. Column 3 includes immigrant settlements from several periods. Columns 4 to 6 present the result of the 2SLS specifications, column 4 with all controls and vote FE, column 5 and 6 without province FE, and column 6 with immigrant settlements from several periods. For the 2SLS specifications, we rely on heteroskedasticity to instrument the number of immigrant settlements between 1870 and 1930. A larger presence of immigrants in a district with a high prevalence of slavery is associated with a higher probability that the district’s representative votes in favor of abolition bills.14 This relationship is robust to the introduction of all controls and of vote FE. It is not robust to the introduction of province FE, although, if anything, it keeps the sign predicted by hypothesis 3. With a p-value of .11, it is not robust either to controlling for other waves of immigrant settlement (as described by Fundação Getúlio Vargas, 2016). However, only the 1870-1930 period of interest comes close to a significant relationship with the abolitionist vote, as predicted by the hypothesis. Earlier settlements have an insignificant and reversed relationship with voting patterns in 1884-1888. If we were to interpret these coefficient, they would be consistent with two alternative historiographical accounts: i) earlier, better educated European immigrants made long-lasting investments in human capital (Rocha et al., 2017) and may have had a positive influence on the intellectual elite in Brazil, as suggested by Conrad (1972); ii) employment of immigrants in agriculture may have made the interests of the slave-holders less salient. If we take the estimate of column 2, an additional settlement in the 1870-1930 wave is associated with −.021 + .239 × 50% = 9.8 percentage points increase in the probability that a representative from

14Table 8 in the Appendix A.1 provides additional speciﬁcations, including versions with no interacted term.

25 Table 4: H3 - Immigration and voting decisions, colonial settlements

1(Abolition vote) OLS 2SLS (1) (2) (3) (4) (5) (6) Share of slaves −0.697∗∗∗ −0.611∗∗∗ −0.580∗∗∗ −0.698∗∗∗ −0.599∗∗∗ −0.576∗∗ (0.245) (0.198) (0.224) (0.242) (0.201) (0.224)

1748-1800 settlements 0.068∗ 0.058 (0.039) (0.043)

1800-1870 settlements 0.006 0.019 (0.031) (0.044)

1870-1930 settlements 0.001 −0.021∗∗ −0.013 −0.005 −0.030∗∗ −0.032 (0.017) (0.011) (0.012) (0.028) (0.012) (0.020)

1930-1970 settlements −0.026 −0.018 (0.031) (0.029)

Share of slaves × 1748- −0.210 −0.146 1800 settlements (0.208) (0.187)

Share of slaves × 1800- −0.052 −0.113 1870 settlements (0.100) (0.159)

Share of slaves × 1870- 0.073 0.239∗∗ 0.159 0.144 0.361∗∗ 0.333∗ 1930 settlements (0.188) (0.120) (0.100) (0.333) (0.158) (0.191)

Share of slaves × 1930- 0.281 0.208 1970 settlements (0.356) (0.325)

Controls & Vote FEs All All All All All All Province FEs Yes No No Yes No No Observations 1,586 1,586 1,586 1,586 1,586 1,586 R2 0.329 0.310 0.313 0.329 0.309 0.313 1st stage F (standard iid) 231.1∗∗∗ 326.6∗∗∗ 350.5∗∗∗ 1st stage F (HAC excl. inst.) 346.1∗∗∗ 219.9∗∗∗ 62.27∗∗∗ Breusch-Pagan p-value < 2.2e-16 < 2.2e-16 < 2.2e-16 Wu-Hausman p-value 0.25 0.33 0.40 Sargan-Hansen p-value 0.59 0.56 0.15 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 For multiple endogenous variables, we always report the most conservative diagnostic tests’ statistics. a 50% prevalence district votes in favor of emancipation. This value is consistent with the historiography of that period: in the 1870-1930, a large share of immigrants settled in the region of São Paulo, whose representatives were instrumental in tipping the scales in favor of abolition. That sponsored settlements seem to have had limited inﬂuence within provinces is not overly surprising. Again, historiographical accounts attribute to the arrival

26 of immigrants at the very end of the 19th century part of the responsibility for the conversion of planters from the province of São Paulo, and for finally isolating representatives from the province of Rio de Janeiro in their unconditional defense of the coercive institution (Conrad, 1972; Klein and Luna, 2009). We may thus expect the lion’s share of the influence of the (prospect of the) arrival of immigrants on voting decisions to occur across provinces and remain relatively homogeneous within provinces. Furthermore, only a limited number of provinces ended up receiving núcleos coloniais (see figure A.11 in Appendix A.3) so that these estimates rely on limited variation and their precision is constrained by available data. Nonetheless, the relationship that we find without province FE is slightly im- proved by instrumenting the location of núcleos coloniais between 1870 and 1930 with heteroskedasticity-based instruments (the previous two sections justify the advantage of working with such instruments). Again, we verify that heteroskedasticity with respect to first stage covariates is high enough to warrant the use of internal instruments. The statistic of the Sargan-Hansen test gets higher when the four waves of settlement are considered, but not high enough to reject over-identifying restrictions. According to column 5, 2SLS estimates imply that an additional settlement in the 1870-1930 wave is associated with a 15 percentage point increase in the probability that a representative from a 50% prevalence district votes in favor of emancipation. Again, the larger magnitude of the effect in the 2SLS estimation suggests possible reverse causation: more abolitionist districts may have had, as a result of culture or institutions, less need for núcleos coloniais (although Wu-Hausman test statistics are never high enough for the null to be rejected under standard thresholds).

7 Robustness

In this section, we brieﬂy address i) additional factors that may explain variations in the abolitionist vote, ii) a possible concern that our results may be driven by a selection of legislators through their presence during the vote, or by a selection of emancipation bills, iii) alternative speciﬁcations of the spatial correlations of standard errors, and iv) a discrete choice model.

7.1 Clubs, newspapers, and industrialization The historiography of abolition in Brazil mentions the importance of interest groups, whether in favor of abolition or pro-slavery, and activist newspapers, in pushing forward (or hindering) the abolitionist cause (Conrad, 1972; Ridings, 1994; Klein and Luna, 2009). Previous works in diﬀerent settings have also emphasized the possible

27 influence of industrialization in the transition to free labor (Ashraf et al., 2018). In table 9 in Appendix A.1, we report ‘naive’ OLS estimates of the relationship between the abolitionist vote and the distance to an abolitionist journal (columns 1 and 2), an abolitionist club (column 3), a pro-slavery club (columns 4 and 5), and the share of capitalists (industrialists) in the district. In all specifications, we include the full set of controls, the prevalence of slavery, and vote and province FE. As all of these new covariates are likely to be ‘bad controls’ if we meant to capture the marginal effect of slavery prevalence on the vote (i.e. to the possible exception of the share of capitalists, they are highly endogenous), some caution is warranted when interpreting the corresponding coefficients. We do find interesting relationships, although mostly driven by low prevalence districts. If anything, the proximity of an abolitionist newspaper is associated with a higher likelihood to vote in favor of abolition. This association is driven by low- prevalence districts. Conversely, the proximity of a pro-slavery clube da lavoura is associated with a higher likelihood to vote against abolition. Again, this association is driven by low-prevalence districts. We find no equivalent association for abolitionist clubs. Perhaps surprisingly, a larger presence of capitalists is associated with a higher likelihood to vote against abolition. This suggests that Brazil was still in the early phases of its industrialization process, so that the existence of a capital-skill complementarity effect in the spirit of Ashraf et al. (2018) is unlikely to have played a significant role. As these relationships are undoubtedly endogenous, we make no claim that these relationships are causal.

7.2 Legislators and emancipation bills We may be worried that the decision to call the roll on an emancipation-related bill was in some way dependent on the anticipated distribution of the votes. However, bills relating to emancipation were particularly sensitive, strongly publicized and surrounded by heated debate, so that the roll was called for every emancipation bill that passed through the Câmara dos Deputados during our sample’s time frame. More pointedly, the bills discussed in parliament were sometimes divided in several sub-parts, which may individually be the object of nominative voting. Thus, our results could still suﬀer from a selection bias if the bills voted in several parts presented systematic idiosyncratic characteristics skewing the results to a particular direction (e.g. this could be the case if bills were systematically decomposed into individual sub-parts when they were particularly divisive in the Chamber). To account for this possibility, in table 10 in Appendix A.1, we report our main speciﬁcations using only the three most important emancipation bills. In doing so, we reduce the size of

28 our sample from 1586 to 366 district × bill observations. Despite some power loss, this does not change substantially our results’ interpretation and suggests that our conclusions are not driven by a specific selection of bills. We may also be worried that legislators from specific districts would display consistent patterns of absence and abstention on (scheduled) bills related to emancipation, thus biasing our results. For example, it is possible that some legislators supporting a given side were systematically absent when a bill was scheduled to be voted and they thought their side had no chance to win (or on the contrary, if they thought the matter was already won). At all times, a considerable number of legislators was absent from the Chamber (close to a fifth of the assembly). Absences were supposed to be justified, and they sometimes were. Most of the time, however, absent legislators did not provide justification. In table 10 (column 1) in Appendix A.1, we report OLS estimates of the relationship between absence and the full vector of our explanatory and control variables. The sign of some coefficients might hint to a particular direction (e.g. westernmost legislators within a province might be more inclined to absenteeism), but overall this does not suggest any obvious pattern in the selection of legislators.

7.3 One-way clustered and HAC standard errors Clustered standard errors non-parametrically allow auto-correlation within groups, under the parametric assumption that observations are uncorrelated across groups. In our context, there are reasons to expect observations to be both spatially and serially correlated, and we therefore believe that clustering at both the district and period levels is the appropriate approach. However, because our time-dimension is ‘shorter’ than our spatial-dimension (i.e. there are 13 bills and 122 districts) and since the asymptotic theory underlying two-way clustering relies on clusters in the smallest dimension, we may be worried that two-way clustered standard errors are too demanding. In column 1 of table 11 in Appendix A.1, we report simple OLS estimates of H1 using standard errors clustered at the level of the district. In columns 2 to 5, we also report the same estimates with HAC standard errors computed using Conley’s (1999; 2010) approach, allowing auto-correlation within a given radius from each municipality’s head town (50, 100, 250, and 1000km). Our results are virtually identical with one-way clustering of the standard errors and, if anything, HAC standard errors improve the precision of our estimates. In- terestingly, notice that observations tend to be negatively spatially correlated after a certain threshold, so that standard errors do not necessarily get larger as the window increases.

29 7.4 Discrete choice model The linear probability model allows for easier aggregation and interpretation and, most importantly, it allows us to implement simply our two-pronged instrumental variables strategy, and to partial out unobservable time-varying determinants of voting patterns as well as time-invariant characteristics of each province. The downside of that approach is that the predicted probabilities are not bound between 0 and 1. We compare our linear probability OLS estimates with the whole set of controls to the estimates of a discrete choice model, typically more appropriate with a dichoto- mous outcome variable. Assume that the legislator representing district j derives a utility Ujk each time she chooses alternative k on an anti-abolition bill. In H1, for 0 example, Ujk = Sjkβk +xjkγ k +εjk. To test a discrete choice version of H1, we would 1884−1888 therefore rewrite equation 1: Pjk = P(∀l 6= k, Ujk > Ujl) = P(∀l 6= k, εjl −εjk < Vjk − Vjl). If we assume that each εjk is i.i.d. (a very strong assumption here) type I Extreme Value, this can be described as a logit model, as the differences between our utility functions’ random portions follow a logistic distribution. In table 12 in the Appendix A.1, we report the results of the regressions using this specification, and we compare them to the corresponding OLS regressions without FE. Between columns 1 and 2, we find comparable marginal effects of the prevalence of slavery on the abolitionist vote, and they are comparably significant. Between columns 3 and 4, we find comparable marginal effects of the quilombo area in high prevalence districts (although this relationship loses significance without province FE). Between columns 5 and 6, we also find comparable marginal effects of immigrant settlements in high prevalence districts, and they are comparably significant. These results suggest that the linear probability model, which is less demanding assumptions-wise and offers more straightforward ways to account for unobservables and endogeneity, was an appropriate choice throughout.15

8 Conclusions

In this paper, we investigated the determinants of the persistence and change of labor coercion institutions. We considered nineteenth-century Brazil, the last Western nation to abolish legal slavery, the largest importer of slaves through the Atlantic slave trade, and a country with large areas of unexplored land. These three observations distinguish the Brazilian experience from any other; they also make it the

15That the linear probability model and the generalized linear model yield highly comparable marginal eﬀects is not surprising. For intermediate values, log odds have a nearly linear relationship with probabilities (and our dependent variable has a .55 mean).

30 ideal case study to unbundle the interests of slave-holding elites in the presence of an open agricultural frontier. This paper also makes an original contribution to the historiography of nineteenth- century Brazil. From the archival records of the Imperial parliament, we built a district-vote-level data set documenting political decision-making on emancipation- related bills during the last decade of the Empire. We relied on census surveys, historical maps, and geo-referenced data sources to identify local variations in slave- holders’ interests. We proposed a two-pronged instrumental variables strategy, leveraging historical and topographic determinants of the location of slaves and maroons across space as well as heteroskedasticity with respect to the regressors. Our first result established the importance of slave-holders’ interests in explaining voting patterns on emancipation bills. Representatives from districts with a high prevalence of slavery were less inclined to vote in favor of abolition. This accords with a large literature on the influence of elite interests and, more specifically, with the labor demand effect, and Engerman’s (1973) argument for the persistence of the institutions of coercion. Our second result established that slave-holders’ interests differed between districts, depending on the local cost of enforcing the institutions of coercion. Representatives from high-prevalence districts where escaping slavery was easier were more likely to vote in favor of abolition. This nuances considerably the Nieboer-Domar hypothesis: even if the elite as a whole favors coercion when land is abundant, it is the frontier planters, most concerned with the abundance of land, that favored switching to free labor. This second result is in line with the outside option effect (Acemoglu and Wolitzky, 2011), and in our opinion, is the main contribution of this paper. As an important corollary, this result emphasizes the role of coerced workers themselves in precipitating the collapse of the legal coercion system in Brazil. Insurrections and flights raised the costs of coercion to the planter class, which contributed to undermining the institution. Our third result established that slave-holders’ opposition to abolition was alleviated by the local availability of immigrant labor as a substitute to slave labor. This supports the importance of the land-to-labor ratio, as emphasized in the Brenner debate. Together, these three hypotheses lay the foundations of a more general theory of institutional change in a democracy dominated by oligarchic interests. Following Stasavage (2014, p. 338), “a political regime results in the provision of property rights for a specific group, accompanied by significant barriers to entry.” Individual elite members compare how they would profit under the two alternative institutional arrangements, that imply different patterns of ownership of productive assets – in this paper, a claim to owning the labor of enslaved workers. This paper revealed the importance of the mobility of labor (the ability for slaves to withdraw themselves

31 from the institutional arrangement) and of the rivalry between slave and immigrant labor. Unbundling the interests of the elite looks like a promising way to expand this analysis to other aspects of institutional – and technological – innovations.

References

Acemoglu, D., García-Jimeno, C., and Robinson, J. A. (2012). Finding Eldorado: Slavery and long-run development in Colombia. Journal of Comparative Eco- nomics, 40(4):534–564. Acemoglu, D. and Wolitzky, A. (2011). The economics of labor coercion. Economet- rica, 79(2):555–600. Acharya, A., Blackwell, M., and Sen, M. (2016). The political legacy of American slavery. Journal of Politics, 78(3):621–641. Allen, T. (2015). The promise of freedom: Fertility decisions and the escape from slavery. Review of Economics and Statistics, 97(2):472–484. Anderson, R. N. (1996). The quilombo of Palmares: a new overview of a maroon state in seventeenth-century Brazil. Journal of Latin American Studies, 28(3):545–566. Ashraf, Q. H., Cinnirella, F., Galor, O., Gershman, B., and Hornung, E. (2018). Capital-skill complementarity and the emergence of labor emancipation. Technical Report 6423, CESifo. Aston, T. H. and Philpin, C., editors (1985). The Brenner Debate: Agrarian Class Structure and Economic Development in Pre-Industrial Europe. Cambridge Uni- versity Press. Bertocchi, G. and Dimico, A. (2014). Slavery, education, and inequality. European Economic Review, 70:197–209. Bethell, L. (1989). Brazil: Empire and Republic, 1822-1930, volume 3. Cambridge University Press. Bethell, L. (2002). A abolição do comércio brasileiro de escravos. Senado Federal. Bobonis, G. J. and Morrow, P. M. (2014). Labor coercion and the accumulation of human capital. Journal of Development Economics, 108:32–53. Brazil (1874). Recenseamento do Brasil em 1872. Typ. G. Leuzinger. Retrieved from https://biblioteca.ibge.gov.br/biblioteca-catalogo?id= 225477&view=detalhes. Last accessed Dec 14, 2019. Brenner, R. (1976). Agrarian class structure and economic development in pre- industrial Europe. Past & Present, 70(1):30–75. Carvalho, J.-P. and Dippel, C. (2020). Elite identity and political accountability: A tale of ten islands. The Economic Journal. (in press).

32 Conley, T. G. (1999). GMM estimation with cross sectional dependence. Journal of econometrics, 92(1):1–45. Conley, T. G. (2010). Spatial econometrics. In Durlauf, S. N. and Blume, L. E., editors, Microeconometrics, pages 303–313. Springer. Conrad, R. E. (1972). The destruction of Brazilian slavery, 1850-1888. University of California Press. Curtin, P. D. (1972). The Atlantic slave trade: a census. University of Wisconsin Press. de Alencastro, L. F. (1979). La traite négrière et l’unité nationale brésilienne. Outre- Mers. Revue d’Histoire, 66(244):395–419. de Souza, B. G. W. (2019). The rationale of sharecropping: immigrant bonded laborers and the transition from slavery in Brazil (1830-1890). Technical Report 239, Ibero America Institute for Econ. Research (IAI). Dean, W. (2012). The Industrialization of São Paulo, 1800-1945, volume 17. Uni- versity of Texas Press. Dell, M. (2010). The persistent effects of Peru’s mining mita. Econometrica, 78(6):1863–1903. Dippel, C., Greif, A., and Trefler, D. (2020). Outside options, coercion, and wages: Removing the sugar coating. Economic Journal. (in press). Domar, E. D. (1970). The causes of slavery or serfdom: A hypothesis. The Journal of Economic History, 30(1):18–32. Eltis, D. (2007). ‘A brief overview of the trans-Atlantic slave trade,’ Slave voyages: The trans-Atlantic slave trade database. Retrieved from http://www. slavevoyages.org/estimates/kCBqlBWr and http://www.slavevoyages.org/ estimates/r3lNeZuD on Feb. 2, 2020. Engerman, S. L. (1973). Some considerations relating to property rights in man. The Journal of Economic History, 33(1):43–65. Fenske, J. (2012). Land abundance and economic institutions: Egba land and slavery, 1830-1914. The Economic History Review, 65(2):527–555. Fogel, R. W. and Engerman, S. L. (1995 [1974]). Time on the cross: The economics of American Negro slavery, volume 1. WW Norton & Company. Reissue edition. Fujiwara, T., Laudares, H., and Valencia Caicedo, F. (2019). Tordesillas, slavery and the origins of Brazilian inequality. unpublished. Fundação Getúlio Vargas (2016). Atlas histórico do Brasil. Centro de Pesquisa e Documentação de História Contemporânea do Brasil. Retrieved from https: //atlas.fgv.br/ on Dec. 14, 2019. Fundação Palmares (2020). Comunidades quilombolas certificadas. http://www. palmares.gov.br/?page_id=37551. Retrieved on June 21, 2020.

33 Graham, R. (1968). Britain and the Onset of Modernization in Brazil 1850-1914. Cambridge University Press. Green, E. (2014). The economics of slavery in the eighteenth-century Cape colony: Revising the Nieboer-Domar hypothesis. International Review of Social History, 59(1):39–70. IBGE (2010). Evolução da divisão territorial do Brasil - conteúdo digital. Retrieved from https://www.ibge.gov.br/geociencias/organizacao-do-territorio/ estruturaterritorial/ on Dec. 14, 2019. IIASA/FAO (2012). Global agro-ecological zones (gaez v3.0). IIASA, Laxenburg, Austria and FAO, Rome, Italy. INCRA (2020). Acervo fundiário: áreas de quilombolas. http://acervofundiario. incra.gov.br/acervo/acv.php. Retrieved on June 24, 2020. Jobim, N. and Porto, W. C. (1996). Legislação eleitoral no Brasil : do século XVI a nossos dias. Brasília : Senado Federal, Subsecretaria de Biblioteca. Retrieved from http://www2.senado.leg.br/bdsf/handle/id/224494 on Dec. 14, 2019. Klein, A. and Ogilvie, S. (2017). Was Domar right? Serfdom and factor endowments in Bohemia. CAGE Online Working Paper Series 344, Competitive Advantage in the Global Economy (CAGE). Klein, H. S. (2010). The Atlantic slave trade. Cambridge University Press. Klein, H. S. and Luna, F. V. (2009). Slavery in Brazil. Cambridge University Press. Le Roy Ladurie, E. (1969). Les paysans de Languedoc. Flammarion. Leff, N. H. (1972). Economic retardation in nineteenth-century Brazil. The Economic History Review, 25(3):489–507. Leff, N. H. (1991). Economic development in Brazil, 1822-1913. Technical Report 92-02, Graduate School of Business, Columbia University. Lewbel, A. (2012). Using heteroskedasticity to identity and estimate mismeasured and endogenous regressor models. Journal of Business & Economic Statistics, 30(1):67–80. Lowrie, S. (1938). O elemento negro na população de São Paulo. Revista do Arquivo Municipal, 48:27–32. Luna, L. (1976). O negro na luta contra a escravidão. Livraria Editora Cátedra. Markevich, A. and Zhuravskaya, E. (2018). The economic effects of the abolition of serfdom: Evidence from the Russian empire. American Economic Review, 108(4- 5):1074–1117. Martins, J. d. S. (1973). A imigração e a crise do Brasil agrário. Pioneira. Milliet, S. (1941). Roteiro do café: e outros ensaios, volume 25. Coleção do Depar- tamento de Cultura, 3rd edition. Digitized on Oct. 2, 2007.

34 Morse, R. M. (1958). From Community to Metropolis: A Biography of São Paulo, Brazil. University of Florida Press. Nabuco, C. (1950). The Life of Joaquim Nabuco. Stanford University Press. Naidu, S. and Yuchtman, N. (2013). Coercive contract enforcement: law and the labor market in nineteenth century industrial Britain. American Economic Review, 103(1):107–144. Naritomi, J., Soares, R. R., and Assunção, J. J. (2012). Institutional development and colonial heritage within Brazil. The journal of economic history, 72(2):393– 422. Nieboer, H. J. (2011 [1900]). Slavery as an industrial system: Ethnological researches. Cambridge University Press. Nogueira, O. and Firmo, J. S. (1973). Parlamentares do Império (obra comemora- tiva do sesquicentenário da instituição parlamentar no Brasil), volume 1. Centro Gráfico do Senado Federal. North, D. C. and Thomas, R. P. (1973). The rise of the western world: A new economic history. Cambridge University Press. Nunn, N. (2008). The long-term effects of Africa’s slave trades. The Quarterly Journal of Economics, 123(1):139–176. Nunn, N. and Puga, D. (2012). Ruggedness: The blessing of bad geography in Africa. Review of Economics and Statistics, 94(1):20–36. Nunn, N. and Wantchekon, L. (2011). The slave trade and the origins of mistrust in Africa. American Economic Review, 101(7):3221–52. Nützenadel, A. and Trentmann, F. (2008). Food and globalization: consumption, markets and politics in the modern world. Berg. Papadia, A. (2019). Slaves, migrants and development in Brazil, 1872-1923. Technical Report 2019/05, European University Institute. Postan, M. (1937). The chronology of labour services. Transactions of the Royal Historical Society, 20:169–193. Reis, E. J. and Reis, E. P. (1988). As elites agrárias e a abolição da escravidão no Brasil. Dados, Revista de Ciéncias Sociais, 31(3):309–341. Repórter Brasil (2015). Dados sobre trabalho escravo no B rasil. https:// reporterbrasil.org.br/dados/trabalhoescravo. Retrieved in Mar. 20, 2020. Ridings, E. (1994). Business interest groups in nineteenth-century Brazil. Cambridge University Press. Riley, S. J., DeGloria, S. D., and Elliot, R. (1999). A terrain ruggedness index that quantifies topographic heterogeneity. Intermountain Journal of sciences, 5(1- 4):23–27.

35 Rocha, R., Ferraz, C., and Soares, R. R. (2017). Human capital persistence and development. American Economic Journal: Applied Economics, 9(4):105–36. Schwartz, S. B. (1978). Indian labor and new world plantations: European demands and indian responses in northeastern brazil. The American Historical Review, 83(1):43–79. Stasavage, D. (2014). Was Weber right? the role of urban autonomy in Europe’s rise. American Political Science Review, 108(2):337–54. Stein, S. J. (1986 [1957]). Vassouras, a Brazilian coﬀee county, 1850-1900: the roles of planter and slave in a plantation society. Princeton University Press. Summerhill, W. (2010). Colonial institutions, slavery, inequality, and development: Evidence from São Paulo, Brazil. MPRA Paper 22162, University Library of Munich, Germany. Sutch, R. C. (2018). The economics of African American slavery: The cliometrics debate. NBER Working Papers 25197, National Bureau of Economic Research, Inc. Tabellini, M. (2020). Gifts of the immigrants, woes of the natives: Lessons from the age of mass migration. The Review of Economic Studies, 87(1):454–486. The International Labour Organization and the Walk Free Foundation (2017). Global Estimates of Modern Slavery: Forced Labour and Forced Marriage. ILO. Retrieved form https://www.ilo.org/wcmsp5/groups/public/---dgreports/ ---dcomm/documents/publication/wcms_575479.pdf. Last accessed Jul 1, 2020. Toplin, R. B. (1969). Upheaval, violence, and the abolition of slavery in Brazil: the case of São Paulo. The Hispanic American Historical Review, 49(4):639–655. Viotti da Costa, E. (1989). 1870-1889. In Bethell, L., editor, Brazil: Empire and Republic, 1822-1930, volume 3. Cambridge University Press. Williams, E. (1994 [1944]). Capitalism and slavery. UNC Press Books. with a new introduction by Colin A. ePalmer. Zemella, M. P. (1990 [1951]). O abastecimento da capitania das Minas Gerais no século XVIII, volume 19 of Coleção Estudos Históricos. Hucitec.

36 A Appendix

A.1 Additional graphs and tables

Figure A.5: 1st-stage plots

0.6

0.4 Share of slaves

0.2

0.0

1000 2000 Distance to nearest gold path (km)

600

400

Dist. to nearest 10−quil.−muni. (km) Dist. to nearest 10−quil.−muni. 200

0 1 2 3 Average TRI

37 Table 5: H1 2SLS – Preferred speciﬁcations details

External only With internal inst. 1st stage 2nd stage 1st stage 2nd stage Sh. slaves 1(Abo) Sh. slaves 1(Abo) (1) (2) (3) (4) Share of slaves −1.261∗ −0.724∗∗ (0.662) (0.304) 16th cent. Indian rep. × 0.0003∗∗∗ −0.00004 Dist. G. P. (0.0001) (0.0001)

Liberal −0.022∗∗∗ 0.365∗∗∗ −0.001 0.364∗∗∗ (0.008) (0.072) (0.006) (0.073) Share of free colored 0.064 0.134 −0.238∗∗∗ 0.025 (0.059) (0.183) (0.057) (0.173) Share of literates −0.008 0.308 −0.338∗∗∗ 0.278 (0.088) (0.253) (0.089) (0.263) Ln(Pop. density) 0.018∗∗∗ −0.031 −0.005 −0.034 (0.006) (0.020) (0.006) (0.021) Coﬀee suit. 0.003∗∗∗ −0.001 0.0001 −0.002 (0.001) (0.003) (0.001) (0.004) Suigar suit. −0.003∗∗∗ 0.001 0.00004 0.003 (0.001) (0.003) (0.001) (0.003) Rainfall −0.00003 0.0001 0.0001∗∗∗ 0.0001 (0.00002) (0.0001) (0.00003) (0.0001) Latitude 0.0005 0.002 0.004 0.011 (0.003) (0.012) (0.004) (0.020) Longitude −0.009∗∗∗ −0.007 −0.004 0.008 (0.001) (0.006) (0.003) (0.012) Dist. coast 0.00002 −0.0001 −0.00004 −0.0001 (0.00005) (0.0001) (0.00004) (0.0002)

IIV(Coﬀee suit.) 0.034∗ (0.019) IIV(Sugar suit.) −0.058∗∗∗ (0.021) IIV(Latitude) 0.009 (0.046) IIV(Longitude) −0.045∗∗ (0.022)

Province FEs No No Yes Yes Vote FEs Yes Yes Yes Yes Observations 1,586 1,586 1,586 1,586 R2 0.481 0.298 0.827 0.327 1st stage F (standard iid) 62.95∗∗∗ 159.7∗∗∗ 1st stage F (HAC excl. inst.) 11.3∗∗∗ 5.716∗∗∗ Breusch-Pagan p-value < 2.2e-16 < 2.2e-16 Wu-Hausman p-value 0.20 0.96 Sargan-Hansen p-value 0.00§ 0.57 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 §Model just-identiﬁed.

38 Table 6: H1 2SLS 2nd stage – alternative external instruments

1(Abolition vote) Instrument 1† Instrument 2‡ Instrument 3†† Instrument 1 alt.k (1) (2) (3) (4) (5) (6) (7) (8) Predicted share of slaves −1.261∗ 4.528 −1.174∗ −0.340 −1.117∗∗ −2.104∗ −1.326∗ 4.487 (0.662) (9.072) (0.645) (1.670) (0.560) (1.218) (0.724) (7.764)

Controls All All All All All All All All Province FEs No Yes No Yes No Yes No Yes Vote FEs Yes Yes Yes Yes Yes Yes Yes Yes

Observations 1,586 1,586 1,586 1,586 1,586 1,586 1,586 1,586 R2 0.298 0.062 0.300 0.326 0.302 0.308 0.296 0.066 1st stage F (standard iid) 62.95∗∗∗ 118.6∗∗∗ 60.88∗∗∗ 118∗∗∗ 57.45∗∗∗ 118.4∗∗∗ 62.77∗∗∗ 118.6∗∗∗ 1st stage F (HAC excl. inst.) 11.3∗∗∗ 0.26 5.85∗∗∗ 1.66 3.03∗∗ 3.24∗∗ 10.07∗∗∗ 0.353 Wu-Hausman p-value 0.20 0.37 0.26 0.82 0.23 0.18 0.20 0.28 Sargan-Hansen p-value 0.00§ 0.00§ 0.46 0.37 0.05 0.06 0.00§ 0.00§ Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 †Baseline instrument: Share of 16th cent. Indian enslavement area × Distance to Gold Paths. ‡Alternative 1: includes an interaction with 17-18th cent. Indian enslavement areas. ††Alternative 2: includes a distance quadratic term. kBaseline instrument controlling for diamond mines. §Model just-identiﬁed.

39 Table 7: H2 2SLS – Area quilombola – Preferred speciﬁcations details

External only With internal inst. 1st stage 1 1st stage 2 2nd stage 1st stage 1 1st stage 2 2nd stage Single Interacted 1(Abo) Single Interacted 1(Abo) (1) (2) (3) (4) (5) (6) Quilombo area −0.0003∗ −0.00003 (0.0001) (0.00003) Share slaves × 0.004∗ 0.001∗∗∗ Quilombo area (0.002) (0.0004)

TRI −741.573 12.181 −63.029 18.492 (724.407) (37.029) (161.290) (15.418) Share of slaves × TRI 940.961 −131.156 107.239 −95.119 (2,597.486) (146.084) (599.730) (59.384) TSI −34.801∗∗ −1.568∗ (16.747) (0.873) Share of slaves × TSI 50.378 4.312 (57.026) (3.881)

Share of slaves −3,603.444 −51.887 −0.997∗∗∗ −122.899 114.497∗∗ −0.810∗∗∗ (4,992.727) (270.858) (0.296) (572.372) (56.689) (0.258) Liberal −10.034 4.554 0.346∗∗∗ −43.527∗ −1.117 0.358∗∗∗ (70.112) (5.272) (0.075) (25.964) (2.128) (0.072) Share of free colored 2,702.943∗∗ 96.523∗∗ 0.327 1,100.190∗∗∗ 50.416∗ 0.003 (1,095.706) (48.505) (0.321) (340.346) (26.626) (0.164) Share of literates 112.068 −13.413 0.340 395.658 26.659 0.308 (1,284.959) (59.369) (0.340) (404.719) (39.068) (0.265) Ln(Pop. density) 136.442 −5.314 0.015 −75.388∗∗ −10.517∗∗∗ −0.025 (126.017) (5.011) (0.028) (34.687) (3.451) (0.022) Coﬀee suit. −20.536∗ −1.553∗∗ 0.001 −2.015 −0.717∗ −0.001 (11.573) (0.771) (0.005) (4.902) (0.379) (0.004) Sugar suit. 6.624 0.721 −0.0005 −2.378 0.210 0.002 (8.244) (0.512) (0.004) (4.352) (0.347) (0.003) Rainfall 0.424 0.052∗∗ 0.00001 0.002 0.019∗ 0.00004 (0.304) (0.022) (0.0001) (0.139) (0.011) (0.0001) Latitude −92.974 −2.841 0.008 16.209 0.606 0.011 (67.906) (3.793) (0.020) (21.394) (1.724) (0.020) Longitude 104.160∗ 7.779∗ −0.002 13.101 3.575 0.004 (55.467) (4.449) (0.017) (27.398) (2.261) (0.013) Dist. coast 2.537∗ 0.113∗∗ 0.0002 0.594∗∗ 0.020 −0.0001 (1.333) (0.054) (0.0004) (0.263) (0.020) (0.0002)

IIV1(Coﬀee suit.) 0.094∗∗∗ 0.002 (0.025) (0.002) IIV2(Coﬀee suit.) −1.077∗∗∗ −0.021 (0.370) (0.029) IIV1(Sugar suit.) −0.082∗∗∗ 0.001 (0.028) (0.002) IIV2(Sugar suit.) 0.694 −0.038 (0.441) (0.039) IIV1(Latitude) −0.042∗∗ 0.002 (0.021) (0.001) IIV1(Longitude) 0.015 −0.001 (0.015) (0.001) IIV2(Latitude) −0.990∗∗ −0.135∗∗∗ (0.478) (0.033) IIV2(Longitude) 0.193 0.061∗∗∗ (0.145) (0.015)

Province FEs Yes Yes Yes Yes Yes Yes Vote FEs Yes Yes Yes Yes Yes Yes Observations 1,586 1,586 1,586 1,586 1,586 1,586 R2 0.969 0.793 0.281 0.996 0.937 0.327 1st stage F (standard iid) 1062∗∗∗ 128.1∗∗∗ 8222∗∗∗ 441.3∗∗∗ 1st stage F (HAC excl. inst.) 2.033∗ 1.666 140.4∗∗∗ 72.89∗∗∗ Breusch-Pagan p-value < 2.2e-16 < 2.2e-16 < 2.2e-16 < 2.2e-16 Wu-Hausman p-value 0.045 0.208 Sargan-Hansen p-value 0.719 0.220 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

40 Table 8: H3 OLS – Preferred speciﬁcations details

1(Abolition vote) (1) (2) (3) (4) (5) (6) Share of slaves −0.694∗∗∗ −0.697∗∗∗ −0.611∗∗∗ −0.682∗∗∗ −0.784∗∗∗ −0.580∗∗∗ (0.246) (0.245) (0.198) (0.245) (0.232) (0.224)

1748-1800 settlements 0.029 0.052 0.068∗ (0.021) (0.041) (0.039)

1800-1870 settlements −0.017 −0.061 0.006 (0.023) (0.051) (0.031)

1870-1930 settlements 0.008∗∗ 0.001 −0.021∗∗ 0.012∗∗ 0.017 −0.013 (0.004) (0.017) (0.011) (0.006) (0.014) (0.012)

1930-1970 settlements −0.005∗ −0.020 −0.026 (0.003) (0.043) (0.031)

Share of slaves × 1748-1800 −0.118 −0.210 settlements (0.158) (0.208)

Share of slaves × 1800-1870 0.267∗∗ −0.052 settlements (0.117) (0.100)

Share of slaves × 1870-1930 0.073 0.239∗∗ −0.047 0.159 settlements (0.188) (0.120) (0.146) (0.100)

Share of slaves × 1930-1970 0.182 0.281 settlements (0.496) (0.356)

Liberal 0.362∗∗∗ 0.363∗∗∗ 0.379∗∗∗ 0.364∗∗∗ 0.364∗∗∗ 0.376∗∗∗ (0.073) (0.073) (0.070) (0.074) (0.075) (0.071)

Share of free colored 0.030 0.035 0.104 0.037 −0.002 0.148 (0.170) (0.171) (0.155) (0.172) (0.178) (0.162)

Share of literates 0.284 0.294 0.386 0.310 0.284 0.342 (0.256) (0.264) (0.268) (0.258) (0.257) (0.297)

Ln(Pop. density) −0.028 −0.028 −0.035∗ −0.028 −0.026 −0.021 (0.021) (0.021) (0.019) (0.021) (0.024) (0.029)

Coﬀee suit. −0.002 −0.002 −0.003 −0.003 −0.004 −0.002 (0.004) (0.004) (0.003) (0.005) (0.006) (0.004)

Sugar suit. 0.003 0.003 0.003 0.005 0.005 0.003 (0.003) (0.004) (0.002) (0.004) (0.006) (0.004)

Rainfall 0.0001 0.0001 0.0001 0.0001 0.0001 0.00002 (0.0001) (0.0001) (0.0001) (0.0001) (0.0001) (0.0001)

Latitude 0.009 0.009 0.005 0.008 0.008 −0.001 (0.019) (0.020) (0.011) (0.020) (0.024) (0.013)

Longitude 0.009 0.009 −0.003 0.009 0.009 −0.00003 (0.012) (0.012) (0.006) (0.012) (0.013) (0.007)

Dist. coast −0.0001 −0.0001 −0.0001 −0.0001 −0.0001 −0.0001 (0.0002) (0.0002) (0.0001) (0.0002) (0.0003) (0.0002)

Province FEs Yes Yes No Yes Yes No Vote FEs Yes Yes Yes Yes Yes Yes Observations 1,586 1,586 1,586 1,586 1,586 1,586 R2 0.329 0.329 0.310 0.331 0.331 0.313 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

41 Table 9: Alternative channels

1(Abolition vote) (1) (2) (3) (4) (5) (6) Share of slaves −0.698∗∗∗ −1.217∗∗∗ −0.707∗∗∗ −0.682∗∗∗ −0.607∗∗∗ −0.698∗∗∗ (0.240) (0.231) (0.248) (0.236) (0.234) (0.246)

Dist. abo. journal −0.0002 −0.001∗∗∗ (0.0002) (0.0002)

Share of slaves × Dist. 0.002∗∗∗ abo. journal (0.001)

Dist. abo. club 0.0001 (0.0002)

Dist. slavery club 0.0002 0.0005∗∗ (0.0002) (0.0002)

Literacy share × Dist. −0.002∗∗ slavery club (0.001)

Share of capitalists −4.963∗ (2.611)

Controls & FEs All All All All All All Observations 1,586 1,586 1,586 1,586 1,586 1,586 R2 0.329 0.334 0.328 0.328 0.332 0.328 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

42 Table 10: Selection – baseline speciﬁcations

1(Absent) 1(Abolition vote) (1) (2) (3) (4) (5) Share of slaves 0.035 −0.905∗∗ −1.006∗∗ −0.292 −0.837 (0.296) (0.415) (0.447) (0.526) (0.594)

Quilombo area 0.00000 (0.00005) Share slaves × Quilombo 0.001 area (0.001)

Ln av. dist. closest 1 0.003∗ quilombo-muni. (0.001) Share slaves × Ln av. dist. −0.020∗∗ closest 1 quilombo-muni. (0.009)

1748-1800 settlements −0.043 (0.128) Share of slaves × 1748-1800 0.449 settlements (0.751) 1800-1870 settlements 0.027 (0.021) Share of slaves × 1800-1870 −0.060 settlements (0.203) 1870-1930 settlements −0.028∗ (0.017) Share of slaves × 1870-1930 0.157 settlements (0.132) 1930-1970 settlements 0.041 (0.029) Share of slaves × 1930-1970 −0.390 settlements (0.367)

Liberal 0.005 0.132 0.129∗∗∗ 0.127∗∗∗ 0.147 (0.043) (0.146) (0.047) (0.047) (0.129) Share of free colored 0.428 −0.055 −0.144 −0.155 0.285 (0.329) (0.252) (0.410) (0.406) (0.236) Share of literates −0.136 0.136 0.157 0.142 0.469 (0.281) (0.248) (0.502) (0.500) (0.566) Ln(Pop. density) 0.014 0.007 0.010 −0.002 −0.044 (0.020) (0.053) (0.035) (0.034) (0.063) Coﬀee suit. −0.002 −0.004 −0.003 −0.002 −0.003 (0.002) (0.008) (0.006) (0.006) (0.006) Sugar suit. 0.001 0.001 −0.0001 −0.001 0.001 (0.002) (0.006) (0.006) (0.006) (0.005) Rainfall 0.0001 −0.0001 −0.0001 −0.0001 0.00002 (0.0001) (0.0003) (0.0002) (0.0001) (0.0002) Latitude 0.011 0.012 0.013 0.009 0.011 (0.019) (0.035) (0.026) (0.026) (0.024) Longitude −0.033∗∗∗ −0.0001 −0.005 0.007 −0.011 (0.012) (0.026) (0.020) (0.021) (0.012) Dist. coast 0.0003∗ −0.00003 −0.0001 −0.0001 −0.0001 (0.0002) (0.0002) (0.0003) (0.0003) (0.0002)

Province FEs Yes Yes Yes Yes Yes Vote FEs Yes Yes Yes Yes Yes Observations 1,586 366 366 366 366 R2 0.085 0.270 0.275 0.281 0.230 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

43 Table 11: One-way clustered and HAC standard errors

1(Abolition vote) (1-way clust.) (HAC 50km) (HAC 100km) (HAC 250km) (HAC 1000) Share of slaves −0.704∗∗∗ −0.704∗∗∗ −0.704∗∗∗ −0.704∗∗∗ −0.704∗∗∗ (0.257) (0.248) (0.182) (0.159) (0.130)

Liberal 0.364∗∗∗ 0.364∗∗∗ 0.364∗∗∗ 0.364∗∗∗ 0.364∗∗∗ (0.036) (0.039) (0.040) (0.048) (0.017)

Share of free colored 0.029 0.029 0.029 0.029 0.029 (0.238) (0.224) (0.220) (0.259) (0.302)

Share of literates 0.285 0.285 0.285 0.285 0.285∗∗∗ (0.305) (0.284) (0.300) (0.286) (0.039)

Ln(Pop. density) −0.034 −0.034∗ −0.034∗ −0.034∗ −0.034∗∗∗ (0.023) (0.022) (0.020) (0.018) (0.05)

Coﬀee suit. −0.002 −0.002 −0.002 −0.002 −0.002 (0.005) (0.004) (0.004) (0.002) (0.003)

Sugar suit. 0.003 0.003 0.003 0.003 0.003 (0.004) (0.004) (0.003) (0.002) (0.002)

Rainfall 0.0001 0.0001 0.0001 0.0001 0.0001 (0.0001) (0.0001) (0.0001) (0.0001) (0.0001)

Latitude 0.011 0.011 0.011 0.011 0.011 (0.020) (0.019) (0.019) (0.017) (0.019)

Longitude 0.009 0.009 0.009 0.009 0.009 (0.013) (0.013) (0.012) (0.011) (0.006)

Dist. coast −0.0001 −0.0001 −0.0001 −0.0001 −0.0001 (0.0002) (0.0002) (0.0002) (0.0002) (0.0001)

Vote & province FEs Yes Yes Yes Yes Yes Observations 1,586 1,586 1,586 1,586 1,586 R2 0.327 0.327 0.327 0.327 0.327 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

44 Table 12: Discrete choice model

1(Abolition vote) H1 H2 H3 OLS GLM OLS GLM OLS GLM (1) (2) (3) (4) (5) (6) Share of Slaves −0.647∗∗∗ −3.224∗∗∗ −0.672∗∗∗ −3.346∗∗∗ −0.653∗∗∗ −3.275∗∗∗ (0.206) (1.026) (0.207) (1.054) (0.218) (1.150) [-0.661] [-0.686] [-0.670]

Quilombo area −2.63e-06 −1.07e-05 (6.65e-06) (3.47e-05) [-2.19e-06]

Share of slaves × Quilombo 2.45e-04 1.15e-03 area (3.96e-04) (1.76e-03) [2.36e-04]

1870-1930 settlements −0.021∗ −0.114∗ (0.011) (0.068) [-0.023]

Share of slaves × 1870- 0.234∗ 1.358∗ 1930 settlements (0.121) (0.693) [0.278]

Controls All All All All All All Vote and province FEs No No No No No No Observations 1,586 1,586 1,586 1,586 1,586 1,586 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 Marginal eﬀects reported between brackets.

45 A.2 Outliers In this subsection, we make sure that some of our results are not driven by the existence of aberrant values. In particular, looking at the distribution of the total area quilombola by district on figure A.6 (and the distribution of truncated data in inset), it is apparent that some districts have—to say the least—an extremely large surface area occupied by certified quilombos compared to most districts. In table 13, we start by dropping all districts with a quilombola area larger than 5000km2 and re-run our OLS specification from hypothesis 2 (column 1). Notice that the coefficients associated to our variables of interest actually become more significant and gain in magnitude. We go further in column 2 and drop all observations above the IQR (141km2). The significance of the interacted term falls just short of the 10% threshold. Finally, column 3 presents the results of a robust regression, using an outliers-robust M-estimator (because no closed-form solution exists, this is fitted using iteratively re-weighted least squares). Again, coefficients associated to variables of interest remain suitably significant.

Table 13: H2 – Outliers

1(Abolition vote) (1) (2) (3) Share of slaves −0.772∗∗∗ −1.050∗∗∗ −0.822∗∗∗ (0.256) (0.383) (0.309)

Quilombo area −0.0001∗ −0.002∗ −0.00002 (0.0001) (0.001) (0.00002)

Share slaves × 0.001∗∗∗ 0.006 0.001∗∗∗ Quilombo area (0.0003) (0.004) (0.0002)

Controls All All All Vote & province FEs Yes Yes Yes Observations 1,560 1,157 1,586 R2 0.328 0.328 n/a Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

We proceed similarly with our hypothesis 3, as it appears (see figure A.7) that one of the districts in our sample received a comparatively large number of sponsored settlements during the 1870-1930 wave (2nd district of Paraná). Dropping this district results in a loss of significance in the specification which does not control for other settlement periods, but in a significance gain when the latter are accounted for (table 14). We conclude that the presence of outliers does not drive any of our results.

46 Figure A.6: Area quilombola by district

Figure A.7: 1870-1930 settlements by district

47 Table 14: H3 – Outliers

1(Abolition vote) (1) (2) Share of slaves −0.638∗∗∗ −0.696∗∗∗ (0.211) (0.261)

1748-1800 settlements 0.140∗∗ (0.061) Share of slaves × 1748- −0.623∗ 1800 settlements (0.335)

1800-1870 settlements −0.041 (0.049) Share of slaves × 1800- 0.140 1870 settlements (0.180)

1870-1930 settlements −0.033 −0.055 (0.030) (0.035) Share of slaves × 1870- 0.309 0.419∗ 1930 settlements (0.214) (0.235)

1930-1970 settlements −0.024 (0.030) Share of slaves × 1930- 0.221 1970 settlements (0.343)

Controls & Vote FEs All All Province FEs No No Observations 1,573 1,573 R2 0.308 0.312 Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01

48 A.3 Additional maps

Figure A.8: Abolition maps – other bills (1)

49 Figure A.9: Abolition maps – other bills (2)

50 Figure A.10: Quilombos (INCRA, 2020) and district-level boundaries. To illustrate some mentions in the text, we display the names of the largest quilombos.

Figure A.11: Núcleos coloniais (Fundação Getúlio Vargas, 2016) and district-level boundaries.

51 Figure A.12: Eighteenth-century mining activities and Indian enslavement (Fundação Getúlio Vargas, 2016).

52 A.4 Database construction and variables creation This section provides additional details on how we proceed to build our legislative database and generate our variables of interest.

Abolitionist voting behavior. Our main dependent variable captures the voting behavior on emancipation-related bills of legislators in the lower house of the Brazilian parliament (the Câmara dos Deputados, Chamber of Deputies). To build this variable, we explored the archival records of parliamentary debates from the onset of the eighteenth legislature (1882) to the end of the twentieth legislature (1889) and identified every occurrence of roll-call vote relative to slavery bills. Across this three legislatures, we identified thirteen such events of nominative voting. In order to make it into the final data set, occurrences of nominative voting had to be i) clearly related to the emancipation of slaves and ii) discussed in the parliamentary records in a way that allowed clear-cut identification of how the vote related to emancipation.16 The votes we retain in our final data set are the following:

• 03/06/1884: First motion of no conﬁdence against the new Presidente do Con- selho, nominated by the Emperor with the known objective to push forward the gradual emancipation of slaves. (Abolitionists vote against.)

• 30/06/1884: The roll is called on the decision to delay discussions on the Empires’ ﬁnances until after the government’s project related to labor emancipation is presented to the Chamber. (Abolitionists vote against.)

• 15/07/1884: Dantas (Presidente do Conselho, in charge to build a proposal for the gradual abolition of slavery) ﬁnally presents his project to the chamber. This occasions a large upheaval, in particular because the proposal fails to mention any compensation to slave owners. This leads the President of the Chamber (Moreira de Barros) to oﬀer his resignation, on which the roll is called. (Abolitionists vote in favor.)

• 28/07/1884: New motion of no conﬁdence against Dantas. The Chamber re- jects the government’s proposal on the emancipation of slaves and withdraws its conﬁdence from Dantas. (Abolitionists vote against.)

• 13/04/1885: New motion of no conﬁdence, this time ending in a tie. (Aboli- tionists vote against.)

16In other words, the annals must have made clear that voting in favour/against the bill meant being in favour/against the emancipation of slaves.

53 • 13/07/1885: The roll is called on a bill designed to increase the State valuation of slaves aged between 60 and 65 years old. (Abolitionists vote against.)

• 14/07/1885: The roll is called on an amendment related to the unconditional abolition of slaves aged 60 years or older. (Abolitionists vote in favor.)

• 27/07/1885 (1): The roll is called on a project regarding slave prices. (Aboli- tionists vote against).

• 27/07/1885 (2): The roll is called on a project regarding the manumission of disabled slaves. (Abolitionists vote in favor).

• 13/08/1885: The roll is called on the ﬁnal version of the bill that would later become the Lei dos Sexagenários. (Abolitionists vote in favor.)

• 04/05/1885: New motion of no conﬁdence. (Abolitionists vote against.)

• 05/05/1887: First project regarding the complete abolition of slavery. The roll is called on allowing it to proceed without hindrance. (Abolitionists vote in favor.)

• 09/05/1888: The roll is called on the project later known as the Lei Áurea. (Abolitionists vote in favor.)

Among these votes, we retain three into our core legislative data set, one by legislature: the motion of no conﬁdence against Dantas on July 28, 1884 (this precipitates the end of the eighteenth legislature and the annals could hardly be clearer about it being related to emancipation), the vote on the Lei dos Sexagenários, and the vote on the Lei Áurea (these two are the only two bills that become laws in their own right). Our goal in building this secondary data set (which we use only to assess selection issues) is to abstain—as much as possible—from any discretionary judgment, and we thus only keep the most important vote within each legislature. We believe that together, the thirteen occurrences of nominative voting outlined above constitute, for the three legislatures considered, the universe of roll call votes with clear cut interests related to emancipation. In each of these cases, the annals of the Câmara dos Deputados provide the nominative list of legislators voting in favor or against (yeas and nays). We record these names and votes and match each legislator with the electoral district she represents. This is mostly done using records of Juntas Veriﬁcadoras de Poderes (special councils that occur during the early months of each legislature), double checked using Nogueira and Firmo (1973). We then link each district with the municipalities and parochias (the lowest administrative unit)

54 that comprise them using the transcript of the 1881 electoral reform in Jobim and Porto (1996), which details the Empire’s electoral division after the 1881 Saraiva law. Ultimately, we are thus left with a mapping of 1881 parochias and municipalities into 1882-1889 district-legislator-vote observations.

Demographic variables. Our demographic variables come from Brazil’s first nation-wide demographic census in 1872 (Brazil, 1874). The main challenge here comes from matching 1872 municipalities with 1881 municipalities. For increased precision, we implement a first matching procedure at the parochia level. Although less than ten years separate the census from the legislation, a significant number of parochias were created in this time frame. From the 1441 parochias the country counted in 1872, there were 1662 enumerated in the 1881 legislation. Most of these 221 parochias are added to existing municipalities, but several municipalities were also created in the meantime and several 1872 parochias had become municipalities by 1881. To improve the precision of our matching, we exploit (whenever possible) the IBGE (2010) Evoluçao da Divisao Territorial to (painstakingly) manually trace the genealogy of municipalities. Matched parochias are then aggregated at the level of the municipality and the district.

Spatial data. We exploit several sources of geocoded data, most notably from IIASA/FAO (2012), Nunn and Puga (2012) and INCRA (2020). Information provided by these sources are used in combination with the IBGE’s municipality-level boundaries for the year 1872. This allows us to compute zonal statistics and distance measures at the municipality and district levels. In particular, for each administrative unit, we compute: land suitability, climatic and topographic measures, major towns and centroids’ geographical coordinates, remoteness measures (distance to the coast, to provincial capitals and to Rio de Janeiro), population density (using Brazil (1874)), and ’frontier openness/distance to freedom’ proxies (e.g. number and size of quilombos, districts-quilombos distances matrices and, in combination with Fun- dação Palmares (2020), average distance to the closest municipalities with at least k quilombos). Linking our matched 1872-1881 municipalities-to-districts data to the 1872 municipality grid allows to draw district-level maps of voting patterns on the bills we consider. Aggregating municipality boundaries into districts oﬀers an additional matching challenge, as 1872 municipalities do not perfectly map into 1881 districts. Two issues may arise: 1) Some municipalities are actually comprised of several districts. This occurs for some large cities, e.g. Salvador (two districts, BA1-2) and Rio de Janeiro (three districts, RJ1-3). 2) Some 1881 districts include municipalities that were created/whose territory was altered after 1872, in which case apparent

55 inconsistencies may occur. For example, the province of Amazonas is comprised of two districts in 1881, but the second district maps into two non-contiguous polygons based on the 1872 territorial division. This most likely occurs because the territorial division of the two (very large) municipalities that comprise the first ditrict of Amazonas (Manaus and Barcellos) shrunk (the former in particular) before 1881 as newer municipalities expanded. Hence, the resulting maps do not offer a perfect representation of the 1881 electoral division by any means, but they do illustrate the geographical configuration of abolition votes. Finally, we also georeference a number of existing maps, in particular from Milliet (1941) and Fundação Getúlio Vargas (2016). Most notably, this allows us to approx- imate the location of state-sponsored settlements in a succession of waves and to assess proximity to Caminhos do Ouro, mining sites, insurrection sites, Tupi-Guarani enslavement areas, slavery/abolitionist interest groups and abolitionist journals.