Factors Related to Internal Migration in Brazil: how does a conditional cash-transfer program contribute to this phenomenon? 1 Luiz Carlos Day Gama 2 Ana Maria Hermeto Camilo de Oliveira 3 Abstract The main objective of this paper is to analyze how changes in the economic and social conditions influence the likelihood of migrating in Brazil, focusing on the crucial role of a conditional cashtransfer program (Bolsa Família Program), given its many social implications, especially on poverty patterns and on the decline of income inequality. We can point out other important social changes occurring in Brazil in the last two decades: the growth of women's participation in the labor market, which influences the economic decisions of households, and the growth of average schooling of the population, which generates social returns; increasing the number of femaleheaded households, which has consequences on poverty, etc. It is assumed that all these changes can contribute to changes in the individual decisions, including the decision to migrate, including return migration. The methodology applies a hierarchical logit model that includes individual characteristics on the first level and municipal characteristics on the second, assuming that the probability of moving varies between different locations due to aggregate determinants. The main results indicate i) a decrease in non-return migration e return migration due to the program, and ii) significant effects of municipal characteristics on migration. INTRODUCTION Studying the phenomenon of migration and its implications is of paramount importance not only for understanding the functioning of the labor market, but also the functioning of society in general. To understand this process, it is important to address its determinants. In developing countries, two causes can be emphasized: regional disparities (Borjas, 2004; Cooper, 1994) and adverse social conditions (Dedecca and Cunha, 2004). In short, people migrate from poor regions to rich regions in order to improve their economic and social conditions. As the Bolsa Família Program (BFP) has contributed to the decline in regional disparities and to improve social conditions, it can be argued that it is possible that it is influencing individuals' decisions to migrate. In the last two decades, Brazil has undergone profound changes, among them the reduction in income inequality and poverty alleviation. Soares (2006) showed that, in the middle of the 2000s, Brazil has achieved the lowest income inequality level since the 1970s, with the bulk of the drop attributed to social programs, especially since the 1990s. Part of the literature indicates that the BFP is mainly responsible for the fall in inequality from 2003, the year of its implementation. The BFP was created from the unification of other social programs, with the aim of eradicating poverty 1 The authors acknowledge the support of CNPq and FAPEMIG to the development of the work. 2 PhD candidate in Economics at Cedeplar/UFMG Brazil. 3 Associate Professor at Cedeplar/UFMG Brazil.
2 in Brazil. According to data from the Brazilian Ministry of Social Development (MDS), the BFP currently serves more than 13 million families. We can highlight three changes that should be taken into account in recent years: first, the increasing number of families headed by women. If in 2000 Brazil had 24.9% of households headed by women, in 2010 the percentage was of 38.7%. Most studies (Azcona, 2009; Gama, 2014) show that men are more propense to migrate. In fact, since most families are headed by men, this is not a surprise, since a good part of the movements is driven by work opportunities. Therefore, the growing in the number of households headed by women can have impact in this result. Gama (2014) showed that men are still more propense to migrate, but the difference in significance felt between 2000 and 2010. Second, the increase in female labor market participation, which affects the economic decisions of households and is also connected to the increase of female-headed households. Last but not least, specifically with regard to migration, the growth of average level of education in Brazil can contribute to a reduction in mismatched migration because most educated individuals observe conditions at the destination in a more well-informed basis, which reduces the cost of migration and the probability that the migration worsens their situation. The increase in the average education of the population also has a social contribution; e.g. more educated parents are more likely to invest in children's education (Appleton and Mackinnon, 1993). The decision to migrate is usually taken for the same reasons: individuals aspire to improve their economic conditions, have greater access to education, or are forced, either by crises or natural disasters (Borjas, 2004; Azcona, 2009; Dustmann and Glitz, 2011 ). Thus, migration can be considered an investment in human capital, and it is expected that migrants are more educated and younger than non-migrants. There is not a consensus whether the decision to migrate is individual or family. Here, similarly as in Mincer (1978) and Borjas (2004), migration is considered a family decision. Within this framework, the concept of tied movers and tied stayers becomes important, since in many cases women move to follow their partners, even if it leads to a drop in their income, i.e., despite individual gains are not sufficient to lead them to migrate, the gains are familiar. Women should not necessarily be the tied mover, but due to persistent gender discrimination in Brazil, men receive higher earnings and are generally are still the breadwinner of the family. Return migration, defined as individuals who left their place of birth (or origin locality) and decided to return after some time, has grown considerably in recent years in Brazil. This decision can be understood as an optimal decision of the migrants in their life cycle: they migrate to acquire knowledge and return to the place of origin, because this new knowledge gained there will be more appreciated (Borjas, 1994). Another reason can lead individuals to return to their place of origin: after migrating, they can conclude that the prediction was inaccurate and the return becomes a way to correct this mistake. In this paper, we introduce a third hypothesis for this phenomenon: for those that receive the benefit or expect to receive in the origin city, income arising from the receipt of the benefit can outweigh the costs associated to the return to hometown, mainly in Northeast, where the purchase power is higher. Therefore, BFP would be contributing to the increase of return migrants, mainly in Northeast.
3 DETERMINANTS OF MIGRATION AND RETURN MIGRATION A pioneering study on migration is due to Sjaastad (1962), in which he states that a worker faces monetary and non-monetary costs while migrating in order to maximize their utility over the life cycle. The author views the migration process similarly to the educational background, i.e. an individual investment. It is largely known the higher the earnings in a given locality, the less likely workers quit their jobs and migrate to another region. However, the higher the total income of the workers, more conditions they have to support their migration costs, which could increase their likelihood of migration (Pereira, 2000). According to Sasaki and Assis (2000), recent research has challenged assumptions and conclusions of neoclassical theory, among them that the decision to migrate would not be an individual decision, but a joint decision, a family decision, not being only related to monetary reasons. In this paper, variables related to family characteristics are used as controls in the estimates. To understand the migration phenomenon in Brazil and what leads someone to take such decision we need to consider some characteristics, as sex, age, condition in the family, etc. For example, Oliveira and Jannuzzi (2005), using a national representative dataset, found that about 35% of men and 12% of women declared the reason for migration was work, about 3% for studying and 40% of men and 63% of women declared the reason was follow the family. So, first of all we need to decide and understand the sample we are looking at before trying to make any predictions about the reasons that leaded than to migrate. If we want to study individuals that migrated for work, the better approached would be use the heads of the families. In recent years, the phenomenon called return migration has grown in Brazil. Return migrants are defined as individuals who left their places of origin, but after a while returned. From the individual point of view, the consolidated literature attests that return migration can be understood as an optimal decision of migrants in their life cycle. They migrate to broaden their knowledge or to work and then returns to his homeland, where this new knowledge is more valued (BORJAS, 1994). Another reason may also lead an individual to migrate back: after migrating the worker can realize the decision was a mistake, and therefore decide to return to the region of origin. In the present work we introduce a third option: for those that receive the BF benefit or expect to receive in the origin city, income arising from the receipt of the benefit can outweigh the costs associated to the return to hometown. As stated for migration, we need first understand the sample or population in focus and then try to figure out the reasons for such a movement. Eggert et al. (2009) argue that regions where wages are higher also have a high proportion of skilled workers and a high demand for these workers, which attracts workers from poorer regions. Thus, these poorer regions may suffer from the problem of brain drain. Dustmann et al. (2010) state that return migration can lead to a decrease in the brain drain, or even to greater attraction of brains, brain gain, occurring when individuals who return to their home countries bring higher skills. According to Mayr and Peri (2008) return migration combined with incentives can be an important factor to transform a brain drain situation in brain gain, especially in less developed countries, which are those most affected by this problem. Encourage individuals to migrate to regions where opportunities are more advantageous can also contribute to reducing unemployment. This phenomenon is more common or more observable in the context of international migration, but in a country with continental dimensions like Brazil is probably applicable, even if in a smaller magnitude.
4 In the Brazilian context, Cunha and Baeninger (2005), comparing the 1970s to the 1990s, found that there was an increase of 221% of return migration, specifically talking about internal migration. In addition, higher return contingents were recorded in the Northeast, a region that is historically a population expeller. Cunha (2006) states that around 30% of migratory movements were of return migration, and, just analyzing the Northeast, this figure rises to 51%. Ramalho and Silveira Neto (2009) found that the probability of return migrants be employed in the formal sector is positively correlated with their level of education. Among them, female migrants, with low education and who are not household heads are more likely to enter in the informal sector. Thus, return migrants are more educated than non-migrants, and, moreover earn higher wages. Also according to the authors, between 2003 and 2007, 11% of interstate migration was of return. Between 1995 and 2000, 41% of return migrants were headed to the Northeast. Contrasting to what occurs in many countries, the majority return migrants were not older people who returned to their places of origin, but people between 30 and 39 years that probably were not successful at the destination location. DATA Data used are from the 2010 Brazilian Demographic Census, conducted by the Brazilian Institute of Geography and Statistics (IBGE), and comprise a sample of individuals (level 1) and municipalities (level 2). The sample was restricted to individuals who were between 21 and 65 years, since the purpose is to assess migrants who made the decision to migrate based on economic incentives. The sample consists of 11,789,858 individuals, consisting of approximately 48.5% white, 51.4% women, mean age is 39.1 years and the majority of individuals (42.9%) have no education or have less than eight years of schooling. Furthermore, only 11.4% of the sample had completed higher education. Altogether, we have 5.565 municipalities in Brazil. In the econometric estimates, municipalities are divided into six different groups based on their population sizes, which are: up to 20,000 inhabitants; between 20,001 and 50,000; between 50,001 and 100,000; between 100,001 and 500,000; between 500.001 and 1,000,000; and population over 1,000,000 inhabitants. Moreover, data from the Atlas of Human Development 2013 are also used, which organizes social and economic information on municipalities. As the aim is to analyze the effect of a cash transfer program on an individual decision, the estimates take into account observations of only those eligible to receive the benefit. This restriction is necessary because it is assumed that the program does not exercise influence on those who do not meet the criteria for receipt. In 2010, where considered eligible families with a monthly per capita income between R$70.00 and R$140.00 4 and with at least one child or adolescent under 17 years in the household. Families with a monthly per capita income less than R$ 70.00 per month entering the program regardless of their composition. Hierarchical Logit Model In order to achieve the objectives proposed in the paper, we apply a hierarchical logit model, under the assumption that the probability of workers migration differ among municipalities. In multilevel or hierarchical analysis, the dependent variable is measured at the lowest level of disaggregation 4 In October 2010 these values where about 41 and 82 dollars respectively.
5 (level 1) and the explanatory variables can be specified on the first level or at higher levels (Fontes, Simões and Hermeto, 2006). Based on Tabachnick and Fidell (2007), we highlight two advantages of applying a hierarchical model: no need for independence of errors, and avoids the so-called ecological fallacy, which occurs when interpreting aggregate results at the individual level, confounding individual effects with aggregate effects. When estimating the probability of migrating in certain locations while controlling for individual characteristics, the methodology has the advantage that variables that are defined at different levels of hierarchy can be combined into a single model. This structure is based on a correlation assumption, or non-independent clustered data, so that individuals within the same municipality will tend to be more alike. This hypothesis is reasonable because municipal characteristics may influence the decision of an individual to migrate for the first time or to migrate back. Multilevel modeling allows investigating the nature of the variation between groups and the effects of the characteristics at the group level on individual results. Two hierarchical logit models are estimated, with two levels, to address the binary dependent variable, which take the value 1 if the individual is a migrant or return migrant and 0 if the individual is not migrant. The first level refers to the individual characteristics and the second to the information of 5565 municipalities available on the 2010 Demographic Census. 5 The model specification 6 consists of a two level structure where a total of i individuals (level 1) are nested within j municipalities (level 2). The logit model, considered in its generalized linear fashion, with a random intercept, can be represented as in equation 1: log ( π ij 1 π ij ) = E(Y ij X ij, μ j ) = β 0 + β 1 X ij + μ j, where μ j ~N(0, σ u 2 ) (1) π ij 1 π ij is the odds of Y=1 to the individual i, πij represents the probability of success for individual i and log ( π ij 1 π ij ) is the log-odds, also known as logit. β 0 can be interpreted as the log-odds of Y=1 when X=0 and μ =0, in other words, it is the overall intercept in the linear relation between the log-odds and X. β 1 has the same interpretation of the one level model. The intercept for a specific group j is β 0 + μ j, and can be bigger or smaller than the overall intercept. μ j is the group effect (random) or level 2 effect. The between-group variance is represented by Var(μ j ) = σ μ 2. Random intercept models assume that the random term only affects the model intercept, in a way that each explanatory variable X has the same effect on Y in all groups. In the present work, we assume that this is a strong assumption, in others words, we do not expect that BFP has the same effect on the probability of migrate in all municipalities. Therefore, a random coefficient model will be tested and applied in the present work. 5 The data are unbalanced, that is, there is not the same number of individuals in each municipality. However, this is not a problem because multilevel models do not require that there is the same number of units at the lowest level in each unit of the highest level (Rasbash, 2008). 6 Based on Steele (2008b).
6 We can extend equation (1) so it can entail the random term in the explanatory variables, as in equation (2): log ( π ij 1 π ij ) = β 0 + β 1 X ij + μ 0j+ μ 1j X ij (2) It is assumed that the random effects μ 0j e μ 1j follow a normal distribution with 0 mean, variances 2 2 σ μ0 and σ μ1 respectively and covariance σ μ01. Now, the slope of the linear relation between X and the log-odds of Y=1 is β 1 + μ 1j for group j. The procedure of estimation is by Maximum Likelihood via Adaptive Quadrature. Independent variables at the first level are: dummies for marital status, gender, race, and education levels, age, squared age, number of children living in the household and the treatment variable - status of beneficiary of Bolsa Familia Program. At the second level, the variables are: geographic region, municipality per capita income, proportion of Bolsa Familia Program beneficiaries, and urbanization rate. Regarding the migration status, we considered migration between municipalities, both within-state and between-state. Migrant is a person who moved to the current place of residence at most five years ago. Return migrants are individuals who in July 31, 2010 lived in the same municipality that July 31, 2005, but had done some migratory movement in the between. Individuals residing for more than five years in the same municipality are considered native. This definition can be viewed in a more clear way in Table 1. Table 1 Definition of migrant and return migrant Questions Possible answers 1 - Time of residence in the municipality 0 a 65 years. Missing for those who never migrated. 2 - Municipality of residence on July 31, 2005 3 - Municipality of residence on July 31, 2010 4- Municipality of previous residence Variables created from the previous questions Answers to previous questions Dummy equal to 1 if different responses to questions 2 and 3 Migrant and equal to 0 if time of residence in the municipality is equal or more than five years, or missing. Dummy equal to 1 if identical responses to questions 2 and 3 and e live for less than five years in the same municipality. Equal Return migrant to 0 if time of residence in the municipality is equal or more than five years, or missing. Source: Own elaboration, based on Gama (2013). Expected results are: first, migrants are positively selected with respect to observable characteristics, in other words, are more educated and receive higher wages. Second, it is expected that the BFP reduce the likelihood of an individual migrate due to social and regional improvements, which mean retention effects. Third, it is expected that the receipt of the benefit increases the probability of return migration, because, due to the improvements brought about by social and regional program, individuals may conclude that the reasons that led them to migrate in the past (search for economic progress) can now be obtained from the place of origin.
7 Endogeneity Dealing with causality between migration and the reception of Bolsa Família benefit, the endogeneity issue arises. Our hypothesis is that the program exerts influence on the odds of someone migrate. However, it is possible that some characteristics of migrants and return migrants change the odds of receiving the benefit. In order to overcome the problem just mentioned, we choose first to do a Logit regression to identify what causes an individual to migrate, taking in consideration the follow variables: marital status dummy, education level, age, squared age, dummy for race, number of children living in the household, per capita household income without the PBF and geographic region of residence. Given the low correlation between family income and the decisions for migrate (since every family in the sample has low income), as can be observed on Table 2 and attested by Gama (2013), we use this variable as instrument. Table 2 Correlation between migrant, return migrant and family income Migrant Return Migrant Family income per capita without Bolsa Familia Migrant 1 Return Migrant. 1 Family income per capita without Bolsa Familia 0.0143-0.0023 1 Fonte: Own elaboration based on the 2010 Census data. RESULTS Descriptive Statistics Before the econometrics analysis it is important to do some descriptive analysis in order to know better the data we are using. In recent years several countries started to spend much of their resources in social assistance programs. Nevertheless, in Brazil and in other developing countries, a significant portion of the poor remains out of access to this assistance, which analysts say could be a result of governments focusing those programs in a not properly way (Fernandes and Pazello, 2001 ). In Brazil, around 52.67% of the individuals included in our sample had access to the Bolsa Família. Therefore, much of the population that need it does not have access to such benefit. In Table 3, it can be seen that the those eligible but not benefit from the program are on average more educated, have higher incomes, the white percentage is higher, and the percentage of residents in urban areas, metropolitan areas and is also higher the percentage of migrants and returning migrants. It is therefore possible that some of the eligible who do not receive the benefits, choose not to apply for benefits, given their better observable characteristics. It may also be an indication that the program is more focused on rural areas and outside metropolitan areas. One can observe the very low percentage of individuals living in urban areas, especially among the beneficiaries, as in Brazil, according to the Census own in 2010, around 85% of the population lives in urban areas. Regarding marital status, it is worth noting that the married percentage is higher among beneficiaries (78.53%) than among non-beneficiaries (72.15%). But the most interesting is the big
8 difference, among the beneficiaries, in the percentage of single women and single men. While single women, who largely are left with the children on a possible separation, represent 20.48% of beneficiaries, single men represent only 1.61%. The fact that for both groups the percentage of men with spouses is higher is due to the fact that our sample was composed mostly of men (60.94%) and individuals living with partners (71.63%). Last but not least, regarding the place of residence, it is observed that the overwhelming majority of beneficiaries is in the Northeast, while among non-beneficiaries the percentage is higher in the Southeast. This finding was expected, given the increased focus of the program in the Northeast. Table 3: Descriptive Statistics: Beneficiaries and non-beneficiaries Beneficiary Non beneficiary age (mean) 38.67 39.08 Education level (% with college degree) 0.22 3.33 Race (% of white) 26.03 36.78 Marital status (%) Single woman 20.48 22.57 Woman with spouse 19.85 15.07 Single man 1.61 12.78 Man with spouse 58.05 49.58 Per capita household income without PBF (mean) 48.82 52.80 Region of residence (%) North 10.85 12.69 Northeast 63.60 35.51 Southeast 16.52 36.54 Southeast 5.60 8.88 Midwest 3.43 6.39 Area of residence (% of residentes in urban areas) 59.9 75.29 % of residents in metropolitan areas 24.00 45.77 Migrants (%) 6.15 8.10 Return migrants (%) 1.46 2.11 Source: Own elaboration. Regarding the migration process we found that 90.78% of the individuals are not migrants, 7.53% are migrants and 1.69% return migrants. Table 4 shows that, as expected, migrants are more educated, younger and get better yields than non-migrants. In other words, they are positively selected with respect to observable attributes. In relation to marital status it is not observed major differences among the three groups. Regarded with region of residence, it is observed that the Northeast has the largest percentage of individuals in all three groups, which is not surprising, given that concentrates the highest percentage of eligible people in the country. However, it is worth noting that the percentage of migrant and return migrants in Northeast (40.21 and 41.98% respectively) are well below the percentage of nonmigrants (51.25%), showing that the Northeast is less attractive to migrants than other regions.
9 Finally, we observe that among the non-migrants the percentage of beneficiaries is higher than among the rest. It is likely that the fact that the individuals living longer in the same location have a greater range of information about access to benefits has contributed to this result. Table 4: Descriptive statistics: migrants, return migrants and non-migrants Migrantes Migrantes de retorno Não migrantes age (mean) 35.77 35.48 39.17 Education level (% with college degree) 2.98 4.32 1.54 Race (% of white) 34.69 35.48 30.76 Marital status (%) Single woman 21.71 23.05 21.42 Woman with spouse 17.39 17.48 17.60 Single man 8.25 8.68 6.76 Man with spouse 52.65 50.79 54.21 Per capita household income without PBF (mean) 53.57 50.50 49.55 Region of residence (%) North 13.66 11.57 11.58 Northeast 40.21 41.98 51.25 Southeast 27.42 28.49 25.83 Southeast 9.87 9.72 6.89 Midwest 8.84 8.23 4.46 Area of residence (% of residentes in urban areas) 72.66 75.32 66.6 % of residents in metropolitan areas 34.81 38.45 34.18 Beneficiaries (%) 45.79 43.53 53.38 Source: Own elaboration. Estimation Results Table 5 presents the results of the ordinary logit model estimated to identify what leads an eligible individual to have greater or lesser chances of receiving the benefit and to correct the endogeneity problem. Considering the variable living arrangement, regardless the size of the municipality, households with couples present higher odds of receiving the benefit. Among those with children, the odds are even higher comparing to singles without kids. As for race, regardless the size of the municipality of residence, blacks and browns are more likely to be benefited by the program, which was expected. Older individuals also are more likely to receive the benefit, however, when age becomes more advanced, fall the chances of being assisted by the benefit, probably because most of them receive others benefits, as pensions and retirement benefits. For all sizes of municipalities, as expected, more educated individuals are less likely to receive the benefit, and this effect is greater the extent that the municipality size grows. In larger municipalities, employment opportunities for more qualified individuals are larger and it is likely that this has led to this result. The per capita income without the BFP affects negatively the odds of receiving the benefit in municipalities with up to 1,000,000 inhabitants and increases in larger municipalities, which was not expected. However, the coefficient is close to 1, so it can be
10 considered that, because the sample contain only eligible individuals, income is not decisive for the receipt, despite the coefficient being significant. Table 5 - Logit Odds of Receiving PBF 1 2 3 4 5 6 Living arrangment (single without kids omitted) Single with kids 9.982*** 9.330*** 8.607*** 9.011*** 7.425*** 8.080*** (0.1652) (0.2448) (0.2982) (0.2930) (0.5532) (0.4642) Couple without kids 3.475*** 3.431*** 3.030*** 2.675*** 1.931*** 2.398*** (0.0591) (0.0968) (0.1174) (0.1044) (0.1827) (0.1749) Couple with kids 11.138*** 10.777*** 9.303*** 8.565*** 6.946*** 7.093*** (0.1634) (0.2615) (0.3041) (0.2712) (0.5078) (0.4004) Race (white omitted) Black 1.312*** 1.229*** 1.256*** 1.426*** 1.464*** 1.461*** (0.0166) (0.0232) (0.0308) (0.0291) (0.0715) (0.0529) Brown 1.298*** 1.209*** 1.250*** 1.326*** 1.350*** 1.364*** (0.0099) (0.0145) (0.0198) (0.0186) (0.0472) (0.0369) Age 1.211*** 1.193*** 1.155*** 1.106*** 1.037** 1.092*** (0.0034) (0.0050) (0.0064) (0.0057) (0.0134) (0.0103) Age squared 0.998*** 0.998*** 0.998*** 0.999*** 0.999*** 0.999*** (0.0000) (0.0001) (0.0001) (0.0001) (0.0002) (0.0001) Level of education (less than 8 years omitted) Between 8 and 10 years of education 0.810*** 0.800*** 0.797*** 0.719*** 0.734*** 0.755*** (0.0090) (0.0130) (0.0161) (0.0117) (0.0286) (0.0213) High School or incomplete college 0.619*** 0.598*** 0.560*** 0.481*** 0.498*** 0.486*** (0.0080) (0.0114) (0.0131) (0.0085) (0.0213) (0.0146) College or more than college 0.201*** 0.173*** 0.140*** 0.095*** 0.091*** 0.051*** (0.0098) (0.0134) (0.0143) (0.0083) (0.0185) (0.0086) Per capita household income without BFP 0.994*** 0.995*** 0.996*** 0.998*** 0.999*** 1.001*** (0.0001) (0.0001) (0.0001) (0.0001) (0.0003) (0.0002) Geographic region of the municipal (northeast omitted) North 0.447*** 0.442*** 0.423*** 0.582***. 0.473*** (0.0053) (0.0065) (0.0076) (0.0114) (.) (0.0167) Southeast 0.470*** 0.407*** 0.361*** 0.340*** 0.337*** 0.233*** (0.0042) (0.0060) (0.0071) (0.0049) (0.0110) (0.0061) South 0.389*** 0.339*** 0.379*** 0.469*** 0.510*** 0.500*** (0.0044) (0.0077) (0.0105) (0.0104) (0.0456) (0.0293) Midwest 0.329*** 0.339*** 0.313*** 0.368*** 0.461*** 0.459*** (0.0048) (0.0093) (0.0119) (0.0116) (0.0311) (0.0212) Number of observations 532001 209452 112021 144938 24218 55779 Note 1: 1: Population up to 20.000 inh.; 2: Population between 20.001 and 50.000 inh.; 3: Population between 50.001 and 100.000 inh.; 4: Population between 100.001 e 500.000 inh.; 5: Population between 5000.001 and 1.000.000 inh.; 6: Population over 1.000.000 inh. Note 2: In region North there is no muncipality with population in the 5 range. * p <0.05, ** p<0.01, *** p <0.001 Source: Own elaboration based on the estimates. In each tables 6, 7, 8, and 9 below are presented six estimations (one for each size of municipality) and the coefficients are reported in odds ratio. Tables 6 and 7 present as dependent variable the migration dummy and interactions between the BFP dummy and the regions dummies. Since one of our hypotheses is that the region is important to explain the movement, the inclusion of interactions can better help us to understand the process. Tables 8 and 9 report the return migration dummy and interactions between the BFP dummy and living arrangement dummies. As explained before, the family composition have changed a lot in
11 Brazil in the last years, so we try to understand how individuals in these new family formations, with access to this benefit, decide or not to migrate in comparison to those who do not receive it. Starting from the random coefficients and the necessity of such modeling, it is important to mention that in all models but model 6 in Table 9 the statistics of test of the Likelihood Ratio test when compared to the qui-squared distribution with three degrees of freedom 7 produce p-values up to 0.03. Therefore, at 5% significance level, it is concluded that the effect of PBF on the probability of migrating and return migrating varies between different municipalities. It is worth mention the effect of the other explanatory variables on the odds of being a migrant or return migrant is the same for all municipalities. In all estimated models, the intercept and random coefficients were allowed to covariate, in other words, it was assumed that they were not independent. The reported coefficient is significate only for municipalities up to 20,000 inhabitants. The negative signal means that municipalities with above average percentage of migrants (residual intercept μ 0j > 0) tend to have effects of BFP below the average (residual coefficient below zero). In another words, the effect of Bolsa Familia on the odds of migrate and return migrate are lower in destiny municipalities with high percentages of migrants and returned migrants. Before analyzing the effect of our variable of interest on the probability of migration, some comments about interactions should be made. The inclusion of interactions in our models allow us to better understand the relationships among the variables in the model and allows more hypotheses to be tested. The inclusion of the interaction between BFP and regions and BFP and living arrangement is to test if the relationship between BFP and the dependents variables is different in different regions and different families. Besides, the inclusion of interactions changes the interpretation of the others coefficients. Now, for interactions with regions, dummy PBF is interpreted as the effect of Bolsa Familia on the variable dependent in region Northeast and the region dummies measure the relationship with the dependent variable for those who are not benefited by the program. For the living arrangement interactions, the coefficient of BFP dummy is interpreted as the effect of Bolsa Familia on singles without children, and the family dummies are for those who do not receive the benefit. One can see in table 6 that for municipalities with population up to 100.000 inhabitants, among residents of Northeast, to receive Bolsa Familia decreases from 13.4 to 33.1% the chances that an eligible individual to receive the benefit be a migrant. In cities with populations between 100,001 and 1,000,000 inhabitants the coefficient was not significant and in cities with population over 1,000,000 inhabitants, receive the benefit increases the chances of the individual be a migrant in about 30%. Bottom line, in major cities migrants seem to have better chances of getting the benefit, with encourage them to migrate. However, can be a result of a difficult experienced by them to find jobs or jobs in the formal sector, which leads them to receive the benefit. Still focusing on Table 6, in small cities the benefit probably lead those that receive it to continue in the city, raising the odds of those that receive it being no migrants. The interactions coefficients show that in municipalities up to 50,000 inhabitants, beneficiaries in the others regions are more likely to be migrants than in Northeast. Again, this result show how in Northeast the BFP is 7 Model 5 in Table 8 could not be estimated with a random coefficient for Bolsa Familia. The variance components were too close to the boundary of the parameter space. When such problem happens likelihood ratio test can no longer be trusted and p-values are showed to be incorrect (citation), so standard deviations are no reported by Stata. So, in this cases we opted for estimate a random intercept model for this subsample.
12 influencing the no mobility of its population. Finally, the coefficients of the variable that defines the beneficiary municipal average show that in municipalities with population bigger than 20,000 and up to five hundred thousand, the percentage of the population receiving the benefit decreases the odds of migrating drastically. This result can be reflecting two scenarios: first, people in this city do not migrate due the fact that they are receiving the benefit and they do not want to risk loose it, so the percentage of non-migrants is higher; second, this can be a reflect that those cities are poor, so most inhabitants need the benefit, and poor cities present less opportunities and consequently attract less migrants. Analyzing the interacted coefficients in Table 7, one can see that independently of the size of the municipality, among those singles individuals without kids, receive the benefit decreases the odds of being a migrant. It is worthy noting that those individuals are the most vulnerable to poverty, so probably they do not want and they do not need to risk losing the benefit by migrating. Those that do not have access to the benefit need to chase other options, and one of them is migration. If they are married or have children, the necessity of look to other options is even bigger, since they are responsible for others. Despite the size of the municipality, as expected, beneficiaries are more likely to be married and to have kids. Relating to return migration, in Table 8, one can see that for those living in municipalities with population up to 50,000 inhabitants, receive the PBF decreases the odds of being return migrant in Northeast. For the rest of the municipalities the coefficient is not significant. So, the hypothesis that PBF could be increasing the return migration in poorest regions seems do not stand. Looking at the interactions terms, in small municipalities (model 1) beneficiaries are more likely to be return migrants in South and Southeast in comparison with Northeast. The percentage of eligible receiving BF does not affect the odds of return. Finally, the models in Table 9 show the same result for the coefficients of PBF dummy and for the coefficients of the variable that defines the percentage of beneficiaries per municipality. For the interactions with family type, the results are similar to those found on Table 6, with heads of the household living with partners and having kids presenting higher odds of return.
13 Table 6 - Hierarchical Logit Model to migration, with region interaction 1 2 3 4 5 6 Fixed Effects Level 1 BFP beneficiary? (No omitted) 0.669*** 0.704*** 0.866** 0.925 1.062 1.301* (0.0136) (0.0205) (0.0398) (0.0491) (0.0984) (0.1387) Living arrangment (single without kids omitted) Single with kids 1.115* 1.101 1.472** 1.519* 0.320 0.024*** (0.0490) (0.0870) (0.1752) (0.2684) (0.2996) (0.0269) Couple without kids 1.334*** 1.298*** 1.375*** 1.434*** 0.758 0.243** (0.0400) (0.0698) (0.1033) (0.1226) (0.2404) (0.1143) Couple with kids 0.984 1.024 1.338* 1.584** 0.395 0.035** (0.0410) (0.0792) (0.1538) (0.2634) (0.3504) (0.0366) Predict BF 0.966* 0.896*** 0.767*** 0.657*** 1.223 3.859* (0.0165) (0.0292) (0.0407) (0.0533) (0.5760) (2.0279) Level 2 Geographic region of the municipal (northeast omitted) North 1.378*** 0.958 0.976 0.708* 1.000 3.569** (0.0685) (0.0716) (0.1189) (0.1044) (.) (1.6697) Southeast 0.730*** 0.535*** 0.496*** 0.400*** 1.387 16.279*** (0.0327) (0.0440) (0.0682) (0.0614) (0.7821) (13.7821) South 0.666*** 0.533*** 0.602*** 0.547*** 1.737 11.672*** (0.0374) (0.0539) (0.0921) (0.0841) (0.7265) (6.1849) Midwest 1.132* 0.753** 0.904 0.694 1.148 13.848*** (0.0646) (0.0786) (0.1644) (0.1293) (0.5208) (7.2540) Percentage of elegible receiving BF 0.922 0.487*** 0.377*** 0.242*** 1.134 0.708 (0.0826) (0.0852) (0.1097) (0.0934) (0.9786) (0.8741) Interactions by Region of Residence BF beneficiary x North 1.160*** 1.115 0.971 1.006 1.000 1.071 (0.0475) (0.0646) (0.0802) (0.1023) (.) (0.1905) BF beneficiary x Southeast 1.305*** 1.406*** 1.140 1.198** 1.036 1.260 (0.0418) (0.0766) (0.0919) (0.0813) (0.1426) (0.1919) BF beneficiary x South 1.465*** 1.390*** 1.087 1.093 1.446 0.595 (0.0546) (0.1055) (0.1155) (0.1030) (0.3743) (0.1582) BF beneficiary x Midwest 1.204*** 1.324*** 1.054 1.055 1.559 0.824 (0.0516) (0.1083) (0.1396) (0.1311) (0.3625) (0.1595) Random Effects Var (Bolsa Família) 0.086*** 0.037*** 0.069*** 0.066*** 0.011* 0.009*** (0.0114) (0.0160) (0.0217) (0.0173) (0.0198) (0.0130) Var (constant) 0.285*** 0.208*** 0.224*** 0.199*** 0.028*** 0.028*** (0.0129) (0.0174) (0.0260) (0.0221) (0.0185) (0.0146) Cov(Bolsa Família, constant) -0.046*** -0.007-0.026-0.010-0.018-0.002 (0.0101) (0.0133) (0.0189) (0.0148) (0.0204) (0.0134) Number of observations 523848 205846 110019 142062 23727 54734 Number of groups 3912 1043 325 245 23 15 loglikelihood -139064-48240.205-26811.595-40266.264-5522.2148-9094.3612 chi2(3) 5152.55 1356.14 969.21 1834.84 9.02 24.85 Note 1: 1: Population up to 20.000 inh.; 2: Population between 20.001 and 50.000 inh.; 3: Population between 50.001 and 100.000 inh.; 4: Population between 100.001 e 500.000 inh.; 5: Population between 5000.001 and 1.000.000 inh.; 6: Population over 1.000.000 inh. Note 2: In region North there is no muncipality with population in the 5 range. Not 4: All models are controlled by race, age, age squared, level of education, percentage of urbanization in the municipality and average municipality household income. * p <0.05, ** p<0.01, *** p <0.001 Source: Own elaboration based on the estimates.
14 Table 7 - Hierarchical Logit Model to migration, with living arrangement interaction 1 2 3 4 5 6 Fixed Effects Level 1 BFP beneficiary? (No omitted) 0.612*** 0.616*** 0.606*** 0.620*** 0.912 0.520* (0.0325) (0.0587) (0.0806) (0.0734) (0.2277) (0.1550) Living arrangment (single without kids omitted) Single with kids 1.110* 1.086 1.342* 1.502* 0.261 0.022*** (0.0523) (0.0893) (0.1632) (0.2660) (0.2455) (0.0246) Couple without kids 1.211*** 1.213** 1.273** 1.336** 0.627 0.224** (0.0407) (0.0716) (0.1036) (0.1184) (0.2037) (0.1059) Couple with kids 0.957 1.019 1.343* 1.509* 0.280 0.031** (0.0407) (0.0798) (0.1558) (0.2530) (0.2506) (0.0327) Predict BF 0.961* 0.886*** 0.758*** 0.659*** 1.433 3.999** (0.0165) (0.0290) (0.0404) (0.0538) (0.6802) (2.1187) Level 2 Geographic region of the municipal (northeast omitted) North 1.534*** 1.028 0.963 0.716* 1.000 3.526** (0.0676) (0.0716) (0.1120) (0.1035) (.) (1.6335) Southeast 0.868*** 0.632*** 0.520*** 0.424*** 1.764 16.584*** (0.0350) (0.0490) (0.0686) (0.0641) (1.0085) (13.9883) South 0.839*** 0.626*** 0.622** 0.566*** 2.446* 9.395*** (0.0435) (0.0600) (0.0909) (0.0854) (0.9970) (5.1210) Midwest 1.294*** 0.868 0.919 0.713 1.710 12.245*** (0.0685) (0.0860) (0.1613) (0.1304) (0.7614) (6.4807) Percentage of elegible receiving BF 1.000 0.536*** 0.388** 0.244*** 1.248 0.450 (0.0892) (0.0935) (0.1124) (0.0944) (1.1676) (0.5409) Interactions by Family type BF beneficiary x Single with kids 1.238*** 1.306* 1.687*** 1.555*** 0.955 2.499** (0.0742) (0.1368) (0.2413) (0.1929) (0.2633) (0.7767) BF beneficiary x Couple without kids 1.501*** 1.484*** 1.723*** 1.963*** 1.887* 3.018** (0.0913) (0.1628) (0.2668) (0.2730) (0.6068) (1.0547) BF beneficiary x Couple with kids 1.276*** 1.279* 1.418** 1.682*** 1.349 2.652** (0.0694) (0.1245) (0.1916) (0.2006) (0.3468) (0.7937) Random Effects Var (Bolsa Família) 0.102*** 0.055*** 0.071*** 0.076*** 0.021** 0.038*** (0.0121) (0.0175) (0.0220) (0.0182) (0.0282) (0.0245) Var (constant) 0.291*** 0.209*** 0.223*** 0.200*** 0.034*** 0.035*** (0.0131) (0.0176) (0.0259) (0.0223) (0.0240) (0.0191) Cov(Bolsa Família, constant) -0.055*** -0.011-0.025-0.012-0.026-0.020 (0.0103) (0.0141) (0.0189) (0.0157) (0.0256) (0.0210) Number of observations 523848 205846 110019 142062 23727 54734 Number of groups 3912 1043 325 245 23 15 loglikelihood -139103.82-48259.286-26803.146-40256.178-5519.542-9091.608 chi2(3) 5176.23 1364.91 973.34 1858.79 10.29 32.12 Note 1: 1: Population up to 20.000 inh.; 2: Population between 20.001 and 50.000 inh.; 3: Population between 50.001 and 100.000 inh.; 4: Population between 100.001 e 500.000 inh.; 5: Population between 5000.001 and 1.000.000 inh.; 6: Population over 1.000.000 inh. Note 2: In region North there is no muncipality with population in the 5 range. Not 4: All models are controlled by race, age, age squared, level of education, percentage of urbanization in the municipality and average municipality household income. * p <0.05, ** p<0.01, *** p <0.001 Source: Own elaboration based on the estimates.
15 Table 8 - Hierarchical Logit Model to return migration, with region interaction 1 2 3 4 5 6 Fixed Effects Level 1 BFP beneficiary? (No omitted) 0.659*** 0.792*** 0.875 0.900 0.868 0.780 (0.0271) (0.0431) (0.0645) (0.0685) (0.1237) (0.1648) Living arrangment (single without kids omitted) Single with kids 1.165 0.986 1.123 0.154*** 2.143 0.144 (0.1121) (0.1482) (0.2574) (0.0554) (3.4041) (0.2416) Couple without kids 1.138 1.055 1.177 0.477*** 1.432 0.489 (0.0773) (0.1100) (0.1719) (0.0832) (0.7752) (0.3470) Couple with kids 0.878 0.879 0.911 0.166*** 1.859 0.182 (0.0807) (0.1296) (0.2025) (0.0563) (2.7960) (0.2886) Predict BF 1.078* 1.003 0.970 2.083*** 0.633 1.819 (0.0400) (0.0621) (0.0982) (0.3418) (0.5053) (1.4429) Level 2 Geographic region of the municipal (northeast omitted) North 1.078 1.183 1.029 1.363 1.000 2.474 (0.0884) (0.1295) (0.1588) (0.2333) (.) (1.5844) Southeast 0.908 0.851 0.851 1.513 0.913 5.455 (0.0657) (0.1022) (0.1492) (0.3273) (0.8748) (6.5380) South 0.898 0.942 1.147 1.479* 1.603 5.495** (0.0814) (0.1387) (0.2186) (0.2833) (1.1170) (3.4634) Midwest 1.134 1.301 1.254 3.118*** 1.407 7.274** (0.1075) (0.1952) (0.2799) (0.7343) (1.0597) (4.8663) Percentage of elegible receiving BF 1.110 1.134 0.845 0.499 5.888 2.320 (0.1518) (0.2669) (0.2612) (0.1838) (8.7375) (2.6101) Interactions by Region of Residence BF beneficiary x North 1.069 0.936 1.038 0.891 1.000 1.410 (0.0924) (0.0986) (0.1402) (0.1335) (.) (0.4878) BF beneficiary x Southeast 1.136* 1.017 0.870 1.119 1.116 1.441 (0.0712) (0.1019) (0.1237) (0.1152) (0.2441) (0.4024) BF beneficiary x South 1.287*** 0.978 0.917 0.814 0.964 0.583 (0.0949) (0.1332) (0.1598) (0.1263) (0.4531) (0.3002) BF beneficiary x Midwest 1.116 0.598** 1.072 0.943 0.614 0.879 (0.1001) (0.1011) (0.2391) (0.1701) (0.2588) (0.3310) Random Effects Var (Bolsa Família) 0.089*** 0.034* 0.046* 0.030** - 0.068** (0.0392) (0.0500) (0.0573) (0.0365) - (0.0666) Var (constant) 0.295*** 0.203*** 0.119*** 0.140*** 0.052*** 0.006** (0.0299) (0.0342) (0.0362) (0.0270) (0.0292) (0.0099) Cov(Bolsa Família, constant) -0.047* -0.009-0.014-0.048-0.020 (0.0283) (0.0343) (0.0381) (0.0259) - (0.0246) Number of observations 489397 195532 104134 132181 22633 53284 Number of groups 3912 1043 325 245 23 15 loglikelihood -39943.887-17395.739-9597.3968-13405.344-2300.4564-4851.4321 chi2(3) - Model 5 chi2(1) 548.97 214.48 53.31 120.13 9.39 4.81 Note 1: 1: Population up to 20.000 inh.; 2: Population between 20.001 and 50.000 inh.; 3: Population between 50.001 and 100.000 inh.; 4: Population between 100.001 e 500.000 inh.; 5: Population between 5000.001 and 1.000.000 inh.; 6: Population over 1.000.000 inh. Note 2: In region North there is no muncipality with population in the 5 range. Not 4: All models are controlled by race, age, age squared, level of education, percentage of urbanization in the municipality and average municipality household income. * p <0.05, ** p<0.01, *** p <0.001 Source: Own elaboration based on the estimates.
16 Table 9 - Hierarchical Logit Model to return migration, with living arrangement interaction 1 2 3 4 5 6 Efeitos Fixos Level 1 BFP beneficiary? (No omitted) 0.609*** 0.700* 0.461** 0.723 0.522 0.600 (0.0752) (0.1246) (0.1349) (0.1618) (0.3121) (0.2596) Living arrangment (single without kids omitted) Single with kids 1.151 0.936 1.010 0.146*** 1.683 0.146 (0.1174) (0.1467) (0.2358) (0.0523) (2.6783) (0.2464) Couple without kids 1.090 1.047 1.038 0.402*** 1.109 0.486 (0.0819) (0.1189) (0.1638) (0.0726) (0.6158) (0.3469) Couple with kids 0.868 0.867 0.868 0.146*** 1.455 0.178 (0.0815) (0.1295) (0.1947) (0.0497) (2.2055) (0.2848) Predict BF 1.070 1.006 0.968 2.197*** 0.716 1.813 (0.0399) (0.0624) (0.0986) (0.3621) (0.5753) (1.4483) Level 2 Geographic region of the municipality (northeast omitted) North 1.133 1.144 1.036 1.324 1.000 2.571 (0.0766) (0.1112) (0.1457) (0.2009) (.) (1.6369) Southeast 0.986 0.853 0.800 1.667* 1.062 5.691 (0.0614) (0.0937) (0.1329) (0.3538) (1.0251) (6.8392) South 1.035 0.930 1.098 1.447* 1.747 5.685** (0.0839) (0.1276) (0.1958) (0.2630) (1.2049) (3.6290) Midwest 1.221* 1.103 1.263 3.217*** 1.383 7.633** (0.1035) (0.1565) (0.2618) (0.7172) (1.0418) (5.1484) Percentage of elegible receiving BF 1.170 1.132 0.822 0.486 5.576 2.952 (0.1590) (0.2653) (0.2527) (0.1804) (8.3179) (3.6819) Interactions by Family type BF beneficiary x Single with kids 1.177 1.152 2.009* 1.064 1.571 1.380 (0.1590) (0.2251) (0.6206) (0.2526) (0.9830) (0.6234) BF beneficiary x Couple without kids 1.261 1.062 2.278* 2.106** 3.701 1.410 (0.1785) (0.2235) (0.7492) (0.5614) (2.5414) (0.7912) BF beneficiary x Couple with kids 1.170 1.084 1.823* 1.273 1.624 1.534 (0.1471) (0.1971) (0.5419) (0.2907) (0.9916) (0.6640) Random Effects Var (Bolsa Família) 0.091*** 0.045** 0.047* 0.035** 0.001 0.119* (0.0395) (0.0506) (0.0573) (0.0383) (0.0059) (0.1229) Var (constant) 0.294*** 0.207*** 0.119*** 0.140*** 0.055*** 0.005* (0.0298) (0.0344) (0.0363) (0.0271) (0.0367) (0.0120) Cov(Bolsa Família, constant) -0.047* -0.015-0.014-0.047-0.005 0.025 (0.0284) (0.0346) (0.0381) (0.0268) (0.0314) (0.0392) Number of observations 489397 195532 104134 132181 22633 53284 Number of groups 3912 1043 325 245 23 15 loglikelihood -39948.921-17400.514-9594.2995-13400.035-2298.6202-4853.3508 chi2(3) 547.45 215.21 53.73 122.57 9.18 6.24 Note 1: 1: Population up to 20.000 inh.; 2: Population between 20.001 and 50.000 inh.; 3: Population between 50.001 and 100.000 inh.; 4: Population between 100.001 e 500.000 inh.; 5: Population between 5000.001 and 1.000.000 inh.; 6: Population over 1.000.000 inh. Note 2: In region North there is no muncipality with population in the 5 range. Not 4: All models are controlled by race, age, age squared, level of education, percentage of urbanization in the municipality and average municipality household * p <0.05, ** p<0.01, *** p <0.001 Source: Own elaboration based on the estimates.