The Value of Brazilian Climate: New Evidence from a Spatial Equilibrium Model with Two Sectors Jaqueline Oliveira Clemson University

∗

Paula Pereda

†

University of Sao Paulo

October 30, 2015

Abstract In Brazil, winter temperatures are expected to increase up to 5.0 degrees Celsius and annual rainfall accumulation is expected to fall by approximately 170mm. As changes in the climate unfold, quantifying the effects of these changes on welfare is crucial in order to inform policy. In this paper, we set up a spatial equilibrium model to estimate the willingness to pay for climate amenities in Brazil. To account for the role of climate on agriculture wages, we adapt the framework and model two local productive sectors - agriculture and non-agriculture. Our estimates indicate that, even in milder climate settings like Brazil, individuals are willing to pay large amounts to increase temperatures during colder seasons and decrease temperatures during warmer seasons, although the willingness to pay for climate amenities is much higher among non-agriculture workers. Counterfactual simulations from different climate scenarios reveal sizable positive effects of climate change on equilibrium migration rates and welfare. This positive affect can be attributed to sharp predicted increases in temperature during the winter season, which seem to have offset the disutility from increased summer temperatures for both types of workers. We find non-negligible differences between partial and general equilibrium counterfactuals. In particular, the positive impact of climate change on migration and welfare is slightly smaller after accounting for changes in housing prices that follow from the spatial resorting of workers. This suggests that not accounting for general equilibrium effects might lead one to overstate the welfare effects of changes in climate. Keywords: Climate Change, Amenity Value, Migration, Spatial Equilibrium, Labor Markets, Structural Estimation JEL Classification: O15, Q54, R13, Q51

∗ †

Email: [email protected] Email: [email protected]

1

Introduction

The climate has important implications for economic activity and human well-being. It impacts the consumption of food, clothing, energy and housing, the desirability of different locations for living and recreation, as well as the outcomes of the agricultural markets. In Brazil, winter temperatures are expected to increase up to 5.0 degrees Celsius and annual rainfall accumulation is expected to fall by approximately 170mm (CPTEC/INPE). As changes in the global climate unfold, quantifying the effects of these changes on welfare is crucial in order to inform policy. A growing literature has utilized the framework developed by Roback (1982) to estimate the value individuals place on climate amenities. The intuition is that a person’s location choice reveals preferences for amenities, as amenities are traded for lower wages/higher living costs. Virtually all the studies in this literature focus on measuring willingness to pay for warmer winters, less snowfall, and cleaner air in developed countries (Blomquist et al. (1988), Cragg and Kahn (1997), Maddison and Bigano (2003), Rehdanz and Maddison (2009), Albouy et al. (2013), and Cropper and Sinha (2013)).1 However, the findings drawn from these studies might not be directly applicable to most developing countries. First, a large fraction of the world’s poor live in more tropical zones, so the preferences for climate amenities might be very different. Second, agriculture is the main source of livelihood for most people living in developing countries and the climate is likely to directly impact the incomes of those employed in the sector. In this paper, we seek to contribute to this literature by estimating the willingness to pay for climate amenities and assessing the welfare impacts of climate change in Brazil. As previous studies, our framework is based in the work by Roback (1982). In this framework, workers choose where to live based on the wages, the cost of living, and the location’s amenity value. Individuals also have idiosyncratic preferences for locations.2 We modify the model in two ways. First, workers have to pay a cost to relocate in response to changes in local attributes.3 Second, and more importantly, because the agriculture sector employs 1 Important exceptions are Timmins (2007) and Mueller (2005), who provide estimates using Brazilian data. 2 Discrete-choice models inspired by McFadden (1973) have been extensively used to compute the value of local amenities. See Timmins (2007) for temperature and rainfall, Bayer et al. (2009) for clean air, Klaiber and Phaneuf (2010) for open space, Cropper and Sinha (2013) for temperature, and Diamond (2013) for endogenous amenities that depend on city skill mix. 3 Bayer et al. (2009) have shown that the presence of migration costs biases the estimates of the willingness to pay for amenities when these costs are not accounted for. Morten and Oliveira (2014) show that originspecific migration costs are important determinants of workers’ choice of location in Brazil, and that these costs are impacted by access to road infrastructure. Timmins (2007) uses a cruder measure of moving costs, as measured by distance to one’s birth state, and finds that costly adaptation to climate change is relevant when evaluating the welfare effects of climate change. Cropper and Sinha (2013) also find relevant moving costs for the U.S., although their estimates of willingness to pay for climate are likely biased due to restricting

1

a large fraction of Brazil’s labor force, we allow for climate to impact location choice not only through its consumption amenity value, but also through its effect on productivity and agricultural wages.4 In our spatial equilibrium model, each local labor market is composed of two sectors, agriculture and non-agriculture. We assume that switching costs are too high to allow workers to change sector of employment. The sorting of workers of each sector among different locations determines the local labor supply. Wages are determined by the equilibrium in each local labor market. Additionally, each location is assumed to have a well-developed housing market that can be accessed by agriculture and non-agriculture workers, which face the same price for a housing unit. Workers’ location choice determine the demand for housing in each location which, along with the housing supply curve, sets the equilibrium housing prices. In the spatial equilibrium, workers do not have incentive to move to a different location, firms maximize profits, and both labor and housing markets clear. We estimate the parameters of the spatial equilibrium model using a two-step estimation procedure, following Berry et al. (2004). Data on wages, rents, and migration come from decennial population censuses spanning the 1980-2010 period. The first step is a discrete-choice model (McFadden (1973)) where we estimate separately the location-specific components of the indirect utility and the components attributed to geographic distance and climate similarities between origin and destination locations. These parameters can be separately identified using information on the individual’s location of residence five years prior to the census. Allowing the location choice to depend on climatic similarities helps shed light on whether preferences for climate amenities are path-dependent. In the second step, we use the indirect utilities recovered from the first step, along with data on wages, rents, and climate amenities, to estimate the marginal utilities of climate amenities and the parameters in the equations that characterize the equilibrium in the labor and housing markets. We rely on the exogeneity of local labor demand shocks (Bartik (1991)), and short-term variation in rainfall and temperature, to generate moment conditions and estimate these parameters using a simultaneous GMM estimator. Consistent with previous studies, the first-step estimates reveal that moving is costly. As individuals use migration as an adaptation strategy, the magnitude of these costs should count towards the overall cost of climate change. We also find that non-agriculture workers are more likely to choose a location that is climatically similar to their previous location. While non-agriculture workers seem to value similarities in climate attributes, the same is not true of those in the agriculture sector. From the parameter estimates obtained in the second step, we conclude that workers their sample to migrants. 4 According to our estimates based on data from the 2010 population census, nearly 26% of the country’s labor force was employed in the agriculture sector.

2

from both sectors value local wages more relatively to housing prices, although agriculture workers seem to be more sensitive to changes in the local labor market conditions. Workers from both sectors dislike increases in January temperatures (Brazil’s summer season) and value increases in July temperatures (Brazil’s winter season). Non-agriculture workers place higher value on climate amenities. Back-of-the-envelope calculations indicate that agriculture workers are willing to pay 484 USD/year in order to avoid a one-standard-deviation increase in January temperatures, while non-agriculture workers are wiling to pay 7,150 USD/year for the same change. Both types of workers place higher value on increasing July temperatures relative to decreasing January temperatures by one standard deviation. Workers in the agriculture sector are willing to pay nearly 1,650 USD/year for a one-standarddeviation increase in July temperatures, whereas non-agriculture workers’ willingness to pay is 11,300 USD/year. Finally, agriculture workers value a one-standard-deviation increase in sunshine hours in 250 USD/year. The value for non-agriculture workers is approximately 1,650 USD/year. Finally, counterfactual simulations from different scenarios reveal sizable positive effects of climate change on equilibrium migration rates and welfare. The positive welfare affect can be attributed to sharp predicted increases in temperature during the winter season, which seems to have offset the disutility from increased temperatures in January for both types of workers. The increases in average utility range from 27 to 65% relative to the baseline. For all years except 1980, agriculture workers benefit more from changes in climate than nonagriculture workers. This is possibly due to the beneficial impact of higher temperatures on equilibrium agriculture wages. It is worth noting that there are considerable differences between partial and general equilibrium effects. The impact on migration and welfare is slightly smaller after accounting for changes in housing prices that follow the initial resorting of workers across space. This suggests that not accounting for general equilibrium effects might lead one to overestimate the welfare effects of changes in climate. This paper relates more closely to Timmins (2007), and more generally to the body of research that estimates the value of climate amenities in developed countries (Blomquist et al. (1988), Cragg and Kahn (1997), Maddison and Bigano (2003), Mueller (2005), Rehdanz and Maddison (2009), Albouy et al. (2013), Cropper and Sinha (2013)). Our main departure from previous studies is the effort to account for both the productive and consumption value of climate amenities when computing the welfare effects of climate change. We also exploit information on workers’ previous residence to assess how climate similarities influence location choice. Finally, by modeling the equilibrium in the housing and labor markets, we show that partial equilibrium counterfactual analyses are likely to overstate the welfare effects of climate change. This paper also relates to a growing literature that has estimated spatial equilibrium

3

models to understand the connection between population sorting and a wide variety of economic phenomena.5 It also contributes to a literature that studies the adaptation strategies to climate change in developing countries.6 Many studies in this literature have focused on the role of migration as a strategy to cope with climate change (Findley (1994), Barrios et al. (2006), Salda˜ na Zorrilla and Sandberg (2009), Drabo and Mbaye (2011), Marchiori and Schumacher (2011), Marchiori et al. (2012)). However, the framework of analysis cannot be readily used to measure the welfare effects of changes in climatic conditions. Additionally, most of these studies focus on extreme events such as natural disasters. Although they provide valuable information, their findings might not be sufficient to provide a complete picture of the repercussions of climate change in poor countries. This paper is organized as follows. The next section presents a review of the literature on the effects of climate on production and human behavior. Section 3 lays out the spatial equilibrium model and the estimation strategy we adopt to estimate the structural parameters. Section 4 describes the dataset we utilize in the estimation, as well as the building of historical climate data. Section 5 presents the parameter estimates, followed by section 6, which discusses the effects of climate change on migration and welfare. Section 7 concludes.

2

Background

Before we proceed with our study of the effects of climate change in Brazil, we review the related literature. The first body of research related to our study is the literature that considers how climate amenities influence households’ location choice. Using the framework developed in Roback (1982), this literature employs hedonic wage and housing price models to value climate amenities (Blomquist et al. (1988), Cragg and Kahn (1999)). Under the assumption that moving is costless, the amenity values are capitalized into local wages and housing prices. In the equilibrium, the individuals select a location to live such that the marginal cost of the amenity – in terms of forgone wages and higher rents – equals the value they place on the amenity. The idea is that desirable locations would be characterized by some mix of higher rents and lower wages due to the people demanding the consumption 5

See Bayer et al. (2007) for a framework to estimate preferences for schools in U.S. neighborhoods; Albouy (2009) for a study of the effects of taxation on employment and wages; Diamond (2013) for a model of endogenous production of amenities that explains the college gap in the U.S.; Morten and Oliveira (2014) for the welfare effects of increased labor market connectivity through road networks; Blair (2014) for an application of these models to the study of neighborhood typing points; Mangum (2015) for a dynamic spatial equilibrium model with barriers to labor adjustment to local labor demand shocks. 6 See Seo and Mendelsohn (2008) and Seo et al. (2010) for how farmers in Africa adjust their choices of livestock species in face of changes in climate; Hornbeck (2012) for a study of the economic adjustments in the early 20th century America to a large environmental shock; Taraz (2013) for a study of Indian farmers’ coping strategies using irrigation investments and crop adaptation; da Cunha et al. (2014) for the role of irrigation in allowing Brazilian farms to avoid some of the adverse effects of climate change.

4

amenities. The overwhelming majority of previous studies were conducted in developed country settings. Blomquist et al. (1988) find that humidity might be considered a negative amenity. Sunshine hours are considered an amenity for both the wage and housing markets in urban U.S.. Rainfall and extreme temperature seem to affect only the production of goods. Albouy et al. (2013) also utilize data on households in the U.S. and find that households prefer mild climates (about 65o F) and pay more on the margin to avoid heat than cold. In Germany, Rehdanz and Maddison (2009) find that temperature and precipitation influence the housing markets more than the labor markets. In general, German households appear to prefer warmer winters with less rainfall. The same analysis for Italian households shows that Italians prefer a drier climate during winter months combined with lower summertime temperatures (Maddison and Bigano (2003)). Bayer et al. (2009) criticize the hypothesis that individuals move across locations without paying a cost, arguing that these costs might change the equilibrium conditions as stated by Roback (1982). In line with this argument, Morten and Oliveira (2014) and Timmins (2007) show that migration costs in Brazil are important determinants of individuals’ location choices and adaptation, respectively. An alternative approach that allows for moving costs and climate amenities valuation is to estimate the individuals’ utility function using discrete-choice models (Cragg and Kahn (1997), Timmins (2007), Bayer et al. (2009), Cropper and Sinha (2013)). These models treat each location as a bundle of attributes (McFadden (1973)). These attributes are, mainly, income opportunities, rental prices, and a set of environmental and non-environmental amenities (Cragg and Kahn (1997)). Therefore, the income-amenity tradeoff implicit in the location choice by migrants is used a measure of the willingness to pay for climate. Moreover, this approach also solves the problem of valuing non-marginal changes in climate, as discussed by Timmins (2007). As such models are derived from a random utility model, welfare measures can also be calculated straightforwardly (Mueller (2005)). Many authors applied this approach to estimate the value of climate, mostly in developed countries. Cragg and Kahn (1997) find that individuals in the U.S. are willing to pay positive amounts for moderate climate. The attribute most valued by American households is the temperature in February. More specifically, college graduates aged 30-40 are willing to pay 3,000 USD for a standarddeviation increase in that month’s temperature, while people aged 50-60 are willing to pay 8,800 USD for the same increase. Cropper and Sinha (2013) also applied the discrete location choice model in the U.S metropolitan areas. The authors find that increases in winter temperatures are worth less than the same increase in summer temperatures at median temperatures, but the reverse is true at winter temperatures below 25o F (approximately -4o C). They estimated a welfare loss of 2.7% and 5% of household income from 2020-2050 under a

5

B1 and A2 climate change scenarios, respectively. In Brazil, Timmins (2007) evaluates the welfare effects from changes in Brazilian climate using data from the 1991 population census. He finds positive effects of rainfall during spring (with negative effects after a threshold) and evidence of a positive effect of temperature for some education-level groups. In regard to the climate change welfare losses, Timmins (2007) shows that the northern regions will be more affected. The aggregate annual cost due to climate change is expected to be about 1.6 billion USD. Mueller (2005) also applies a location equilibrium model to measure the marginal values of climate amenities in some selected Brazilian cities. She shows that climate amenities are significant only in the rental equation. One of her explanations relates to the fact that unions, labor laws, and individuals skills play a more significant role on wage variation than climate. In Brazil, due to the size of the agricultural sector in the labor market, we also need to take into account the effect of climate on this economic activity. Many studies investigate the climate impacts on the agricultural outputs by using the Ricardian or hedonic approach (Mendelsohn and Nordhaus (1994), Sanghi et al. (1997))7 or a production function approach (Adams (1989)).8 Mendelsohn and Nordhaus (1994) measure the economic impact of climate on land prices in the U.S.. They find that more rainfall outside the fall season increases farm values, and higher temperatures in all seasons – except for fall – reduce the farm values. When considering the climate change forecasts, they find that global warming might have economic benefits for agriculture in the U.S.. Greenstone and Deschenes (2012) propose a fixed-effects model, exploiting the year-to-year variation in temperature and precipitation, presumably random, to estimate their effect on changes in agricultural profits controlling for regional fixed effects by year. The authors also find that climate change will increase the U.S. agricultural profits in 4%. In Brazil, Sanghi et al. (1997) estimate that for a 2.5o C increase of temperature and a 7% increase in rainfall, the net impact on land values will be negative from 2.1 to 7.4% of mean land values. Most of the studies using Brazilian data indicate that the net impact of climate change on agriculture is negative, although it varies regionally. The northern and central areas of the country seem to be more vulnerable to climate change (Evenson and 7

This approach intends to measure the direct effect of climate on land values. If land markets are perfect, the long-term productivity is reflected in land values and implicitly incorporates producers change of behavior from new climate regimes into the analysis (adaptation measures). The disadvantage of this approach is related to the unobserved factors (e.g. soil quality), that might be confounders to the climate effect. 8 This approach uses farmers’ production structure to measure the optimal allocation of different crops to inputs. It also considers that the local agro-climatic conditions are known by farmers and therefore should not be treated as random, since they influence producers’ choices. As a result, changes in average climatic conditions can modify the behavior of farmers as they take into account local climate patterns (temperature and precipitation) in deciding on the output-input mix. Thus, assuming that farmers observe the past climate conditions, changes to climate conditions influence the farmers productivity (small changes and, mainly, extreme weather events that may occur during growing and harvesting seasons).

6

Alves (1998), Sanghi and Mendelsohn (2008), and F´eres et al. (2008)).

3

Spatial Equilibrium Model with Consumption and Productive Climate Amenities

This section lays out the spatial equilibrium model we utilize in order to quantify workers’ willingness to pay for climate amenities. The basic framework is based on work by Roback (1982). Each location provides individuals with a bundle of wages, housing cost, and climate and non-climate amenities. The utility agents derive from a given location is also affected by idiosyncratic preference shocks, which we assume are i.i.d across individuals and location. We also allow for the utility to live in a given location to depend on the agents’ initial location. More specifically, the location choice can be impacted by the geographic distance and the “climatic” distance between origin and destination location. Individuals’ location choice determines the amount of labor that is supplied to each location at a given period. Amenities are treated as exogenous location-specific traits, whereas wages and rents are endogenously determined by the equilibrium in the labor and housing markets. In equilibrium, the marginal mover is indifferent between staying in his or her current location or moving to another location, and both labor and housing markets are cleared. Because the focus of our empirical exercise is to measure the welfare effects of predicted changes in the Brazilian climate, we extend Roback’s framework and allow for climate to impact the spatial distribution of workers through its effect on agriculture productivity. Next, we present our framework in more detail.

3.1

The Setup

Labor Demand Consider a country with K regions, each indexed by k = 1, 2, ..., K. In each period t, a region produces a manufactured good, Yktm , and an agricultural good, Yktf , according to the following production functions: m γm Yktm = Am kt (Lkt )

(1) f

Yktf = Afkt (Lfkt )γf (H k )1−γf ,

(2)

where Askt is the location-specific total factor productivity in sector s = m, f , and Lskt is the amount of labor in sector s. We assume that production of manufactured goods depends solely on labor, while the production of agricultural goods is a function of a location’s enf dowment of productive land, H k . We make the following assumption about the productivity

7

parameters: m Am kt = exp(kt )

(3)

c

Afkt = (Ckt )α exp(fkt ),

(4)

where Ckt is a vector of local climate factors that affect agricultural productivity in k in time t (more specifically, temperature and precipitation). We assume that the climate does not affect productivity in the manufacturing sector. Productivity in each sector is also affected by other types of idiosyncratic productivity shocks, skt . At the firm’s optimum, labor is paid its marginal product, leading to the following labor demand equations for each sector: m m log wkt = log γm + (γm − 1) log Lm kt + kt

(5) f

f log wkt = log γf + (γf − 1) log Lfkt + (1 − γf ) log H k + αc log Ckt + fkt

(6)

Housing Supply The price of housing units is set by the equilibrium in the local housing market. Developers produce housing units using construction materials and land as inputs. If we assume they are price takers and housing units are homogeneous, the price of housing units will be set equal to the marginal cost: phkt = mchkt . In the asset market, prices are set equal to the present value of rents, so that: rkt = it mchkt , where it is the interest rate. We assume further that housing is owned by absentee landlords who rent the houses to local residents. Developers’ marginal costs depend on the cost of land, which in turn is a function of the population size.9 A housing supply curve can be derived for each region in a given time period as follows: f r log rkt = log it + ρ log(Lm kt + Lkt ) + kt ,

(7)

where ρ is the elasticity of housing price to the size of population, and rkt is the idiosyncratic component of housing prices. 9

Furthermore, the elasticity of housing prices to changes in population depends on the extent of the land that is proper for housing development. Saiz (2010) consider geographic characteristics that make the land in a location undevelopable.

8

Labor Supply Workers consume manufactured and agricultural goods (which are nationally produced) and housing (which is locally produced). They gain utility from local amenities, represented by a vector Zikt . Workers have Cobb-Douglas preferences over manufactured goods, agricultural goods, and housing, which they maximize subject to a budget constraint: max

m ,Y f ,R Ykt kt kt

s.t.

f

m

(Yktm )λ (Yktf )λ (Rkt )1−λ

m −λf

exp z(Zikt ) (8)

m pm t Ykt

+

pft Yktf

+ rkt Rkt ≤

s wkt

This optimization problem leads to the indirect utility for worker i in sector s from living in k: s s wkt w s λf m s m f wkt 1−λm −λf − λ ) )λ (λf kt ) ((1 − λ ) exp z(Zikt ) (9) U ikt = (λm m pt rkt pft Taking the natural logarithm from both sides of equation (9) yields:10 s s Vikt = λv + λvt + (1 − λr ) log wkt − λr log rkt + z(Zikt ),

(10)

λv = λm log λm + λf log λf + (1 − λm − λf )(log(1 − λm − λf ))

(11)

λvt = −λf log pft − λm log pm t

(12)

λr = 1 − λm − λf

(13)

where

The utility value of local amenities is represented by the function z(.). We extend this function to allow for the amenity value of a given location to depend on the migrant’s origin, indexed by j: j z(Zikt ) = λc Ckt + λnc N Ckt + λj Xkj + ikt , (14) where Ckt are climate amenities (average temperature during January and July, and sunshine hours), N Ckt are non-climate amenities (which by assumption grow over time at a constant rate), 11 and Xkj are attributes of location k that vary depending on migrant’s origin but are time invariant. The first two components of this vector are moving costs. We assume that these costs are comprised of a fixed component, f ixed cost, and a component that 10

To simplify the notation, we suppress the super-script s from the parameters in the utility function. However, in our empirical analysis, we allow these parameters to be different for agriculture and nonagriculture workers. 11 This assumption is not crucial. Even if non-climate amenities grow at different rates for different locations, the requirement is that these amenities are not correlated with temperature and sunshine hours in the current period.

9

depends on the euclidean distance between origin-destination pairs, distancejk .12 The other components are meant to capture the similarities between origin and destination. These include an indicator for whether the two locations are adjacent, neighborsjk , whether they belong to the same state, samestatejk , and an indicator for whether origin and destination are located in the same biome, samebiomejk .13 Finally, ikt is an idiosyncratic taste for location k, which we assume ikt ∼ Type I Extreme Value. It follows from the assumption on idiosyncratic taste shocks that the probability that worker i in sector s currently living in j will choose to live in location k is: s s s πijkt = P (Vijkt ≥ Vijlt ∀l 6= k)

= PK

(15)

s − λr log rkt + λc Ckt + λnc N Ckt + λj Xkj ) exp(λv + λvt + (1 − λr ) log wkt

m=1

j s exp(λv + λvt + (1 − λr ) log wmt − λr log rmt + λc Cmt + λnc N Cmt + λj Xm ) (16)

Given the initial distribution of workers from each sector across space (Lsj,t−1 , j = 1, 2, ..., K), the total supply of workers from each sector to location k in time t is: Lskt =

PK j∈K

!

s − λr log rkt + λc Ckt + λnc N Ckt + λj Xkj ) exp(λv + λvt + (1 − λr ) log wkt

X

m=1

exp(λv

+

λvt

+ (1 −

s λr ) log wmt

−

λr

log rmt +

λc C

mt

+

λnc N C

mt

+

j λj Xm )

(17) s , r1 , r2 , ..., rK ) The equilibrium is characterized by a vector of local prices (w1s , w2s , ..., wK and labor allocation (Ls1 , Ls2 , ..., LsK ), s = f, m, such that:

1. Workers choose the location that maximizes their utility, satisfying equation (15); 2. Firms maximize profits according to labor demand equations (5) and (6); 3. Housing markets clear according to equation (7).

3.2

Estimation

Our estimation strategy follows a well-established literature which focuses in estimating demand for differentiated products. In particular, we adopt the procedure developed by Berry et al. (2004) and applied in several other studies that estimate spatial equilibrium 12 13

See Morten and Oliveira (2014) for the role of migration costs on labor markets and welfare. Biomes are large areas characterized by their climate and dominant vegetation.

10

Lsj,t−1

models (Timmins (2007), Bayer et al. (2009), Cropper and Sinha (2013), Diamond (2013), Morten and Oliveira (2014)). Denote the location-specific component of the indirect utility for a worker in sector s by: s s δkt = (1 − λr ) log wkt − λr log rkt + λc Ckt + λnc N Ckt

The log-likelihood function is as follows: s , λv , λvt ) L(δkt

=

n X i=1

log

!

s exp(λv + λvt + δkt + λj Xkj )

PK

m=1

j s exp(λv + λvt + δmt + λj Xm )

× ((ji = j, ki = k))

(18)

The model in equation (18) is a conditional logit model, which we estimate separately for each census year. Therefore, we allow for the coefficients on the components of origin-specific variables (λj ) to vary over time and across workers of different sectors. The component of s workers’ utility that depends on climate amenities will be included in δkt . In the second step, we estimate the parameters of the labor demand equations, the housing supply curve, and the marginal utilities of wages and climate amenities by simultaneous GMM. The moments conditions used in the estimation are based on measures of local labor demand shocks, known in the literature as Bartik shocks (Bartik (1991)). We take advantage of the availability of multiple censuses and use the variables in first differences. This procedure has the advantage of allowing us to purge the potential correlation between the error term in the structural equations and unobserved time-invariant location-specific attributes. Recall the equations for productivity in each sector: m Am kt = exp(kt )

(19)

αc

Afkt = (Ckt ) exp(fkt ).

(20)

We redefine the change in (log) productivity shocks as: b,m ∆ log Am ∆Bkt + ∆m kt = α kt

(21)

∆ log Afkt = αc ∆Ckt + αb,f ∆Bkt + ∆fkt ,

(22)

where ∆Bkt are the Bartik shocks, constructed as follows: ∆Bkt =

X

(wind,−k,t − wind,−k,t−10 )

ind

Lind,k,t0 . Lk,t0

The subscript ind indexes the industry, t0 is the baseline census year (in our case, 1980), wind,−k,t is the average national (log) wage in industry ind and period t, excluding location 11

k from the average, Lind,k,t0 is total employment in industry ind, in location k, and time t0 , and Lk,t0 is total employment in location k and time t0 . Denote the vector of exogenous variables by: ∆Ikt = {∆Bkt , ∆Ckt } . Our estimating equations are as follows: m b,m ∆ log wkt = (γm − 1)∆ log Lm ∆Bkt + ∆m kt + α kt

(23)

f ∆ log wkt = (γf − 1)∆ log Lfkt + αc ∆Ckt + αb,f ∆Bkt + ∆fkt

(24)

f r ∆ log rkt = ρ∆ log(Lm kt + Lkt ) + ∆kt

(25)

m m = (1 − λr )∆ log wkt − λr ∆ log rkt + λc ∆Ckt + λnc ∆N Ckt + ∆vktm ∆δkt v

f f ∆δkt = (1 − λr )∆ log wkt − λr ∆ log rkt + λc ∆Ckt + λnc ∆N Ckt + ∆ktf ,

(26) (27)

and the exclusion restrictions are: E (∆m kt |∆Ikt ) = 0   f E ∆kt |∆Ikt = 0

(29)

E (∆rkt |∆Ikt ) = 0

(30)

E (∆vktm |∆Ikt ) = 0
v E ∆ktf |∆Ikt = 0.

(31)

(28)

(32)

In practice it is not possible to observe all the non-climate amenities provided by a given location. However, since our focus in on climate amenities, we circumvent this lack of data by estimating the value of non-climate amenities as the constant term in equations (26) and (27). Additionally, to make the model more tractable, we assume that that γm and γf are equal to one. This implies that the only source of general equilibrium effects is the adjustment of rents in the local housing markets.

4 4.1

Data Census data

The main source of information on workers’ geographic location, wages, and rents is the micro-level data from Brazil’s decennial population censuses spanning the 1980-2010 period. The censuses are the largest available datasets which provide information on individuals and households that are representative at a fine geographic level, the municipalities. For the 12

purpose of estimating the spatial equilibrium model, we aggregate the municipalities into more homogenous areas, known as meso-regions. The classification of Brazil’s municipalities into meso-regions was developed by the Brazilian Institute of Geography and Statistics (IBGE) and was based on the municipalities’ shared cultural history, natural resources, and connectivity. To compute the meso-region-level variables, we select all males aged 20 to 65, who reported positive earnings in the week prior to the census. The wage rate for each individual was computed using information on earnings from the main occupation, divided by the total hours worked in that occupation. All nominal variables were deflated and differences in currency were accounted for, so that wages and rents are converted to 2010 BRL (Brazilian Reais). The censuses also collect information on rental prices for those households who rented their homes, which we use as measure of housing cost. Finally, due to the availability of data on individuals’ geographic location 5 years prior to the census, we are able to include characteristics of origin location (geographic and climatic distance) as additional determinants of location choice. Information on individuals’ sector of employment allows us to construct all variables for agriculture and non-agriculture sector separately. Table 1 presents summary statistics for the individual-level samples. The share of the labor force in agriculture has been declining over the years, dropping from 33% in 1980 to 26% in 2010. Migration is less prevalent among workers in the agriculture sector. While 4.3% of those in the agriculture sector had migrated in 5-year period preceding the 2010 census, 6.6% of those in non-agriculture sectors had moved in the past 5 years. Wages are more than two orders of magnitude larger in the non-agriculture sector compared to agriculture. Rental prices are on average 65 BRL. It is worth mentioning that a small fraction of the population reported paying rents in each census year.

4.2

Historical climate data

Historical climate data are available by weather station. These data were collected by the Brazilian Institute of Meteorology (Portuguese acronym: INMET) and consist of daily average temperature (in degrees Celsius), accumulated rainfall (in millimeters/day), and sunshine hours (insolation hours/day). Brazil has a network of weather stations that covers much of the coast, but has a low density in the country’s interior, especially in the North and Midwest regions. The daily station data were converted to meso-region-level data using a kriging technique of interpolation (Haas (1990)). This method assumes that each geographical coordinate is a realization of a spatial process. It allows the interpolation of data with flexibility to specify the covariance between the outputs.14 14

The interpolation technique might cause spatial autocorrelation among the errors in the regressions. However, this issue is corrected by using robust estimators or bootstrapped standard errors.

13

Table 1 describes the climate variables pertaining to the 135 meso-regions used in the estimation. These climate variables correspond to the average over each census year. Average temperature in January is in the 24-25 degrees Celsius. Winters in Brazil are mild, with July temperatures averaging 20 degree Celsius. Out of the total 8,765 hours within a given year, Brazilians enjoyed on average 2,400 hours of sunshine in 2010. Accumulated rainfall varies considerably over the years, and also across seasons within a given year. The first semester is wetter compared to the second semester. The last panel in Table 1 shows descriptive statistics for the decennial differences in climate variables. This is important to the extent that identification of the marginal utilities of climate variables relies on temporal variation in temperature and rainfall. On average, temperatures in January increased 0.5 degrees Celsius between 1991 and 2000, and 0.61 degrees between 2000 and 2010. The standard deviations associated with these differences are also large. As expected, there is a great deal of variation in accumulated rainfall across census years. Table 2 reports climate data by macro-region (North, Northeast, Southeast, South, and Midwest). The data correspond to the historical average covering the 1961-2010 period. Average temperatures are typically very high, especially in the northern regions. The south of Brazil has lower temperatures during winter (June to September). There is little spatial variation in average temperatures. The northern region, which includes the Amazon Forest, is very rainy, reaching accumulated rainfall levels of more than 2,000 mm per year. The rainy season also lasts longer in this region, contrasting with the climate of the neighboring region, the Northeast, which accounts for the highest temperatures and driest seasons in the country. The semi-desert vegetation is explained by the dry climate pattern of the region. For sunshine hours, measured by the hours of insolation during the year, the Northeast region has the higher hours of sunshine of all the regions. However, all the Brazilian regions present high levels of sunshine hours during the year. Figures (1) and (2) illustrate further the spatial variation in average temperatures and rainfall. The more economically developed areas of the country (South and Southeast) have milder average temperatures, whereas the poorest regions in the country (North, Northeast, and Midwest) are hotter on average. This correlation between temperature and economic development suggests that estimating the model in first differences may be warranted to the extent that location-specific unobservable traits would be controlled for. Finally, figure (3) displays the Brazilian biomes. There are six types of biome in Brazil, which are defined by the set of fauna, flora, and soil on a regional scale with similar geoclimatic conditions. The largest biome is the Amazon (Amazonia), which covers 49.3% of the Brazilian area. Most of the Brazilian wildlife is found in this biome. The Amazon is defined by the climatic region (warm and wet during almost all the year), forest physiognomy and geographic location. The second largest biome is the Cerrado or Savannah (23.9%), which

14

is a flat grassland of tropical or subtropical regions. The Caatinga is a composition of stunted trees and thorny bushes, found in areas of little rainfall and high temperatures in the northeastern parts of Brazil (9.9%). The Atlantic Forest (Mata Atlantica) occupies the entire Brazilian continental east Atlantic coast and inland of southern and southeastern Brazil (13%). The mountainous region is characterized by a vegetation similar to tropical forests. This biome is one of the most biodiverse areas in the world. The Pampa is found in open areas covered with grass - mainly located in the southern Brazil (2.1%). The climate is subtropical with warm temperatures and rains almost year round. The soil is good for agriculture and cattle raising. Finally, the Pantanal biome is the smaller biome of the country (1.8%), composed by swampy areas enclosed by mountains where water runs down and causes swampy ground. We use the information of Brazilian biomes to classify pairs of origin-destination locations in terms of their climatic similarities and estimate the role of these characteristics on location choice. It is worth mentioning that we use the average climate information of the meso-regions for the census years (or previous year) to measure the climate individuals experience. Thus, we are unable to capture climate amenities actually experienced by the households, so the climate variables could be measured with error.

4.3

Background: Brazil’s agriculture

The Brazilian crop and livestock production is distributed between two regions: the temperate region, covering the South and Southeast regions, and the tropical Midwest region. There are different farming and livestock patterns within each of these regions. Table 3 describes the sowing and harvesting seasons per crop and macro-region, as well as the most suitable climate conditions for each crop (temperature and rainfall). It is worth mentioning that in most parts of the country, the sowing and growing of the main crops occur during the second semester. The agricultural year in Brazil normally begins at the second semester and ends in May of the following year. This is important for modeling the climate effects on agricultural productivity. The table also shows that for most part of the crops, low average temperatures decrease the crops growth rate. For instance, soybeans, which accounted for 29% of the Brazilian agricultural production in 2013, do not grow properly under temperatures lower than 10o C.15 The threshold for maize (another important annual crop in the country) is 15.5o C. Higher temperatures can also be harmful for many other crops such as soybeans, beans, rice, grapes. Regarding rainfall, for the majority of crops, lack of rain is harmful during the growing season. Regular rainfalls ranging from 2 to 5 mm per day are ideal conditions for most crops, such as maize, soybeans, some fruits, manioc, 15

IBGE - Municipal Agricultural Production (annual + permanent crops).

15

cotton, coffee, tobacco.

5 5.1

Parameter Estimates First step: migration choice

The first set of results is estimates of the structural parameters associated with geographic and climatic distance in the indirect utility. We use the euclidean distance between workers’ origin and the destinations as a proxy for moving costs associated with each choice (more specifically the euclidean distance – in kilometers – between the meso-region centroids). Additionally, we include indicators for whether a pair of regions shares a border, as well as dummies for being in the same state. Finally, we allow for preferences for living in the same climatic zone by including a dummy for whether the origin and destination are in the same biome. Tables 4 and 5 present the first-step estimates from a conditional logit model. These estimates were obtained separately by sector. The average migration rate and distance travelled are considerably larger among those in non-agriculture sectors.16 The estimated parameters suggest there are substantial costs of moving across locations.17 First, migration entails large fixed utility costs. These costs alone are important when assessing the welfare effects of climate change, as it relates to people’s ability to adapt to changes in climate by moving to more preferred locations. Additionally, moving costs also depend on geographic location. Workers in both sectors are less likely to migrate to a destination that is further away from the origin. Controlling for distance, individuals employed in both sectors prefer to move to neighboring meso-regions. There is also a strong preference for moving to a meso-region which is located in the same state as the origin. These additional estimates also have implications for measuring the costs of climate change since climate is highly spatially correlated. Individuals located in places where the climate is supposed to change the most would likely have to move longer distances in order to be able to enjoy more desirable climate. Finally, we find that being in the same climatic zone as the origin makes a meso-region more likely to be chosen as a new location, but only for those in the non-agriculture sector. Figures 4 and 5 plot meso-region actual out-migration rates against those predicted by the conditional logit using the 1991 and 2010 censuses. The graphs include a 45 degree line and the data points are separated according to the macro-region to which the meso-regions 16

Notice that these migration rates are slightly different from the rates presented in Table 1. This is so because, in the latter, the rate is computed using the current population in the mesa-region as reference, whereas the migration rates computed in Table 4 and 5 use the previous population (5 years ago) as reference. 17 Other studies have modeled moving costs when estimating the value of climate (Timmins (2007), Cropper and Sinha (2013)) and air pollution (Bayer et al. (2009)). For a more comprehensive study of the role of transport costs on migration and welfare, see Morten and Oliveira (2014).

16

belong. The closer the data points are to the line, the better the model fit. The R-squared ranges from 0.35 to 0.56, indicating that the model performs reasonably well. However, a careful inspection of Figure 5 shows that out-migration rates are somewhat overestimated for meso-regions in the southeast of Brazil.

5.2

Second step: the value of climate amenities

This subsection presents the final step in estimating the parameters of the spatial equilibrium model. We use the location-specific components of indirect utilities (estimated in the first step), data on wages, rents, employment, and climate amenities to estimate the remaining structural parameters in equations (23)-(27). The estimates are displayed in Table 6. The first panel shows estimates associated with productivity in the agriculture sector. The elasticity of agricultural wages to rainfall is positive and statistically significant, meaning that increases in rainfall affect positively wages in agriculture through its effect on productivity. The elasticity of agricultural wages to average temperature is also positive and significant, corroborating the notion that Brazilian crops are more suited to warmer temperatures and that lower temperatures during the growing season adversely affect agricultural productivity for many crops. Both weather variables were included in the first lag, as the agricultural year in Brazil begins in June of the previous year. The local labor demand shifters, known as Bartik shocks, are also good predictors of wages in both sectors. These shifters are meant to capture unexpected but permanent changes in total factor productivity. Finally, the last two panels display the elasticities of utility to real wages and climate. Workers from both sectors value local wages more relatively to housing prices, although agriculture workers seem to be more sensitive to changes in the local labor market conditions, as seen by the larger coefficient on real wages in the indirect utility function. Workers from both sectors dislike increases in temperature in January (Brazil’s summer season) and like increases in temperature in July (Brazil’s winter season). Interestingly, our results are consistent with the findings in the study conducted by Cropper and Sinha (2013) for the U.S., where temperatures are much colder during winters. It also emerges from our estimates that non-agriculture workers place higher value on climate amenities compared to other workers. In order to compute the willingness to pay for climate, we totally differentiate the indirect utility with respect to wages and the amenity, while holding all other factors constant, yielding:    w w α × × ∆C, ∆W = αc C where αw is the estimate of the marginal utility to (log) wages relative to rents, and αc is the marginal utility of (log) climate. We use the average hourly wage in each sector, 17

as well as the mean values of the climate amenities, measured in 2010. We consider a onestandard-deviation change in the climate variables, which corresponds to a 1.5-degree Celsius decrease in January temperature, a 4.5-degree Celsius increase in July temperature, and a 350-hour increase in total hours of sunshine. These back-of-the-envelope calculations indicate that agriculture workers are willing to pay 484 USD/year in order to avoid a one-standarddeviation increase in January temperatures, while non-agriculture workers are wiling to pay 7,150 USD/year for the same change.18 Both types of workers place higher value on increasing July temperatures relative to decreasing January temperatures by one standard deviation. Workers in the agriculture sector are willing to pay nearly 1,650 USD/year for a one-standard-deviation increase in July temperatures, whereas non-agriculture workers’ willingness to pay is 11,300 USD/year. Finally, agriculture workers value a one-standarddeviation increase in sunshine hours in 250 USD/year. The value for non-agriculture workers is approximately 1,650 USD/year.

6

Policy Evaluation: The Effect of Climate Change on Welfare

Next, we consider the effects of predicted changes in climate on workers’ well-being while accounting for the general equilibrium effects generated by population resorting. We begin by describing our main source of information on climate change forecast for Brazil. Then, we present the counterfactual migration rates and indirect utilities under each of the climate change scenarios.

6.1

Climate Data Forecasts

The climate change forecast for the Brazilian territory is produced by CPTEC/INPE, the Department for Weather Forecasting and Climate Studies affiliated with the National Institute for Space Research. Their forecasts cover all South American countries and are based on the boundaries of the global model (HadCM3), which was developed by the Met Office Hadley Centre in the UK. As the model is in a regional scale, it allows for a better understanding of the climate impacts under global warming conditions (Chou et al. (2012) and Marengo et al. (2012)). CPTEC/INPE considered the A1B emission scenario, which is a family of the A1 scenario from Intergovernamental Panel on Climate Change (2014). The A1 emissions scenario assume the following hypotheses for the future socioeconomic conditions: rapid economic growth, global population that peaks in mid-century and declines thereafter, and rapid introduction of more efficient technologies. In addition to those, the 18

In 2010, 1 BRL = 0.55 USD.

18

A1B scenario assumes a balance use of fossil and non-fossil energy sources in the future (the underlying assumption is that the new efficient technologies apply to all energy supply). (Intergovernamental Panel on Climate Change (2014)). The CPTEC/INPE selected the A1B scenario as this is the only one available at HadCM3 model (Marengo et al. (2012)). In order to improve the sensitivity analysis, they developed a set of forecasts from different sensitivity models in order to account for climate uncertainty, as well as simulate closely the Brazilian current climate. Therefore, four sensitivity models were used: High sensitivity (high), Average sensitivity (midi), Low sensitivity (low), and a not disturbed model, called control model (ctrl).19 Table 7 reports the average climate variables predicted by the CPTEC/INPE models for the 2040-2070 period, using 2010 climate average as reference. All the forecasts predict a high increase (2.4 to 5.0o C) in July temperatures (Brazilian winter). Except for the low sensitivity scenario, the temperatures of January are also expected to increase (1.8 to 3o C). For rainfall, all the sensitivity models (except for the low model) predict considerable decreases in average rainfall accumulation (about 170mm per year). We consider all these scenarios when computing the welfare effects of changes in Brazilian climate in the next subsection.

6.2

Counterfactual Analysis

Tables 8 and 9 display the partial and general equilibrium counterfactual migration rates and utilities under four climate change scenarios. The partial equilibrium counterfactuals are computed assuming that the resorting of workers in response to climate change does not affect local housing costs. As such, counterfactual migration rates are expected to be higher. General equilibrium counterfactuals, on the other hand, account for adjustments in the housing markets in response to migration inflows (outflows). Regardless of the scenario, we estimate large and positive effects on migration in both sectors, indicating that individuals seek migration as a way to adapt to changes in climate. The increase in migration rate ranges from 33 to 60% of the baseline migration rates, depending on the year of the estimating sample. As expected, general equilibrium migration rates are smaller, suggesting that increases in the cost of living prevent further migration from taking place. 19

According to Marengo et al. (2012), the sensitivity scenarios consist of perturbations on the physical parametrization schemes of the HadCM3 model. By accounting for such perturbations on the model’s boundary conditions, it is possible to replicate some of the IPCC SRES (Special Report on Emissions Scenarios). The high sensitivity scenario corresponds to the A1FI emissions scenario (main hypotheses: rapid growth and fossil intensive use of energy sources), and the low sensitivity scenario approximates to the B1 emissions scenario (main hypotheses: same global population that peaks in mid-century and declines thereafter, change towards a service and information economy, with clean and efficient technologies). The medium sensitivity is an intermediate scenario and the undisturbed model corresponds to the A1B scenario.

19

Our analysis indicates that there are positive welfare effects from changes in Brazilian climate. This positive affect can be attributed to sharp predicted increases in temperature during the winter season, which seems to have offset the disutility from increased temperatures in January for both types of workers. The increases in average utility range from 27 to 65% relative to the baseline. For all years except 1980, agriculture workers benefit more from changes in climate than non-agriculture workers. This is possibly due to the beneficial impact of higher temperatures on agriculture productivity which raises equilibrium agriculture wages (see estimates in Table 6). General equilibrium counterfactual welfare changes are usually lower compared to partial equilibrium changes, reflecting the loss of utility due to increased housing prices. These differences suggest that not accounting for general equilibrium effects could lead to overstating the welfare effects of climate change.

7

Final Considerations

As debates over the magnitudes and economic impacts of climate change gain significance in the political arena, more research aiming at quantifying these impacts is warranted in order to inform public policy. While there has been a growing number of studies on how individuals in developed countries value the climate around them and how predicted changes in climate are expected to affect well-being, less is known about these matters in poor country settings. Our study seeks to fill this gap in the current literature. To estimate willingness to pay for climate amenities in Brazil, we set up a spatial equilibrium model that accounts for both the consumption and productive amenity values of climate. We find that non-agriculture workers place higher value in increasing winter temperatures and decreasing summer temperatures relative to agriculture workers, but both types of workers benefit from predicted climate changes. This is so because climate change scenarios predict large increases in winter temperatures in Brazil, offsetting the negative impact of increases in summer temperature. Additionally, agriculture workers are expected to benefit more from climate change than non-agriculture workers due to predicted increases in agriculture productivity due to warming. A large body of research has been focused at measuring the impacts of climate change on agriculture profits, land value, and forced migration in poor countries in face of extreme weather events and natural disasters. While these studies are undoubtedly valuable, not considering the potential positive effects on welfare due to the value individuals place on certain climate amenities might paint an overly pessimistic picture when it comes to predicting climate change impact in the developing world

20

References Adams, R. M., 1989. Global Climate Change and Agriculture An Economic Perspective.pdf. American Agricultural Economics Association 71 (5), 1272–79. Albouy, D., 2009. The Unequal Geographic Burden of Federal Taxation. Journal of Political Economy 117 (4), 635–667. Albouy, D., Graf, W., Kellogg, R., Hendrik, W., 2013. Climate Amenities, Climate Change, and American Quality of Life. NBER working paper #18925. Barrios, S., Bertinelli, L., Strobl, E., 2006. Climatic change and rural-urban migration: The case of sub-Saharan Africa. Journal of Urban Economics 60 (3), 357–371. Bartik, T. J., 1991. Who Benefits from State and Local Economic Development Policies? Books from Upjohn Press. Bayer, P., Ferreira, F., McMillan, R., 2007. A Unified Framework for Measuring Preferences for Schools and Neighborhoods. Journal of Political Economy 115 (4), 588–638. Bayer, P., Keohane, N., Timmins, C., 2009. Migration and hedonic valuation: The case of air quality. Journal of Environmental Economics and Management 58 (1), 1–14. Berry, S., Levinsohn, J., Pakes, A., 2004. Differentiated Products Demand Systems from a Combination of Micro and Macro Data: The New Car Market. Journal of Political Economy 112 (1), 68–105. Blair, P. Q., 2014. The Effect of Outside Options on Neighborhood Tipping Points. Ph.D. thesis, University of Pennsylvania. Blomquist, G., Berger, M., Hoehn, J., 1988. New estimates of quality of life in urban areas. The American Economic Review 78 (1), 89–107. Chou, S. C., Marengo, J. a., Lyra, A. a., Sueiro, G., Pesquero, J. F., Alves, L. M., Kay, G., Betts, R., Chagas, D. J., Gomes, J. L., Bustamante, J. F., Tavares, P., 2012. Downscaling of South America present climate driven by 4-member HadCM3 runs. Climate Dynamics 38 (3-4), 635–653. Cragg, M., Kahn, M., 1997. New Estimates of Climate Demand: Evidence from Location Choice. Journal of Urban Economics 42 (2), 261–284. Cragg, M. I., Kahn, M. E., 1999. Climate consumption and climate pricing from 1940 to 1990. Regional Science and Urban Economics 29 (4), 519–539. Cropper, M. L., Sinha, P., 2013. The Value of Climate Amenities: Evidence from U.S. Migration Decisions. NBER working paper #18756. da Cunha, D. A., Coelho, A. B., F´eres, J. G., 2014. Irrigation as an adaptive strategy to climate change: an economic perspective on Brazilian agriculture. Environment and Development Economics 20 (01), 57–79. 21

Diamond, R., 2013. The Determinants and Welfare Implications of US Workers Diverging Location Choices by Skill: 1980-2000. Manuscript, 1980–2000. Drabo, A., Mbaye, L. M., 2011. Climate Change , Natural Disasters and Migration : An Empirical Analysis in Developing Countries. IZA Discussion Papers (5927). Evenson, R. E., Alves, D. C. O., 1998. Technology, climate change, productivity and land use in Brazilian agriculture. Planejamento e Pol´ıticas P´ ublicas, 223–260. F´eres, J., Reis, E., Speranza, J., 2008. Assessing the Impact of Climate Change on the Brazilian Agricultural Sector. Findley, S. E., 1994. Does drought increase migration? A study of migration from rural Mali during the 1983-1985 drought. The International Migration Review 28 (3), 539–553. Greenstone, M., Deschenes, O., 2012. The Economic Impacts of Climate Change: Evidence from in Weather Output and Random Fluctuations Agricultural. American Economic Review 97, 354–385. Haas, T. C., 1990. Kriging and automated variogram modeling within a moving window. Atmospheric Environment. Part A. General Topics 24 (7), 1759–1769. Hornbeck, R., 2012. The enduring impact of the American Dust Bowl: Short- and long-run adjustments to environmental catastrophe. American Economic Review 102 (4), 1477– 1507. Intergovernamental Panel on Climate Change, I., 2014. Climate Change 2014: Migration of Climate Change. Contribution of Working Group III to the Fiflth Assessment Report of the Intergovernmental Panel on Climate Change [Edenhofer, O., R. Pichs-Madruga, Y. Sokona, E. Farahani, S. Kadner, K. Seyboth, A. Adler]. Tech. rep., Cambridge University Press, Cambridge, UK and New York, NY, USA. Klaiber, H. A., Phaneuf, D. J., 2010. Valuing open space in a residential sorting model of the Twin Cities. Journal of Environmental Economics and Management 60 (2), 57–77. Maddison, D., Bigano, A., 2003. The Amenity Value of the Italian Climate. Journal of Environmental Economics and Management 45 (2), 319–332. Mangum, K., 2015. Cities and Labor Market Dynamics. Ph.D. thesis, Duke University. Marchiori, L., Maystadt, J. F., Schumacher, I., 2012. The impact of weather anomalies on migration in sub-Saharan Africa. Journal of Environmental Economics and Management 63 (3), 355–374. Marchiori, L., Schumacher, I., 2011. When nature rebels: International migration, climate change, and inequality. Journal of Population Economics 24 (2), 569–600.

22

Marengo, J. a., Chou, S. C., Kay, G., Alves, L. M., Pesquero, J. F., Soares, W. R., Santos, D. C., Lyra, a. a., Sueiro, G., Betts, R., Chagas, D. J., Gomes, J. L., Bustamante, J. F., Tavares, P., 2012. Development of regional future climate change scenarios in South America using the Eta CPTEC/HadCM3 climate change projections: Climatology and regional analyses for the Amazon, Sao Francisco and the Parana River basins. Climate Dynamics 38 (9-10), 1829–1848. McFadden, D., 1973. Conditional Logit Analyisis of Qualitative Choice Behavior. In: P. Zarembka (Ed.), In Frontiers in Econometrics. Academic Press, pp. 105–142. Mendelsohn, R., Nordhaus, W., 1994. The Impact of Global Warming on Agriculture: A Ricardian Analysis. The American Economic Review 86 (5), 1312–1315. Morten, M., Oliveira, J., 2014. Migration, Roads and Labor Market Integration: Evidence from a Planned Capital City. Unpublished Manuscript. Mueller, V. A., 2005. Valuing Climate Amenities in Brazil Using a Hedonic Pricing Framework. Ph.D. thesis, University of Maryland. Rehdanz, K., Maddison, D., 2009. The Amenity Value of Climate to Households in Germany. Oxford Economic Papers 61 (1), 150–167. Roback, J., 1982. Wages, Rents, and the Quality of Life. The Journal of Political Economy 90 (6), 1257–1278. Saiz, A., 2010. The Geographic Determinants of Housing Supply. Quarterly Journal of Economics 125 (August), 1253–1296. Salda˜ na Zorrilla, S. O., Sandberg, K., 2009. Impact of climate-related disasters on human migration in Mexico: A spatial model. Climatic Change 96 (1), 97–118. Sanghi, A., Alves, D., Evenson, R., Mendelsohn, R., 1997. Global Warming Impacts on Brazilian Agriculture: Estimates of the Ricardian Model. Economia Aplicada 1 (1), 1–15. Sanghi, A., Mendelsohn, R., 2008. The impacts of global warming on farmers in Brazil and India. Global Environmental Change 18 (4), 655–665. Seo, S. N., McCarl, B. a., Mendelsohn, R., 2010. From beef cattle to sheep under global warming? An analysis of adaptation by livestock species choice in South America. Ecological Economics 69 (12), 2486–2494. Seo, S. N., Mendelsohn, R., 2008. Measuring impacts and adaptations to climate change: A structural Ricardian model of African livestock management. Agricultural Economics 38 (2), 151–165. Taraz, V., 2013. Adaptation to Climate Change: Historical Evidence from the Indian Monsoon. Ph.D. thesis, Yale University. Timmins, C., 2007. If you cannot take the heat, get out of the cerrado... Recovering the equilibrium amenity cost of nonmarginal climate change in Brazil. Journal of Regional Science 47 (1), 1–25. 23

Figure 1: Average temperature, in o C Notes: Figure shows a map of Brazil’s meso-regions classified according to average temperature over the 1961-2010 span.

24

Figure 2: Average rainfall, in mm/year Notes: Figure shows a map of Brazil’s meso-regions classified according to average rainfall over the 19612010 span.

25

Figure 3: Brazilian biomes Notes: Figure shows a map of Brazil’s meso-regions classified according to their biomes.

26

Figure 4: Predicted vs actual out-migration rates, agriculture Notes: Figure shows predicted out-migration rates again actual out-migration rates. These rates are computed as one minus the probability of choosing the current meso-region as destination. Each dot is a meso-region and the data are presented by five macro-regions, which are N (North), NE (Northeast), SE (Southeast), S (South), and MW (Midwest). Source: Authors’ calculations based on census data.

27

Figure 5: Predicted vs actual out-migration rates, non-agriculture Notes: Figure shows predicted out-migration rates again actual out-migration rates. These rates are computed as one minus the probability of choosing the current meso-region as destination. Each dot is a meso-region and the data are presented by five macro-regions, which are N (North), NE (Northeast), SE (Southeast), S (South), and MW (Midwest). Source: Authors’ calculations based on census data.

28

Table 1: Summary statistics, by census year

Mean/sd

(1) 1980

(2) 1991

(3) 2000

(4) 2010

0.33 (0.47) 0.060 (0.24) 0.11 (0.31) 2.77 (3.73) 6.81 (6.14) 0.21 (0.41) 68.7 (58.7)

0.32 (0.47) 0.049 (0.22) 0.080 (0.27) 2.43 (3.68) 6.47 (7.43) 0.14 (0.35) 61.0 (50.8)

0.27 (0.44) 0.045 (0.21) 0.067 (0.25) 3.27 (5.62) 7.55 (9.92) 0.12 (0.33)

0.26 (0.44) 0.043 (0.20) 0.066 (0.25) 4.52 (6.99) 8.23 (11.3) 0.16 (0.36) 63.7 (39.9)

24.4 (1.77) 20.8 (4.02) 2241.8 (272.4) 866.9 (325.4) 586.2 (267.3)

24.4 (1.68) 20.6 (4.66) 2168.4 (317.8) 1071.4 (418.0) 615.2 (218.0)

24.9 (1.47) 19.9 (5.25) 2344.2 (289.1) 932.6 (367.7) 551.8 (205.0)

25.5 (1.56) 21.5 (4.32) 2420.3 (345.4) 849.2 (300.3) 735.5 (363.6)

0.042 (0.68) -0.26 (1.00) -73.4 (190.0) 204.6 (236.9) 28.9 (119.2)

0.49 (0.99) -0.65 (1.22) 175.8 (200.7) -138.8 (209.4) -63.4 (181.2)

0.61 (0.72) 1.57 (1.35) 76.2 (141.6) -83.5 (242.9) 183.8 (320.3)

1803967 135

2025411 135

2235274 135

Economic Variables Share of employment in agriculture Agriculture migration rate Non-agriculture migration rate Agriculture wages (BRL per hour) Non-agriculture wages (BRL per hour) Share of renters Rental rates (BRL per room) Climate Variables January temperature o C July temperature o C Sunshine hours Rainfall 1st quarter (mm) Rainfall 2nd quarter (mm) Climate Variables (in Differences) January temperature o C July temperature o C Sunshine hours Rainfall 1st quarter (mm) Rainfall 2nd quarter (mm) Number observations Number meso-regions

3030130 135

Notes: Summary statistics calculated from29 Census microdata. Sample is 20-65 year old males with non-zero earnings in main occupation. Monetary variables expressed in 2000 Brazilian reais (BRL).

30

2,208 2,743 2,395 2,059 2,415

1,964 2,498 2,216 2,141 2,114

North Northeast Southeast South Midwest

1,672 2,453 2,173 2,203 1,968

Sunshine (hours/year) 1980 1991 2010

Region 25.90 25.87 23.28 22.42 23.98

25.62 26.07 23.29 22.71 23.74

26.87 26.49 25.05 23.57 25.30

Mean Temp. - Jan (o C) 1980 1991 2010 25.24 24.09 18.53 14.90 20.50

25.95 24.42 17.69 13.81 19.98

26.28 24.71 19.43 14.86 21.23

Mean Temp. - Jul (o C) 1980 1991 2010

2,132 1,043 1,267 1,621 1,698

2,397 1,354 1,422 1,447 1,789

2,076 1,043 1,313 1,801 1,491

Rainfall (mm/year) 1980 1991 2010

Table 2: Temperature, precipitation, and sunshine hours, by region and Census years.

31

Feb-May Feb-Jun Feb-May After 6 months Dec-Mar, Jun-Jul Jun-Dec Jun-Oct Sep-Apr After 6 months After 1 year Jan-May Jan-Jun Mar-Jul Apr-Jun Jan-Apr Feb-Jun Jan-May Apr-Dec Mar-Aug Mar-Jun Jan-Apr Feb-Jun Apr-Sep May-Aug Mar-Jul After 3-4 months After 1 year Dec-Feb Nov-Apr Oct-Mar Dec-Jun Jan-May Sep-Dec May-Nov Jan-Apr

Oct-Jan Oct-Dec wet season wet season Dec-Feb, Jun-Jul Feb-Apr Mar-Apr Nov-Jan wet season all year Oct-Jan Aug-Dec Sep-Apr Nov-Jan Sep-Dec Oct-Dec Oct-Jan Oct-Dec Sep-Dec Oct-Dec Oct-Dec Oct-Dec Oct-Mar Oct-Dec Sep-Dec Apr-May all year Aug-Dec Sep-Dec Aug-Dec Aug-Dec Sep-Jan May-Jul Sep-May Dec-Mar

Soybean Maize Fruits (cocoa) Fruits (coconut) Fruits (grape) Fruits (mango) Manioc Sugarcane Cotton Fruits (coconut) Fruits (passion fruit) Rice Maize Manioc Cotton Rice Maize Soybean Sugarcane Manioc Cotton Beans Maize Sugarcane Coffee Manioc Potato Fruits (banana) Tabacco Rice Beans Maize Soybean Wheat Coffee Fruits (Apple)

Harvesting Period

Sowing Period

Crop

4

3

2

1

Low growth <10C. Flourishing problem >40C Very low growth <15.5C T>35C are very harmful Low growth<20C.

1 2 3 4

20C21C Average 27C Resistent to low temp.3 High temperatures 20C20C2 Average 27C 21C
1

Temperature Conditions

Source: EMBRAPA. Available at http://sistemasdeproducao.cnptia.embrapa.br/.

South

Southeast

Midwest

North

Northeast

Region

Avg 3-5mm/day Avg 2-5mm/day Regular rainfall Avg 3-7mm/day

Avg 2-5mm/day Avg 2-5mm/day Avg 2-5mm/day Not excess of rainfall Avg 3-7mm/day Avg 3-5mm/day Avg 2-5mm/day Not excess of rainfall Avg 3-7mm/day Regular rainfall

Regular rainfall Avg 3-7mm/day Avg 5mm/day Avg 2-5mm/day Resistent to droughts Irregular rainfall Avg 3-5mm/day

Rainfall Conditions

Avg 5mm/day Avg 3-5mm/day Water abundance Regular rainfall Regular rainfall Harmful: Excess of rainfall Regular rainfall Avg 3-7mm/day Regular rainfall Not excess of rainfall Regular rainfall (5mm/day) Regular rainfall

Table 3: Ideal climate by crop and region of production

Table 4: First-step estimates: migration choice, agriculture sector

Fixed migration cost Bilateral distance (km) Adjacent region Same state Same biome No. individuals Mean migration rate Mean distance migrated

(1) 1980

(2) 1991

(3) 2000

(4) 2010

-4.43*** (0.092) -0.0021*** (0.00026) 1.65*** (0.11) 0.64*** (0.089) 0.074 (0.092)

-4.84*** (0.097) -0.0012*** (0.00025) 1.90*** (0.16) 0.69*** (0.087) -0.077 (0.075)

-4.99*** (0.095) -0.0010*** (0.00020) 1.91*** (0.15) 0.81*** (0.094) -0.048 (0.078)

-5.15*** (0.095) -0.00073*** (0.00018) 1.97*** (0.16) 0.75*** (0.11) -0.017 (0.083)

995,763 0.056 27.2

560,806 0.046 25.5

527,764 0.042 24.5

564,467 0.042 28.6

Notes: Source: Brazilian Census data, 1980-2010. Table reports estimates from a conditional logit model of migration. Standard errors are clustered at the meso region level. Location fixed effects estimated but not reported.

Table 5: First-step estimates: migration choice, non-agriculture sector

Fixed migration cost Bilateral distance (km) Adjacent region Same state Same biome No. individuals Mean migration rate Mean distance migrated

(1) 1980

(2) 1991

(3) 2000

(4) 2010

-3.93*** (0.092) -0.00088*** (0.00016) 0.97*** (0.19) 1.32*** (0.11) 0.31*** (0.10)

-4.36*** (0.066) -0.00071*** (0.00015) 1.04*** (0.16) 1.24*** (0.12) 0.21*** (0.083)

-4.57*** (0.064) -0.00072*** (0.00014) 1.03*** (0.12) 1.20*** (0.11) 0.24*** (0.075)

-4.65*** (0.068) -0.00083*** (0.00015) 1.21*** (0.087) 1.02*** (0.081) 0.18*** (0.070)

2,008,212 0.11 72.7

1,215,523 0.077 53.6

1,461,115 0.065 45.0

1,627,688 0.065 43.5

Notes: Source: Brazilian Census data, 1980-2010. Table reports estimates from a conditional logit model of migration. Standard errors are clustered at the meso region level. Location fixed effects estimated but not reported.

32

Table 6: Structural coefficient estimates Parameter estimates b/se Labor demand agriculture Current rainfall

0.14*** (0.028) 0.17*** (0.019) 1.62*** (0.28) 1.38*** (0.036)

Lagged rainfall Lagged temperature Bartik Shock Labor demand non-agriculture Bartik Shock

0.97*** (0.019)

Housing supply Housing elasticity

0.57*** (0.034)

Indirect utility agriculture Utility from wages Utility from temp. in Jan Utility from temp. in July Utility from sunshine hours

4.59*** (1.34) -7.09*** (1.54) 6.63*** (1.06) 1.64** (0.66)

Indirect utility non-agriculture Utility from wages Utility from temp. in Jan Utility from temp. in July Utility from sunshine hours Year FE

2.61** (1.28) -32.5*** (3.42) 14.5*** (2.08) 3.21*** (0.71) Yes

Notes: Estimated using 1980-2010 data. Coefficients calculated using two-step GMM. Elasticity of utility to rent normalized to -1. Robust standard errors provided.

33

Table 7: Current climate (2010) and climate change projections (average from 2040-70), Brazil Climate Variable

2010

Low

Acc Rainfall (mm/year) 1,584.68 o Avg Temperature in Jan ( C) 25.54 o Avg Temperature in Jul ( C) 21.48

1,609.29 25.96 23.89

Scenario A1B Midi Control 1,403.03 28.38 26.37

High

1,417.88 27.33 25.19

1,415.61 28.51 26.48

Table 8: Counterfactual simulations from climate change: agriculture 1980 (1) (2) PE GE

1991 (3) (4) PE GE

0.12

0.13

2010 (5) PE

(6) GE

0.12

0.099

0.089

Migration rate agriculture 0.13 0.11 0.17 0.15 Change in utility agriculture (%) 39.2 38.9 52.9 51.6

0.17 60.8

0.15 59.0

0.17 60.8

0.15 59.0

0.18 59.4

0.15 57.3

0.17 65.6

0.15 63.8

Baseline Migration rate agriculture

0.10

Scenario I: Controlled

Scenario II: High Migration rate agriculture 0.13 0.11 0.17 0.15 Change in utility agriculture (%) 39.2 38.9 52.9 51.6 Scenario II: Low Migration rate agriculture 0.13 0.12 0.18 0.15 Change in utility agriculture (%) 38.3 37.9 52.1 50.5 Scenario IV: Medium Migration rate agriculture 0.13 0.11 0.17 0.15 Change in utility agriculture (%) 42.0 41.7 57.1 55.7

Notes: Source: Brazilian Census data, 1980-2000. PE = partial equilibrium. GE = general equilibrium. Sample only male individuals aged 20 to 65 who reported positive earnings. Table reports estimates derived from structural estimation. Utility is relative to baseline; only observed components (wage, rent and amenities) of utility included in calculation.

34

Table 9: Counterfactual simulations from climate change: non-agriculture 1980 (1) (2) PE GE

1991 (3) (4) PE GE

0.21

0.13

2010 (5) PE

(6) GE

0.13

0.078

0.078

Migration rate non-agriculture 0.30 0.28 0.22 0.21 Change in utility non-agriculture (%) 47.3 44.8 43.7 41.5

0.14 28.9

0.13 27.4

0.14 28.9

0.13 27.4

0.14 29.2

0.13 27.7

0.14 29.6

0.13 28.1

Baseline Migration rate non-agriculture

0.21

Scenario I: Controlled

Scenario II: High Migration rate non-agriculture 0.30 0.28 0.22 0.21 Change in utility non-agriculture (%) 47.3 44.8 43.7 41.5 Scenario II: Low Migration rate non-agriculture 0.29 0.28 0.22 0.21 Change in utility non-agriculture (%) 47.2 44.7 43.8 41.6 Scenario IV: Medium Migration rate non-agriculture 0.30 0.28 0.23 0.21 Change in utility non-agriculture (%) 48.8 46.4 44.9 42.7

Notes: Source: Brazilian Census data, 1980-2000. PE = partial equilibrium. GE = general equilibrium. Sample only male individuals aged 20 to 65 who reported positive earnings. Table reports estimates derived from structural estimation. Utility is relative to baseline; only observed components (wage, rent and amenities) of utility included in calculation.

35