Returns to Electricity: Evidence from the Quasi-Random Placement of Hydropower Plants in Brazil Preliminary Draft, not for general circulation: This version 7/30/08

Molly Lipscomb University of Colorado, Boulder A. Mushfiq Mobarak Yale School of Management Tania Barham University of Colorado, Boulder Abstract We exploit quasi-random variation in hydro-power generation and transmission in Brazil in order to isolate of the causal effects of electricity grid expansion on changes in population density and GDP. Since hydro-power generation requires intercepting water at high velocity, there is a random component to households’ access to electricity in a country that relies heavily on hydro-power, as that access depends on the household’s proximity to rivers with a gradient suitable for hydro-electricity generation. We predict hydropower plant placement based on exogenous geologic characteristics (river gradient and water flow) of locations throughout Brazil (while simultaneously controlling for land slope in surrounding areas), and then develop a cost-minimizing “engineering model” to predict the expansions of transmission lines from each of those predicted hypothetical stations every decade. We then examine the effects of electricity availability using only the portion of availability that is attributable to purely supply-side geology-driven cost considerations. We find that in the cross-section, grid expansions strongly induce people to move to areas suitable for hydropower dam construction, although this may be driven by a cross-sectional correlation between population density and water. In fixed effect regressions, this effect of electricity on density is much smaller, and fixed effects IV estimates show that even that smaller effect is likely a result of electricity grid expansions following population projections. Electricity is estimated to increase GDP per capita under the fixed effects IV. This is probably due to a true causal effect of electricity on some aspect of productivity, since the population density results suggest that some other ‘general equilibrium’ effect such as changes in the composition of population and skills following inmigration to electrified areas might not be such a big part of this story once location fixed effects and the endogenous expansions of the electricity grid are accounted for.

*We thank the University of Colorado NICHD Population Center, Corporación Andina de Fomento, Center for Advancement in Research and Teaching in the Social Sciences at the University of Colorado and the Macmillan Center at Yale University for the financial support that made this data collection possible. We also thank Daniel Ortega and seminar participants at a Corporación Andina de Fomento research workshop and a Yale School of Management faculty lunch for comments, and Vanessa Empinotti and Steven Li for excellent research assistance.

1

I. Introduction A fundamental role of governments in developing countries is to provide public services such as health, education and infrastructure for its citizens. To this end, developing countries devote about a third of their budget to health, education, and infrastructure programs (World Bank, 2004).

Assisting developing countries in providing services that enhance human

development (such as education, health, water, and electricity) is a primary responsibility of multi-lateral development agencies, such as the World Bank which spent over 20 billion dollars in development projects in 2004 alone (World Bank, 2005a). It is important for social scientists to inform policy-makers about the returns to each type of public investment so that money is spent effectively to reduce poverty and stimulate economic growth. Our knowledge about the returns to many health and education initiatives are well developed, since it has been possible to design small-scale randomized experiments in order to measure the effectiveness of isolated social interventions, for example, increasing teachers’ salaries or paying parents to take children for preventive health checkups.

Designing

randomized experiments for large-scale infrastructure projects (such as building electricity grids) is beyond researchers’ capabilities, which has limited our understanding of the true impacts of such projects. The need for research on the returns to infrastructure is particularly pressing because there is now renewed support from the development community for large infrastructure projects as a means to poverty reduction (World Bank, 2003; Ali and Pernia, 2003). This renewed focus on infrastructure is partly a response to large unmet needs. There are 3 billion people who lack access to modern energy, 1.1 billion people who lack access to clean water supply, 2.4 billion 2

who live without adequate sanitation, 20 percent of the rural population who live more than 2 km from an all-weather road, and a third of the world’s population which has never used a telephone (World Bank, 2005b). This paper exploits quasi-random variation in hydro-power generation and transmission in Brazil to determine the causal impact of electricity provision on income and population density. River gradient (controlling for the slope of the land in nearby areas) is an exogenous source of variation in the potential for hydro-power generation, and the cost-minimizing expansion paths of transmission lines from the hydropower generation plants create panel variation in this exogenous component of electricity generation.

We will use these exogenous

geographic components of electricity provision to examine the effects of electricity on income and population density using county-level data for Brazil from 1960 to 2000. We focus on both population density and GDP per capita as outcome variables to better understand the precise mechanisms through which electricity affects socio-economic outcomes – is it that electrification induces in-migration of highly productive people and firms, or does it enhance the productivity of existing firms and workers. Cross-sectional instrumental variables estimates show a very large population density response to electrification, although this may be due to a positive cross-sectional correlation between density and rivers (which is required for hydropower, but also attracts people for other reasons). In location fixed effects estimates, electrification induces a 10-15% increase in density, but this may be driven by electric grid expansion plans following population projections.

Finally, in a fixed effects instrumental

variables estimate, the effect of electricity on population density is statistically indistinguishable from zero. This suggests that the large positive effect of electrification on subsequent GDP per capita increases that we observe in our fixed effects IV estimates is perhaps not entirely be driven by the in-migration of more productive workers and firms. 3

Section II reviews the relevant literature on the effects of infrastructure on development, section III explains our estimation strategy, section IV presents an overview of the data which we are using and we are continuing to collect, section V presents estimates on the role of local GDP and population in the expansion of the electricity networks, and section VI concludes and presents a framework for the continuing process we will follow in order to strengthen the instrument and improve the accuracy of our coefficient estimates.

II.

Background and Literature Brazil is a particularly relevant setting for our project because the country relies almost

exclusively on hydro-power to meet its electricity needs. Eighty-seven percent of the electricity in Brazil is generated from hydro-electric plants as opposed to 23 percent across the developing world (World Bank, 2005d). This dependence on hydro-electricity suggests that electricity provision in Brazil is more closely tied to quasi-random geographic variation associated with river gradients and flow accumulation than electricity provision elsewhere.

Twenty-seven

percent of rural Brazilians still lack access to electricity (World Bank, 2005c). Furthermore, over the course of the study, 1950-2000, electricity networks in Brazil have exploded in size. The transmission network in Brazil has grown at an average rate of 8.9 percent per year, increasing in size from 2,359 kilometers in 1950 to 174,806 kilometers in 2000 (Siese, 1991, 2002). In addition, the Brazilian Congress is currently debating the construction of what would be the world’s second largest hydro-electric dam in the Amazon, and its current president is committed to increasing hydro-electric energy in the country (New York Times, 2005). President Luiz Inácio Lula da Silva was recently quoted in the New York Times as saying: “There was a dereliction in not building hydroelectric projects in the previous government. With

4

the projects that are underway, we can permanently guarantee [supplies of energy] for 5, 6 or even 10 years down the line.” Despite the growth in multi-lateral lending for infrastructure in the early 2000s (World Bank, 2003; Ali and Pernia, 2003), there has been very limited academic research on the impacts of electricity. In the little evidence on electricity that does exist (Balisacan and Pernia, 2002; Fan et al., 2002; Balisacan et al., 2002; Taylor, 2005; Escobal et al., 2005), it is difficult to make causal inferences about its impacts, since studies fail to account for the fact the electricity is often expanded first in areas with the greatest potential for economic growth. The problem is particularly acute when researchers simply contrast areas with electricity against areas without using cross-sectional data (Escobal et al., 2005), since the two types of areas are likely to be different in important but unobservable ways, and people and firms who expect to benefit most from electricity access are likely to migrate in response to provision. In studies that use panel data, it is possible to compare changes in outcomes over time in ‘treatment’ (with electricity) and ‘control’ (without) areas, but the treatment areas have not been randomly selected.1 The lack of clear empirical evidence on the consequences of electricity is in sharp contrast to the knowledge social scientists have developed on the returns to investments in health and education. This imbalance is mostly attributable to the nature of statistical inquiry. Since the provision of government services typically responds to existing conditions or expectations of benefits, to clearly measure the returns to any public project, we need to observe provision in an “experimental” setting, where the allocation is randomized either by design or by nature. While it is feasible for social scientists to design small-scale randomized interventions in health or

1

Fan et al. (2002) use Chinese provincial-level data from 1970-97 and show that for every 10,000 yuan spent on electricity development, 2.3 persons are brought out of poverty. Balisacan and Pernia (2002) use Filipino provincial level data from 19851997 to argue that the rich tend to benefit from access to electricity. Balisacan et al. (2002) shows that in Indonesia in 1990, a 10% improvement in access to a composite technology measure (presence of public phones, TV postal office, and electricity in the village) raises the income of the poor by roughly 2 percent.

5

education,2 large-scale infrastructure projects such as roads or electricity networks are rarely, if ever, randomly placed. In a simple observational study, we would over-estimate the returns to building electric grids if electricity is more likely to be provided in areas where the government expects it to lead to larger benefits. The U.N. commissioned Canambra Engineering feasibility study reports (1968) drew up expansion plans for the electricity grid in the southern and south-eastern states of Brazil based on forecasts of load factors (the key demand side factor) that were primarily linked to local GDP and projected GDP growth. Expansion of networks is prioritized where forecast loads are largest. Even in a panel data study that controls for location fixed effects, expansion plans based on forecasts may result in an upward bias in the estimated effect of electricity provision on various growth measures (e.g. economic or population growth), which are directly related to the demand side factors for electricity used for grid expansion plans. Conversely, we may under-estimate returns in an observational study if the government’s objective is to provide electricity to areas deemed in greater need or if it emphasizes the expansion of coverage to new regions or states that were previously left uncovered.

This paper attempts to develop a

methodology for estimating the returns to electricity that accounts for the non-random placement of power grids. Geographic variation has been used as a source of identification in a few empirical studies of the effects of public investment in new infrastructure projects. Michaels (2007) uses whether a county is connected to the interstate highway network to estimate the effects of lowered trade barriers on the demand for skilled labor. Duflo and Pande (2005) use slope as an instrument for the placement of dams in Indian districts. Since slope does not vary over time, they create an interaction of slope and state-wide dam construction in each time period to create 2

See Kremer (2003) for a review of some randomized experiments in education. Other examples include: Miguel and Kremer (2003); Banerjee et al., 2004; Gertler and Boyce, 2001; Shultz, 2001; Kremer et al., 2002; Thomas et al., 2003.

6

panel variation in their instrument. Dinkelman (2008) is the most closely related paper to ours, and uses land gradient as an instrument to show that there are significant cooking technology and female labor participation impacts of communities’ connection to the electricity grid after 1996 in Kwa-Zulu Natal province of South Africa. In her design, slope increases the cost of providing electricity, making a connection less likely. In our study we use river gradient as a positive predictor of hydropower dam placement, and are careful to always include land slope as a control variable since - as Dinkelman (2008) notes - land slope may be correlated with agricultural outcomes, the cost of providing other public services, and other characteristics of the population. In Kwa-zulu Natal the correlation between slope and other socio-economic variables Dinkelman (2008) gathers data on is low, but this may not be more generally true across Brazil. Also, since land or river gradient would only have cross-sectional variation, we rely on a simplified costminimizing engineering model of the grid expansion to examine the cross-sectional time series effects of electricity grid expansion across Brazil over the time period 1960-2000.

III.

Estimation Strategy Even though electricity is not randomly allocated, households’ access to electricity in a

country that relies heavily on hydro-power may have an exogenous geographic component to it because it would depend on proximity to rivers with a gradient suitable for hydro-electricity generation. Our research design exploits the quasi-random placement of hydro-electric plants in Brazil based on variation in geography that affects the possibility of hydro-power generation. Generating hydro-electricity requires high-velocity water flow, which is contingent on intercepting water at a steep gradient. The ability to construct reservoirs and dams and the availability of sufficient water flow are crucial to guaranteeing consistent electricity supply from a given power plant. Hydroelectric plants must be constructed where river depth and flow is 7

sufficiently large to maintain the plant throughout the year (Canambra, 1968). Since geography (particularly river gradient and flow) affects the suitability of hydro-electric plant construction, this creates some natural variation in the source points for electricity across Brazil. To the extent that the physical slope of a river is uncorrelated with socio-economic characteristics of the local population, this creates a nice natural experiment where observationally similar households are either more or less likely to be endowed with electricity access depending on the their distance to an appropriately sloped river. One possible drawback of this methodology is that geography (and in particular the slope of the land where a household resides) could indeed be correlated with socio-economic outcomes.

Slope can affect land use, agricultural outcomes, and

occupational choices.3 However, river gradient (which affects hydro-power generation and suitability of dam construction) is typically different from the slope of the land, so we can separately control for geographic characteristics that are related to socio-economic outcomes, including the slope of the land, while simultaneously using river gradient as an instrument for hydropower plant construction. Measures of river gradient and water flow only vary in the cross-section and are constant over time at a given location. To generate panel variation in the predicted expansion of electricity grids, we develop a simplified engineering model of the least-cost placement of transmission lines and sub-stations (which distribute the electricity generated by our predicted generation plants). Electricity transmission networks transport electricity from the generation plant to the regional hub which provides electricity to local distribution networks. Construction of new transmission lines is expensive and requires extensive planning in order to ensure that the new lines do not cause system shortages. Our simplified engineering model is decidedly a-behavioral 3

Dinkelman (2008) uses slope as an instrument for electricity provision in South Africa, arguing that it’s more costly to extend coverage to sloped areas. Apart from the relationship of land gradient to land use and agriculture, gradient might also make it costlier to supply other public services,

8

in that transmission lines are chosen to simply minimize costs without any demand side considerations. The actual placement of transmission lines in the real world is based on both cost and demand-side behavioral and human considerations, but since we consider pull factors from demand to be endogenous, the idea here is to extract only that portion of the variability in electricity availability that is attributable to variation in river gradient, water flow, and distance to the locations that possess those geographic characteristics suitable for generating hydropower.

IV.

Data Brazil has experienced almost exponential growth in its electricity network since 1950

(figure 1), largely fueled by hydropower. The transmission network expanded from 2,359 kilometers in 1950 to 167,443 kilometers in 2000 (SIESE, 2000). We have assembled a database of the locations of all power plants and all electricity transmission substations across Brazil from the 1960s until 2000, using the feasibility studies and inventories that major electricity companies in Brazil undertake prior to planning expansion of their networks. The power plant and transmission line data come in two forms – (1) tables with inventories of all transmission lines that typically specify the county of origin, the destination county, length and voltage (for transmission lines) and location (for hydropower generation plants, or (2) large paper maps of generation plants and lines by region of Brazil. We digitize and combine all this information into five different GIS maps of the Brazilian electricity network for the 1960s, 70s, 80s, 90s and 2000.4 Power plants were placed on the digital map according to their reported latitude and longitude, while transmission substations were assumed to be located at the centroid of their

4

The 1960s network is based on the comprehensive inventory taken by Canambra (1967) and Canambra (1969) for 1965 and 1967. The 1970s network is pieced together from various maps and tables from the different regions of Brazil from a variety of sources (e.g. see Figure 5). The 1980s network is based on another comprehensive inventory by SIESE (1987). The 1990s is again pieced together from various sources (e.g. Furnas 1993), and the 2000 network is based on SIESE (2000).

9

county of record. Please see Figures 2-5 for examples of the maps and tables which were used to construct the data set. 775 major electricity plants have been constructed in Brazil since 1910, and 546 of which are hydroelectric plants. A. Unit of Observation: Constructing a digital map database of the evolution of the electricity network allows us to overlay GIS maps of elevation and rivers and merge information on slope and water flow to the power plant and transmission data by location. We define “location” (i.e. our spatial unit of observation) to be a grid of 32,500 evenly spaced points for all of Brazil set 16km apart from each other. Our task then is to create measures of electricity availability, land slope, river gradient, water availability and water flow for each of these grid points. Choosing grid points as our unit of observation rather than states or counties helps diffuse potential bias arising from political economy factors that might determine the shapes and sizes of political jurisdictions. B. Constructing measures of electricity availability: We have data on generation plants and transmission lines, but not on distribution networks, which is the final component of the electricity grid that connects users to the substations. Transmission lines transfer electricity from the generation plants to the regions which are being supplied, while distribution networks transport electricity from the major local transmission substation to household, industrial and agricultural consumers of electricity. It was not possible to map these distribution networks going back to the 1960s because electricity distribution in Brazil is decentralized across 64 privatized electricity companies, and there is no central clearinghouse for data on their operations. We do however need to account for the distribution network since assuming that only areas with substations have electricity would be unnatural. Based on our conversations with electricity sector professionals in Brazil, we assume that distribution networks stretch 100 km across, so all grid points within a 50 kilometer radius 10

of the centroid of a county containing a transmission substation have access to electricity. Figure 6 illustrates how these ideas are implemented in South Brazil for the 1960s. The dark blue points are counties which have transmission substations and the light blue circles surrounding them are assumed to be the distribution networks associated with those substations. All grid points (black dots) which fall within light blue areas are assumed to have access to electricity. Note that in our data electricity availability is only a direct function of the location of substations (i.e. the endpoints of transmission lines), and not the placement of hydropower plants. Building long transmission lines is costly, so proximity to power plants does indirectly affect a location’s likelihood of having access to power. Figures 7-11 map the evolution of the electricity network in Brazil from the 1960s through 2000. As one might expect, the early development of the electricity network was focused in relatively affluent and industrial south and from 1970s onwards the grid was expanded to the populous (but poorer) south-east and north-east. The network has expanded westward every decade since the 1970s, and by 2000, the coastal areas of the south-east and north-east had almost universal coverage. The Amazon and Pantanal areas have remained largely uncovered. C. Geographic data: Figure 12 demonstrates our construction of the river gradient instrument. We draw circles of radius 10km around the evenly spaced grid points throughout Brazil and measure average land slope in those circles to use as a control variable. We then overlay a map of water bodies, create 2 km buffers on either side of each river, and compute the gradient along the river using elevation maps. We use this river gradient along with measures of water flow to help predict plant location (while simultaneously controlling for land slope in the surrounding areas). The specific indicators of water availability used are average and maximum flow accumulation 11

(which measures the amount of water flowing into each point on the river), calculated based on GIS maps from U.S. Geological Survey’s Hydro1k program. D. Constructing the instrument: While the exogenous geographic factors such as river gradient and flow accumulation help predict hydropower plant location, we need to construct an instrument for “predicted electricity availability in each decade”, which accounts for the evolution of transmission lines and substations and varies both over time and in the cross section. We construct a very simple engineering model of electricity grid expansion in which decisions are made solely based on geography-induced cost considerations in order to generate a prediction for whether each of the 32,500 evenly spaced grid points has electricity access in each of the five time periods of data between 1960 and 2000. Our objective is to generate matched predictions for the five specific time periods of data on the actual electricity grid available to us – 1960s based on inventories conducted in 1965 and 1967, 1970s based on maps from 1973 of the network, 1980s based on a comprehensive inventory conducted in 1987, 1990s and 2000 from the National Agency for Electricity’s data on recent construction of major transmission lines. From the perspective of the engineering model, the specific dates for which we have data (i.e. the years that inventories were conducted) is essentially arbitrary, which implies that the scale of expansion between two periods – i.e. the number of new power plants and transmission lines built since the last inventory - would remain indeterminate. We therefore match the scale of expansion between two periods to match the scale of investment in hydropower plants observed in the actual data. In other words, we allocate a budget of 240 power plants to the model in the 1960s because that is the number of hydropower plants in existence in Brazil by 1967 - our inventory date for that decade. By similar reasoning, the budget for 1970s was 53 additional power plants, 36 additional plants for 1980s, 12

25 additional plants for 1990s, and 24 additional plants for 2000. The model takes these budgets as given, and chooses the optimal location of hydropower plants and transmission lines based purely on exogenous geological factors. The first step in the model is to choose the locations of the budgeted hydropower plants. The model uses as inputs (i.e. “instruments”) whether the location (a circle of radius 10km around each grid point) has a river, the average and maximum gradient of the river, maximum water flow accumulation anywhere within that circle, and an indicator for whether that location falls in the Amazon. We let the data guide us on the relative importance of each of these factors in determining plant location. Specifically, we run a probit regression of actual hydropower plant locations on these instruments to generate predicted probabilities for hydropower plant placement as a function of geologic conditions (see table 1). As expected, hydropower plants are most likely to be located in areas with steep river gradient, with greater water flow, and away from the Amazon. Armed with this prediction, our model places the first 240 power plants in the highest probability locations based on geologic considerations in the 1960s. Figure 13 maps these 240 predicted plants as red dots and the entire set of hydropower plants used to generate the probit predictions as yellow dots against a background of elevation (darker colors are closer to sea level) and rivers. Our model predicts a large number of power plants along the south-east to north-east corridor (Sao Francisco river basin) where elevation suddenly moves to high from the low-lying areas of the coast, implying a steep increase in slope. One drawback is that the model predicts excessively dense clustering of hydropower plants due to the high spatial correlation in geologic factors conducive to hydropower plant construction. In a future version of this model, we plan to introduce a slightly more sophisticated decision-making rule where plants cannot be constructed too close together along the same river.

13

The next step in the model is to predict the locations of substations (i.e. directions of transmission lines) which deliver the electricity generated at each plant predicted from the previous step. We make the simplifying assumptions that all power plants have the same generation capacity, and - guided by the data - that each plant has exactly two transmission substations which are connected through a single line (the average number of transmission substations per power plant as of 2000 was 2.5 (SINDAT, 2008)). The electricity network is fully durable, and new substations and power plants cannot be placed in locations which have already received electricity in prior decades. The model arrives at the optimal lowest cost electricity network in each decade by computing costs for all possible arrangements of transmission lines. There are a finite but arbitrarily large number of possible permutations of transmission lines, and the numeric method we use to arrive at the lowest-cost grid in equilibrium is detailed in the Appendix. The model assumes that cost increases with distance and is very high when building substations in the Amazon (due to high material transport costs). In a future version, we plan to introduce other natural geographic barriers (e.g. crossing a mountain) as cost-enhancing. Once the equilibrium set of transmission lines are determined, we assume that all grid points within a 50 kilometer radius of any substation will receive access to electricity, accounting for the distribution network for that substation. In other words, we purposefully remain agnostic about the direction in which the distribution networks are expanded because that choice is governed by demand-side factors which are ‘endogenous’ to measures of economic performance. The chosen 50km radius is also data-driven (based on the average size of actual distribution networks in Brazil), and mirrors our treatment of distribution networks in the actual electricity data. In subsequent decades, new power plants are placed in the highest probability circles among those which have not yet

14

received electricity, and alternative grid points for transmission substations are proposed from among the points which remain without electricity from the previous decades. Figures 14-18 plot the areas predicted to receive electricity by this model by decade. There is a reasonably good cross-sectional spatial correlation with the actual electricity network for Brazil (figures 7-11), and there are encouraging signs of correlation in terms of direction of expansion. Geologic considerations induce the model to cover the south and south-east region first, and then move towards the northeast and slowly west-ward, which is a pattern we observe in the actual data. However, overall the model chooses to cover a much greater portion of Brazil with electricity in the first period (1960s) than we see in the actual data. As a result, there is much lower panel variation in coverage in the time-series dimension than we see in the actual data, as highlighted in figure 19. As we will see, this adversely affects our predictive power in the panel regressions. This problem can be traced back to our simplifying assumption that all hydropower plants have the same power generation capacity which cannot be expanded. In allocating a generation plant budget to the model for each decade, we assumed that the scale of expansion of the electricity grid has to match actual number of power plants built between time periods. This ignores the fact that many power plants were initially built small and had their capacity expanded in later decades, with an associated increase in the number of substations allocated to the plant. The constant capacity (and constant number of 2 substations per plant) assumed therefore forces the model to over-predict electricity availability in earlier years and under-predict it in later years. We are currently putting together data on the evolution of the generation capacity of each plant and the cost of expanding capacity relative to the cost of building new plants, so that we incorporate a decision on whether to expand capacity on existing plants versus build new plants into the engineering model. E. Dependent Variables 15

This study examines the effect of electricity on changes in population density, GDP per capita and industrial GDP per capita from the 1960s until today. County level population counts are based on the decennial demographic censuses, while county GDP number are estimated by the Instituto de Pesquisa Economica Aplicada (IPEA) based on various sources of data including the census of manufacturing and of agriculture. Brazilian counties have split numerous times over the period of our analysis. Since we do not have digital county maps from 1960, we are unable to assign the grid points (our spatial unit of observation) to the exact county they fall into in every decade. However, we do know the area of each county for every decade. In order to approximate the boundaries of each county over time, we digitally identify the location of all counties existing in a given decade from the centroid of the county as of 2001 and draw circles around the centroid of those counties such that the area of the circle equals the area of the county in the decade of interest. We assume all grid points which fall in the circles belong to the county in that year. In cases where the grid point falls into more than one county circle, the average value of the dependent variable is used. Grid points which fall outside of all circles are assigned to the nearest county circle in that decade. Figure 20 plots the 1960 county proxies, which is the year with the fewest number of counties. Although our unit of observation is grid points, the underlying source of data is at the county level, so we cluster all errors by 1960 counties, which is the most conservative assumption on clustering (for later decades we actually have more counties and therefore more data). Table 2 reports the summary statistics for these dependent variables. There is much greater cross-sectional variation in population density than there is panel variation holding locations fixed. Population density is also the variable that we feel is measured most accurately over time.

Given the way county-level GDP is estimated by IPEA, the cross-sectional

differences in these series are of higher quality than changes in the panel within locations. In 16

other words, after the addition of location fixed effects, it is likely that the signal to noise ratios in the GDP series decrease dramatically. So for GDP, one has to be more careful making inferences with fixed effects estimates.

V.

Results There are a number of distinct mechanisms through which electricity may affect

economic outcomes. The most direct mechanism, and the one most closely related the research questions motivated in the introduction, is the productivity-enhancing effects of, say, lighting a school room or refrigerating perishable vaccines, or introducing electricity-powered capital into a firm’s production process. A less direct but equally plausible mechanism is that electrification induces the movement of people and firms. And if more highly skilled workers and firms respond differentially to the availability of electricity, the resulting in-migration to electrified areas can also lead to a positive correlation between electricity and socio-economic outcomes. For example, if electricity and worker skill are complementary inputs in the production process and if workers are mobile, then we may observe a large positive effect of electricity on firm productivity or area GDP per capita even if electricity is not itself highly productive as an input. In that case, electrification only leads to a spatial reallocation of resources within a country rather than an overall productivity boost at the macro level. And in that case our estimates will not speak directly to the comparison between the returns to spending an extra dollar of funds on educational investments or electricity investments.5

5

Other ‘general equilibrium’ connections between electricity investments and socio-economic outcomes can also prevent us from isolating direct causal effect of electricity on productivity. For example, imagine that with limited budgets, simple cost-benefit calculus induces governments to allocate a greater share of the budget to electricity in areas where electricity is low-cost to generate (e.g. sloped areas with water, as in our instruments), whereas in highcost areas a greater share of the budget is allocated to other public services such as health, sanitation or education. In that case, even in the instrumental variables regression we would be comparing electrified areas with low levels of other public services to non-electrified areas with high levels of other public services, which, in this case, would lead

17

The multiple potential mechanisms at play suggest that it would be useful to examine electricity’s effect on a variety of outcome variables that can help isolate specific mechanisms. For example, we could examine the effect of electricity on migration patterns, especially as it relates to movements of firms and worker types. This may help identify whether a productivityenhancing effect of electricity is due to the in-migration of productive firms to electrified areas or to increases in productivity at existing firms. Unfortunately, we do not yet have access to such specific variables over our long panel stretching back to the 1960s. So we begin by exploring the effects of electricity on two summary outcomes for which it is possible to get long panels of data – population density and GDP per capita. These two variables are related to the two main hypothesized mechanisms at play – the movement of people and direct production function effects. Population density changes likely reflect migration since the magnitudes of the short-term effects of electricity on mortality and fertility – which are the two other important sources of change in population – are unlikely to be large enough to affect population density numbers greatly. GDP per capita changes may reflect both differential migration of skilled workers and productive firms, or improvements in the productivity of existing firms and workers. We will thus begin with an examination of population density to see whether the migration mechanism appears empirically important, which will in turn help us better interpret the sources of any changes in GDP and industrial GDP per capita. Our regression analysis of the effect of electricity on population density and GDP use different types of variation in the data in turn. We begin with a purely cross-sectional approach and take advantage of variation across all regions and states of Brazil (while always clustering standard errors by the large 1960 counties). As one would anticipate based on a cursory glance

to an under-estimate of the returns to electricity if those other public services are productive. The solution to this type of a problem is conceptually straightforward – we would need to add control variables for such other productive public services in our regressions.

18

of the maps of the actual electricity network in figures 7-11 and the maps of the predicted (modeled) electricity network in figures 14-18, our instrument’s fit to the endogenous electricity variable in this case is excellent. Areas predicted by our geologic instruments and model to have electricity have a 43 percentage point greater chance of actually having electricity available. However, this statistically significant correlation is somewhat driven by the Amazon dummy in our predicted hydro-power plant locations, and the large effect of the instrument reflects the large differences in electrification between the Amazon and non-Amazon regions of Brazil. These regions are different in other economically-relevant dimensions as well, so we next show estimates that include region fixed effects, so that the empirical inference is not based on differences across the disparate regions.

The estimated effect in the first-stage regression

immediately jumps down: Areas predicted by our geologic instruments are 8 percentage points more likely to have electricity than other areas within the same region. The third column in table 3 shows that state dummies have a very similar effect on the first-stage estimates as region dummies (predicted locations are also 8 percentage points more likely to actually have electricity as other locations within the same state). In column 5 we add random effects by grid point in addition to region dummies, and in the last column fixed effects by grid point. The instrument remains positive and strongly predictive of the actual electricity distribution. Land slope is also strongly correlated with hydropower plant placement, which both (a) works to reduce the power of our instrument (due to the correlation between land slope and river gradient, and that land slope is a separate control variable in the second stage), and (b) makes it important to separately control for slope in the second stage if there are other possible links between slope and socioeconomic or demographic outcomes. The first three columns of table 4 report OLS estimates of electricity availability (lagged one period) on population density. All specifications always separately control for land slope 19

and standard errors are always (conservatively) clustered by the large county areas in 1960. The second and third columns add region and state dummies respectively, which expectedly lowers the magnitude of the effect of electricity on population density. The estimated effects are quite large – electricity availability in an area is associated with its population density being twice as large the following decade (at the mean), even relative to other areas within that same state. The average population density in the non-Amazon regions of Brazil is 23 people per square kilometer during the entire sample 1960-2000, and this average rises to 33 people in the nonAmazon and non-Pantanal regions.

The instrumental variables estimates of the effect of

electricity on population density (in the last 3 columns) are always larger than the corresponding OLS estimates, which implies that areas suitable for hydropower plant construction (and transmission line expansions from those plants) have an even an even larger correlation with dense population. Moreover, the IV estimates get even larger when region and state dummies are added, indicating strong population clustering near hydropower suitable sites within each state. If we were to ascribe causality to these IV estimates, then the results would be consistent with electricity being targeted more to areas that were sparsely populated. Such targeting would appear to be most pronounced within states. In other words, under-populated parts within states were targeted more than denser parts, as opposed to some states being targeted more than others. However, since these are cross-sectional estimates, then there is another plausible interpretation of this data. In the cross-section, our instrumental variables estimates are essentially comparing areas that have rivers with a steep gradient and voluminous water flow against all other areas that do not. This control group is actually comprised of three different types of areas: those with flat rivers, those with narrow, shallow rivers, and those with no river at all. Given the presence of no or low water areas in the control group, our IV estimates may be conflating the effects of 20

electricity with the effects of having water available. And water may draw dense population concentrations for reasons completely independent of electricity generation. Note that this inference problem would be completely mitigated in panel regressions where we control for location fixed effects, since water availability at flow at each location is essentially a fixed factor. Thus it is important to present fixed effects instrumental variables estimates in this context in order to isolate the causal effect of electricity availability. In summary, we find a strong positive cross-sectional (across Brazil, within-state and within-region) correlation between electrification and dense population using OLS, and an even stronger correlation between hydropower generation (and associated grid expansion) and subsequent population density using IV. It is quite possible that the former is driven by reverse causality (grid expansions follow population projections, which makes those projections selffulfilling), and the latter is driven by people’s attraction to both electricity and water. Whatever the nature or direction of the causality, there is certainly a relationship between electrification and in-migration, as evidenced by the very large population density gains. Such large gains in density are unlikely to be driven by changing fertility and mortality within a decade. Table 5 examines the effect of electricity on GDP per capita, again in the cross-section. The OLS estimates indicate a strong positive effect, but that effect is understandably smaller once region dummies or state dummies are added. Electrifying a location is associated with a 15% larger GDP per capita there the following decade relative to other areas within the same region. The effect becomes smaller and loses statistical significance under the state dummies specification, but with the caveat that the methods to estimate county-level GDP by IPEA using underlying sources that were designed to be reported only at the state level probably carries a lot of measurement error when we focus on its within-state variation. The instrumental variables estimates of electricity on GDP per capita are less precisely estimated and quite variable 21

depending on nature of the data variation being used, ranging from a strong positive effect across Brazil to a negative point estimate (but an effect that is statistically indistinguishable from zero) within states. Estimates of electricity’s effect on just the industrial component of GDP are much more stable. The OLS estimates are positive and statistically significant and indicate that within a region, electricity increases industrial GDP by 45% at the mean year-2000 value. The IV estimates are of comparable magnitude, but the standard errors are much larger. The IV point estimate suggests that electricity increases industrial GDP by 20% at the mean within regions, but this is imprecisely estimated and we cannot rule out that the true effect is zero. As with the population density specifications, the effect on GDP in the cross-sectional IV estimates may reflect the effect of both water and hydropower. We estimate random effects and fixed effects models in tables 7 and 8, taking full advantage of the panel dimension of our dataset. In table 7 we add random effects by grid point in addition to the region dummies in the IV model, which assumes that location specific characteristics are uncorrelated with our predicted electricity instrument. Table 8 relaxes this questionable assumption and reports a results from a “within” estimator where we effectively regress the decadal change in the outcome variable (GDP or population density) on the lagged change in our model’s prediction for whether that particular location has electricity. The random effects IV model shows that electrification increases GDP per capita by about 20% within region. The estimate for industrial GDP has a comparable magnitude but the effect is not statistically significant. Electrification has a large positive effect on subsequent GDP per capita changes in the fixed effects IV regressions. The most notable difference between the panel and pooled cross-sectional models is that the effect of electricity on population density is estimated to be much smaller in the panel. Once we add location (i.e. grid point) fixed effects without instrumenting in column 3 of table 8, we 22

find that electrification is associated with a 10-15% increase in population density the following decade. This effect is an order of magnitude smaller than all cross-sectional OLS and IV estimates. Thus the cross-sectional IV estimates do indeed appear to be biased upward due to a positive correlation between rivers and density. Moreover, even the remaining 10-15% positive effect we uncover in this fixed effects regression may be driven by the electricity grid being expanded to areas where planners expect future population growth (and acting on that expectation in turn fulfills that growth promise). To account for this, we report within-region random effects IV and grid point fixed effects IV estimates in the last columns of tables 7 and 8, and find that the effect of electrification on population density is statistically indistinguishable from zero in both specifications.

Any upward bias from grid expansion plans based on

population projections appears to be small. If the fixed effects IV results are successful in isolating the causal effect of electricity, then the GDP and population density results indicate that there is a large GDP per capita increase following electrification which cannot be fully explained by large-scale movements of people. Thus electricity likely has a direct positive production function effect in stimulating greater economic productivity. We are concerned about the poor fit of the predictive engineering model in the panel dimension, particularly in the earlier two decades when we vastly over-predict electricity availability and under-predict its growth.

Table 10 therefore re-visits the fixed effects

instrumental variables models on a restricted sample of 1980s-2000, since our model’s predictions of electricity growth during this period matches reality (although we still systematically under-predict electricity level during this period). The instrumental variables estimates with grid-point fixed effects produce incredibly large estimates of electricity on GDP. Both GDP per capita and industrial GDP are predicted to more than double following 23

electrification. Electricity induces a population density increase of about 50-60% at the mean, but the standard error associated with this estimate is again very large.

VI.

Conclusion Understanding the role of access to electricity in economic development is crucial to

planning long term investments in infrastructure improvements in developing countries. Taking advantage of the close link between geology and hydropower plant construction in Brazil, we produce quasi-experimental estimates of the long-term effects of electrification on local income and population density over four decades. Our methodology is able to predict quite well the cross-sectional variation in electricity availability across regions of Brazil based only on geologic factors and cost-minimizing engineering considerations. Our model still requires refinements in order to improve its prediction of the panel variation in electrification, particularly in the intertemporal dimension. While this geography-based instrumental variables methodology can help isolate the causal effect of electrification on economic development, there may be multiple mechanisms that mediate this relationship that are interesting to distinguish. Electricity may induce the inmigration of people and firms, or the in-migration of a certain type of (e.g. disproportionately higher quality) worker or firm, and it may simply improve the productivity of existing workers or firms. From a national-level policy perspective, the last effect is a pure net gain while the migration-related effects are in part a reallocation of already productive resources. Our focus on two different types of dependent variables – GDP and population density – helps to distinguish between the two types of mechanisms to a certain extent, since population density changes are closely related to migration, while GDP changes reflect an overall effect that aggregates both mechanisms. 24

Using cross-sectional variation in the data we find very strong effects of electricity on population density, in the sense that areas with geologic characteristics conducive to building hydropower plants are predicted to be more than twice as dense on average than areas that are not. The panel (fixed and random effects) estimates suggest that that electrification during a decade induces about a 10-15% increase in population density at those locations in the following decade. In the fixed effects IV estimates (where we simultaneously add the location fixed effects and use our instrumental variables strategy to address the non-random placement of the electricity grid), the effect of electricity on population density is statistically indistinguishable from zero. Although the GDP estimates are less stable, electricity appears to have a strong positive effect on subsequent GDP per capita even in the fixed effects instrumental variables specification. The accuracy of our first-stage predictions of the panel variation in location of electricity provision still needs to improve. New data on capacity of each generation plant should help in this regard. Further research remains to be done on the mechanisms through which the provision of electricity affects socio-economic outcomes. We plan to use a broader set of outcome variables from household and firm surveys to more precisely understand these mechanisms.

25

References Ali, I. and E. Pernia (2003). “Infrastructure and Poverty Reduction: What is the Connection?” ERD Policy Brief Series No. 13, Economics and Research Department, Asian Development Bank, Manila, Philippines. Arellano, Manuel and Stephen Bond (1991). “Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations.” Review of Economic Studies. V.58 No.2. p.277-297. Balisacan, A., Banerjee, A., Cole , S., E. Pernia. (2002). “Probing Beneath Cross-National Averages: Poverty, Inequality, and Growth in the Philippines.” ERD Working Paper Series No. 7, Economics and Research Department, Asian Development Bank, Manila. Balisacan, A., E. Pernia, A. Asra. (2002). “Revisiting Growth and Poverty Reduction in Indonesia: What Do Sub-National Data Show?” ERD Working Paper Series No. 25, Economics and Research Department, Asian Development Bank, Manila. Banerjee, A., S. Cole, E. Duflo, and L. Linden (2003). “Remedying Education: Evidence from Two Randomized Experiments in India,” mimeo, MIT, November 2004. Baldacci, E. N. Guin-Siu, L. de Mello (2003). “More on the Effectiveness of Public Spending on Health and Education: A Covariance Structure Model,” Journal of International Development, 15(6), 709-725. Bloom, D., D. Canning, and J. Sevilla (2001). Health, Human Capital, and Economic Growth Randomized Evaluations of Educational Programs in Developing Countries: Some Lessons. Commission on Macroeconomics and Health Working Paper #WG1:4. Briceno-Garmendia, C. Stache, A. Shafik. (2004). “Infrastructure Services in Developing Countries: Access, Quality, Cost and Policy Reform.” World Bank Policy Research Working Paper No 3468, Washington D.C. Cachapuz, Paulo Brandi de Barros (2002). O Planejamento da expansão do setor de energia elétrica : a atuação da Eletrobrás e do Grupo Coordenador do Planejamento dos Sistemas Elétricos. Eletrobras, Rio de Janeiro. Canambra engineering consultants, Power Study of South Brazil. 1969. Dinkelman, Taryn (2008). “The Effects of Rural Electrification on Employment: New Evidence from South Africa.” Duflo, E. and R. Pande (2005). “Dams”, mimeo MIT, July. Duflo, E. (2001). “Schooling and Labor Market Consequences of School Construction in Indonesia: Evidence from an Unusual Policy Experiment,” American Economic Review, 91(4), 795-813. Oster, E. (2005). Sexually Transmitted Infections, Sexual Behavior and the HIV/AIDS Epidemic, mimeo, Harvard. Escobal, G. and M. Torero (2005). “Measuring the Impact of Asset Complementarities: The Case of Rural Peru,” Cuadernos de Economia, 42(May), 137-164. Fan, S., L. Zhang, and X. Chang (2002). Growth, Inequality and Poverty in Rural China: The Role of Public Investments, Research Report 125, International Food Policy Research Institute, Washington D.C. Gertler P. and S. Boyce (2001). “An Experiment in Incentive-Based Welfare: The Impact of Progresa on Health in Mexico,” mimeo, University of California at Berkeley.

26

Glewwe, P., M. Kremer, and S. Moulin (1998). “Textbooks and Test Scores: Evidence from a Randomized Evaluation in Kenya,” Washington, DC: World Bank Development Research Group. Gramlich, E. (1994). “Infrastructure Investment: A Review Essay,” Journal of Economic Literature. 32(30), 1176-1196. International Monetary Fund (IMF) (2004). International Financial Statistics, Washington D.C. Kremer, M. (2003). “Randomized Evaluations of Educational Programs in Developing Countries: Some Lessons,” American Economic Review 93(2) , May 2003, pp. 102-106. Kremer, Michael, Josh Angrist, Eric Bettinger, Erik Bloom, and Elizabeth King (2002). “Vouchers for Private Schooling in Colombia: Evidence from a Randomized Natural Experiment,” American Economic Review, 92(5), 1535-1558. Linden, L., A. Banerjee, and E. Duflo (2003). “Computer-Assisted Learning: Evidence from a Randomized Experiment,” Poverty Action Lab Paper #5. Londono, J. (1996). “Poverty, Inequality, and Human Capital Development in Latin America 1950-2025,” The World Bank, Washington D.C. Mehrotra, S. and E. Delamonica (2002). “Public Spending for Children: An Empirical Note,” Journal of International Development, 14 (8), 1105-1116. Michaels, Guy. (2007). “The Effect of Trade on the Demand for Skill—Evidence from the Interstate Highway System.” Forthcoming, Review of Economics and Statistics. Miguel, E. and M. Kremer (2004). “Worms: Identifying Impacts on Education and Health in the Presence of Treatment Externalities,” Econometrica, 72(1), 159-217. Mingat, A. and J. Tan (1996). “The Full Social Returns to Education: Estimates Based on Countries’ Economic Growth,” Human Capital Working Paper 16131, World Bank, Washington D.C. New York Times (2005). “Brazil Weighs Costs and Benefits of Alliance With China,” September 20. Schultz, T. Paul (2004). “School Subsidies for the Poor: Evaluating the Mexican Progresa Poverty Program,” Journal of Development Economics, 74(1), 199-250. Summers, L. (1994). Investing in all the People: Educating Women in Developing Countries, The World Bank, Washington D.C. Taylor, M. (2005). “Electrifying Rural Guatemala: Central Policy and Rural Reality,” Environment and Planning C: Government and Policy, 23, 173-189. Thomas, D. et al. (2003) “Iron Deficiency and the Well-Being of Older Adults: Early Results from a Randomized Nutrition Intervention,” mimeo, UCLA. The World Bank (2005a). Ten Things You Never Knew About the World Bank, http://www.worldbank.org/tenthings/, accessed November 14, 2005. The World Bank (2005b) Infrastructure, http://www.worldbank.org/infrastructure/, accessed November 11, 2005. The World Bank (2005c). Brazil: Background Study for a National Rural Electrification Strategy: Aiming for Universal Access, The World Bank, Washington, D.C. The World Bank (2005d). World Development Indicators, Washington, D.C. The World Bank (2005e). Energy Working Notes: May 2005, Energy and Mining Sector Board, The World Bank, Washington D.C. The World Bank (2004). World Development Report 2004: Making Services Work for Poor People, World Bank, Washington D.C. The World Bank (2003). Infrastructure Action Plan. The World Bank, Washington D.C., July The World Bank (1994) World Development Report 1994: Infrastructure. The World Bank, D.C. 27

The World Bank (1993) World Development Report 1993: Investing in Health. The World Bank, D.C.

28

Data Appendix: There is no comprehensive source for electricity data over our period of analysis across Brazil. Most of the network was privatized in the 1990s, and the overseeing agencies, Operador Nacional do Sistema Eletrico (ONS) and Agencia Nacional de Energia Eletrica (ANEEL) were formed in the early 1990s and have little institutional memory for the period prior to their charters. In order to assemble our data set, we traveled to Brazil and met with professionals in the field at major electricity companies and government agencies in Brasilia, Sao Paulo, Rio de Janeiro, Curitiba, Salvador, and Foz do Iguaçu (Itaipu). The meetings with local professionals were informative not only in terms of data with which they provided, but also for the broader understanding of the development of the electricity network in Brazil. We collected data from the Ministério de Minas e Energia and ANEEL in Brasilia, Eletrobras, ONS, the Memoria de Eletricidade, and Furnas in Rio de Janeiro, Compania Hidroeletrica de Sao Francisco (CHESF) in Salvador, Copel in Curitiba, and Itaipu Binacional in Foz do Iguaçu. Data on the location and year of creation of plants was assembled from a database of important power plants from Sistema de Informações Georreferenciadas do Setor Elétrico (SIGEL) and a historical study of hydroelectricity in Brazil by the Memoria de Eletricidade. Data on the state of the network in each decade was assembled from a combination of sources. Data from the 1960s was procured primarily through feasibility studies conducted by the Canambra Engineering Consultants who did a comprehensive survey of electricity in Brazil in the 1960s, focusing on the South and South-Central Brazil. Inventories of the network as of the publication dates in 1965 and 1967 were included as part of the surveys, and maps were also included to show the placement of the network. CHESF also provided limited information about the state of the network in North Eastern Brazil from 1960 through the present. The 1970s data was put together primarily through maps which were drawn by Furnas and Eletrobras in 1973. Data from the 1980s is from a comprehensive inventory done by SIESE in 1987. The survey includes detailed data on both transmission lines and generation plants. Data from the 1990s is from a listing by SIGEL which is a survey of the current electricity network done by ANEEL. Data from 2000 is from both SIGEL and SINDAT, the database of the current electricity network done by ONS. We had the data from the inventories in each period entered into Excel spreadsheets by firms in India and Bangladesh. Data on line voltage was used to ensure comparability of 29

inventories conducted by different sources—only lines of at least 69 kilovolts were included as transmission lines—13 kilo volt lines were considered part of the distribution networks. Data from the tables were transferred into GIS maps which we then compared against maps which were drawn of the electricity network in each decade in order to insure the accuracy of our final decade-by-decade networks.

30

Appendix: Modeling Electricity Networks in Brazil: An Engineering Cost Approach

The model begins by selecting the locations of generation plants based on probabilities estimated in a probit equation based on geographic factors. An indicator for whether or not a hydropower plant occurs within a grid circle is regressed on an Amazon indicator which accounts for the lower probability of building power plants in the Amazon due to high materials costs, the average and maximum slope of the river within the grid circle, the log of the maximum flow accumulation (the number of raster grids which flow into each raster), and a water indicator which takes the value of one if the circle has a river or stream in it, zero otherwise.

The initial placement of transmission substations is randomly selected from the remaining grid points across Brazil. Two substations are allocated to each power plant, and a single line is assumed to pass through the three points.

Slope=.29

Slope=.15

Slope=.20

Slope=.35

Slope=.48

Slope=.61

Slope=.54

The cost of the transmission lines is calculated based on distance and an Amazon indicator. (High materials costs are assumed for transmission stations within the Amazon as transport costs are high). Costs are calculated by dividing each transmission line into 100 equally spaced points. 31

Points occurring within a perimeter of a grid point are assigned the average slope which has been calculated for the region surrounding the closest grid point.

The program randomly chooses alternative points and calculates the cost of the transmission lines through those points. If the line cost for the alternative points is lower, the alternative points are retained.

The process is repeated until a lowest cost allocation is achieved for each decade (30,000 iterations).

Distribution networks surrounding each transmission substation and power plant are generated. All points within a 50 km radius are allocated electricity.

32

The process is repeated for subsequent decades. In each decade, power plants are allocated to the highest probability points which have not yet received electricity. Initial locations for transmission substations are randomly chosen from the points across Brazil which have not yet received substations or power plants.

Alternative points are proposed for the new substations. Substations and plants from previous decades are not altered.

The program is again iterated until the lowest cost equilibrium is reached. Distribution networks within a 50 kilometer radius of the chosen points are again assigned. Grid points assigned to receive electricity in a given decade are given a value of 1, while those which are not projected to receive electricity in that decade are given a value of 0. The vector of projected indicator values for each grid point is then used as our instrument for electricity provision in the instrumental variable regressions.

33

Figures:

Figure 1: Increase in kilometers of transmission lines from 1950 through 2000 Source: SIESE

Figure 2: Inventory from Canambra report 1969

Figure 3: Inventory from SIESE report 1987

Source: Canambra (1969)

Source: SIESE (1987)

34

Figure 4: South Brazil Transmission as of 1967

Figure 5: South Brazil Transmission as of 1979

Source: Canambra (1967)

Source: PAREE (1979)

Figure 6: South Brazil Electricity Grid as of 1960

35

Figure 7: 1960s Transmission with Distribution

Figure 8: 1970s Transmission with Distribution

Figure 9: 1980s Transmission with Distribution

Figure 10: 1990s Transmission with Distribution

36

Figure 11: 2000 Transmission with Distribution

37

Figure 12: Construction of the river gradient instrument

Figure 13: Predicted Locations of Hydropower plants, actual plants, rivers and elevation

38

Figure 14: 1960’s modeled power allocation

Figure 15: 1970’s modeled power allocation

Figure 16: 1980’s modeled power allocation

Figure 17: 1990’s modeled power allocation

39

Figure 18: 2000 modeled power allocation

Figure 19: Low level of panel variation in model

Figure 20: 1960 county proxies

40

Table 1:  Probit Forecast of Placement of Generation  Plants  (step 1 of the engineering model that creates the  instrument)  Dependent variable:  Indicator for location has a  hydropower plant  Log of Maximum Flow  0.029**  Accumulation  (0.014)  0.044  Average Slope in the river  (0.030)  0.062***  Maximum Slope in the river  (0.012)  ‐0.753***  Amazon Indicator  (0.066)  Indicator for location has a river in  ‐0.030  it  (0.063)  Observations  33342  R‐squared  .  Standard errors in parentheses  *** p<0.01, ** p<0.05, * p<0.1 

41

Table 2:  Summary Statistics  GDP per capita  1970  1980  1990  2000  Industrial GDP pc  1970  1980  1990  2000 

Population Density  1970  1980  1990  2000 

(thousands of 2000 R$ per person)  Obs  Mean  Std. Dev.  Min  35257  1.383992 1.192956 0.045881 32528  3.416521 7.051991 0.060664 32528  2.90737 3.221777 0.088616 32528  3.671918 4.05901 0.461927

Max  47.90996  455.9149  91.90031  184.9774 

(thousands of 2000 R$ per person)  Obs  Mean  Std. Dev.  Min  32527  0.206988 0.442533 0 32528  1.055195 2.576341 0.000513 32528  0.603734 2.001518 7.78E‐06 32528  0.8449 2.735928 0.004097

Max  27.54974  211.3002  89.60116  112.1848 

(people per square  kilometer)  Obs  Mean  32527  11.0904 32528  14.28085 32528  17.15497 32528  20.3986

Max  8893.941  11729.97  12199.77  12915.98 

Std. Dev.  99.17917 139.1132 152.0696 181.7498

42

Min  0.015955 0.042236 0.089627 0.131608

Table 3:  First Stage for Instrumental Variables Regressions  Dependent Variable is Indicator for whether a grid point actually has electricity in the data 

Modeled Electricity  Indicator  Land Slope 

Fixed  Pooled Cross‐sectional   Random Effects  Effects  0.432***  0.082*** 0.080*** 0.355*** 0.118***  0.198***  (0.018)  (0.014)  (0.013)  (0.024)  (0.016)  (0.025)  0.021***  0.004*** 0.002*  0.027*** 0.003*  (0.002)  (0.001)  (0.001)  (0.003)  (0.002) 

Region Fixed  Effects?  N  Y  N  N  Y  .  State Fixed Effects?  N  N  Y  N  N  .  Fixed Effects?  N  N  N  N  N  Y  Observations  130111  130023  130023  130108  130020  130820  Number of  Locations for Fixed  Effects  32528  32506  32705  R‐squared  0.257  0.475  0.543  .  .  0.164  Robust standard errors clustered by 1960 counties in parentheses  *** p<0.01, ** p<0.05, * p<0.1  Notes:  Observations are by 16km spaced grid points.  Standard Errors are clustered by county. 

43

Table 4: Effect of Electricity on Population Density     OLS 

Instrumental Variables 

47.023***  36.511*** 20.956*** 67.663*** 72.703***  (4.061)  (3.710)  (2.340)  (8.164)  (28.211)  1.657  0.707  ‐0.401  0.548  0.472  Land Slope  (1.092)  (1.074)  (1.019)  (1.088)  (1.089)  Region Fixed Effects?  N  Y  N  N  Y  State Fixed Effects?  N  N  Y  N  N  Observations  130111  130023  130023  130111  130023  R‐squared  0.020  0.025  0.048  0.017  0.019  Robust standard errors clustered by 1960 counties in parentheses  *** p<0.01, ** p<0.05, * p<0.1  Notes:  Observations are by 16km spaced grid points.  Errors are clustered by county.    Lagged Electricity 

44

98.326*** (37.689)  ‐0.700  (1.049)  N  Y  130023  0.025 

Table 5:  Effect of Electricity on GDP per Capita  OLS  Instrumental Variables  0.667***  0.451*** 0.165  2.084*** 0.455  ‐0.601  Lagged Electricity  (0.170)  (0.122)  (0.105)  (0.443)  (1.450)  (1.482)  ‐ 0.043**  0.003  0.028** ‐0.033  0.003  ‐0.025*  Land Slope  (0.017)  (0.019)  (0.014)  (0.024)  (0.021)  (0.015)  Region Fixed Effects?  N  Y  N  N  Y  N  State Fixed Effects?  N  N  Y  N  N  Y  Observations  130111  130023  130023  130111  130023  130023  R‐squared  0.043  0.095  0.118  0.027  0.095  0.116  *** p<0.01, ** p<0.05, * p<0.1  Robust standard errors clustered by 1960 counties in parentheses  Notes:  Observations are by 16km spaced grid points.  Errors are clustered by county.   

45

Table 6:  Effect of Electricity on Industrial GDP per Capita  OLS 

Instrumental Variables 

0.318**  0.383*** 0.253*** 0.463  (0.138)  (0.051)  (0.052)  (0.302)  0.016  0.016  0.000  0.008  Land Slope  (0.011)  (0.014)  (0.011)  (0.015)  Region Fixed Effects?  N  Y  N  N  State Fixed Effects?  N  N  Y  N  Observations  130111  130023  130023  130111 R‐squared  0.025  0.041  0.071  0.025  *** p<0.01, ** p<0.05, * p<0.1  Robust standard errors clustered by 1960 counties in parentheses  Lagged Electricity 

0.158  (0.544)  0.018  (0.015)  Y  N  130023  0.040 

0.287  (0.567)  ‐0.000  (0.011)  N  Y  130023  0.071 

Notes:  Observations are by 16km spaced grid points.  Errors are clustered by county.   

46

47

Table 7:  Random Effects Instrumental Variable Regressions with Bootstrapped Errors  Industrial GDP per  GDP per Capita  Capita  Population Density  2.096***  0.790**  0.458***  0.229  30.088***  3.865  (0.313)  (0.054)  (0.209)  (4.428)  (6.287)  Lagged Electricity Provision  (0.124)  ‐0.034**  0.001  0.008  0.017**  2.567**  0.920  Land Slope  (0.014)  (0.012)  (0.007)  (0.007)  (1.085)  (1.101)  Region Fixed Effects?  N  Y  N  Y  N  Y  Observations  130111  130023  130111  130023  130111  130023  Number of fid  32528  32506  32528  32506  32528  32506  R‐squared  .  .  .  .  .  .  *** p<0.01, ** p<0.05, * p<0.1  Standard errors in parentheses  Notes:  Observations are by 16 km spaced grid points by decade.  Errors are bootstrapped with 50 repetitions.  Lagged  predicted electricity has been used as the instrument. 

48

Table 8:  Fixed Effects Regressions     OLS Panel Regressions  GDP per  Industrial  Population  Capita     GDP pc     Density  ‐0.056  0.051  4.856***  Lagged Electricity  (0.129)  (0.074)  (1.396)  Fixed Effects?  Y  Y  Y  Observations  130108  130108  130108  Number of fid  32527  32527  32527  R‐squared  0.063  0.040  0.008  Robust standard errors clustered by 1960 counties in parentheses  *** p<0.01, ** p<0.05, * p<0.1 

Instrumental Variable Regressions  GDP per  Industrial  Population  Capita     GDP pc     Density  2.453*** 0.382  ‐3.572  (0.917)  (0.473)  (11.348)  Y  Y  Y  130820  130820  130820  32705  32705  32705  0.044  0.038  0.006 

Notes:  Observations are by 16 km spaced grid points across Brazil by decade.    Errors are clustered by county as  of 1960.  Where a grid point is overlapped by more than one county circle, the assignment choice of clusters  between adjacent counties is random. 

49

50

Table 9:  Fixed Effects Regressions for 1980s‐2000  Industrial GDP  GDP pc  pc 

Population  Density 

8.312***  4.276***  10.487  Lagged Electricity  Access Indicator  (2.622)  (1.653)  (26.090)  Observations  98115    98115    98115  R‐squared  ‐0.090  ‐0.106  0.005  Number of fid  32705  32705  32705  *** p<0.01, ** p<0.05, * p<0.1  Robust standard errors clustered by 1960 counties in parentheses 

51

Returns to Electricity: Evidence from the Quasi-Random ...

access to modern energy, 1.1 billion people who lack access to clean water ..... companies in Brazil undertake prior to planning expansion of their networks.

3MB Sizes 3 Downloads 215 Views

Recommend Documents

The Returns to Nursing: Evidence from a Parental ...
Oct 11, 2017 - IL 60208, phone: 847-491-1908, [email protected]. ‡ ..... the average age and share of elderly people living in nursing homes are very similar in Denmark .... a large share of educational leave among nurses

The Returns to Nursing: Evidence from a Parental ... - WordPress.com
Jan 22, 2018 - on health care delivery and patient health outcomes. .... delivery across providers and across patient groups. ...... C.2 Proof of Proposition 1.

Returns to experience and seniority: Evidence from ...
Nov 30, 2017 - in the US for workers with 10 years of labor market experience, depending on the education group. The returns to tenure are also lower in Spain, with returns after 10 years in the same job being 30-54% of the returns for workers in the

Education Quality and Returns to Schooling: Evidence ...
Feb 4, 2017 - Keywords: education quality, returns to schooling, development accounting. ... states. Regional means range from 3.4% in the Northeast to 9.7% in the ...... application to estimating the effect of schooling quality on earnings.

Education Quality and Returns to Schooling: Evidence ... - CAEN/UFC
Feb 4, 2017 - We use migrant data to estimate returns to schooling of individuals who stud- .... quality variables are scarce, whereas data on earnings and ...

Returns to Quality and Location of College - Evidence ...
Jan 30, 2015 - quality, they find about a 0.04 increase in wages for a one standard ... education evaluated ten years after individuals' high school senior year, using ... The trade-off in all these studies is that they, by definition, measure local 

Evidence from Head Start
Sep 30, 2013 - Portuguesa, Banco de Portugal, 2008 RES Conference, 2008 SOLE meetings, 2008 ESPE ... Opponents call for the outright termination of ..... We construct each child's income eligibility status in the following way (a detailed.

Disaggregating the Returns to College
Nov 17, 2013 - postsecondary choices and option values which includes community college and estimates rates of return to vocational school, 2-year college, ...

Striking Evidence from the London Underground Network
May 16, 2017 - 3 The strike. On January 10, 2014, the Rail Maritime Transport union, the largest trade union in the British transport sector, announced a 48-hour strike of London Tube workers. The strike was scheduled to begin on Tuesday evening (21:

Striking Evidence from the London Underground Network
May 16, 2017 - We present evidence that a significant fraction of commuters on the London under- ground do not travel on their optimal route. We show that a strike on the underground, which forced many commuters to experiment with new routes, brought

evidence from the eurosystem's ltro
Jun 18, 2017 - (2016) further argue that the Eurosystem's liquidity ..... businesses, corporations, and sole proprietors engaged in ...... credit growth trends.

Evidence from the Bangladesh Garment
probability that children age 6-12 are currently enrolled in school. However ... Keywords: labor market opportunities, decision-making power, trade liberalization, ...

Evidence from the Great Depression - Vanderbilt University
Mar 30, 2011 - University of California, Davis and NBER. Abstract: A large body of cross-country .... Austria and spread to Germany and the United Kingdom eventually led to speculative attacks on those countries remaining on gold. ... The list of obs

RETURNS TO SCALE.pdf
Connect more apps... Try one of the apps below to open or edit this item. RETURNS TO SCALE.pdf. RETURNS TO SCALE.pdf. Open. Extract. Open with. Sign In.

Evidence from Goa
hardly any opportunity for business, less opportunity to enhance human ... labour market, his continuance in Goa or his duration of residence depends not only.

Evidence from Ethiopia
of school fees in Ethiopia led to an increase of over two years of schooling for women impacted by the reform .... education to each of nine newly formed regional authorities and two independent administrations located in ...... Technical report,.

Exploiting evidence from unstructured data to enhance master data ...
reports, emails, call-center transcripts, and chat logs. How-. ever, those ...... with master records in IBM InfoSphere MDM Advanced. Edition repository.

Testing Limits to Policy Reversal: Evidence from Indian ...
Sep 6, 2007 - privatization fell by 7.5 percent relative to private firms. .... number of factors that cloud the interpretation of these long&run results as the.

Testing Limits to Policy Reversal: Evidence from Indian ...
Sep 6, 2007 - Further analysis suggests that layoffs, combined .... changes in a companyms state of incorporation is predictive of market reaction to ... incorporate data on labor intensiveness and layoffs in companies under consider& ... would not c

Learning to Participate in Politics: Evidence From ...
Keywords: Political Behaviour, Impressionable Years, Jewish Expulsions, ... secondary and primary school teachers, because the Jewish population at ..... professional associations, trade guilds and many other occupations ..... Table 2 reports the dif