Brazil’s Bolsa Família: Neighborhood and Racial Group Networks

Abstract Are families that live in the same neighborhood and share similar characteristics more likely to participate in welfare programs? Using a unique administrative data set, we study beneficiaries of the Bolsa Família – a Brazilian cash transfer program – from 2013 to 2015. We analyze data containing information on the living conditions of the most vulnerable families, such as income, household characteristics, schooling, and disability. An eight-digit zip code defines a neighborhood. Families form a network if they live in the same neighborhood and belong to the same racial group. We provide evidence that place of residence and racial group networks are important determinants of the family participation in the program. Individuals in a neighborhood-racial group network are more likely to to participate in the Bolsa Família than not to participate. In areas where program coverage is low, families of the same racial composition and zip code are more likely to be beneficiaries. For a given neighborhood-racial group network, the presence of one additional beneficiary implies that the probability that a non-beneficiary family will become a beneficiary is, on average, 6.5% higher than otherwise. We conduct several robustness checks, e.g., controlling for network density and coverage.


Introduction
In this paper, we empirically investigate the relevance of an individual's network on her participation in the Bolsa Fam ılia Program, the most important cash transfer program in Brazil and one of the largest conditional welfare programs in the world. 1 We define an individual's network based on her neighborhood and racial group. We combine 182,811 eight digits zip-codes and five (self-declared) racial groups (White, Black, Yellow, Brown, and Indigenous) to construct our key measure of neighborhood-racial group network, i.e., individuals that belong to the same racial group and live in the same neighborhood. 2 We find evidence that an individual's neighborhood-racial group network positively affected the probability participation in the program. Being connected to the same racial group of peers in a given neighborhood increased the likelihood that an individual participates in the Bolsa Fam ılia program. Our results also show that the more individuals of a given racial group live in a neighborhood (contact density), the higher the odds of program participation. However, being in a neighborhood predominantly of one's own racial group has a positive effect on an individuals' participation in the Bolsa Fam ılia only for White individuals.
We use data from the CadUnico, a comprehensive system that integrates various social programs administered by the Brazilian federal government and it is the largest governmental administrative database in Brazil containing information of poor and extremely poor families targeted by social programs (e.g., Bolsa Fam ılia). Launched in 2003, the Bolsa Fam ılia was a cash transfer program aimed at reducing poverty and improving the socio-economic conditions of families living in poverty and extreme poverty in Brazil. In 2015, the program reached about 14 million Brazilian families, representing approximately 60 million poor people (30% of the Brazilian population) with an annual budget of more than 24 billion reais (US$11 billion, 0.5% of GDP). To provide some perspective, the Bolsa Fam ılia reached almost three times more people, was about three times larger in terms of the budget than the well-known conditional cash transfer (CCT) program Progresa/Oportunidades in M exico (Brollo et al., 2020).
Our sample size is set based on a sampling fraction of 5% of the population size in each geographic stratum defined by a zip code, resulting in a panel of 2,797,477 observations (data from the CadUnico for the years 2013-2015 available at the household level and at the individual level). There are significantly more women in our sample (89%) and a higher proportion of individuals of the (self-declared) Brown racial group (62.36%). The average education level is relatively low (32% of the families have incomplete primary school) and 59% of household heads either do not work or are employed in the informal sector.
The empirical strategy and model specification follow Aizer and Currie (2004) and Deri (2005). Aizer and Currie (2004) considers women eligible for the prenatal care program in California (United States), who live in the same 5-digit postal code and belong to the same ethnic group. We study eligible family households (which includes beneficiaries and non-beneficiaries of the program), who live in the same 8-digit zip-code and belong to the same racial group. Following Deri (2005), we construct a variable to represent the individuals' contact density. This is defined as the proportion of families of a given racial group to the total number of families in a particular zip code and it controls for the fact that individuals may live in a predominantly racial group neighborhood, which might influence their participation in the social program (unobservable family characteristics). As in Deri (2005), we also consider the interaction between the variables network and contact density.
Social networks represent important groups of individuals in any social and welfare programs. These groups can be engaged to provide feedback, identify priorities and opportunities, establish positions on issues and approaches, and plan strategies for intervention. Moreover, social networks are important because they allow people to develop relationships with others with whom they might not otherwise be able to connect. However, the sharing of knowledge and attitudes about policy interventions among networks of potential beneficiaries is one set of social interaction that remains under-documented in the setting of social policies in developing countries (Bobba & Gignoux, 2019). 3 The conditional cash transfer (CCT) programs that have been recently implemented in developing countries create many opportunities for knowledge spillovers between beneficiaries. This is the case of the Bolsa Fam ılia program. For instance, the recipients of the Bolsa Fam ılia benefit, notably women and mothers, regularly encounter each other during program operations, for instance in meetings of beneficiaries or during activities of complementary interventions, such as visits to health centers. The targeting of those interventions implies that participants often have similar socioeconomic backgrounds and are thus likely to identify with each other (Akerlof, 1997). Hence, CCT program interventions are likely to both enhance the existing interactions among groups of beneficiaries and to further shape those groups, thus producing externalities that would not occur were individuals treated in isolation. In fact, in the Bolsa Fam ılia setting, many beneficiary neighborhood are very close to each other, thus spillovers may occur not only within, but also across neighborhoods through, for instance, an individual's racial group.
In this paper we use an individual's zip-code and racial group to proxy for social links between individuals within a neighborhood and racial group. Our underlying assumption is that individuals living in the same zip-code and belonging to the same racial group will have a larger pool of available contacts who can facilitate knowledge of and participation in the Bolsa Fam ılia program. In a world of imperfect information, social ties can provide an individual with useful information about opportunities and rights otherwise not available. This information advantage of social relations has typically been used to illustrate the importance of social networks in the job market (Granovetter, 1995). Its relevance for social programs is straightforward. Individuals that belong to neighborhood-racial group have a higher probability of receiving information about potential social programs compared to more excluded households. Implications of these characteristics and their interactions are of particular interest for social programs targeted to reduce poverty as it is the case of the Bolsa Fam ılia in Brazil. We use the number of people in one's neighborhood who belong to the same racial group to measure the quantity of networks (contact availability). Contacts drawn from high Bolsa Fam ılia-using groups will likely exert a stronger influence on program participation.
We find evidence that a network of Bolsa Fam ılia eligible individuals positively affects the probability of an individual's participation in the program. Individuals in a neighborhoodracial group network are 6.5% more likely to participate in the Bolsa Fam ılia than not participate in it. This result suggests that being in a neighborhood predominantly of one's own race has a positive effect on the individuals' participation in the social program. Moreover, the more individuals of a given racial group live in the neighborhood (contact density), the higher the odds individuals (all racial groups) will participate in the program. The odds of participation are, however, smaller. A male individual has a lower probability of participation in the program than a female individual, which reflects an important feature of the CadUnico data (in our sample, 89.8% of the household heads are women).
We also investigate whether there are differences across racial groups and participation in the Bolsa Fam ılia. We focus on the three most representative (self-declared) racial groups in Brazil, i.e., White, Black, and Brown individuals. When we control for the density of contacts in an individual's network and the interaction between this variable and the neighborhood-racial group network variable for particular racial groups, we find that Black and Brown individuals are, respectively, 11% and 11.5% more likely to participate in the Bolsa Fam ılia than not to participate in it (versus 6.5% for all racial groups). The coefficient for White individuals is not statistically significant.
Two additional exercises are conducted to investigate the robustness of our results regarding the effect of neighborhood and racial group networks on individuals' participation in the Bolsa Fam ılia. We define Bolsa Fam ılia coverage as the ratio of beneficiaries to eligible families registered in the CadUnico per municipality. We find that for individuals in the same neighborhoodracial group network the likelihood of participation in the program of those at the bottom of the coverage distribution is about one percentage point lower, i.e., 6.5% (odds ratio equals 1.065), than the likelihood of those at the top of the distribution (7.3% odds ratio equals 1.073). That is, individuals are less likely to participate in the program in municipalities with relative less eligible beneficiaries.
Finally, we consider two alternative definitions of networks. First, a neighborhood network is formed by individuals that live in the same neighborhood (zip code), regardless their racial group. Second, a racial group network considers individuals of the same racial group, regardless their neighborhoods. Results suggest that individuals in a neighborhood-network are 3.3% more likely to participate in the Bolsa Fam ılia than not to participate in it (a result 2.6 p.p. higher than in the case of a racial group network; 0.7% odds ratio). This can be interpreted as evidence that being connected to the same zip code peers increases the odds that an individual Neighborhood and racial group networks 595 will participate in a social program more than simply being in the same neighborhood and belonging to the same racial group network.
While the effect of social networks on individual behavior (e.g., social pressure and information spillovers) has long been emphasized by theoretical economists, empirical work has found it difficult to demonstrate network effects. 4 In fact, social networks provide information on institutional details of a social program and it can reduce the search costs of joining and complying with program regulations. 5 Networks can even alter the demand for the program transfers by affecting the perceived efficacy or program benefit desirability. In the context of this study, social network structures may contribute to allow eligible people who are not Bolsa Fam ılia beneficiaries to learn about program eligibility criteria (e.g., documents required for registration, etc.) from their neighbors of the same racial group.
The empirical evidence we put forward that place of residence and racial group networks are important determinants of the family participation in the program sheds light on the mechanisms behind program externalities. Interactions within (neighborhood-racial group) networks of potential beneficiaries seem to have contributed to increase program participation. We argue that, while interactions through preexisting social networks should affect all households that share local resources, social interactions that are restricted to program beneficiaries are likely to be associated with knowledge and attitudes toward the program. Accordingly, there is also evidence that participation in the Bolsa Fam ılia is associated with increased knowledge among eligible households about the different components of the program-notably education and health indicators (Camargo et al., 2013).
Finally, we illustrate our approach by adapting Bertrand et al. (2000)'s simple example. Imagine an individual of race X moves to a neighborhood Z. In order to participate and receive the Bolsa Fam ılia, she would need help in understanding the rules and procedures. As the number of Bolsa Fam ılia beneficiaries in her area (neighborhood) increases, so too does the number of people who could potentially help her. Moreover, the familiarity that Bolsa Fam ılia beneficiaries have with the social program and requirements affects the kind of help they could provide. At one extreme, if the all individuals of race X living in neighborhood Z purposely avoid the Bolsa Fam ılia program and were quite unfamiliar with it, they may even discourage her from participating. At the other extreme, if they all knew a great deal about it, this may actively encourage her to participate. Therefore, the "return," in terms of program participation, to being surrounded by individuals of the same race rises with the familiarity these individuals have of the Bolsa Fam ılia program.

Related literature
Our paper contributes to two main strands of literature. First, our analysis is related to the literature on social interactions and learning about social programs. 6 Two papers are closely related to ours and we discuss them in details. Aizer and Currie (2004) study the potential network effects in the use of publicly-funded prenatal care using Vital Statistics data from California for 1989 to 2000. They find that the use of public programs is highly correlated within groups defined using race/ethnicity and neighborhoods. The probability that public prenatal care is chosen increases with both contact availability and with the share of one's own group using the service. Deri (2005) focuses on individuals living in Canada whose mother tongues are not English, French or any other official languages and investigates the effects of social networks on health service use in Canada. A measure of contact density is constructed and used to show that social networks affect people's behavior regarding the use of services, as language groups living in high concentration areas of the same language group can increase access to such services.
Second, there is extensive literature on conditional cash transfer (CCT) programs and their direct and indirect (spillover) effects on beneficiaries and their peers. Bobba and Gignoux, (2019) examine the role of spillover effects in the form of information sharing within networks of potential beneficiaries and in shaping the take-up of the schooling subsidy component of the Progresa-Oportunidades CCT program (see, e.g., Paul Schultz (2004) and Parker et al. (2008)). The authors find evidence of positive spillovers within networks of beneficiaries, and they suggest that spillovers stem partially from the sharing of information about the program among eligible households. They argue that ignoring peer effects leads to a biased assessment of the program's real impact. Brollo et al. (2020) analyze how individuals respond to peers experiencing enforcement of Bolsa Fam ılia conditions (e.g., when families fail to comply with program requirements, they receive a series of warnings and financial penalties). They show that enforcement creates spillover effects. In particular, individuals respond to notifications received by classmates or siblings (in other schools and of another gender) and neighbors.
The remainder of the article is organized as follows. Section 2 presents the institutional framework and our sample data. Our empirical strategy is discussed in Sections 3 and 4 presents our empirical results. Section 5 concludes the paper.

Brazil's Bolsa Fam ılia and CadUnico
In the year 2000, the Brazilian government established a comprehensive system, the Cadastro Unico (CadUnico), to integrate data on the various social programs managed by the federal government. The CadUnico administrative dataset has information of approximately 40% of the Brazilian population. It is considered the biggest representative dataset of the poorest and most vulnerable families in Brazil. Data on poor and extremely poor families are included in the CadUnico and are collected at the city level. Through the CadUnico, public officials can keep track of relevant data on social program beneficiaries, such as family household composition and unique identification numbers, addresses and telephone numbers, family size, monthly income and expenses, education level, and labor market status. In 2013, there were 27 million families registered in the CadUnico.
Among the several social programs covered by the CadUnico, the Bolsa Fam ılia program was the most important one. The Bolsa Fam ılia was a conditional-cash transfer (CCT) program established in 2003 to provide additional resources to low-income families through direct cash transfers. Transfers were conditional on education, health, and social assistance criteria. 7 Among other factors, the amount received depended on the family composition (e.g., family size, age, presence of pregnant women and/or teenagers in the household) and the family monthly income. In particular, the Bolsa Fam ılia targeted families living in poverty and extreme poverty conditions (families with monthly per capita income of up to US$20.00 (R$85.00) and families with children with age below 17 years-old and monthly income of up to US$40.00 (R$170.00) (Brasil, 2015;Soares et al., 2009).
The Bolsa Fam ılia screening and registration was based on basically four administrative steps. First, in order to apply for the Bolsa Fam ılia benefit families must be registered in the CadUnico. Then, three administrative steps (stages) would follow, namely, qualification, selection, and benefit-granting. In the qualification stage, it was verified whether families registered in the CadUnico met the eligibility criteria for the Bolsa Fam ılia benefit, a process based on the self-declared socioeconomic information of the families, a method called Unverified Means Testing (UMT). Next, the maximum number of families per municipality that could join the Bolsa Fam ılia in a given payroll cycle was defined (selection stage). This was done because each municipality had quotas of beneficiaries based on governmental estimates of the number of Neighborhood and racial group networks 597 poor people living in the municipalities (Camargo et al., 2013). In the final stage (benefit-granting), individual families were identified and granted the Bolsa Fam ılia benefit. 8 We observe a decline in the Bolsa Fam ılia participation rate, with variations among racial groups and Brazilian regions in the years 2013-2014. For example, White individuals had a substantial decrease in the program participation compared to the other racial group's. In the years 2014-2015, the program increased the number of beneficiaries, in which White and Black racial groups reached higher participation rates than Brown, Yellow and Indigenous groups-over the 2014-2015 period, the increases were as follows: White (þ19%), Black (þ14%), Brown (þ10%), Yellow (þ12%), and Indigenous (þ6%). Geographically, the Brazilian South region had a noticeable decline in the Bolsa Fam ılia participation rate compared to other regions (Southeast, North, Northeast, and Midwest).

Our sample data
To empirically investigate the relevance of an individual's network on the Bolsa Fam ılia participation, we explore the uniqueness of the CadUnico. We consider data from the CadUnico for the years 2013 to 2015, available at the household level and at the individual level (i.e., for the individual responsible for providing information about the household and CadUnico) registration. Data at the municipality level (city, town, village, township) are also available. Household and individual data are linked through a family code included in both datasets. Individual-level variables contain information about gender, place of work, place of birth, education, and (self-declared) racial group. Regarding the individual's racial group, we consider five (self-declared) racial groups: White, Black, Yellow, Brown, and Indigenous, as defined in the CadUnico database.
Family household variables include information regarding neighborhood, per capita income, per capita expenditure, house floor and wall materials, access to running water, bathroom or toilet in the house, availability of sewage system, garbage collection, water supply, street pavement, and electric power grid. Our sample size is set based on a sampling fraction of 5% of the population size in each geographic stratum defined by zip code, resulting in a panel of 2,797,477 observations. 9 There are significantly more women in our sample (89%) and a higher proportion of individuals of the (self-declared) Brown racial group (62.36%). The average education level is relatively low (32% of the families have incomplete primary school). Fifty-nine percent of household heads either do not work or are employed in the informal sector. Individuals eligible to the Bolsa Fam ılia live in regions of poor infrastructure; for instance, 50.21% of Bolsa Fam ılia eligible families live in streets without pavement, and 54% do not have access to a sewerage system. Regarding the educational and health requirements, our data shows that more families tend to abide more by the former than the latter.
We define an individual's network based on her neighborhood and racial group. Zip code geographical areas at the city level in Brazil are currently composed of eight digits (00000-000). The first five numbers represent the geographical region, sub-region, sector, sub-sector, and sub-sector digits. The last three digits represent additional local identifiers. Hence, we combine 182,811 zip-codes and five (self-declared) racial groups to construct our key measure of neighborhood and racial group network, i.e., individuals that belong to the same racial group and live in the same neighborhood. In the period 2013À2015, the groups Indigenous and Brown have higher average of beneficiaries in their neighborhood-racial group networks (greater that 60%), while the racial group White has the lowest average number of beneficiaries in the network (below than 50%).
Descriptive statistics of continuous and categorical variables are presented in Tables 1 and 2, respectively. More detailed and additional description of our variables and descriptive statistics are presented in Appendix.

Empirical strategy
In our empirical analysis, we consider beneficiaries and non-beneficiaries of the Bolsa Fam ılia in the context of the neighborhood and (self-declared) racial group networks, following Aizer and Currie (2004) and Deri (2005). Let a neighborhood-racial group network z ¼ jk be formed by individuals that live in the same neighborhood (zip code j) and belong to the same racial group k. Let the variable N izt denote that an individual i belongs to a network z ¼ jk in the period t. To construct this variable and to make the estimation computationally feasible, we consider the average of the relevant characteristics, with the underlying assumption that the individuals' contacts are randomly distributed in a neighborhood (Durlauf, 2004;Jackson, 2003)). That is, the variable N izt is a continuous variable constructed from the combination of 182,811 zip codes with five racial groups, given by the average of family representatives i who live in the same zip code j and are of the same racial group k ¼ fWhite, Black, Yellow, Brown, Indigenousg in each year t ¼ f2013, 2014, 2015g: We investigate the likelihood an individual is a Bolsa Fam ılia beneficiary based on the following Logit model:  Neighborhood and racial group networks 599 where the variable y izt represents whether a family i is a Bolsa Fam ılia beneficiary: if y izt ¼ 1, a family i of racial group k that lives in the zip code j receives the Bolsa Fam ılia benefit; and y izt ¼ 0, otherwise. 10 Before we present and discuss other variables of Equation (1) we must acknowledge that our strategy of averaging out all families beneficiaries of the Bolsa Fam ılia based on their (selfdeclared) racial group and zip code characteristics (variable N iztÀ1 ) might be poised by omitted bias variable problems. For instance, the participation of a neighborhood-racial group network jk individual in the Bolsa Fam ılia may be correlated with an unobserved characteristic that she shares with other individuals in the neighborhood j. To address this issue, we use geographical and racial group fixed effects following Bertrand et al. (2000).
To control for unobservable family characteristics that may be correlated with the probability of belonging to a given racial group k in a given zip code j, we construct a variable D zt , z ¼ jk, which represents the contact density of the network and captures the fraction of families of the same racial characteristic jk living in a given location j. That is, D zt is the ratio of the number of households of the same racial group k, living in the same neighborhood j at time t to the total number of household is a zip code j at time t.
Since the social network variable N iztÀ1 alone might only capture the quantitative aspects of contact among members of a given network z ¼ jk, we follow Deri (2005) to address the qualitative features of the network through the interaction between N iztÀ1 and D zt . In accounting for the interaction of these two variables, we intend to capture the "quality" of the contacts in the individual's network (Deri, 2005). Moreover, the network contact density variable facilitates the comparison of networks of different sizes as it captures the deviations from the average of the entire sample in all groups, normalized by the size of the groups formed in each network.
The vector X izt , Equation (1), includes individual characteristics of the head of the family unit i who belongs to the racial group k and lives in the zip code j at time t (e.g., schooling, gender, place of work, income, and place of birth). The head of the family is the person who filled out the application for the social benefit, is responsible for the family unit (household care) and for any updates regarding changes in the family status. The vector V izt includes variables at the family level, such as income, expenditures, house construction type, location, characteristics of the neighborhood (street pavement, lighting, sanitation, among others). And, the vector M jt includes variables at the municipal level, such as specific characteristics of the municipalities, number of social reference centers (SARC), a Decentralized Management Index (DMI), among others. We also consider a year dummy variable (Q t ) and u izt is a random error term.

Bolsa Fam ılia and neighborhood-racial group networks
In this section, we present our empirical results and discussion regarding the effects of neighborhood and racial group networks on individuals' participation in the Bolsa Fam ılia program. In particular, we investigate whether individuals that live in the same geographical area (zip code) and belong to the same racial group, i.e., belong to the same neighborhood-racial group network, are more likely to receive the Bolsa Fam ılia benefit. For presentation purposes, our discussion will focus on the main variables of the study, but complete estimation results are reported in the Supplemental Material. Table 3 presents the results for the estimation of Equation (1), our fixed effects logit model. 11 Controlling for family, individual and municipal characteristics, the results of Model I (Table 3) show that the variable Neighborhood-Racial Group Network is positive and statistically significant. This means that an individual that belongs to a specific (self-declared) racial group and lives in a particular zip code is more likely to participate in the Bolsa Fam ılia and receive its benefit than otherwise. Quantitatively, we find that individuals in a neighborhood-racial group network are 27.4% more likely to participate in the Bolsa Fam ılia than not participate in it. This result suggests that being connected to peers of the same racial group in a given neighborhood increases the odds of an individual to participate in a social program.
We also control for the density of contacts in a given network and the interaction between the variables Neighborhood-Racial Group Network and Contact Density (Model II, Table 3). Recall that the variable Contact Density, defined as the proportion of families of a given racial group to the total number of families in a particular zip code, is meant to control for the fact that individuals might live in a predominantly racial group neighborhood, which might influence their participation in the social program (unobservable family characteristics). Our results show that both variables have positive and statistically significant coefficients, i.e., being in a neighborhood predominantly of one's racial group has a positive effect on the individuals' (of the same racial group) participation in the Bolsa Fam ılia. The odds of participation are, however, smaller. The more individuals of a given racial group live in a particular neighborhood (contact density), the higher the odds that individuals (all racial groups) will participate in the program (5.4% higher). If the individual belongs to the predominant racial group, the participation in the program is one percentage point higher, i.e., 6.5% (an odds ratio equals 1.065).
We find that the coefficients for family, individual and municipal controls have the expected signs and are significant (Table 14, Supplemental Material). Our results also reveal that a family unit which the household head is a male is less likely to participate in the Bolsa Fam ılia (81.1% lower odds ratio) than a household in which the head is a female. Families with higher per capita income and higher per capita expenditures are less likely to participate in the program. We also find that the higher the number of individuals meeting the education qualification criteria of the Bolsa Fam ılia and the higher the number of Social Assistance Reference Centers (SARC) in the municipality, the lower the likelihood individuals will participate in the Bolsa Fam ılia. And, we identify a positive relationship between schooling and the probability of participation in the program Bolsa Fam ılia.
In order to further understand the interaction between race and neighborhood and their effects for the participation in the Bolsa Fam ılia, we focus on the most representative (selfselected) racial groups in Brazil, i.e., White, Black and Brown individuals.Recall that we adopt the same terminology used by the Brazilian Institute of Geography and Statistics, which is based on the individual's self-declared race (racial group). We conduct two exercises: first, we do not control for the density of contacts (Table 4) and, second, an exercise where the variable Contact Density is included as a control variable (Table 5). 12 The results for all racial groups are also presented in Tables 4 and 5 to facilitate comparisons. When the variable Contact Density is not included as a control variable (Table 4), our results reveal that the Neighborhood-Racial Group Network coefficients remain positive and statistically significant for the racial groups White, Black and Brown. In particular, Black and Brown's individuals are 37.3% and 30.7% more likely to participate in the Bolsa Fam ılia than otherwise, respectively. While these likelihoods are higher than the ones observed for all racial groups (27.4%), individuals in a White neighborhood-racial group network are only 17.2% more likely to participate in the Bolsa Fam ılia than not participate in it.
Controlling for the density of contacts in a neighborhood-racial group network for specific racial groups (Table 5), we find that being in a neighborhood predominantly of one's racial   Table 5) suggest that if an individual belongs to the predominant racial group in a neighborhood-racial group network, the odds that a Black or Brown individual will participate in the program is about 5 p.p. higher than for all other racial groups (the odds ratio is 6.5% for all races and about 11% for Black and Brown individuals). In summary, Tables 3-5 present suggestive evidence of endogenous (neighborhood-racial group network, contact density), exogenous (education and health indicators, SARC and DMI), and correlated (sewage system, garbage collection, type of street lighting and paving) network effects in the Bolsa Fam ılia. See Tables 13-16, Supplemental Material, for complete results.

Additional results and robustness checks
In this section, we investigate the robustness of our results regarding the effect of neighborhood and racial group networks on individuals' participation in the Bolsa Fam ılia. First, we define the Bolsa Fam ılia coverage as the ratio of beneficiaries to eligible families registered in the CadUnico per municipality. We observe that in regions with high coverage, there is also a low average of non-beneficiaries per networki.e., there is a negative correlation between Bolsa Fam ılia coverage and the average number of non-beneficiaries in the neighborhood-racial group network. We find that the Bolsa Fam ılia coverage is 56% at the 25 th percentile threshold, i.e., 25% of the municipalities have up to 56% coverage of their eligible families. When considering the 33 th percentile, the Bolsa Fam ılia coverage increases to 60.16%. With this variable, we can study the role of neighborhood-racial group networks in regions with high and low Bolsa Fam ılia coverage. Next, we use this coverage criterion to investigate whether the effect of neighborhood and racial group networks on the participation in the Bolsa Fam ılia is affected by the program coverage at the municipality level. Recall that results presented in Tables 3-5 suggest that being in a neighborhood predominantly of one's race has a positive effect on the individuals' participation in the Bolsa Fam ılia. However, this effect is smaller for those at the bottom (first two columns of Table 6) vis-a-vis individuals at the top 25 th percentile of the Bolsa Fam ılia program coverage (last two columns of Table 6). The results of Table 6 suggest that the household position in the coverage distribution matters. For individuals in the same neighborhood-racial group network, the likelihood of participation in the program of those at the bottom of the coverage distribution is one percentage point lower, i.e., 6.5% (odds ratio equals 1.065), than the likelihood of those at the top of the distribution (7.3% odds ratio equals 1.073). Regarding the interaction between the variables Neighborhood-Racial Group network and Contact Density, the coefficients are positive and statistically significant; however, of similar magnitude for both Bottom 25 th and Top 25 th groups (5.8% versus 5.7%, respectively). 13 Next, we consider two alternative definitions of networks. First, a neighborhood (j) network is formed by individuals that live in the same neighborhood (zip code), regardless of their racial group. That is, the variable N iztÀ1 , Equation (1), denotes that an individual i belongs to a neighborhood network z ¼ j. Second, a racial group (k) network is broader as it considers individuals of the same racial group, regardless of their neighborhood, i.e., N iztÀ1 for z ¼ k. Controlling for contact density and household, individual and municipal characteristics, coefficients for the neighborhood-network are positive and statistically significant (Table 7). Results suggest individuals in a neighborhood-network are 3.3% more likely to participate in the Bolsa Fam ılia than not to participate in it (a result 2.6 p.p. higher than in the case of a racial group network; 0.7% odds ratio equals 1.007). This can be interpreted as evidence that being connected to the same zip code peers increases the odds that an individual will participate in a social program more than just being in the same neighborhood and belonging to the same racial group network. In the specific case of racial group-networks, we find that the coefficient for the variable Contact Density is statistically significant only when we consider networks formed by individuals of the same racial group (see Tables 19-20, Supplemental Material, for additional details).

Conclusion
We study the importance of an individual's racial group and neighborhood on the participation in the Programa Bolsa Fam ılia (Bolsa Fam ılia), Brazil. We combine 182,811 zip codes and five racial groups to construct neighborhood, racial group, and neighborhood-racial group networks. We find evidence that the network of an individual eligible for the Bolsa Fam ılia positively affects the probability that the individual participates in the program. Being connected to the same racial group of peers in a given neighborhood increases the odds (6.5%) that an individual will participate in the Bolsa Fam ılia. Our results show that the density of contacts (defined as the proportion of families of a given race to the total number of families in a particular zip code) and the coefficient for the interaction between the variables neighborhoodracial group network and contact density are positive and statistically significant. The more individuals of a given race in the neighborhood (contact density), the higher the odds individuals (all racial groups) will participate in the program. However, being in a neighborhood predominantly of one's own racial group has a positive effect on the individuals' participation in the Bolsa Fam ılia only for White (not Black or Brown) individuals. Several robustness exercises were conducted. For instance, if the individual belongs to the predominant racial group at the bottom of the coverage distribution, the likelihood of participation in the program is one percentage point lower than at the top of the distribution. In sum, this study highlights the importance of (neighborhood and racial groups) networks for individuals' participation in social programs and its potential relevance for the design of public policies. are no studies or additional information that would allow us to address how well validated self-reported racial categories are in Brazil. Hence, in this study the definition of "racial group" relies solely on the Instituto Brasileiro de Geografia e Estatistica (IBGE) approach to collect information regarding an individual's selfdeclared racial group. In the CadUnico questionnaire, individuals are asked about their race according to the following (self-declared) options: white, black, brown, indigenous, or yellow (Brasil, 2017). 3. The role of information-sharing and initial preferences and prejudices in determining program participation has been emphasized in the context of social policies in the United States. For instance, Bertrand et al. (2000) and Aizer and Currie (2004) find evidence of networks effects, that is, correlations in program take-up decisions within neighborhoods and ethnic groups. In the case of the Food Stamp Program, Daponte et al. (1999) find that ignorance about the program contributes to non-participation. 4. Game theorists have studied the importance of learning from neighbors and information spillovers in the emergence of equilibrium. Macroeconomists have stressed the importance of human capital spillovers as determinants of growth and inequality, and labor and public economists have used stigma and information spillovers to explain a range of outcomes including program participation, fertility, crime and education. See Bertrand et al. (2000) for additional references. The existing empirical work reveals that many individual outcomes, ranging, for instance, from crime activities to educational attainment, are indeed positively correlated with friends', neighbors', and ethnic group's outcomes. While suggestive of network effects, these correlations may result from unobserved factors about individuals, neighborhoods, and ethnic groups (Bertrand et al., 2000). For example, some areas may have better schools and health clinics, making both individuals and their neighbors less likely to use welfare program such as the Bolsa Fam ılia. 5. Knowledge and actions regarding policy interventions and social programs among potential beneficiaries are issues that remain under-documented in developing countries (Bobba & Gignoux, 2019). A better understanding of the potential interactions between policy and individuals' decisions can contribute to improve access to land, labor, and credit transactions for specific groups, such as Indigenous people (Phan et al., 2020) and access to micro-credit by a network of families (Abay et al., 2018). 6. See also Alatas et al. (2016) on network structure and the aggregation of information in a community-based targeting program; Banerjee et al. (2013Banerjee et al. ( , 2019 on information transmission in the context of a micro-finance program; Duflo and Saez (2003) on the transmission of information on retirement plans; and Dahl et al. (2014) on peer effects via learning in the context of paternity leave. Bertrand et al. (2000) also find evidence of networks effects, that is, correlations in program take-up decisions within neighborhoods and ethnic groups. 7. Beneficiary families must meet health and education conditions. Six to seventeen-year-old children must be enrolled in school with a monthly attendance requirement of at least 85% for those 6-15-year-old and 75% for those 16-17-year-old. Health requirements include up-to-date vaccination for children 7-year-old and younger and prenatal care and attendance to pregnancy follow-ups (Brasil, 2015). 8. Notice that if an individual met the program criteria it was not guaranteed she would be selected for the program as there was a maximum amount of resources available per municipality in a given fiscal year. Each municipality had its own quota of beneficiaries, which was based on an estimate of the number of people living in poverty in that particular location. Priority was given to those families who had the lowest monthly per capita income and more children age 0-17 years-old (Camargo et al., 2013). 9. Our sample was selected by a probability design to allow for inferences about the population defined by CadUnico. A simple stratified sampling procedure with a proportional allocation is adopted. This process divides the population into strata and simple random samples are selected independently (Cochran, 1977). 10. In the Supplemental Materials, we present Linear Probability Model (LPM) fixed effects estimation results. The results based on the LPM fixed effects are qualitatively consistent with those based on the Logit fixed effects and corroborate the main results presented in Section 4. 11. We perform the estimations of Overall, Between and Within effects of all variables used in the fixed effects regression and their respective results are presented Table 6, Supplemental Material. 12. See Tables 15-16, Supplemental Material, for complete results. 13. We also split the sample based on the 33 th percentile for the Bolsa Fam ılia coverage and the results are very similar to the ones based on the 25 th percentile (see Tables 17-18, Supplemental Material, for details). Neighborhood and racial group networks 609