The Heterogeneous Local Labor Effects of Mining Booms

Transcription

1 The Heterogeneous Local Labor Effects of Mining Booms Edgar Salgado * March 9, 2016 Abstract Using two rounds of population census for 1043 districts in Peru I document that large-scale mining activity had a positive effect on local employment over 14 years. The effect is differentiated by industry, skill and migration status. Employment grew by 0.23 percentage points for skilled workers in the non-tradeable sector percentage points were explained by an increase in high skilled employment for workers born in the district and were living there five years ago, while 0.02 percentage points by workers born in the district that returned from other district. Using data from 10 annual household surveys I find that, consistent with a model of heterogeneous firms and labor, wages for low skilled workers increased by 0.47% with proximity to the mining activity. Together these findings suggest that large-scale mining activity increases the demand for non-tradeable employment, which is filled by high skilled locals who decide to stay or return, while low skilled workers enjoy a wage gain. 1 Introduction Over the past decade commodity prices boomed to the benefit of many countries endowed with natural resources. When a producing country experiences a price boom, revenues increase from a macro perspective and GDP also booms, mainly fueled by *Univeristy of Sussex, E.Salgado-Chavez@sussex.ac.uk. I am indebted to Andy McKay and Paolo Masella for their supervision and guidance. I also thank Sebastian Sotelo for valuable comments and discussions through different stages of this project. I would also like to thank Nemera Mamo and Sambit Bhattacharyya for sharing their mineral data; and the participants of the 2015 PhD Conference at Sussex for comments. The views expressed in this paper are mine alone, as are all errors. 1

2 higher revenues. However, from a micro perspective, who gains from the mining boom? Far from being easy to answer, this question poses certain difficulties when the analysis of the effect takes a detailed perspective. Large scale mining is generally regarded as an industry with little contribution to the labor market due to its high dependency on capital. This is, of course, of even more relevance in a small country whose main activity is mining. In concrete, one could formulate the following questions: which industries? What type of workers? Are the residents benefited or the newcomers? This paper aims to address these three questions. Although for several years the academic and policy debate revolved around the Natural Resource Curse, or Dutch Disease ideas championed by Jeffrey Sachs and Andrew Warner in the nineties (Sachs and Warner (1995), Sachs and Warner (1999)), recent developments in the literature have adopted a within country perspective to re-visit this old idea (Alcott and Keniston (2015)). Beyond the question whether natural resources abundance is good or not for long-run growth lays the question about the mechanisms operating within the country. What are the local effects of a surge in oil or mineral production? Evidence on the effect of local oil production in United States cities suggest an increase in local wages and employment while no evidence of a Dutch Disease (Alcott and Keniston (2015), Black et al. (2005)). Interestingly, the effect of oil or coal has longer lasting effects on the local economy than agricultural booms (as found by Hornbeck and Keskin (2015)). Be it natural resource or agriculture, there has been also a surge on the analysis at the level of the local economy. In particular, several works attempt to understand the local effects of a local demand shock (Moretti (2010), Notowidigdo (2013)) and its interaction with policy intervention (Kline and Moretti (2014)). Therefore, this paper lies in the intersection of those two streams of literature: the theoretical and empirical investigation on the effects of natural resource abundance, and local labor markets. Evidently, the large availability of data in the US allowed many of these investigations, however it remains an issue that the US does not rely almost completely on natural resources to promote economic growth. Countries like Peru or Chile, on the other hand, heavily depend on the performance of the natural resource sector to promote growth at the aggregate level. And it is precisely in those contexts where the question about the Natural Resource Curse or Dutch Disease becomes more relevant. The new availability of rich dataset provides a unique opportunity to explore this idea. Therefore, in answering the three questions mentioned above, this paper contributes in two ways. First, I build an spatial equilirium model that summarizes some of the early empirical evidence on the effect of mining booms in the local econ- 2

3 omy. Second, I propose an empirical strategy to answer the main predictions derived from the model. Theoretically, this paper relates to several works in recent literature. In Alcott and Keniston (2015) the authors develop a general equilibrium model with three types of firms: natural resources, monopolistic intermediaries and local firms; and just one type of labor. Notowidigdo (2013) presents a model that differentiates between high and low skilled workers, but labor demand is derived from just one type of firms in the tradeable sector. Kline and Moretti (2014) extend the basic setting presented in Moretti (2011) to examine the effect of government transfers in the local economy. Moretti and Thulin (2013) provide a conceptual framework to understand local multipliers in a context of firms in the traded and non traded sectors and different types of labor. The model presented here assumes three types of firms and two types of labor, which constitutes a novelty in the literature. The literature on local labor market has boomed in recent years. It is probably the work of Moretti (2011) the one that ignited a re-visit to the work of Roback (1982), Topel (1986), Bartik (1991) and Blanchard and Katz (1992). From that point in time the literature has sought to understand the impact of different local shocks. Moretti (2011) provides the theoretical framework to understand the effects of a productivity shock in local markets. The effects come as a consequence of the interaction of labor and housing markets in a context where individuals take decisions based on the wage differences across cities. Empirically, Moretti (2010) proposes an estimation strategy based on the Bartik type of shocks (Bartik (1991)) to understand the creation of jobs in the non-traded sector as a consequence of a demand shock in the traded sector. Notowidigdo (2013) develops a theoretical model to explain the effect of local demand shocks conditional on the degree of mobility attributed to different types of labor. His estimates suggest that low skilled workers react less to local demand shocks. Kline and Moretti (2014) also develop a theoretical model to understand the impact of local based interventions. One extension of their theoretical model relates the effect of productivity shocks to agglomeration. More recently, Alcott and Keniston (2015) develop a theoretical model to evaluate the local impact of a shock in the natural resource industry given its links to intermediate industries and local firms. Their model is tested with data from the United States and assumes, however, only one type of labor. Moretti and Thulin (2013) provide a conceptual framework to estimate the effect of a demand shock in the traded sector and its effect on the non-traded industry for the United States and Sweden. Monras (2015) provides a theoretical model useful to evaluate the response of migration as a consequence of the financial crisis in the United States. His results 3

4 suggest that in-migration flows were smaller in the location that were severely hit by the crisis. Empirically, this paper is close to the literature on the effect of natural resources booms within a country. Recently, there have been efforts to re-evaluate the Natural Resource Curse through the lenses of within country variation. For the United States, James and Aadland (2011), James and James (2011) and Papyrakis and Gerlagh (2007) examine the effects of natural resource booms on economic outcomes. The empirical literature has, however, taken a broader approach by analyzing the effect on natural resource boom on conflict (Dube and Vargas (2013)), or corruption (Brollo et al. (2013)). Closer to the purpose of this paper, Black et al. (2005) and Aragón and Rud (2013) evaluate the impact of specific mine exploitation on labor variables. Black et al. (2005) evaluate the impact the coal mining boom, peak and bust for a group of producing cities in the United States. In the developing country setting, much effort has been placed on identifying short-term impacts, but there is a lack of evidence on the long-term impact or the heterogeneity of the response by the type of worker, as well as a better documentation of the channels. Aragón and Rud (2013) explore the backward linkages of a demand shock from a gold mine. The authors build upon the framework in Moretti (2010) to explore local multipliers. Fafchamps et al. (2015) propose a broad empirical model to understand the booming gold activity in Ghana. Using night lights data, the authors conclude that mining shocks can predict protourbanization in the area surrounding the mine. Interestingly, they also find that the results are not reversed once the mine is closed. The literature has usually found a positive impact on income and a reduction in poverty for regions close to a mine (Aragón and Rud (2013), Black et al. (2005), Loayza et al. (2013), Fafchamps et al. (2015)). However, little is known about the processes explaining that outcome. The main contribution of this investigation it to provide an explanation for that result. In particular, I frame the analysis in a model for industries selling traded and non-traded goods, that also differentiate by skill groups in the labor force. Additionally, the analysis tries to disentangle whether locals or migrants benefit from the labor impact brought by the booming mining activity. In concrete, I use a sample of 1043 districts from Peru that were close to a mine in I construct a comparable set of districts and evaluate their performance 14 years later. The empirical section focuses on the Andean region of Peru. This allows to reduce the effect of unobservable factors that are important, for instance, in the main economic region: the wealthy coast. In that regard, the estimation of the results addresses a common belief in countries like Peru: that the aggregate prosperity brought 4

5 by large-scale mines does not spill into the local economies. At the core of the estimation strategy, this paper exploits geographical variation in the exposition to mining activity measured by the proximity to the mine. Other things equal, districts close to the mine and those in the surrounding area can be comparable. The challenge consist, precisely, on gathering the appropriate control group for comparison. In this regard, the choice of the period helps to establish an appropriate baseline period. The empirical section details the design, but at this moment is worth mentioning that this approach is not new in the literature. Aragón and Rud (2013) also exploit geographical distance to a gold mine in northern Peru to evaluate the effects of the mine activity in the local economy. Black et al. (2005) are also a reference methodology, as they create a sample of comparable cities in the US to understand the effects of coal booms and boost periods. Not related to natural resources, Hornbeck and Keskin (2015) define a similar methodology to explain agricultural spillovers in the long-term. Analogous methodologies rely on regression discontinuity (Dell (2010)) to evaluate the impact of districts exposed to mining activity during the colonial period. As a preview, results suggest that large-scale mining activity created high skilled employment in the non-traded sector of the localities close to the mine. The employment increase was filled by people born in the district. Wages also increased, focused on low skilled workers, and in the agricultural sector. For the remainder, this paper is organized as follows: section 2 defines the theoretical framework to guide the analysis. Section3 details the data used in the analysis, explains the selection of the districts and specifies the identification strategy. Section 4presents the results while section 5 discusses the results in terms of the elasticity of substitution of labor and mobility costs. Finally section 6 concludes. 2 Conceptual Framework: Model The theoretical basis for the analysis is grounded on the local labor markets literature (Moretti (2011)). In this section I develop a spatial equilibrium model with heterogenous firms (tradeble, non tradeable, agriculture and natural resources), and two types of labor (high skilled and low skilled). 2.1 Environment The unit of analysis is the district or locality i, where four types of firms, indexed by j, operate. Natural resource firms (R) produce Yi R to sell in the international markets 5

6 for mineral commodities. Firms in the tradeable sector (T) produce Yi T to sell in the national, local or international markets. tradable sector is comprised by manufacture and agrculture firms (I model them together theoretically, but empirically I look at different effects). Firms in the local economy, or non-tradeable sector (S) produce Y S i to sell locally. Non-tradeable sector is comprised by all other firms with the exception of government. These firms compete for two types of workers in the locality: high skilled workers (H i ) and low skilled workers (L i ), who receive different wages: w H and w L, where w H > w L. The population, N is fixed. So do the stock of high skilled H and low skilled L workers. The model does not account for agglomeration, but it could be incorporated by assuming, for instance, that productivity shocks are magnified by the change in population. 2.2 Workers A worker c in location i maximizes his utility depending on aggregate wages for his type, amenities and a random component that defines his location preference (also depending on his type). Both types of workers have the same preference and enjoy the city amenity similarly 1. Therefore, high skilled workers exhibit the following preference: while low skilled workers: U Hci = ln w Hi + A i + e Hci (1) U Lci = ln w Li + A i + e Lci (2) I assume each worker supplies one unit of labor. A i defines any type of local amenity, and e Hci and e Lci are the random preference for city i, idiosyncratic to the individual and drawn from an type I extreme value distribution with mean zero and scaling parameter S H for high skilled and S L for low skilled workers. Notowidigdo (2013) points out that this could represent not only the location preference but general tastes and mobility cost. If there are two cities a and b, the difference between the two random components divided by the scaling parameter can be expressed as a standard logistic distribution with mean zero and variance 1. For example, high skilled individual have the 1 Moretti (2011) differentiates the utility derived from amenities by each type of labor, however this simplification does not alter the main conclusions of the model. 6

7 following preferences for location: e Hca e Hcb S H logistic(0, 1) (3) The marginal high skilled worker will prefer city a over b if e Hca e Hcb > (ln w Hb ln w Ha ) + (A b A a ). Therefore, the proportion of high skilled workers that locate in city a is: H a H = Λ ( ) (ln wha ln w Hb ) + (A a A b ) Where Λ(.) = e(.) 1+e (.), is the cumulative density of the standard logistic function. Solving (4), the labor supply of high skilled workers is: S H ln H a = 1 S H [ln H b + (ln w Ha ln w Hb ) + (A a A b )] (5) From equation (5) it is clear that large values of S H (strong location preferences, or mobility costs) imply a labor supply that is inelastic to changes in wages, while small values S H imply a labor supply that is very elastic. Totally differentiating equation (5) for both cities and assuming amenities are exogenous, the labor response of high skilled workers in city a is: H a = (4) 1 S H + 1 [ w Ha w Hb ] (6) Analogously, low skilled worker s migration response is: L a = 1 S L + 1 [ w La w Lb ] (7) Where represents percentage change. From equation 6 is is clear that the migration response of the workers in negatively associated with the location preference (or migration cost). High preferences for one location, or high migration costs imply a less responsive labor supply. If location preferences are irrelevant or there are no mobility costs, labor supply is very sensitive to wage differences between cities. 2.3 Labor Demand There is no distinction in the use of factors of production, they all use capital (K), high skilled workers (H) and low skilled workers (L), but with different shares. Each sector j, therefore combines capital and a Z function for labor using a Cobb-Douglas 7

8 aggregation function: Y j i = X φ j K (1 α j) ij Z α j ij (8) X = PX is the idiosyncratic productivity shock in the natural resource sector. This shock has two components: P, which reflects changes in the international prices of mineral commodities, and X, that represents the technology changes in the sector. For simplicity of exposition, I refer to the combined X as a productivity shock, but empirically I will evaluate it as a price shock. It affects the other sectors through φ j. For instance, in the natural resource sector, a productivity shock is fully absorbed, therefore, φ R = 1. While for the other sectors, the effect is possibly smaller. For instance: 0 < φ T < φ S < 1. Z ij takes into consideration the substitution between high and low skilled workers, using a Constant Elasticity of Substitution technology: Z ij = [(1 λ j )L ρ ij + λ jh ρ ij ] 1 ρ (9) Where ρ is a parameter associated to the elasticity of substitution between high and low skill workers, σ = 1 ρ 1. The wage of each type of worker is aggregated at the city level: w Hi and w Li, while the price of capital is determined internationally, W K. In consequence, firms maximize profits by equating marginal productivity of each factor to its price (omitting j and i for simplicity): w K = (1 α) Xφ K α [(1 λ)lρ + λh ρ ] α ρ (10) w H = αx φ K (1 α) [(1 λ)l ρ + λh ρ ] α ρ ρ λh ρ 1 (11) w L = αx φ K (1 α) [(1 λ)l ρ + λh ρ ] α ρ ρ (1 λ)l ρ 1 (12) Replacing the first order condition of capital (10) into the conditions for high skilled workers (11) and low skilled workers (12), and totally differentiating: w H = φ α X + 1 [(1 π) L π H] (13) σ w L = φ α X + 1 [π H π L] (14) σ λh Where represents percentage change and π = ρ (1 λ)l ρ +λh ρ. Equations (13) and (14) are the wage response functions for the two types of workers. They depend on 8

9 the exogenous response to the productivity shock in the natural resource sector, and the endogenous response of the labor supply of both high and low skilled workers. It is clear that for each worker type the wage increase due to a productivity shock in the natural resource sector is mitigated by the labor supply response. In the extreme case that labor is fully mobile, no wage differences should prevail. Additionally, the substitution between high and low skilled workers provides an additional source of mitigation on the final effect on wages. 2.4 Equilibrium and Predictions Equilibrium in the labor market takes place when labor demand and supply intersect. This implies interacting equations (6) and (7) with equations (15) and (16). I directly evaluate the equilibrium response when there is a productivity shock in city a, i.e. X a = 0 while X b = 0. H a = φ[(s L +1)σ+2] (S H +1)(S L +1)σ+(S H +1)(π+π )+(S L +1)(2 π π ) X a (15) While the response for low skilled workers in city a is: L a = φ[(s H +1)σ+2] (S H +1)(S L +1)σ+(S H +1)(π+π )+(S L +1)(2 π π ) X a (16) Where π is the corresponding share of high skilled workers drawn from the CES function in city b. In both cases, the location preferences are negatively associated with the magnitude of the response to the shock. Higher values of S H in equation (15) indicate smaller mobility for the high skilled type of worker; while the relation to the elasticity of substitution of labor, σ, is less clear. Figure (5) simulates the response of high skilled labor to different values of the location parameter (or mobility cost) and the elasticity of substitution. As explained, labor response is highly mobile for low values of S H. The continuous line assumes a value of σ = 0.5, which is an indication of high skilled and low skilled workers being complementary. In such case, for low values of S H, high skilled workers are highly mobile. When σ > 1, which is an indication of high substitution between high and low skilled workers, the response of labor supply is smaller, even for small values of S H. Intuitively, although location preferences and mobility costs are important determinants of labor mobility, a large degree of substitution between high and low skilled type of workers attenuates the labor supply response. Each type of worker s wage responds: 9

10 w Ha = φ α X a + 1 σ [(1 π) L a (1 π) H a ] (17) w La = φ α X a + 1 σ [π H a π L a ] (18) In each equation the first element on the right is the exogenous response of wages to the mining shock, while the second element is the endogenous response due to migration. replacing equations (15) and (16) in (17) and (18), the response of wages to the mining shock as a function of only exogenous parameters is: w Ha = w La = [ ] φ α + φ(1 π)(s H S L ) σα[(s H +1)(S L +1)σ+(S H +1)(π+π )+(S L +1)(2 π π X )] a (19) [ ] φ α + φπ(s L S H ) σα[(s H +1)(S L +1)σ+(S H +1)(π+π )+(S L +1)(2 π π X )] a (20) These equations indicate that the increase in wages in the case that both types of workers are the same (or have the same location parameters, S H = S L ) is φ α. Depending on the differences in the location preferences, the impact on wages varies. If high skilled workers are more mobile than low skilled workers, S H < S L, the initial push on wages, φ α, is offset by the differential labor response. However, this balancing force is attenuated by the elasticity of substitution of labor σ. This is clearly depicted in figure (6). The green continuous line represents the independent effect of any mining boom on high skilled wages, φ α, the same as the case in which workers have no different location preferences or mobility costs (blue crossed line). Interestingly, the wage response tends to quickly converge to the independent effect, even for small values of σ 2. In consequence, wages are more responsive than labor to the mining shocks. Depending on the difference between the location preferences, and the elasticity of substitution of labor, the model also predicts different responses in wages by type. The difference between equation (19) and (20) captures this idea: w Ha w La = φ(s H S L ) (S H +1)(S L +1)σ+(S H +1)(π+π )+(S L +1)(2 π π ) X a (21) Again, the difference in wage response by type is determined by the difference between S H and S L, and the magnitude of σ. Figure (7) indicates that if workers are equally mobile (S H = S L ), there is no difference in the wage response, and both type of workers gain. Assuming S H < S L, the model predicts w Ha w La < 0. For 2 When workers have a large degree of substitution (σ > 1), wages do not experience a reduction due to labor response. 10

11 larger mobility costs in general (S H = 20 and S L = 30) the wage gap is smaller. Also, the larger the degree of substitution between workers, the smaller the wage differential. 2.5 Migration Status There are two migration categories in the census: (i) residency in the district five years ago, and (ii) birth. These two dimensions define four categories I define in matrix form: { } n n i = 11 n 12 n 21 n 22 Columns are indicators of being a resident in the district 5 years ago, while rows indicate whether the worker was born in the district. Therefore, a random worker n who was born in the district i and resided there five years ago has n 22 = 1, and I term them as locals. n 11 = 1 indicates that the worker was not living in the district five years ago, neither was born there, which I term as newcomers. n 12 = 1 for workers who were not born in the district but were residing there five years ago, which I term new locals. Finally, n 21 = 1 for workers that were born in the district but did not reside in it five years ago, that can be called returners. In the empirical part, I explore the change in population and labor explained by the change in any of these four categories. For instance, the total number of workers in district i can be understood as: N i = N i11 + N i12 + N i21 + N i22 (22) 3 Data and Identification Strategy 3.1 Data I combine different data sets and construct novel datasets at district level previously not available. For employment I use the population censuses of 1993 and 2007 in the estimations, and the population censuses of 1993 and 1981 to check the comparability of the districts in In consequence, the period of analysis is 1993 to To look at wages I utilize data contained in the employment module from the National Household Survey. This survey is collected annually from 1997 by the National In- 11

12 stitute of Statistics of Peru. I also use this survey to estimate the elasticity of substitution of labor, in particular, the periods and Data on mine location and production are available from InterraRMD (2013), which provides the GPS location of all large-scale mines in Peru, as well as physical production from the beginning of the 20th century in some cases. I discuss all of them in detail. Mineral Data Data about production and geographic location of mines are available from the Raw Material Data (RMD) from InterraRMD (2013). This dataset has records for 633 mines through the period of Each mine has a GPS location that I use to measure the distance from the capital of the district to the mine. Mineral production is available on an annual basis from 1975, and by 2013 there were 367 active projects (some mines are involved in several projects). The panel corresponds to large scale mining projects for a variety of minerals. I constraint the sample of mining projects of five minerals: gold, silver, copper, lead and zinc. Data on prices for these five minerals are available from the U.S. Geological survey (see Kelly and Matos (2013)). All prices have 1994 as base year. The annual mining production from 1950 to 2013 and its composition is presented in figure 2. Mineral production boomed from early 2000 s. Copper was by large the main mineral extracted from Peruvian mines. Gold production gained importance from middle 90 s. To give an idea magnitude of the boom, from 1985 to 1995 mining production for these five minerals grew at an annual average of 3.3%. For the second half of the 90 s the average annual growth rate was 7.3%. During the 2000 s the annual growth rate increased to 16%. Census and household Data Population censuses are collected by the National Institute of Statistics (INEI in Spanish) on an irregular basis. The last population census dates to 2007, while the closest one dates to Both are de facto type of census. INEI collected a de jure type of census in 2005, which I do not use to avoid problems of comparability. The number of districts increased from 1,791 in 1993 to 1,831 in 2007, therefore I match 2007 district borders to 1993 borders. In the empirical design I end up using 1043 districts. I was also able to recover district data for 22 of the 25 regions in Peru recorded in the 1981 census. Unfortunatelly one of the missing regions falls within the geographical space I use in the empirical design, which means a reduction of the number of 12

13 districts to 931. I use these 931 districts only for balance checks and the falsification test. In consequence, the period of analysis is from 1993 to With these data I estimate employment variables at the district level by industry and skill type; as well as demographic variables. INEI also collects the National Household Surveys (ENAHO in Spanish) every year since The survey has changed through time, however it retains many of its important modules set from the beginning. For the estimation of the effect of mining activity on wages I use data from the modules on education, employment and income, household characteristics, and individual characteristics. These modules have not changed much through the years, thus I am able to construct a set of pooled data from 1997 to For the estimation of the elasticity of substitution I also use ENAHO. However, I rely on different periods for this due to changes and additions to previous versions of the survey. As a consequence I use two periods independently: and In section 5.1 I discuss the details. Other data Through the analysis I also make use of data on altitude, rainfall and land extension. Data on altitude and land extension were available from the Ministry of Energy and the agricultural census (that covers only cities with agricultural land, and therefore excludes a few districts in the coastal region with no agricultural production). The source for the historical rainfall data is the Climatic Research Unit from the University of East Anglia 3. I use the version TS3.20 that covers the period from January 1901 to December 2011 and provides precipitation estimates at the 5 5 grid resolution, that I match to the district borders. 3.2 Empirical Framework The unit of observation for the mining activity is the district. The dependent variables are at the district level, for employment and migration; while at the individual level, for wages. The identification strategy selects a set of districts that are comparable at the baseline, Then it compares their response to the boom in commodity prices from 1993 to In 1993 there were 27 mines extracting the following minerals: gold, silver, copper, zinc and lead. One mine extracted the 5 minerals, 5 mines extracted 4 minerals, 11 3 Available at 13

14 mines extracted three minerals, 4 mines extracted 2 minerals, while 6 mines extracted only one mineral. This feature is going to be exploited in the identification strategy. In 1993, a district was under the influence of any of the 27 mines if the distance between its capital and any of the mines was at least 100km. This threshold is assumed in the literature and I opt for it 4. For the estimation of the distance I use the length of the shortest curve between the coordinates of the capital of the district, and the coordinates of the closest mine 5. Then I use districts beyond 100km but not far from 200km as comparable districts. The idea is that by 1993 there were no meaningful differences between districts with a mine at a 100km distance and the closest neighboring districts in the 200km radius of the mine. The choice of 100km can also be argued by the fact that the average district in Peru has an area of 71,866 Km2 while the median district has an area of 20,852 Km2. This means that the 100 km threshold is slightly below the square root of the median area. By 2007 the number of mines increased to 68. Although new mines offer an interesting analysis, many of them started operation between 2004 and 2006, which puts them closer in time to the endline of the period, and would be more suitable for a short-term analysis. In this paper I am interested in the long-term differences. Therefore, in order to rule out any interference coming from these 41 new mines, I exclude those districts that were within the 101k-200km distance interval in 1993 but later were spilled by a new mine. Also, I exclude any district with an altitude below 1,800 meters above the sea level. The result of this is the exclusion of coast cities, and cities located in the rainforest. There are two reasons for this. First, Peru s geography is very heterogeneous and exerts an important influence in the markets. The Andes mountains divide the country into two parts, the coast on the West and the rain-forest on the East, leaving the Andes as a middle region. This partition influences mainly transport costs, which ultimately explains the differences in productivity as documented by Sotelo (2015) 6. Second, there are no large-scale mines operating below 2,000 meters above the sea level. Therefore, the 1,800 threshold allows the inclusion of downstream cities that are still under the influence of the mining activity but somehow isolated from the coast or rain-forest dynamics. In total, 823 districts were within the 100km distance from 4 For instance, in a similar setting, Aragón and Rud (2013) define 100 km as the influence distance. Fafchamps et al. (2015) assume no effect of mining activity beyond 100 km. 5 I implement this by using the STATA command geodist. 6 Former Peruvian president Manuel Pardo, during the early days of the Peruvian Republic in 1862, coined an illustration that has survived to this day: the freight cost from Jauja (a small city in the central highlands) to the capital, Lima was four times higher than the maritime freight from the capital harbor, Callao, to Liverpool. See Pardo (1996) 14

15 a mine, while 220 districts were within the 101km-200km distance interval. Figure 3 illustrates the region under analysis. The general strategy is close to the spatial discontinuity used by Dell (2010) who defines a regression discontinuity framework for districts within a radius of 50 km within the historical boundary of mining district. It is also related to the strategy used by Aragón and Rud (2013) to identify the effect of the expansion of mine purchases in a local economy within a radius of 100 km, with comparison districts located in the km radius. It also resembles the methodology applied by Fafchamps et al. (2015) to analyze the proto-urbanization produced by gold mining in Ghana. These methodologies bear some resemblance to Black et al. (2005) and Loayza et al. (2013). The first identifies a group of treated and control counties for a group of states producing coal in the United States. The second constraints the group of mining and non-mining districts to the same province that reports mining activity. This methodology is also related to Hornbeck and Keskin (2015) who study the long-term impact of agricultural activity for a group of counties in the United States that are close to the Ogallala aquifier. The proximity to the mine is not the only difference between districts. As stated above, I actually exploit the fact that mines had different portfolios of minerals to construct a measure of the price boom experienced by districts. The change in prices for each mine g is defined as: P g = κ m P m (23) m Where m={gold, silver, copper, zinc, lead}. κ m is the mineral weight, estimated as the share in the total mineral production in 1994 valued at 1994 prices. P m is the price change (increase in percentage points) in the mineral from 1993 to Then, I match every district in 1993 to the closest mines. Some districts were under the influence of many mines, therefore, the final measure of price boom adds one additional source of variation: the number of mines influencing the district. In concrete, the district measure that I construct is: P i = g P ig G i (24) Where each district i was under the influence of G i mines, each with a mineral portfolio that yielded a price change P g. The measure P i is simply the average price boom over all mines that influenced the district. Linking this to the theoretical 15

16 model, P i is the proxy for X i, and P i > 0 for all of the 823 districts that had a mine within 100km distance in Figure 4 plots the distribution of P i. The empirical equation, therefore, I estimate for the effect on labor is: ( ) N j i = α + β P i + θ k Γ ki + η p + ε it (25) Where (N j i ) = (Nj i2007 /N i2007) (N j i1993 /N i1993). The strategy evaluates the change in the employment rates in the city i and sector j. I also evaluate the change in employment by skill type.the coefficient of interest is β which captures the effect of the boom in mineral prices on the city, as defined above. θ k is a vector with district geographical controls: altitude, the historical coefficient of variation of rainfall, distance to Lima, land extension of the district and the initial level of population (in logs). η p are province fixed effects. ε it is the empirical disturbance. The empirical equation for migration status is the same. In this case, the dependent variable is the change in the share of any of the 4 categories stated in equation (22), also by industry and skill. The analysis of wages requires some changes. The design of the census does not include any question regarding income or salary, thus I cannot use these data to infer the effect of mining booms on wages. Instead, I use data from the ENAHO. The measure of wages corresponds to all individual earnings from main and secondary occupation expressed on a monthly basis at 2007 prices. All individuals with a dependent job or self-employed were considered. The analysis excludes individuals who report working as unpaid workers. From ENAHO I can also include several individual characteristics as controls. To measure the effect of the boom in mineral prices I try three specifications. The first is simply an individual level application of equation (25): ln w cit = α + β P i + Ω a Π act + θ k Γ ki + η p + t t + ε cit (26) Where ln w cit is the natural log wage of individual c living in district i at time t. The right hand side of the equation is similar to the employment equation, with the exception that I also control for a set of a individual characteristics Π act : gender (male), age and its square, the number of years of schooling, household size, number of income earners in the household, dummies for water and electricity in the household, fixed effects for industry (two digits ISIC), job type (owner, self-employed, white collar, blue collar), year, t t, and province, η p. The variable P i is the same across time for each district and measures, as usual, the effect of the price change from 1993 to

17 Although informative, the previous equation does not exploit annual variation in the cross section. It does not control for district fixed effects either. To address this, I re-define the mineral price variable as a price index (base 1994) that uses as weights the mineral production of 1994, and evaluated lagged one period to allow for adjustment: PI t 1. The empirical equation, in this case, take the following form: ln w cit = α + βpi it 1 + Ω a Π act + η i + t t + ε cit (27) I have kept the individual controls, but now geographical fixed effects are at the district level. Therefore, any change on wages steams from the time variation in the mineral price index. Finally, as a check, I also try with mining production instead of the mineral price index. ln w cit = α + βproduction it 1 + Ω a Π act + η i + t t + ε cit (28) 4 Results 4.1 Baseline results: employment Table 3 presents the baseline results for equation 25. Columns 4, 5 and 6 present the main results. Employment rate (column 4) increased 41 percentage points from 1993 to 2007 by every one standard deviation increase in the mineral prices. This effect is concentrated in high skilled workers: the increase was of 0.23 percentage points. High skilled workers are all those with at least with technical education leaving those with secondary or less as low skilled workers. Low skilled workers did not experience an significant increase in employment. Columns 1, 2 and 3 provide additional results about population changes. Column 1 indicates that there is no statistically significant effect on the population size in the district, however, the composition of the population did change. The share of individuals (not all of them workers) with high skill increased by 0.16 percentage points, while the share of low skilled population decreased by the same amount. This set of initial results already point to one of the main results: the availability of high skilled individuals increased in the districts influenced by the mining activity, and so did the their employment. Employment for low skilled workers did not change, but the availability of low skilled individuals did decrease. In table 1 I conduct a falsification test to check the robustness of this result. In 17

18 panel a. I simply regress the level of employment and population shares in 1993 to the change in mineral prices from 1993 to None of the variables are affected by future changes in the commodity prices. In panel b. I use information from the 1981 census to check whether the change from 1981 to 1993 can be explained by the future change in prices. Results indicate that this is not the case. However, as stated above, I can only recover information for 931 districts in the 1981 census. Looking at the sector response, table 4 evaluates whether industries and the combinations of industry-skill reacted differently to the boom in prices. Column 8 indicates that the increase in high skilled employment was explained by an increase in high skilled employment in the non-tradeable sector. Tradeable and agriculture did not respond to the boom in prices. The mining sector shows no response either. In table 2 I check again whether the boom in prices from 1993 to 2007 explained the change in employment between 1981 and No result is statistically significant at 5%. Again, I can only use 931 districts for this test, and that may explain the 10% significant reduction in low skilled employment in the non-tradeable sector. 4.2 Effect on migration The increase in employment found in the previous section could have different origins. This section sheds light on whether locals or newcomers fill the new employment opportunities. Table 5 shows the results for the effect of the boom in mineral prices on the composition of the population and employment by migration category. Panel a. looks at the population composition and column 2 indicates that the proportion of high skilled individuals born and who were residing in the district 5 years increased by 0.27 percentage points with the boom in prices. There is no meaningful effect in any of the other migration categories. Panel b. of the same table evaluates the employment share according to these categories. Column 1 shows that there is a significant effect on the proportion of employed local people: the mining boom produced an increase of 0.43 percentage points, which is higher compared to the whole change in employment due to the mineral price boom (0.41; column 4 of table 3). More importantly, the share of high skilled locals who are working in the district increased by 0.19 percentage points (column 2), while the proportion of returners grew by 0.02 percentage points. These two coefficients together suggest that out of the 0.23 percentage point increase in high skilled employment, 0.21 were explained by workers born in the district. The other migration categories did not respond to the boom in prices. 18

19 Table 6 looks at the industry detail. Again, the 0.11 percentage points increase in non-tradeable employment found in table 4 (column 8) is mainly explained by an increase in the employment of high skilled locals and returners. There is also an statistically significant increase in the share of low skilled returners in the agriculture sector. Mining and tradable sectors show no response in their employment composition to the boom in mineral prices. 4.3 Effect on wages Results so far indicate that high skilled workers found more jobs in districts under the influence of the mineral price boom. This also meant that the share of individuals with high skill increased while the proportion of low skilled individuals decreased. The theoretical model predicts that under this circumstance the wage for low skilled workers should experience a higher increase than the wage for high skilled workers. Table 7 presents the results of estimating equations (26) in panel a., (27) in panel b; and (28) in panel c. Column one makes no differentiation in the type of worker and estimates the wage response of all individuals. In the three cases, log wage increases with the boom in mineral prices. The effect is quite large in panel a, which controls only for province fixed effects and does not take into account the annual variation of mineral prices but the , only. Through the paper, the definition of high skill has been technical education at least. Estimations considering this definition are in columns 3 and 6. Focusing on them, results in panels a and b indicate that the increase in wage reached only low skilled workers. In panel c both types of workers gain, but the estimated coefficients are not statistically different. If the definition of high skill is extended to individuals with at least secondary education (this compares columns 4 and 7), the results indicate that low skilled workers (in this case, with primary or no education), experience an even greater wage gain in panels a and b. In panel c, both experience a wage increase again, but the coefficients are not statistically different. If, on the contrary, the definition of high skill consider only workers with university degree at least, results still indicate that low skilled workers gained from the mining boom. In panel b, high skilled workers actually experience a decline in their wages. Among these estimations, panel b, which estimates equation (27), seems more reasonable. Equation (28), estimated in panel c links the wage response to mining production every year, which is less exogenous than the mineral price change as esti- 19

20 mated in panel b. Is there a different response by industry? With ENAHO data I was able to estimate the wage response by sectors. Agriculture and non-tradable industries have similar number of observations while the number of observations for mining and tradable is rather small. Panels a and b in table 8 show that the agricultural sector captures the positive response of wages to the boom in mineral prices. Wages in the mining sector have a negative response in panel a., however, the sample size is rather small. Panel c indicates that wages in mining, non-tradable and agriculture sectors increased with the boom in mineral prices. 5 Discussion Results so far indicate that the boom in mineral prices increased local labor for high skilled workers in the tradable sector. This result is in line with the idea that during a booming period in the prices of commodities, mines create local opportunities in the sector that trades goods locally. I do not find any effect on employment in the tradable industry or the agricultural sector. The increase in employment is basically explained by a higher employment share for people born in the district, mainly locals who were born in the district and were living there five years ago. Wages, in general, increase. However, this increase reaches mainly low skilled workers, and workers in the agricultural sector. Under the context of the theoretical model, this is consistent with the idea that high skilled workers are more responsive to changes in wages than low skilled workers. If both types of workers were equally responsive, wages and employment would have a similar reaction to the mining boom (which is the result in panel c of table 7). However, not only the idiosyncratic migration response is at play here. If the elasticity of substitution between high and low skilled workers were too high (which implies a high degree of substitution), the difference in the wage response would be very similar, regardless of the degree of mobility as outlined in figure 7. In this section I shed some light on these two issues. 5.1 The estimation of σ Combining equations (8) and (9), the production function of the firms is: Y j = X φ j K (1 α) j [ ] (1 λ)l ρ α j + ρ λhρ j (29) 20

21 From the estimate of ρ, I can recover the elasticity of substitution between high and low skilled labor, σ = 1. I estimate equation (29) with non-linear leas squares (1 ρ) using different samples. The design of ENAHO included some modules in 2002 to survey the independent worker. The questions remained similar until The module inquired the entrepreneur about output in three categories: production, retail and services. These three categories can broadly be defined as non-agricultural activities, mainly related to the non-tradable sector. Besides information on output, there is information about the number of people employed in the firm and their education level. With this information I am able to classify workers as high skilled and low skilled using the same threshold: technical education. There is also information about assets that I use as a measure of capital. The sample goes from 2002 to 2006 and a deflate prices, first, geographically to prices of Lima, and secondly, temporally, to The unit of observation is the business individual who owns a firm. ENAHO also has a module on agricultural production that I use to retrieve information about labor and output in the agricultural sector. This module surveys individuals with a plot and inquires about plot size, as well as agricultural production, number of hours of labor for the owner and any other participant. With the information of the owner of the plot and any additional person involved in the labor of the plot I am able to calculate a measure of high skilled and low skilled labor. I need an assumption here, which is that the production unit is the household, not the individual. With this assumption I am able to calculate the number of hours each plot owner used in the household, and therefore calculate the number of hours high skilled and low skilled individuals devoted to the agricultural works. As measure of capital, I use the land size at household level. Production is the aggregate production of the household. The period for which I can assemble this information is In a literature review about the elasticity of substitution of labor, Behar (2009) states that the there are very different estimates, some with very high values, depending on the country and disaggregation. The widely accepted estimate comes from Katz and Murphy (1992), and it is around 1.4. In a more recent review, Caselli and Coleman (2006) argues that the elasticity of substitution is very unlikely to fall outside the [1,2] range. Table (9) reports the estimated parameters of equation 29 using non-linear least squares. Columns 1 and 3 report the estimation without including any control variable. The elasticity of substitution for non-agriculture is estimated around 2.16 while for agriculture it is approximately Columns 2 and 4 include two simple controls 21

22 in the estimation: a trend and a dummy variable for firms or plots in the region that includes the capital, Lima. The estimated elasticity of substitution for non-agriculture drops to 1.58 while for agriculture decreases to These two results are close to 2, which suggests that if anything, this is not the main explanation behind the results. However, if the differences between the two estimates are an indication that the elasticity of substitution between labor is smaller in the non-agriculture (mainly non tradable) sector than in the agricultural, labor in the first would be more responsive to a boom in the mineral prices (as depicted in figure 5). 5.2 An idea of S H and S L The theoretical model also predicted smaller labor supply response for higher levels of locational preferences or mobility costs. Although I do not directly measure these parameters, their influence can be deducted through migration differences. Figure (8) plots a common measure of migration: the proportion of individuals who five years ago were living in a district different than their current district of residence. This is done for all individuals, but also by skill (university degree) definition in The left panel shows the nationwide non-parametric estimates. Skilled individual tend to be more mobile than less skilled ones. The right panel does the same estimation but using only the 1043 districts included in the empirical analysis (those within 200km of the mine). Skilled individuals in the sample districts are as mobile as those in the whole country. However, less skilled individuals in the sample are even less mobile than the national average. Another way to approach this is by estimating the proportion of locals by age. This is, the proportion of individuals who were born in the district, and five years ago were also living there. Assuming they never moved, this is a measure of location preferences (which is n 22 in section 2.5). Figure (9) shows that for all ages, individuals in the districts used in the empirical section have stronger location preferences than the national average. To sump up, this adds up to the idea that individuals in the region under analysis are less mobile 7. 7 In terms of the figure 7 this implies an scenario depicted by the dashed blue line around a value of sigma of 2. 22

23 6 Conclusions Mining activity has the potential to affect other industries through labor demand. This paper explores the effects of large-scale mining activity on local labor markets. It proposes a simple spatial model with heterogeneous firms and labor that formalizes some of the empirical results found in recent literature. The model predicts a larger labor response of high skilled workers for a given elasticity of substitution between high skilled and low skilled labor. As a consequence, the wage response for this type of workers is smaller than the wage response of the low skilled workers. Additionally, the paper accounts for changes in the population and employment composition by migration category in order to determine whether locals or newcomers benefit from the surge in employment. The empirical design uses data from 1043 districts from Peru during the period of 1993 to 2007 to determine the effect on employment and population composition. Data on individual wages from household surveys help to determine the effect on wages. Empirically, I confirm that high skilled employment grew faster in districts under the influence of mining activity. Such effect was concentrated in employment for high skilled workers in the non-tradable sector. There is no employment effect in the tradable or agricultural sector. In concrete, during the period, high skilled employment grew by 0.23 percentage points due to the increase in the price of minerals. Out of this 0.23 increase, 0.19 percentage points are explained by an increase in the share of employed locals and 0.02 percentage points by returners. Locals are those employed individuals born in the district and who were living there five years ago. Returners are those individuals born in the district who were not residing in the district five years ago. There is no statistically significant effect for the other two migration categories: new locals (individuals not born in the district but were living there since five years ago), and newcomers (individuals born in a different district who were also living in a different district five years ago). High skilled employment in the non-tradable sector grew by 0.11 percentage points. Out of these 0.11 percentage points 0.08 percentage points are explained by an increase in the employment of high skilled local, and 0.01 percentage points of high skilled returners. Wages also increased with the boom in mineral prices. However, such increase was heterogeneous based on the skill and industry. Low skilled workers gained the 23

24 wage increase while high skilled workers did not. This result is consistent with the prediction of the model when high skilled individuals are more responsive to changes in wages than low skilled individuals. By industry, agricultural workers gained the wage increase, while other for other sectors, the result is indicative of no effect. The paper also attempts an explanation of the results based on the discussion of the elasticity of substitution between high and low skilled labor. An estimation of such elasticity for non-agricultural activities (mainly non-tradable) is about 1.58, while for agricultural activities is around Although the range falls within the interval agreed in literature, the lower elasticity of substitution in the non-agricultural sector may explain the higher labor response in the non-tradable sector. Finally, descriptive evidence on migration rates in 1994 indicates that the region under analysis in the empirical section exhibits larger location preferences, which are an indication of low mobility. Low skilled individuals are by far less mobile than high skilled individuals. Ultimately, this explains the large response in wages for this labor type. 24

25 References ALCOTT, H. and KENISTON, D. (2015). Dutch disease or agglomeration? the local economic effects of natural resource booms in modern america. National Bureau of Economic Research, (20508). ARAGÓN, F. and RUD, J. (2013). Natural resources and local communities: Evidence from a peruvian gold mine. American Economic Journal: Economic Policy, 5 (2), BARTIK, T. (1991). Who benefits from state and local economic development policies. W.E. Upjohn Insitute For Employment Research. BEHAR, A. (2009). Directed technical change, the elasticity of substitution and wage inequality in developing countries. Discussion Paper Series, Economics Department, Oxford University. BLACK, D., MCKINNISH, T. and SANDERS, S. (2005). The economic impact of the coal boom and bust. The Economic Journal, 115, BLANCHARD, O. and KATZ, L. (1992). Regional evolutions. Brookings Papers on Economic Activities, pp BROLLO, F., NANNICINI, T., PEROTTI, R. and TABELLINI, G. (2013). The political resource curse. American Economic Review, 103 (5), CASELLI, F. and COLEMAN, W. (2006). The world technology frontier. American Economic Review, 96 (3), DELL, M. (2010). The persistent effects of peru s mining Mita. Econometrica, 78 (6), DUBE, O. and VARGAS, J. (2013). Commodity price shocks and civil conflict: Evidence from colombia. Review of Economic Studies, 80, FAFCHAMPS, M., KOELLE, M. and SHILPI, F. (2015). Gold mining and protourbanization: Recent evidence from ghana. World Bank, Working Papers, (7347). HORNBECK, R. and KESKIN, P. (2015). Does agriculture generate local economic spillovers? short-run and long-run evidence from the ogallala aquifer. American Economic Journal: Economic Policy, 2 (7),

26 INEI (2015). Instituto de Estadísticas e Informática. INTERRARMD (2013). JAMES, A. and AADLAND, D. (2011). The curse of natural resources: An empirical investigation of u.s. counties. Resource and Energy Economics, 33, and JAMES, R. (2011). Do resource dependent regions grow slower than they should? Economic Letters, 111, KATZ, L. and MURPHY, K. (1992). Changes in relative wages: : Supply and demand factors. Quarterly Journal of Economics, 107 (1), KELLY, T. and MATOS, G. (2013). Historical statistics for mineral and material commodities in the united states (2013 version). U.S. Geological Survey Data Series, 140. KLINE, P. and MORETTI, E. (2014). People, places and public policy: Some simple welfare economics of local economic development program. Annual Review of Economics, 6, LOAYZA, N., MIER Y TERAN, A. and RIGOLINI, J. (2013). Poverty, inequality and the local natural resource curse. World Bank, Working Papers, (6366). MONRAS, J. (2015). Economic shocks and internal migration. IZA Discussion Paper, (8840). MORETTI, E. (2010). Local multipliers. American Economic Review: Papers & Proceedings, 100, (2011). Local labor markets. Handbook of Labor Economics, 4, and THULIN, P. (2013). Local multipliers and human capital in the united states and sweden. Industrial and Corporate Change, 22 (1), NOTOWIDIGDO, M. (2013). The incidence of local labor demand shocks. Northwestern University. Mimeo. PAPYRAKIS, E. and GERLAGH, R. (2007). Resource abundance and economic growth in the united states. European Economic Review, 51, PARDO, M. (1996). Estudios sobre la provincia de Jauja. Colección Populibros regionales, Ediciones José María Arguedas. 26

27 ROBACK, J. (1982). Wages, rents, and the quality of life. Journal of Political Economy, 90 (6), SACHS, J. and WARNER, A. (1995). Natural resource abundance and economic growth. National Bureau of Economic Research, (5398). and (1999). The big push, natural resource boom and growth. Journal of Development Economics, 59, SOTELO, S. (2015). Trade frictions and agricultural productivity: Theory and evidence from peru. University of Michigan. Mimeo. TOPEL, R. (1986). Local labor markets. Journal of Political Economy, 94 (3),

28 7 Tables and Figures 7.1 Tables Table 1: Balance: Demographics in 1993 Population Employment Ln. N. High Skilled Low Skilled Ln. N. Rate High Skilled Low Skilled (1) (2) (3) (4) (5) (6) (7) a. Dependent variables in levels 1993 P (1.51) (0.07) (0.07) (1.53) (0.19) (0.05) (0.19) R Observations Clusters b. Dependent variables in difference P (0.87) (0.06) (0.06) (1.17) (0.20) (0.04) (0.18) R Observations Clusters Notes: [1] Data source: (i) population census of 1993 for panel a. (ii) population censuses of 1993 and 1981 for panel b. [2] 1981 census has fewer districts because, according to the public information provided by INEI, it was not possible to recover data from three regions (Apurimac, Loreto and San Martin). Population of these three regions represent 6 % of total population in Moreover Loreto region, and partially San Martin, fall outside the sample of districts used in this analysis. see INEI (2015). [3] All regressions include as controls: altitude, historical coefficient of variation of rainfall, distance to Lima (not included in the sample) and province fixed effects. [4] Column 1 is the natural log of district population (above 16 years old). Columns 2 and 3 are the share of high skilled and low skilled individuals. Columns 4 to 7 refer to employment. Column 4 is the natural log of the number of workers in the district, column 5 is the employment rate, columns 6 and 7 are the number of high and low skilled workers ove the population. [5] Errors clustered at province level, and coefficients that are statistically different from zero are denoted by the following system: *10%, **5% and ***1% 28

29 Table 2: Balance: Pre-Trends in Employment: Tradable Non-Tradable Agriculture N H L N H L N H L (1) (2) (3) (4) (5) (6) (7) (8) (9) P * * (0.05) (0.01) (0.05) (0.10) (0.03) (0.09) (0.21) (0.01) (0.20) R Observations Clusters Notes: [1] Data source: population censuses of 1993 and [2] 1981 census has fewer districts because, according to the public information provided by INEI, it was not possible to recover data from three regions (Apurimac, Loreto and San Martin). Population of these three regions represent 6 % of total population in Moreover Loreto region, and partially San Martin, fall outside the sample of districts used in this analysis. see INEI (2015). [3] All regressions include as controls: altitude, historical coefficient of variation of rainfall, distance to Lima (not included in the sample) and province fixed effects. [4] All dependent variables are the change in the employment rate of the industry, by skill type. In all cases, the denominator is the total population whle the numerator is the number of workers in the corresponding industry and industry-skill pair. N holds for total, H for high skilled workers and L for low skilled workers. [5] Errors clustered at province level, and coefficients that are statistically different from zero are denoted by the following system: *10%, **5% and ***1% Table 3: Change in Employment: Population Employment Ln. N. High Skilled Low Skilled Rate High Skilled Low Skilled (1) (2) (3) (4) (5) (6) P *** -0.16*** 0.41** 0.23*** 0.18 (0.59) (0.04) (0.04) (0.20) (0.09) (0.17) R Observations Clusters Notes: [1] Data source: population censuses of 1993 and [3] All regressions include as controls: altitude, historical coefficient of variation of rainfall, distance to Lima (not included in the sample), the log of district population in 1993 and province fixed effects. [3] Column 1 is the natural log of district population (above 16 years old). Columns 2 and 3 are the share of high skilled and low skilled individuals. Columns 4 to 7 refer to employment. Column 4 is the natural log of the number of workers in the district, column 5 is the employment rate, columns 6 and 7 are the number of high and low skilled workers ove the population. [4] Errors clustered at province level, and coefficients that are statistically different from zero are denoted by the following system: *10%, **5% and ***1% Table 4: Change in Employment: , By Industry and Skill Mining Tradable Non-Tradable Agriculture N H L N H L N H L N H L (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) P ** (0.10) (0.06) (0.06) (0.03) (0.01) (0.04) (0.08) (0.04) (0.05) (0.19) (0.03) (0.18) R Observations Clusters Notes: [1] Data source: population censuses of 1993 and [2] All regressions include as controls: altitude, historical coefficient of variation of rainfall, distance to Lima (not included in the sample), the log of district population in 1993 and province fixed effects. [3] All dependent variables are the change in the employment rate of the industry, by skill type. In all cases, the denominator is the total population whle the numerator is the number of workers in the corresponding industry and industryskill pair. N holds for total, H for high skilled workers and L for low skilled workers. [4] Errors clustered at province level, and coefficients that are statistically different from zero are denoted by the following system: *10%, **5% and ***1% 29

30 Table 5: Change in Migration Rates a. Population N22 N21 N12 N11 Locals Returners New Locals New Comers N H L N H L N H L N H L (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) P *** (0.20) (0.07) (0.19) (0.05) (0.02) (0.04) (0.16) (0.03) (0.14) (0.11) (0.05) (0.08) R Observations Clusters b. Employment P 0.43** 0.19*** ** (0.19) (0.05) (0.16) (0.02) (0.01) (0.02) (0.07) (0.02) (0.07) (0.11) (0.05) (0.07) R Observations Clusters Notes: [1] Data source: population censuses of 1993 and [2] All regressions include as controls: altitude, historical coefficient of variation of rainfall, distance to Lima (not included in the sample), the log of district population in 1993 and province fixed effects. [3] The dependent variables are the current population composition (share) of the district according to migration categories. Locals are those individuals born in the district that were living there five years ago. Returners are individuals born in the district who were not living there five years ago. New Locals are individuals not born in the district but were living there five years ago. New Comers are individuals who were not neither born or residents of the districts five years ago. For each category N holds for total, H for high skilled workers and L for low skilled workers. [4] Errors clustered at province level, and coefficients that are statistically different from zero are denoted by the following system: *10%, **5% and ***1% 30

31 Table 6: Change in Migration - Degree and Industry a. Mining N22 N21 N12 N11 Locals Returners New Locals New Comers N H L N H L N H L N H L (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) P (0.06) (0.04) (0.03) (0.00) (0.00) (0.00) (0.04) (0.01) (0.04) (0.07) (0.03) (0.04) R Observations Clusters b. Tradable P (0.02) (0.00) (0.02) (0.00) (0.00) (0.00) (0.01) (0.00) (0.01) (0.01) (0.00) (0.01) R Observations Clusters c. Non-Tradable P *** ** (0.06) (0.03) (0.04) (0.01) (0.01) (0.01) (0.03) (0.01) (0.02) (0.05) (0.02) (0.03) R Observations Clusters d. Agriculture P ** ** (0.17) (0.03) (0.16) (0.01) (0.00) (0.01) (0.04) (0.01) (0.04) (0.03) (0.01) (0.03) R Observations Clusters Notes: [1] Data source: population censuses of 1993 and [2] All regressions include as controls: altitude, historical coefficient of variation of rainfall, distance to Lima (not included in the sample), the log of district population in 1993 and province fixed effects. [3] The dependent variables are the current population composition (share) of the district according to migration categories, by industry as indicated in panels. Locals are those individuals born in the district that were living there five years ago. Returners are individuals born in the district who were not living there five years ago. New Locals are individuals not born in the district but were living there five years ago. New Comers are individuals who were not neither born or residents of the districts five years ago. For each category N holds for total, H for high skilled workers and L for low skilled workers. [4] Errors clustered at province level, and coefficients that are statistically different from zero are denoted by the following system: *10%, **5% and ***1% 31

32 Table 7: Wages by Skill: a. P Both High Skilled Low Skilled >= Uni. >= Tec. >= Sec. < Uni. < Tec. < Sec. (1) (2) (3) (4) (5) (6) (7) P 1.50*** -1.65** *** 1.78*** 2.05*** (0.47) (0.83) (0.76) (0.53) (0.49) (0.51) (0.59) R Observations Clusters b. PIt PI it ** ** 0.57*** 0.69*** (0.18) (0.38) (0.26) (0.18) (0.19) (0.20) (0.22) R Observations Clusters c. Production Production it *** 0.13** 0.07** 0.13*** 0.10*** 0.11*** 0.08** (0.03) (0.06) (0.03) (0.03) (0.03) (0.03) (0.03) R Observations Clusters Notes: [1] Data source: ENAHO [2] Dependent variable is the log of monthly wage of main and secondary ocupation at 2007 prices. Column 1 considers all individuals. Columns 2, 3 and 4 use data from high skilled workers. High skilled are those individuals with university education (column 2), or at leat technical education (column 3), or at least secondary education (column 4). Columns 5, 6 and 7 refer to low skilled wages as the alternative to columns 1, 2 and 4, respectively. [3] Panel a. replicates the empirical equation used in the district level regression. Panel b. estimates the effect on wages using a yearly commodity price index. Panel c. uses annual mining production. Individual controls included in panels b. and c.: gender (male), age and its square, the number of years of schooling, household size, number of income earners in the household, dummies for water and electricity in the household, fixed effects for industry (two digits ISIC), job type (owner, self-employed, white collar, blue collar), year and district. [4] Errors clustered at province level in panel a. Errors clustered at district level in panels b. and c. Coefficients that are statistically different from zero are denoted by the following system: *10%, **5% and ***1% 32

33 Table 8: Wages by Industry: a. P Mining Tradable Non-Tradable Agriculture (1) (2) (3) (4) P -4.34*** ** (1.51) (1.02) (0.51) (0.75) R Observations Clusters b. PIt PI it ** (0.47) (0.51) (0.15) (0.30) R Observations Clusters c. Xt Production it *** *** 0.13*** (0.11) (0.07) (0.03) (0.05) R Observations Clusters Notes: [1] Data source: ENAHO [2] Dependent variable is the log of monthly wage of main and secondary ocupation at 2007 prices, by indistry type. [3] Panel a. replicates the empirical equation used in the district level regression. Panel b. estimates the effect on wages using a yearly commodity price index. Panel c. uses annual mining production. Individual controls included in panels b. and c.: gender (male), age and its square, the number of years of schooling, household size, number of income earners in the household, dummies for water and electricity in the household, fixed effects for industry (two digits ISIC), job type (owner, self-employed, white collar, blue collar), year and district. [4] Errors clustered at province level in panel a. Errors clustered at district level in panels b. and c. Coefficients that are statistically different from zero are denoted by the following system: *10%, **5% and ***1% Table 9: Production Function - ENAHO Non-Agriculture Agriculture (1) (2) (3) (4) α 0.57*** 0.45*** 0.64*** 0.64*** ( 0.09) ( 0.14) ( 0.03) ( 0.03) λ 0.52*** 0.50*** 0.67*** 0.68*** ( 0.04) ( 0.04) ( 0.08) ( 0.08) ρ 0.54*** 0.37** 0.66*** 0.62*** ( 0.14) ( 0.16) ( 0.20) ( 0.19) σ R Obs Controls No Yes No Yes Notes: [1] Data source: ENAHO for Non-Agriculture; and for Agriculture. [2] Production function in equation (29) estimated with non-linear least squares. [3] Columns 2 and 4 include as controls: time trend and a dummy for the capital Lima [4] Robust standard errors in parenthesis; and coefficients that are statistically different from zero are denoted by the following system: *10%, **5% and ***1% 33

34 7.2 Figures Figure 1: National and Mining GDP (Thousands of 2007 Soles) Notes: [1] Data source: Peruvian Central Bank. [2] Red line marks the beginning of the period of analysis. Figure 2: Mining Production (Thousands of 1994 US Dollars) Notes: [1] Data source for production: InterraRMD (2013). Data source for prices: Kelly and Matos (2013). 34

35 Figure 3: Mining and Non-Mining Districts Notes: [1] Red coluring intensifies with proximity to any of the 27 mines operating in Light red represents the districts not affected by mining activity (within 101km and 200km). [2] Grey area represents districts that are not considered in the analysis. 35

36 Figure 4: Distribution of P i Figure 5: Simulated Labor Response H a This Figure simulates the response of H a to a one unit change in mining productivity, X a (equation 15), for different values of S H and σ. Other parameters set at: α = 0.5, π = 0.3, π = 0.4, φ =

37 Figure 6: Simulated Wage Response: w Ha This Figure simulates the response of w Ha to a one unit change in mining productivity, X a (equation 19), for different values of S H and σ. Other parameters set at: α = 0.5, π = 0.3, π = 0.4, φ = 0.2 Figure 7: Simulated wage differential: w Ha w La This Figure simulates the response of ( w Ha w La ) to a one unit change in mining productivity, X a (equation 21), for different values of S H and σ. Other parameters set at: α = 0.5, π = 0.3, π = 0.4, φ =

38 Figure 8: Proportion of Migrant Individual (y-axis) by Age (x-axis), in 1994 Notes: [1] Data source: 1994 population census. [2] Skilled: individuals with secondary education or higher. [3] Y-axis it the density for the share of individuals who were not living in the district of current residency fiver years ago. X-axis is age. 38

39 Figure 9: Proportion of individuals born and living in the district since five years ago (y-axis) by Age (x-axis), in 1994 Notes: [1] Data source: 1994 population census. [2] Skilled: individuals with secondary education or higher. [3] Y-axis it the density for the share of individuals born and living in the district since five years ago (locals). X-axis is age. 39