Rainfall, Internal Migration and Local Labor Markets in Brazil

Rainfall, Internal Migration and Local Labor Markets in Brazil Raphael Corbi * and Tiago Ferraz Preliminar version: July 21, 2017 Abstract We investigate the labor market impacts of internal migration in Brazil by following a two-step approach. First, we explore the variation of out-migration flows from the semiarid driven by deviations from historical rainfall averages. Second, we distribute the predicted out-migration flow based on the pre-existing support network in each destination based on the migrant s region of origin. Our results indicate that increasing in-migration rate by 1p.p. (one standard deviation) reduces overall native wages in 0.6% mostly in the formal sector and has no effect on overall employment. These results mask considerable heterogeneity. Migration hampers the prospects of low-skilled natives while improving employability of the high-skilled. On the other hand, the negative effect on wages is concentrated on high-skilled formal workers. This can be reconciled with the view that low-skilled migrants are substitutes for low-skilled natives and complements for high-skilled natives. Moreover, it is consistent with a model in which (more regulated) formal labor markets have higher wage rigidity compared to the more competitive informal sector. Keywords: Migration, labor supply, wage effects, rainfall JEL Codes: J21, J22, J61, R23 * University of São Paulo, Department of Economics, email: rcorbi@usp.br University of São Paulo, Department of Economics, email: tiago.ferraz@usp.br 1

1 Introduction Recent discussions regarding border control and climate change has brought the effect of immigration on labor markets to the front stage of the public policy debate. The Intergovernmental Panel on Climate Change (IPCC) noted in 1990 that the greatest single impact of climate change could be on human migration. Existing support networks in the destination areas also exert an important effect on the decision to migrate. A statement often repeated is that migration hurts natives due to an increase in competition, stealing jobs and lowering wages. In a very simple framework, an in-migration shock represents an outward shift in the labor supply curve leading to lower wages. Alternatively, it can also increase demand for local production and local labor, hence rendering the quantification of the resulting net effect an empirical question. There is a vast and growing literature on this matter 1, focusing particularly in international migration from developing to developed countries, with little consensus about the true impact of immigration on labor markets. Much less effort has been made to understand migration within countries, especially in developing ones. This is an important distinction because moving internally is probably much less costly than migrating internationally - it is no accident that United Nations Population Division estimated that about 750 million people moved within countries around the world in 2005 (Lucas, 2016) while there was approximately 214 million international migrants in 2013 (Kleemans and Magruder, 2017); and because labor markets in developing countries are very different than those in the rich world. In this paper we focus on climate-driven internal migration in Brazil, during the years from 1996 to 2010. In this period about 36 million people or 25% of the population moved within the country. Such a large movement is very likely to affect labor markets in the destination. In particular, we use internal migration from the Brazilian Northeast s semiarid region as historically this is where most of climate-driven out-migration takes place. Indeed over 300, 000 people out of a population total of around 22 million in the semiarid leave their hometown every year. Typically they chose to migrate to more economically dynamic areas of the country in the Southeast and the Mid-West. In comparison to the native population in the destination, these migrants tend to be less educated, less white and poorer. Estimating the effects of internal migration can be a challenge because the flow of migrants to a destination is potentially correlated to unobserved characteristics of local labor markets. For instance, individuals deciding to move could choose a specific destination because the local labor market offers more job opportunities or higher wages, in which case OLS estimates would be upward biased. 2 In order to deal with this endogeneity problem, we follow a two-step approach. First, we 1 Dustmann et al. (2016) provides a good literature review and discuss the differences among those studies. 2 This is known in the literature as moving to opportunity bias. 2

explore the variation of out-migration flows from the semiarid driven by deviations from historical rainfall averages. Second, we distribute the predicted out-migration flow based on the pre-existing support network in each destination based on the migrant s region of origin. By adding the in-migration flow from each area of origin in each destination, we are able to calculate the predicted flow of migrants into each destination driven by exogenous shocks to rainfall in the origin and the pre-existing support network. We use this exogenous shock to estimate the causal effects of in-migration on wages and employment for natives in destination municipalities. Our main data source for this final step is a sample of about 1.1 million individuals, in 1,354 destination municipalities, from two rounds of Census microdata - 2000 and 2010. Our results indicate that one percentage point increase in the predicted in-migration rate has no significant impact on native employment in the destination, but decreases wages in 0.6%. These overall estimates hide significant heterogeneity in the labor market structure and worker skill groups. While there is no significant effects on overall employment, our estimates imply that in a destination municipality receiving a one percentage point increase in predicted in-migration the probability of a low-skilled native hold a informal job reduces by 0.26 percentage points, while for a high-skilled worker this probability will increase by 0.43 percentage points. Also, we find negative effects on wages for both groups but stronger for the unskilled. These findings are consistent with a theoretical model that allows for different labor supply elasticities among groups 3. Most of the estimated effect comes from wages in the formal sector and employment in the informal sector. Since we are using Census data to identify migrants there is not much we can say about them before moving. Most of the characteristics we know relate to their lives after arriving at the places of destination. Therefore, we are not able to deal with selection issues and our estimates are local average treatment effects (LATE). In other words, we are estimating labor market impacts of climate-driven migration. So far the literature on the effects of immigration on labor markets outcomes had found very mixed results. For instance, the impact of immigration on wages varies from strong negative effects as in Borjas (2015), who finds an elasticity of wages with respect to the number of workers between -0.5 and -1.5, to positive impacts as in Foged and Peri (2016) who estimate an elasticity of 1.8. Dustmann et al. (2016) argue that the reason for those contradictory findings is that different specifications estimate different structural parameters of the same canonical model. Our paper relates to a branch of the literature that focus on the impacts of migration on labor markets (Borjas, 2003; Altonji and Card, 1991; Card, 2001) and more specifically is focused on the effects of internal migration as in Boustan et al. (2010) that study how internal migrant flows affect labor markets in U.S. during the Great Depression. More closely, our work is related to Kleemans and Magruder (2017) that investigate the labor market impact of internal migration in Indonesia and find that an inflow of migrants decreases employment in the formal sector and wages in the informal sector especially among low-skilled natives. Our work differ from theirs because in our case Semiarid s migrants are mostly low-skilled as opposed to the high-skilled 3 As the one shown in Dustmann et al. (2017) 3

migration inflow they analyze in the Indonesian context. 2 Background and Data In this section, we first describe the economic background and climate conditions at the Brazilian semiarid in an effort to contextualize our analysis. We then discuss the main sources of data regarding labor market outcomes, migration flows and climate, and present some descriptive statistics. 2.1 Brazilian Semiarid According to the official definition provided by the Ministry of National Integration, the Brazilian semiarid encompasses 1,133 municipalities distributed by 9 states, covering an area of around 976,000km 2 (roughly 11 percent of the country s territory). 4 A municipality officially belongs to the semiarid region if attends at least one of three criteria: (i) yearly average precipitation below 800 mm (in the period 1961-1990); (ii) aridity index up to 0.5 (measured by Thornthwaite Index, which combines humidity and aridity for a given area, in the same period); (iii) has an index of risk of drought above 60% (defined as the share of days under hydric deficit, using the period 1970-1990). The region s climate conditions are characterized by high temperatures and low precipitation. Average historical precipitation is about 740mm, as opposed to around 1,300 mm for the rest of the country, while temperature typically ranges from 24 C to 27 C. The rainy season occurs between November and April, with the highest levels of precipitation after February, when the sowing typically starts. 5. Since Colonial Era, Brazilian Semiarid faced a great number of episodes of severe droughts. Between 1877 and 1879, the event known as The Great Drought took the life of around half million people either by starvation or epidemics (Sousa and Pearson, 2009) and drove away hundreds of thousands more (Greenfield, 1986). Three episodes in early 19 th century 6 had led the imperial authorities to start planning the construction of large dams and groundwater wells (Rebouças, 1997). Despite that, water supply is always at risk because of the region s poor infrastructure. There is a relatively small network of occasional shallow rivers due to low rainfall and soil characteristics. Most of the households use water from dams and rainwater reservoirs. During the rainy season, these reservoirs accumulate water to be used throughout 4 It includes all Northeastern states, except Maranhão, plus the southeastern area of Minas Gerais. 5 Our main measure of rainfall shocks uses these months as baseline, but our results are robust to this particular choice. 6 Those episodes occurred in 1825, 1827 and 1830. 4

the year, but evaporation causes great losses. There are over 70,000 dams and ponds with volumes varying from 10 to 200,000 m 3, but hydrological efficiency is very low - only about 1/5 of the estimated reserves are used (Rebouças, 1997). The Semiarid is one of least developed regions in the country with 80% of the children in households below the poverty line and infant mortality reached 31 per 1000 births in 1996, compared to a national average of 25% and 15 per 1000 births, respectively. 7 Municipalities are typically small (population median is around 20,000) and their economies are based on low productivity subsistence agriculture and cattle raising, both activities highly susceptible to suffer from climate shocks. Average human capital level is relatively low as 60% of the adult population have less than 8 years of schooling, as opposed to a national average of 45%. 2.2 Migration, Labor Market and Climate Data Our main sources of migration data are the 2000 and 2010 Brazilian Census. Using information regarding the number of years in the destination for all migrants, we construct a measure of yearly out-migration flow from each origin municipality in the semiarid and a measure of in-migration to each destination municipality during 1996-2010. 8 In addition, we build a measure of pre-existing networks by associating the share of pre-existing migrant communities from each semiarid origin municipality in each destination municipality based on microdata from the 1991 Census. 9 This is especially relevant for our identification strategy discussed in the next section as a means to resolve the endogeneity problem that arises when migrants choose where to migrate to. Climate data is retrieved from the Climatic Research Unit at University of East Anglia (Harris et al., 2014). The CRU Time Series provides worldwide monthly gridded data of precipitation and temperature, at the 0.5 0.5 level (0.5 is around 56km on the equator). We construct municipality-level monthly precipitation and temperature measures based on grid-level raw data as the weighted average of the municipality grid s four nodes using linear distances to the centroid as weights. 10 Then we define the rainfall shocks as monthly rainfall deviations from the historical mean. More specifically ( 6 ) Rainfall oy = ln r oτy ln( r o ) (1) 7 See Rocha and Soares (2015). 8 We cannot go back further than 1996 because the 2000 Census only asked respondents where they lived 5 years before. The 2010 round allows us to track individuals throughout the entire decade. 9 During our period several new municipalities were split into two in Brazil. In order to avoid potential miscoding regarding migration status or municipality of origin, we use Minimum Comparable Areas (AMC) as unit of analysis to preserve the original municipal boundaries (Reis et al., 2008). Throughout this paper we refer to these AMCs simply as municipalities. 10 This methodology is also used by Rocha and Soares (2015). τ=1 5

where r oτy is the rainfall in municipality of origin o in month τ of year y, and r o is the municipality s historical average precipitation for the same months. We use only the rainy season months, because this is the most important period for growing crops. Historical averages are calculated over the period from 1927 to 2010. Temperature shocks are built in a similar fashion, using average temperatures instead of the sum during the rainy season. Table 1 describes municipality-level data for origin (Panel A) and destination (Panel B) municipalities. Semiarid s areas show lower levels of rainfall and slightly higher temperatures than destination municipalities. On average, 2% of Semiarid s population move to larger cities within the country every year. Our main sample consists of individual-level observations of labor market outcomes at destination municipalities and are also retrieved from the 2000 and 2010 Census. We take a 10 percent sample of each Census, restrict to individuals between age 18 and 65. Individuals born at the destination municipality or who migrated in the previous decade are defined as native. We restrict attention to destination municipalities that receive non-zero in-migration flows in both periods located outside states that (partially) belong to the semiarid as rainfall shocks in nearby regions may be correlated. Our final sample has aproximately 1.1 million individuals from 1,354 municipalities. Table 2 describes the main data. In our sample, 61.3% of individuals are employed, 30.6% have a formal job, 27.3% have a informal job or are self-employed. The average monthly wage 11 is R$ 1,338 and the median R$ 777. Compared to natives in our sample, migrants are younger, less educated, more likely to be employed (especially in the formal sector) despite with lower wages. 12 Their average monthly wage is R$ 830, roughly 60% of the monthly wage for natives. Figure 2 illustrates the distribution of occupations at the destination according to migration status. After leaving their hometowns, semiarid female migrants are more likely to be employed in domestic work and male migrants in construction. 3 Empirical Strategy and Identifying Assumptions In this section, we first describe the empirical framework that allows us to (i) isolate the observed variation in out-migration induced by random shocks in rainfall, and (ii) the in-migration flows into destination municipalities predicted by pre-existing migrant networks. We then discuss and present supportive evidence on the validity of this procedure that is key to isolate the causal effect of in-migration on labor market outcomes of native workers. 11 Wages are measured in R$ 2010. 12 We define individuals as low-skilled (high-skilled) if they have up to (more than) 10 years of schooling. 6

3.1 Empirical Framework. Rainfall-induced Migration. We specify a regression model of labor market outcomes of native individuals as a function of internal migration flows. Specifically we assume that Y idt = α + βm dt + γ X idt + Π d + ε idt (2) where Y idt is a vector of labor outcomes for individual i in census year t and destination municipality d, m dt is the municipality-level within-census cumulated migrant flow, X idt is a vector of individual-level controls, Π d is a destination fixed effect. As discussed before, error term ε idt may include unobserved local labor market characteristics that are correlated with migrant flows which would potentially bias OLS estimates. In particular, migrants could choose a specific destination municipality due to higher wages or job prospects. In order to account for this endogeneity problem we follow a two-step procedure to construct the instrument for cumulated migration in the destination. First we project m oy, the out-migration rate from origin municipality o, onto weather shocks in the previous year 13 m oy = α + β Z oy 1 + φ o + δ y + γt rend sy + ε oy (3) where Z is a vector of rainfall and temperature shocks at the origin municipality o in the previous year, φ o is a municipality fixed effect, δ y is a time fixed effect, T rend sy is a state-specific time trend and ε oy is a random error term. Then we use the predicted rates m oy to calculate the number of migrants expected to leave each municipality by multiplying it by the lagged population level as given by M oy = m oy population oy 1 (4) In the second step we use the pre-existent network of migrants from municipality o to d in order to distribute them throughout the destination areas. For instance, if the municipality of São Paulo concentrated 10% of the migrants from Campina Grande in 1991, we allocate this share of the migrants predicted to leave Campina Grande in 1998 to São Paulo. 14 This allocation procedure is similar to those developed by Munshi (2003) and Kleemans and Magruder (2017). It is important to note there are two different time notations. Predicted out-migration is calculated for each year y while labor market outcomes are available only in census years t. Hence we aggregate the number of migrants entering each destination municipality d in year y into decade t M dt = o y M oy network od dt 1 (5) 13 This approach is widely used in the literature. For instance, see Boustan et al. (2010). 14 For each decade we use the predetermined share of migrants in the previous census to allocate them. We tested another specification fixating the network using only the 1991 census. Our first stage remains robust. 7

Finally we obtain the instrument for in-migration rate m dt by dividing the number of migrants predicted to enter each destination M dt by its population in the previous census and plug it into our baseline specification 2. Hence the specification Y idt = α + β m dt + γ X idt + Π d + ε idt (6) can be thought as a reduced-form relationship that associates labor market outcomes and weather-induced predicted cumulated migrant flow at the destination. 3.2 Identifying Assumptions. A sine qua non requirement implicit in our empirical framework is that both predicted out-migration and in-migration rates, m oy and m dt respectively, are strongly correlated with their observed counterparts. We begin by estimating variations of specification 3 using semiarid municipalitylevel data reported in Table 3. Columns (2)-(4) include additional lags of rainfall shocks. All specifications include municipality and state-year fixed effects in order to account for any time-invariant municipality characteristics that may be correlated to migration and state-specific time trends. Standard errors are clustered at the grid level. It is important to stress that we exploit only the rainy season rainfall data, because this is the most important period for growing crops. 15 Hence Rainfall y 1 captures log-deviation from historical average of rainfall at year y 1, which encompasses the period between November of year y 2 to April of year y 1. As expected, rainfall shocks in the previous year are negatively correlated with migration rates indicating that semiarid inhabitants are induced to leave the region during drought periods. Coefficient estimates are remarkably stable across specifications. Contemporaneous rainfall y also exhibits a negatively effect on outmigration albeit with smaller magnitude as reported in columns (2)-(4). This is consistent with the idea that individuals take some time to react to weather shocks. Interestingly adding more lags in specifications (3)-(4) do not change these estimates and they mostly exhibit insignificant effects. More important to our identification, the coefficient on rainfall y+1 reported in column (5) is insignificant and small in magnitude, while the coefficient for rainfall y 1 remains significant. Our prefered specification is (1) as it yields a F-statistic of joint significance of 14.50. 16 In the second step, we distribute the predicted out-migration shock using the pre-existent network of migrants from origin municipality o to destination d. Figure 1 plots observed and predict out-migration flows from origin municipalities on Panel (a). Most observations float around the 45 o degree line. A similar picture arises as we focus on in-migration flows on Panel (b). All in all we conclude that weather shocks and pre-existing network predict well migration flows. 15 See Rocha and Soares (2015) for a more precise discussion on that. 16 Staiger and Stock (1997) suggests that the F-statistic should be greater than 10 when using one endogenous variable. 8

4 Migration Flows and Labor Market Outcomes Now we turn our attention to the labor markets and try to answer the question: how does internal migration impacts wages and employment prospects of native workers? We begin by reporting estimates in Table 4 of the reduced-form relationship that associates labor market outcomes and predicted in-migration rate as in 6. All specifications include controls for individual covariates and predetermined municipality-level characteristics. 17 In order to account for the fact that our estimates may be driven by other time-varying dimensions, e.g., a change in the wage premium for schooling, we also include interaction terms of state- and schooling-year. Following Boustan et al. (2010), we also weighted all regressions by the inverse of the number of observations in the municipality-year to ensure that every local labor market has the same importance and clustered standard errors at the municipality-year level. 18 Columns (1)-(3) report the effect of predicted migration rate on overall, formal and informal employment, respectively, while columns (4)-(6) report estimates on log wages. Rows (1) reports the overall effect of migration and rows (2) and (3) explore whether such effect differs across skill levels. At a first glance, the effect of migration on overall employment is close to zero. This estimates masks considerable heterogeneity. Low-skilled natives are negatively affected by migration while highskilled workers are positively affected. Moreover, columns (2) and (3) indicate that all the employment effect comes essentially from the informal sector. This result is expected since the migrant inflow is composed mostly by low-skilled workers and this group have a higher probability of holding an informal job, conditional on being employed. Our estimates imply that in a destination municipality receiving a one percentage point increase in predicted in-migration the probability of a low-skilled native hold a informal job reduces by 0.26 percentage points, while for a high-skilled worker this probability will increase by 0.43 percentage points. This is consistent with the view that (more regulated) formal labor markets have higher wage rigidity compared to the more competitive informal sector, so we would expect different patterns to emerge. Indeed Kleemans and Magruder (2017) discuss how the effects of migration are shaped by the dual-sector structure of Indonesian labor markets. The same idea applies to Brazil, a developing country where the more regulated formal sector coexists with a large informal sector. Changes in predicted migration affect wages in a different way. As shown in Column (4), an increase in the predicted inflow of migrants by 1p.p. reduces wages in 0.6%. Moreover, this results are driven by the formal sector as evidenced by columns (5)-(6). As opposed to the employment effects, high-skilled natives suffer a larger impact than low-skilled workers. 17 Individual covariates include indicator variables for male and literate, a categorical variable for race and cubic polynomial for age. Municipal characteristics include share of white population, share of illiterate population, share of migrants, share of population below 10 years old, share of population above 65 years old and log income per capita. 18 This is particular important in our case as there is considerable variation in the size of municipalities and not using weights would imply favoring larger cities in our average effect estimates. 9

All these results can be reconciled with the predictions of a model that allows for a larger response by low-skilled workers at the employment margin when facing a low-skilled migration shock. 19 More intuitively, labor supply could be more inelastic for high-skilled workers because they may have higher opportunity costs of being unemployed and could accept a lower wage to avoid the risk of being out of work. In this framework, a low-skilled migrant inflow leads to a negative employment impact for low-skilled natives while their wages could even increase relative to those of the high-skilled. As we can see in Table 2, migrants from semiarid are relatively less skilled than natives. Also, it seems reasonable to assume that high-skilled labor supply in formal sector should be more inelastic, because a formal job has several indirect benefits. 20 5 Robustness In this section we perform some robustness checks to establish the validity of our findings. First, our identification relies on the assumption that rainfall at origin municipalities affects destination labor markets only through migration flows. This assumption would be violated if, due to a low rainfall, a negative income shock at origin had reduced trade flows with some of the destination areas, for instance. We believe that this concern is somehow addressed by our empirical strategy when we decided to restrict the analysis to destination areas from states outside the semiarid region. But, if we admit this channel to be important we would expect those impacts to be harder on industries producing tradable goods. In order to ensure that this channel is not the relevant one, we run again the reduced-form regression on wages splitting the sample by aggregated categories of industry. As we can see in Table 5, our results are mostly driven by the non-tradable Services category. This is unsurprising as the migrant/native employment ratio is greater in occupations like construction, retail and domestic services. Another assumption we made is that unobserved destination-specific shocks affecting labor markets are orthogonal to in-migration flows. This assumption would be violated if there is any differential time trend in labor markets outcomes across destination municipalities. To address this problem we run the same reduced-form regressions including time dummies interacted with 1991 municipality-level characteristics. 21. Table 6 shows that our findings are robust to these concerns. The point estimates are very similar to our main results and confirm the pattern we described before. 19 Later versions of the paper will include a working version of such model. See for instance Dustmann et al. (2017). 20 For instance, access to pension system or employer-sponsored health insurance. 21 This set includes log-income per capita, share of white population, share of illiterate population, share of migrants, share of population aged between 14 and 65 10

6 Conclusion In this paper we investigate the labor market impacts of climate-driven internal migration in Brazil. We exploit exogenous variation in the number of migrants entering each destination municipality by following a two-step approach. First, we explore the variation of out-migration flows from the semiarid driven by deviations from historical rainfall averages. Second, we distribute the predicted out-migration flow based on the pre-existing support network in each destination based on the migrant s region of origin. By adding the in-migration flow from each area of origin in each destination, we are able to calculate the predicted flow of migrants into each destination driven by exogenous shocks to rainfall in the origin and the pre-existing support network. Our first-stage estimates indicate that a municipality in the semiarid in which monthly rainfall is 10% below its historical average will experience an increase of 3p.p in the out-migration rate. By distributing these predicted migration flows across destinations using the pre-existent support network, we find that increasing in-migration rate by 1p.p. (one standard deviation) reduces overall native wages in 0.6% mostly in the formal sector and has no effect on overall employment. These results mask considerable heterogeneity. Migration hampers the prospects of lowskilled natives while improving employability of the high-skilled. On the other hand, the negative effect on wages is concentrated on high-skilled formal workers. This can be reconciled with the view that low-skilled migrants are substitutes for low-skilled natives and complements for high-skilled natives. Moreover, it is consistent with a model in which (more regulated) formal labor markets have higher wage rigidity compared to the more competitive informal sector as in Kleemans and Magruder (2017). 11

References Altonji, J. G. and Card, D. (1991). The effects of immigration on the labor market outcomes of less-skilled natives. In Immigration, trade, and the labor market, pages 201 234. University of Chicago Press. Borjas, G. J. (2003). The labor demand curve is downward sloping: Reexamining the impact of immigration on the labor market. The quarterly journal of economics, 118(4):1335 1374. Borjas, G. J. (2015). The wage impact of the marielitos: A reappraisal. ILR Review, page 0019793917692945. Boustan, L. P., Fishback, P. V., and Kantor, S. (2010). The effect of internal migration on local labor markets: American cities during the great depression. Journal of Labor Economics, 28(4):719 746. Card, D. (2001). Immigrant inflows, native outflows, and the local labor market impacts of higher immigration. Journal of Labor Economics, 19(1):22 64. Correia, S. (2016). Linear models with high-dimensional fixed effects: An efficient and feasible estimator. Technical report. Working Paper. Dustmann, C., Schönberg, U., and Stuhler, J. (2016). The impact of immigration: Why do studies reach such different results? The Journal of Economic Perspectives, 30(4):31 56. Dustmann, C., Schönberg, U., and Stuhler, J. (2017). Labor supply shocks, native wages, and the adjustment of local employment. The Quarterly Journal of Economics, 132(1):435 483. Foged, M. and Peri, G. (2016). Immigrants effect on native workers: New analysis on longitudinal data. American Economic Journal: Applied Economics, 8(2):1 34. Greenfield, G. M. (1986). Migrant behavior and elite attitudes: Brazil s great drought, 1877-1879. The Americas, 43(1):69 85. Harris, I., Jones, P., Osborn, T., and Lister, D. (2014). Updated high-resolution grids of monthly climatic observations the cru ts3. 10 dataset. International Journal of Climatology, 34(3):623 642. Kleemans, M. and Magruder, J. (2017). Labour market responses to immigration: Evidence from internal migration driven by weather shocks. The Economic Journal. Lucas, R. E. (2016). Internal migration in developing economies: an overview of recent evidence. Geopolitics, History and International Relations, 8(2):159. Munshi, K. (2003). Networks in the modern economy: Mexican migrants in the us labor market. The Quarterly Journal of Economics, 118(2):549 599. 12

Rebouças, A. d. C. (1997). Água na região nordeste: desperdício e escassez. Estudos avançados, 11(29):127 154. Reis, E., Pimentel, M., and Alvarenga, A. I. (2008). Áreas mínimas comparáveis para os períodos intercensitários de 1872 a 2000. Rocha, R. and Soares, R. R. (2015). Water scarcity and birth outcomes in the brazilian semiarid. Journal of Development Economics, 112:72 91. Sousa, A. Q. and Pearson, R. (2009). Drought, smallpox, and emergence of leishmania braziliensis in northeastern brazil. Emerging infectious diseases, 15(6):916. Staiger, D. and Stock, J. H. (1997). Instrumental Variables Regression with Weak Instruments. Econometrica, 65(3):557 586. 13

Figure 1 In- and Out-migration: predicted x observed 14 (a) Out-migration from Semiarid (b) In-migration into Non-Semiarid

Figure 2 Distribution of occupations at destination areas 15

Table 1 Summary statistics: Climate and migration data Panel A: Semiarid Region Variables Mean Std. Deviation Min Max Observations Municipalities Rainfall (level) 437.66 228.65 0.00 1,939.09 27,008 844 Rainfall (shock) 0.01 0.35-1.52 1.15 27,008 844 Temperature (level) 26.01 1.05 22.40 29.27 26,144 844 Temperature (shock) 0.01 0.02-0.06 0.06 26,144 844 Out-migration 448.29 774.68 0.00 17,013.00 27,008 844 Out-migration rate 0.02 0.01 0.00 0.43 24,149 844 Population 23,498.40 43,155.3 935 932,697 26,625 844 16 Panel B: Non-Semiarid Region Variables Mean Std. Deviation Min Max Observations Municipalities Rainfall 616.81 347.81 0.00 3,410.72 87,680 2740 Temperature 24.06 1.79 18.40 28.68 76,480 2740 In-migration 835.36 3,848.35 0.00 550,256 87,789 2740 In-migration rate 0.02 0.02 0.00 0.47 78,465 2740 Population 42,638.73 218,887.20 528 11,200,000 86,529 2740 Notes: Rainfall and temperature shocks are measured as log-deviations from municipality s historical average. Out(in)- migration rate calculated as the ratio between out(in) migration and the population in previous year.

Table 2 Summary statistics: Natives and migrants in destination municipalities All Natives Migrants Means test t-stat Male (%) 0.4897 0.4894 0.5177-6.16 (.4999) (.4999) (.4997) Black (%) 0.0679 0.0678 0.0706-1.19 (.2515) (.2514) (.2561) Age 37.0262 37.1079 29.2647 66.63 (12.8465) (12.8505) (9.7254) Illiterate (%) 0.0274 0.0271 0.0536-17.64 (.1632) (.1624) (.2251) Low Skill (%) 0.6457 0.6443 0.7744-29.61 (.4783) (.4787) (.418) High Skill (%) 0.3543 0.3557 0.2256 29.61 (.4783) (.4787) (.418) Employment rate (%) 0.6133 0.6127 0.6627-11.18 (.487) (.4871) (.4728) Formal job rate (%) 0.3059 0.3047 0.4239-28.19 (.4608) (.4603) (.4942) Informal job rate (%) 0.2732 0.2735 0.2397 8.25 (.4456) (.4458) (.4269) Log wage 6.7134 6.7157 6.5093 21.00 (.8822) (.8845) (.6093) Observations 1.151.376 1.139.388 11.988 Notes: Mean is shown with standard deviations in parentheses. Low skill indicates individuals up to 10 years of schooling. High skill indicates individuals with 11 or more years of schooling. Log wage measured in R$ 2010. 17

Table 3 Out-migration response to weather shocks Dependent variable: out-migration rate (1) (2) (3) (4) (5) Rainfall y+1-0.001 (0.001) Rainfall y -0.001*** -0.001** -0.002** (0.001) (0.001) (0.001) Rainfall y 1-0.003*** -0.003*** -0.003*** -0.003*** -0.001* (0.001) (0.001) (0.001) (0.001) (0.001) Rainfall y 2-0.000-0.001 (0.001) (0.001) Rainfall y 3 0.001* (0.001) F statistic of joint significance 14.50 8.575 5.807 5.332 2.183 Observations 23,381 23,381 20,988 18,582 20,945 Municipalities 817 817 817 817 817 R-squared 0.461 0.461 0.470 0.480 0.413 Notes: Regressions use municipality-level data and standard errors are clustered at grid level. Dependent variable is the out-migration rate. Rainfall is measured as log-deviation from historical average. All specifications include controls for temperature and year, municipality and state-year fixed effects. *** significant at the 1% level. ** significant at the 5% level. * significant at the 10% level 18

Table 4 Reduced Form: Labor market impacts of internal migration, by skill level and sector Employment Log wages Overall Formal Informal Overall Formal Informal (1) (2) (3) (4) (5) (6) In-migration rate -0.005 0.039 0.000-0.583** -0.776*** -0.532 (0.131) (0.129) (0.141) (0.291) (0.284) (0.489) 19 In-migration rate low skill -0.247* 0.046-0.259* -0.125-0.316-0.474 (0.148) (0.140) (0.155) (0.319) (0.308) (0.502) In-migration rate high skill 0.399** 0.029 0.433** -1.202*** -1.275*** -0.662 (0.164) (0.173) (0.172) (0.343) (0.315) (0.638) Observations 1,139,388 1,139,388 1,139,388 722,688 347,144 308,820 Municipalities 1,354 1,354 1,354 1,354 1,354 1,354 R-squared 0.145 0.0536 0.0368 0.183 0.201 0.185 Notes: Standard errors clustered at municipality-year level in parentheses. In columns (1)-(3), dependent variable is a dummy equal to 1 if individual is employed, 0 otherwise. Informal sector indicates unregistered work or self-employment. In columns (4)-(6), dependent variable is log-income from work. All specifications use predicted in-migration rate as instrument, include individual- and municipality-level controls as described in section 4 as well as interaction terms of state- and schooling-year. Also, all regressions are weighted by the inverse of the number of observations in the municipality-year. *** significant at the 1% level. ** significant at the 5% level. * significant at the 10% level

Table 5 Reduced-Form: Labor market impacts of internal migration, by industry Log wages Overall Agriculture Manufacturing Services (1) (2) (3) (4) In-migration rate -0.583** 0.525-0.477-1.330*** (0.291) (0.583) (0.549) (0.435) In-migration rate low skill -0.125 0.549-0.278-1.131** (0.319) (0.602) (0.599) (0.460) In-migration rate high skill -1.202*** 0.441-0.769-1.545*** (0.343) (0.813) (0.586) (0.491) Observations 722,688 100,971 166,944 441,132 Municipalities 1,354 1,354 1,354 1,354 R-squared 0.185 0.091 0.171 0.228 Notes: Standard errors clustered at municipality-year level in parentheses. Dependent variable is log-income from work. All specifications use predicted in-migration rate as instrument, include the same controls listed in Table 4, a set of 1991 municipality-level characteristics interacted with time dummies and are weighted by the inverse of the number of observations in the municipality-year. *** significant at the 1% level. ** significant at the 5% level. * significant at the 10% level 20

Table 6 Reduced Form: Labor market impacts of internal migration with time trends, by skill level and sector Employment Log wages Overall Formal Informal Overall Formal Informal (1) (2) (3) (4) (5) (6) In-migration rate 0.032 0.012 0.082-0.685** -0.748** -0.681 (0.141) (0.133) (0.150) (0.297) (0.291) (0.503) 21 In-migration rate x low skill -0.214 0.011-0.174-0.231-0.293-0.625 (0.157) (0.142) (0.163) (0.321) (0.310) (0.512) In-migration rate x high skill 0.452*** 0.013 0.519*** -1.312*** -1.259*** -0.807 (0.173) (0.176) (0.180) (0.353) (0.325) (0.655) Observations 1,139,388 1.139.388 1.139.388 722,688 347,144 308,820 Municipalities 1.354 1.354 1.354 1.354 1.354 1.354 R-squared 0.146 0.055 0.038 0.185 0.206 0.189 Notes: Standard errors clustered at municipality-year level in parentheses. In columns (1)-(3), dependent variable is a dummy equal to 1 if individual is employed, 0 otherwise. Informal sector indicates unregistered work or self-employment. In columns (4)-(6), dependent variable is log-income from work. All specifications use predicted in-migration rate as instrument, include individual- and municipality-level controls as described in section 4 as well as interaction terms of state- and schooling-year. Also, all regressions are weighted by the inverse of the number of observations in the municipality-year. *** significant at the 1% level. ** significant at the 5% level. * significant at the 10% level