Migrating out of mega-cities: Evidence from Brazil

Every 10 years, the Brazilian National Institute for Geography and Statistics (IBGE) conducts a 10% nationally representative household survey, the Census survey (Censo Demografico 2010, IBGE (2012)). The survey of 2010 comprises around 20 million individual observations in all municipalities of Brazil. It contains information on household composition, living conditions, labor market, education, geographic location, and migration.

The Census survey from 2010 allows to identify migrants in the sample using the following questions: “Were you born in this municipality?” – to know whether people are living in their birthplace; “When did you move to this municipality?” – provides the year of migration; and “In which municipality (in which state) did you live before you moved to this municipality that you are currently living in?” – provides the exact origin of migrants. It further asks for the municipality of the current job and the previous job. Migrants are individuals who used to live and work in a different location than the one they are living in at the time of the survey.

The sample comprises working-age migrants and non-migrant residents. The legal working age in Brazil is 16 years, and the retirement age for men is 65 years. The age group for the sample has been restricted from 25 years to 65 years. This way it can be assumed that students are excluded. All individuals in the sample are currently not in school and are participating in the labor market, which means that they are either employed or unemployed but looking for work. The author restricts the sample of migrants to those who moved within the past year, between 2009 and 2010, to minimize recall bias.

Migration is measured as a change in living and working location at the level of a microregião. Microregiões are geographic and administrative agglomerations of municipalities sharing labor market and economic activities, a bit larger than counties in the US. The author defines 22 of these microregiões as metropolitan based on their population size of 1 million and above.

This definition follows that of the United Nations’ World Urbanization Prospects (UNWUP) (Christiaensen et al., 2013).

The author obtains average wages in 2009 from the National Employment Registry (RAIS). Other information on local characteristics is obtained from Ipeadata. This is an online data pool provided by Instituto de Pesquisa Economica Aplicada (Ipea), a Brazilian public research institute that collects data from several ministries and other public sources. It contains information at the microregião level on quality of education and health provision, and homicide rates as a measure of crime.

Quality of education and health is measured using an index, which is constructed and annually updated by the Industrial Federation of the federal state of Rio de Janeiro (FIRJAN). The index for education provision combines information about the subscription rate of pre-school children, dropout rate, rate of teachers with higher education, average daily teaching hours, and the results of a national education development score. The index for health provision quality comprises the number of pre-natal consultations, deaths due to badly defined causes, and child deaths due to evitable causes. These are measures commonly used in health economics to capture the quality of health care provision (Hanefeld et al., 2017; Donabedian, 1988). Access to health care provision is measured by the number of health care facilities per 100,000 inhabitants.

Health care facilities included are general hospitals, day hospitals, polyclinics, health point, general emergency, and pharmacies and basic health centers.

All variables used and their source is specified in Table A10 in Appendix.

Figure 1 shows a map of Brazil. The map shows 5 greater regions and the 22 metropolitan cities of Brazil, which are the focus of this analysis.

Some of the metropolitan cities comprise more than one microregião so that the number of cities is slightly lower than the number of metropolitan microregiões.

Labor migration within Brazil is historically very common and is mainly attributed to socioeconomic differences between regions, underdeveloped rural areas, and several large urban centers (Yap, 1976). Recently, migration patterns in Brazil have been changing. Of all Brazilian internal migrants in the year before the Census of 2010, 47% moved between non-metropolitan areas (Table 1). The second-largest movement is into and out of metropolitan cities as well as from and to non-metropolitan microregiões comprising around 20% each of all recent migrants, a substantial share of migration in the country. The remaining 12% of migrants move between the metropolises.

Figure 1

The rate of people leaving big cities has been increasing in the past decade as illustrated in Figure 2. The graph plots the out-migration rate from cities with over 1 million inhabitants in Brazil from 2004 to 2009.

Figure 2

Table 2 compares metropolitan and non-metropolitan microregiões in terms of socioeconomic characteristics. In the second and fourth columns in Table 2, the author also includes the coefficient of variation for the metropolitan and non-metropolitan microregiões to illustrate how diverse non-metropolitan areas are.

The skill level of workers is based on the occupation classification by the International Labor Organization (ILO). Industries include extractive industry, processing industry, electricity/gas, sanitation/sewage, and construction. Services include commerce, transport, housing/food, information/communication, financial services, real estate, professional consulting, science and technology, administrative services, arts/culture/sports, domestic services, and other services. Public services include public administration, security, education, health, and social services, international organizations/foreign institutions. Six of the microregiões had missing values for homicide rates. The author replaced them with 0 due to the way homicides are reported.

Metropolitan cities have on average around 10 times more inhabitants than non-metropolitan microregiões. In terms of prices, metropolitan residents face room rents that are >50% higher than in non-metropolitan areas. At the same time, they earn similarly higher wages. As expected, high-skilled occupations are concentrated in the metropolitan areas, and labor markets are much more formalized in these big cities. The employment share of various sectors is higher for services with 53% in the metropolitan areas and 35% in non-metropolitan microregiões. Yet, agriculture in the non-metropolitan areas employs around 30% compared to only 9% in metropolitan cities. While GDP is higher in metropolitan regions, it is growing faster in the non-metropolitan regions. In terms of living standards, almost 60% of metropolitan residents live with adequate sewage, water, and electricity provision compared to only 36% outside of these cities.

The definition of which type of sewage, water, and electricity provision is adequate comes from the report on sub-normal agglomerations in Brazil (IBGE, 2010).

This gap is similar for the indices for the quality of health and education provision in non-metropolitan areas in contrast to higher standards in the big cities. In contrast, crime is concentrated in big cities with a homicide rate of 38 homicides per 100,000 inhabitants compared to around 19 in non-metropolitan areas.

The variation in these characteristics among non-metropolitan microregiões is large. The second and fourth columns in Table 1 show the coefficients of variation for metropolitan and non-metropolitan microregiões, respectively. It is relatively larger for non-metropolitan areas in almost all categories, such as population, public service worker share, and education quality. This motivates the analysis of the metropolitan out-migrants’ destination choice. Labor mobility is expected to respond to this spatial variation of real wages and other socioeconomic characteristics.

Metropolitan out-migrants are unlikely to be representative of the population of metropolitan cities. In Tables 3 and 4, the author compares the characteristics of non-metropolitan residents, metropolitan out-migrants, and metropolitan residents. The comparison of these residents with metropolitan out-migrants allows shedding light on the differences between migrants and residents at the origin and destination.

For simplification, in the descriptive analysis, this comparison does not account for the fact that migrants could all be concentrated in a specific sub-set of non-metropolitan microregiões so that residents in all non-metropolitan microregiões might not represent the exact comparison group of that specific subset.

On average, migrants are slightly younger than residents and relatively more of them are male. Overall, they are much better educated than the average resident at the non-metropolitan destination and their education is very similar to that of metropolitan residents. A slightly larger share of migrants has a tertiary education compared with metropolitan residents. From this comparison, it does not seem that low- or high-educated workers are more likely to leave metropolitan cities than the respective other groups.

Table 4 documents the labor market characteristics of out-migrants and residents. Around 12% of workers who left the metropolitan cities for non-metropolitan destinations are unemployed in contrast to an unemployment rate of only 5% among non-metropolitan residents. This indicates that metropolitan out-migrants are a heterogeneous group and some lose out at their new destination. However, the high unemployment rate might just capture a period of adjustment for very recent migrants who have not found a job yet at their new destination.

In terms of wages, migrants earn on average more than their non-migrant counterparts at non-metropolitan destinations, and they earn almost as much as residents in metropolitan areas. This might only reflect differences in the productivity of locations where migrants live as well as different observable and unobservable characteristics of migrants in contrast to residents. The regression analysis in this article aims to disentangle these factors.

More than 60% of non-migrants in metropolitan cities are employed either in the public formal or in the private formal sector, whereas only around 46% of non-migrants in non-metropolitan towns work in the formal sector. Migrants appear to find relatively more formal employment at the destinations outside of the big cities compared with the residents there.

Most migrants work in service sectors. Only a few work in agriculture at the non-metropolitan destinations even though it is the second-largest sector after services in these locations. In metropolitan cities, services are the main sector of employment. This suggests that most migrants are unlikely to change their sector of activity when they move out of metropolitan areas.

These observations highlight three findings: first, there is a significant difference in economic and social characteristics between metropolitan and non-metropolitan microregiões that are likely to determine migration between these. Prices of housing, non-tradable living costs, are much higher in the metropolitan cities and the non-metropolitan areas are catching up economically. Second, there is a large spatial variation in the characteristics of non-metropolitan microregiões across the country. Hence, metropolitan out-migrants are unlikely to be indifferent between destinations in their choice of where to move. Third, migrants are not a random draw of the population and they are a heterogeneous group. It is important to account for migrants’ characteristics and underlying selection in the econometric analysis of this article.

The empirical analysis focuses on the estimation of the effect of various local attributes on the destination choice of migrants. The analysis is based on a multiple-choice setting presented by McFadden (1974). The empirical application is restricted to those who migrated.

The model allows to include also residents in the analysis and assume that they chose not to move. In the empirical application, this would result in a sample so large that it is not feasible to handle. The decision to migrate itself yields a selection bias distinct from the location choice. Costs of moving are heterogeneous for workers so that some of those who did not move might have done so due to high costs or risk which gives rise to a selection bias in the decision to migrate. By excluding the choice to stay at one's origin, and estimate the destination choice model with migrants only, this specific selection bias does not arise.

Migrants are assumed to choose their location to maximize their utility. Motivated by a random utility model, a migrant i residing in the current location chooses among all possible destinations. Let z_ij be a vector of destination attributes that vary across alternatives and can vary by migrant i and let c_j be the cost of moving to destination j from the current location o. Therefore, the author defines c_j = 0 if j = o. The utility of moving to destination j is assumed to have the following form: (1) Uij=β′zij−cj+ɛij {U_{ij}} = \beta '{z_{ij}} - {c_j} + { \epsilon _{ij}}

The utility of migrant i from moving to destination j depends on the destination attributes, moving costs, and an idiosyncratic random component _ij. The observed choice by the migrant is assumed to reflect the maximum utility of all J utilities. The probability that migrant i chooses destination j is therefore (2) Prob(Uij>Uik)for all otherk≠j \matrix{ {\Pr {\rm{ob}}\left( {{U_{ij}} > {U_{ik}}} \right)} & {{\rm{for}}\,{\rm{all}}\,{\rm{other}}} & {k \ne j} \cr }

It is assumed that the error terms are distributed independently and identically with Weibull distribution as in McFadden (1974): (3) F(ɛij)=exp(−e−ɛij) F\left( {{ \epsilon _{ij}}} \right) = {exp}\left( { - {e^{ - \epsilon_{ij}}}} \right)

The probability of moving to destination j is now modeled conditional on migration (i.e., leaving location o). If Y_i represents a random variable indicating the destination choice of migrant i, the probability that this choice is destination j conditional on migration can then be expressed as: (4) Prob(Yi=j|Yi≠o)=eβ′zij−cj[∑j=1Jeβ′zij−cj]−eβ′zio−co \Pr {\rm{ob}}\left( {{Y_i} = j|{Y_i} \ne o} \right) = {{{e^{\beta '{z_{ij}} - {c_j}}}} \over {\left[ {\sum\nolimits_{j = 1}^J {{e^{\beta '{z_{ij}} - {c_j}}}} } \right] - {e^{\beta '{z_{io}} - {c_o}}}}}

This is equal to: (5) Prob(Yi=j|Yi≠o)=eβ′zij−cj∑j≠oJeβ′zij−cj \Pr {\rm{ob}}\left( {{Y_i} = j|{Y_i} \ne o} \right) = {{{e^{\beta '{z_{ij}} - {c_j}}}} \over { {\sum\nolimits_{j \ne o}^J {{e^{\beta '{z_{ij}} - {c_j}}}} }}}

Eq. (5) represents a conditional logit model. The vector z_ij may comprise individual-specific but destination-invariant characteristics w_i and the attributes of each destination x_ij can vary across destinations and also vary across individuals: (6) zij=g(wi,xij) {z_{ij}} = g\left( {{w_i},{x_{ij}}} \right)

In this analysis, the interest lies in the attributes of destinations and not on migrants’ characteristics. Greene (2000) shows how w_i drops out of the probability in Eq. (5) so that this model automatically controls for any individual-specific factors in the destination choice.

In some applications, e.g., Fafchamps and Shilpi (2013), this is called individual fixed effect alternative specific conditional logit. It is, however, not to be confused with the inclusion of fixed effects as in a panel model.

This model can be estimated by the method of maximum likelihood. Let d_ij = 1 if Y_i = j and 0 otherwise. Then the log-Likelihood function is: (8) log L=∑i=1N∑j=1JdijlogProb(Yi=j|Yi≠o) {log}\,L = \sum\limits_{i = 1}^N {\sum\limits_{j = 1}^J {{d_{ij}}\log \Pr {\rm{ob}}\left( {{Y_i} = j|{Y_i} \ne o} \right)} }

For the main estimation, there are N observations and regressors for each of the 14,509 metropolitan out-migrants.

The sample of metropolitan out-migrants is slightly reduced as the author only includes those in the analysis who are matched so that the results are comparable. Those dropped were not matched.

Based on the human capital migration model (Sjaastad, 1962), the destination attributes of interest in this analysis are wages and prices. Observed average wages and prices at microregiões level in the 2010 data are equilibrium results from local economic dynamics, among other factors people's location choices. The simultaneous inclusion of these variables in the model, therefore, raises endogeneity concerns.

To address this issue, the author computes expected wages instead of applying the observed wages realized by migrants at their destination. Average wages in a location need not reflect the wages a migrant can expect to earn. The author, therefore, predicts expected wages for migrants based on their characteristics and the coefficients from a wage estimation of residents at each location.

First, the author estimates a wage regression separately for all 6.9 million resident observations in each microregião. The wage regression takes the following form: (9) Wij=αj(aij−a¯j)+βj(Eij−E¯j)+γj(Sij−S¯j)+χHij+δj+νij W_i^j = {\alpha _j}\left( {a_i^j - {{\bar a}_j}} \right) + {\beta _j}\left( {E_i^j - {{\bar E}_j}} \right) + {\gamma _j}\left( {S_i^j - {{\bar S}_j}} \right) + \chi H_i^j + {\delta _j} + \nu _i^j

Log hourly wages of individual i in location j are determined by the individual characteristics aij a_i^j , Eij E_i^j , and Sij S_i^j ; household characteristics Hij H_i^j ; a dummy for the microregião δ_j; and an idiosyncratic error term νij \nu _i^j .

The variable aij a_i^j summarizes age and age-squared, Eij E_i^j the education level, and Sij S_i^j measures gender and race (white vs. non-white). Each of these variables is demeaned at the level of the microregião, so that the coefficients α_j, β_j, and γ_j capture the return to these characteristics specific to each location. Additionally, this implies that δ_j measures the unconditional microregião-specific average wages. Household characteristics, Hij H_i^j , include the proportion of children and a dummy for whether the partner works, as these might vary by region, e.g., in more rural areas, households tend to be larger and female labor force participation lower so that wages would be overestimated in these areas if this was not controlled for. The author uses the survey weights in these regressions to make the estimates representative of the population.

For each microregião, the coefficients αj^ \widehat {{\alpha _j}} , βj^ \widehat {{\beta _j}} , γj^ \widehat {{\gamma _j}} from this regression are then used to predict a measure of expected wages for each migrant. This predicted wage reflects what each migrant can expect to earn in each microregião conditional on her characteristics a_i, E_i and S_i, and the unconditional local wage level δj^ \widehat {{\delta _j}} : (10) E[Wij|^Xi]=δj^+αj^(ai−a¯j)+βj^(Ei−E¯j)+γj^(Si−S¯j) E[\widehat {W_i^j|}{X_i}] = \widehat {{\delta _j}} + \widehat {{\alpha _j}}\left( {{a_i} - {{\bar a}_j}} \right) + \widehat {{\beta _j}}\left( {{E_i} - {{\bar E}_j}} \right) + \widehat {{\gamma _j}}\left( {{S_i} - {{\bar S}_j}} \right)

The coefficients of the wage predictions corresponding to Eq. (10) (see Table A8 in Appendix) confirm the relationships documented in the literature: Age has a positive, but diminishing effect on expected wages, women earn less than men, white Brazilians more than non-whites, and wages increase with the level of education.

This approach assumes that migrants are a random draw from the resident population so that the returns to individual characteristics should be the same for migrants and residents. In the descriptive statistics, the author showed that migrants differ from the resident population in several observable characteristics. This implies that the expected wage measures used in the analysis so far could be biased by unobservable characteristics. The author thus estimates another measure for expected wages that should reduce the selection bias. The author predicts expected wages in non-metropolitan destinations from a sample of previous migrants from the same origin as the migrant. These migrants have moved more than a year ago to the destinations. They are assumed to be more comparable to migrants than residents in terms of unobservable characteristics specific to migrants, for example, risk-taking preferences.

For the metropolitan origins, the author predicts expected wages based on a matched sample of residents at the origin. The author applies CEM to use only those residents that look most similar to the migrants. CEM bounds the degree of model dependence in the main analysis and the data are automatically restricted to common support. The large dataset of the Census at hand is very suitable for this matching method, without facing the trade-off of conventional matching methods between bias and variance. Migrants and non-migrants are matched on sex, age (for migrants the year at migration), race, education level, marital status, sector of activity, and city of origin. Balance statistics are presented in Tables A1–A3 in Appendix.

With this approach, expected wages are allowed to vary across individuals and locations in the individual choice model. In this setting, the endogeneity of wages would only arise if we believed that the destination choice of an individual systematically determines the expected wage level at all possible destinations also for other individuals. This is not the case as the model controls for individual-specific effects and employs individual-specific expected wages.

This approach does not apply to prices. Living costs vary by destination only and thus could be driven by location choice. The use of an instrumental variable is required that would predict differences in living costs across locations, but not destination choices. The alternative specific conditional logit model does not allow for the application of an instrumental variable approach as it does not have normally distributed error terms. The author thus proposes a reduced form regression using the mean of average monthly municipality wages in the contingent microregiões surrounding a given microregião as an instrument for local living costs. This approach follows Morten and Oliveira (2018) who used trade-cost weighted Bartik shocks in neighboring locations as an instrument for local prices. The rationale is simply that if productivity in neighboring locations increases, prices for traded goods increase, too. While the authors employed the dynamic approach of Bartik-shocks in wage growth, the author uses simply the average of monthly mean wages of the year before migration (2009). Figure A1 in Appendix shows a strong and positive correlation between these indicators (coeff.: 1.59; SE = 0.06) confirming that this is a valid approach in this setting.

Migration costs are measured by the Euclidean distance between origin and destination in kilometers.

The author also used road travel time based on OpenStreetMap. This measure is positively correlated with Euclidean distance (0.87, significant at 1% level) and results do not change in their sign nor significance. Results are available upon request.

In the specific application of this paper, metropolitan out-migrants choose their destination not only based on destination attributes but also these attributes relative to the attributes of migrants’ metropolitan origins. For each location attribute, the author thus computes the difference between the destination and origin, e.g., the expected wages in destination j minus expected wages in origin o. Table 5 gives an overview of the differences between destinations and origins of all variables of interest and how these differences vary between the destinations that migrants chose to those that they did not choose.

Table 5 already indicates some patterns of destination choice. Similar to what we observed earlier in the descriptive part, nominal wages are on average always lower in non-metropolitan areas. Migrants tend to choose locations, where this gap is relatively smaller, −0.53 compared to −0.61. This also applies for selection-corrected expected wages (Matched wages). Similarly, housing prices, measured in rent per room, are on average higher in the big cities. Migrants settle in locations where this price gap is not as big as in other possible destinations. This could indicate a trade-off between higher wages and lower prices at destinations. The proposed instrument for prices, wages in neighboring microregiões, reflects the pattern of actual living costs (rent per room).

Chosen destinations are on average larger in population size than their alternatives. Based on their average size, they are however not the smallest locations, but still medium-sized microregiões. In terms of amenities, chosen destinations have relatively lower levels of homicide rates in contrast to alternative destinations and metropolitan origins. There are also relatively more health facilities in chosen destinations and health care provision is of better quality. Education service provision is also on average better in the chosen microregião than in alternative destinations. These averages are all statistically different between chosen and alternative destinations.

Another clear pattern is revealed with regards to the geographic and social distance of chosen destinations. About 45% of chosen destinations are in a different state than the origin contrasting 92% of the destination alternatives. This is also reflected in a lower average distance of chosen destinations to the migrants’ origin. Many of these factors are highly correlated with each other which makes it necessary to apply multivariate analysis to disentangle their influence on the metropolitan out-migrants’ destination choice.

Many of these factors are highly correlated with each other which makes it necessary to apply multivariate analysis to disentangle their influence on the metropolitan out-migrants’ destination choice.

Table 6 reports the results of the alternative specific conditional logit model that estimates the probability for destination choice conditional on migration as specified in Eq. (8). The interpretation of coefficients in the alternative specific conditional logit model is not straightforward. It is not possible to compare the coefficient size directly, but only in relative terms which the author will do later in Section 4.3.

The specifications in columns 1 and 2 of Table 6 are that of only wage and price differences, migration costs in form of distance and controlling for population size. In column 1 of Table 6, the author presents the naive estimators of expected wages and local living costs not accounting for selection bias in wages and endogeneity of prices. Both coefficients are insignificant. Only distance and whether a destination is in the same state significantly predict migrants’ destination choice. Migrants prefer destinations that are closer to their origin and within their state of birth. These results remain unchanged across specifications.

Column 2 of Table 6 introduces the selection-corrected expected wages and the instrument for prices, wages in neighboring locations. The coefficients remain insignificant, but the sign for wages switches. This indicates that observable and unobserved characteristics of migrants led to an upward bias in expected wage differences. In column 3 of Table 6, the author then further includes other local amenities. In terms of public service provision, it appears that migrants show preferences for locations with relatively more health facilities, but they accept a relatively lower quality of health service provision. They also significantly prefer locations with relatively higher levels of education provision quality.

In summary, the results of the multivariate analysis did not confirm that wages and prices matter significantly for the destination choice of metropolitan out-migrants. However, physical and social moving costs appear to matter. Finally, the author documents that amenities in terms of public service provision are significantly correlated with the destination choice.

One reason why wages and living costs do not enter significantly could be that they matter only for specific groups of migrants. The author, therefore, has run the analysis for sub-samples of migrants with different levels of education. Results are presented in Table 7. In each column, the author used selection-corrected expected wages and the instrument for prices. Column 1 shows the results for migrants with no or only primary education, column 2 for those with lower secondary education, column 3 for upper secondary, and column 4 for those with higher education.

Three results can be highlighted in Table 7. First, expected wages do not matter significantly for the destination choice of any migrant group, but living costs do for those migrants with lower skill levels. Low-educated migrants (columns 1 and 2) significantly prefer destinations with relatively lower living costs than alternative locations and their metropolitan origins. If the price difference between a destination option and the origin increases, implying that prices are closer to those of the expensive origins, then migrants are less likely to choose such a destination. This does not apply to migrants with secondary schooling or higher education. Second, lower educated migrants give significant importance to the availability of health facilities (access and quantity) but not to the quality of health service provision or education. More educated migrants show a significant preference for education quality (column 3) but are willing to accept lower health provision quality. Third, migration costs in terms of geographical and social distance enter the destination choice equally for all education groups.

Marginal effects of the alternative-specific logit model can be computed for each possible location choice, but this is computationally burdensome and ineffective in presenting the results. To illustrate and compare the effect sizes across variables and sub-samples, we can look at one destination alternative, e.g., the one with a price difference very close to the average price difference to metropolitan origins. The author takes the significant regressors from the full specification with selectivity robust wage measures and the price instrument as in Table 7 by education level and computes their elasticities for this example location. Let the probability of choosing destination j be P_j, then the elasticity of P_j with respect to an attribute x_ij evaluated at the mean x¯ij {\bar x_{ij}} can be written as: (11) δlog(Pj)δlogxij=x¯ij(1−Pj)βx {{\delta \log \left( {{P_j}} \right)} \over {\delta \log {x_{ij}}}}={\bar x_{ij}}\left( {1 - {P_j}} \right){\beta _x} where β_x is the coefficient of the destination attribute from the conditional logit estimation (Greene, 2000). The elasticities for significant covariates for each education group are presented in Table 8:

The elasticities reveal that by far the largest effect on migrants’ destination choice is that of distance and a migrant leaving her/his state of birth. This captures the physical and social costs of moving and confirms that migration costs in Brazil are still high and a significant factor in labor mobility (Morten and Oliveira, 2018). Table 7 further showed that lower educated migrants value the presence of health facilities, but the elasticity of this amenity is relatively small compared to those of distance. Of similar size is the elasticity of living costs, which was also only significant for the sample of less educated migrants.

A 10% increase in the price difference between the metropolitan origin of a migrant and this specific location makes it on average 2% more likely to be chosen as a destination by migrants with no or only primary education. The range of living costs differences is around 148% of its mean so that this is not a negligible effect. Effect sizes for education and health quality are very small.

The wage return to migration is defined as the difference between income at destination, y_d, and at the origin, y_o: (12) r=yd−yo r = {y_d} - {y_o}

In the empirical application, income is proxied by the log of hourly wages, W.

The author chooses to look at hourly wages earned in the main job instead of total income as hourly wages in the main job should mostly reflect the return to individual characteristics based on location whereas total income also depends on household composition and other factors.

Thus, the wage difference due to migration can be identified for migrants as average treatment effect on the treated (ATT): (15) ATT−E(Y1−Y0|M−1)−E(Y1|M−1)−E(Y0|M−1) ATT - E\left( {{Y^1} - {Y^0}|M - 1} \right) - E\left( {{Y^1}|M - 1} \right) - E\left( {{Y^0}|M - 1} \right)

The first term on the right-hand side is observable in the data at hand, the wages of migrants at their destination. The second term represents the counterfactual outcome, what migrants would have earned had they not migrated which cannot be observed. The author only observes wages for migrants at their destination and non-migrants at the origin. If wages were estimated using OLS and then compared, the wage differences would be biased due to selection into migration arising from individual-specific unobservable characteristics.

It is necessary to account for this potential bias in the empirical estimation of migrants’ counterfactual wages. This is especially important in a context where spatial wage differences have been found to reflect variation in labor force composition and industry concentration. In Brazil, the labor force is distributed unequally across space, concentrating better-educated workers in metropolitan areas and economically stronger regions. Thus, the returns to education based on observable characteristics explain around half of the spatial wage differences (Almeida dos Reis and Paes de Barros, 1991; Foguel et al., 2015; Ferreira et al., 2006). Furthermore, Brazilian workers have shown little mobility across industries so that it seems reasonable to focus on self-selection by location and not by sector (Menezes-Filho and Muendler, 2011; Hering and Paillacar, 2015).

The author, therefore, uses the predicted wages from the matched sample of residents in metropolitan origins of migrants as described in Section 4.1. The difference between the actual observed wages at the destination and the predicted counterfactual wages at the origin is the return to migration out of metropolitan cities. Real wages are computed using the local average rent per room as a denominator of actual and predicted nominal wages.

This section presents the results of the counterfactual analysis. Table 9 presents the average return to migration as the difference between average actual and counterfactual wages for migrants moving out of metropolitan areas. These migrants earn significantly lower wages at their non-metropolitan destinations. Once the author accounts for the local living costs by using real wages, the difference becomes positive. This indicates that metropolitan out-migrants lose in nominal terms, but gain in real wages due to lower living costs in non-metropolitan destinations.

The results without matching for nominal wage differences are around 0.1 log points larger than when matching is applied (see Table A4 in Appendix). This indicates an overestimation of wages at origin when not accounting for selection and it suggests that out-migrants are negatively selected from the metropolitan working population.

Table 10 documents heterogeneity in wage returns along with the education level of migrants. The author defines high-educated workers as those who completed high school or any higher level of education. Low-educated workers are those who did not complete high school or any lower level of education. Results show that real wages are higher at the destination than origin for both groups. For highly educated individuals leaving the big cities, the real wage gains are larger because their loss in nominal wages is relatively small. In contrast, low-educated workers see a large loss in nominal terms and a relatively smaller gain in real wages. For both groups, the nominal and real wage differences are statistically significant.

So far, the counterfactual wage comparison has focused on the average wage return. However, the distributional graphs of actual and counterfactual wages document the return to metropolitan out-migration along with the wage distribution. Figures 3 and 4 show the wage distributions of workers who have moved out of metropolitan areas. They compare the observed wages of migrants at their destination and the predicted counterfactual wages at the origin. As suggested from the results in Table 9, for nominal wages, the distribution of observed wages lies left of the predicted earnings in metropolitan origins. Wages are generally higher in origins and out-migration implies a loss in nominal terms.

For real wages (see Figure 4), the distribution of observed wages lies now a bit to the right of counterfactual wages in metropolitan origins reflecting the positive return in real terms to leaving expensive cities.

The distributions are tested to be significantly different with a Kolmogorov-Smirnov test for equality of distributions. Both, the nominal and real wage distributions are significantly different.

In the analysis of the destination choice of migrants, we learned that low-educated metropolitan out-migrants choose between destinations where their gain in lower living costs is the largest. This can explain why some individuals do not experience a positive return to migrating out of metropolitan areas across the income distribution. Some might fail to successfully evaluate their destination alternatives, some might lack the information about wages and prices at all destinations, and others might just not be successful in acquiring the wage employment they had expected or they migrate for other reasons such as family. In this way, some metropolitan out-migrants lose out, while on average they gain in real wage returns.

Figure 3

Figure 4

To understand better how migrants evaluate possible destinations, the author contrasts the previous results with those of people moving into metropolitan cities applying the same methodology. Table A9 in Appendix confirms that in nominal terms, metropolitan cities present a huge income gain, but especially low-skilled migrants experience significant and large losses in real terms. This points to the importance of experience and information. Metropolitan residents face high prices and this variable thus dominates their evaluation of possible destinations. Workers in non-metropolitan areas might have never considered living costs as a significant factor so that they fail to account for them in their destination choice.

In the Census data, households are asked to state the monthly rent they pay if they live in a rented apartment or house and the number of rooms of the unit. The author used these data to aggregate the average room rent at the microregião level. This measure ignores the possibility that the price differences might just reflect differences in housing quality. Similar to Li and Gibson (2014), the author constructs a hedonic housing index that measures the differences in housing costs based on location-specific amenities rather than housing-specific characteristics.

Households are asked to provide information on the quality of walls, floors, and the presence of toilets. Additional questions inform about the quality of sewage, wastewater, and electricity access. With these variables, the author can estimate a hedonic housing price for each location. The author regresses the rent per room on these characteristics weighted by the household survey weight and the author includes a dummy for each microregião. These regression results are presented in Table A7 in Appendix. The coefficients of the microregião dummies capture any location-specific amenities that contribute to spatial price differences. The author extracts these estimates to construct a location-specific hedonic living cost measure. This variable is independent of differences in housing quality.

The author, therefore, also computes the real wages using the hedonic price as a denominator. Results are presented in Table 11. The wage difference remains positive and statistically significant, but it is smaller by around a third than in the initial results.

Further concerns regarding the measure of living costs could arise from the fact, that low and high-educated workers probably face different housing markets with different average prices. The author, therefore, conducts the counterfactual analysis using education-group specific living costs as a denominator when computing real wages (see Tables A5 and A6 in Appendix). The author also uses the median prices in the microregião instead of the mean. The signs and significance of wage differences remain the same as in the initial results, but their size changes. For high-educated workers, the estimates yield larger positive wage differences when the author uses education-group specific living costs. The differences are 0.37 and 0.3 log points for average and median rents specific to high-educated workers. For low-educated workers, the education-specific wage differences are smaller than unadjusted ones. Applying the median prices faced by low-educated workers yields the smallest difference of 0.07 log points compared to 0.13 in the initial results. Rents are on average lower for low-educated workers so that their gain from leaving metropolitan cities becomes smaller than when the author did not account for the education-group specific rents. The opposite applies to high-educated workers.