WPS7274 Policy Research Working Paper 7274 Road Improvement and Deforestation in the Congo Basin Countries Richard Damania David Wheeler Agriculture Global Practice Group May 2015 Policy Research Working Paper 7274 Abstract Road construction has often been viewed as the precursor demonstrates how the detailed analysis can identify areas to deforestation, especially in tropical forests. Traditional of high ecological priority as well as areas at high risk of responses to such threats have been reactive, with attempts forest loss. The paper contributes to several aspects of the to mitigate impacts through physical measures, or the estab- literature. First, it provides the most recent and reliable lishment of protected areas. These approaches often have estimates of the drivers of deforestation in the Congo Basin, not been entirely successful, especially in areas where eco- with the latest high-resolution satellite data on forest cover nomic potential is significant. This paper seeks to mitigate changes. Second, it presents novel estimates of biodiversity such conflicts by proposing a proactive approach to devel- threats by creating an index that combines and synthesizes opment planning and environmental policy. It develops a several measures of biodiversity loss and impacts. It then high-resolution spatial model of road improvement impacts develops an empirical framework for estimating the eco- that includes ecological risks and the economics of forest logical impacts of road improvement. Finally, the paper clearing. The approach is implemented by estimating the illustrates how the empirical framework can be used to potential impact of road upgrading on forest clearing and preempt impacts and avoid potential ecological damage. biodiversity in eight Congo Basin countries. The paper This paper is a product of the Agriculture Global Practice Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at rdamania@ worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Road Improvement and Deforestation in the Congo Basin Countries Richard Damania* David Wheeler Keywords: Deforestation, biodiversity, transport JEL Codes: Q2, Q3, O1 * Authors’ names in alphabetical order. The authors are respectively Lead Economist, World Bank and Senior Fellow, World Resources Institute. Road Improvement and Deforestation in the Congo Basin Countries 1. Introduction Conservation management in tropical forests has traditionally focused on demarcation and protection of relatively large areas that are deemed critical for biodiversity conservation. This strategy seeks to minimize ecological damage by preventing or severely restricting road improvements that increase the profitability of forest clearing within protected areas. In contrast, road improvements often play a central role in agricultural development strategies for poor regions. Potential conflict over the desirability of road improvements is particularly high in forested regions with significant agricultural potential. When protected-area strategies confront this conflict, they may fail to protect critical biodiversity for two reasons. First, governments may seek to minimize economic opportunity costs by siting protected areas in remote regions with low agricultural potential that may not coincide with the areas of highest ecological value. Second, attempts to restrict road improvements in protected areas with strong agricultural potential may fail because economic interests overwhelm the limited resources and political support of conservation managers. This paper seeks to mitigate such conflicts by developing a high-resolution spatial model of road improvement impacts that includes both ecological risks and the economics of forest clearing. We implement our approach by estimating the potential impact of road upgrading on forest clearing and biodiversity in eight Congo Basin countries: Cameroon, Central African Republic, Equatorial Guinea, Gabon, Republic of Congo, Democratic Republic of the Congo, Rwanda and Burundi. The remainder of the paper is organized as follows. Section 2 reviews prior empirical research on the economics of forest clearing. In Section 3, we motivate our exercise with a theoretical economic model of road improvement and deforestation. Section 4 describes our 2 spatially-formatted database, while Section 5 specifies and estimates a deforestation model that incorporates the impact of road improvement. In Section 6, we explore the implications of our results for local, regional and national forest clearing. Section 7 develops an empirical framework for estimating the ecological impacts of road improvement. Section 8 applies our empirical framework to an assessment of local and regional impacts in the Congo Basin countries. Section 9 summarizes and concludes the paper. 2. Prior Research Empirical research has provided many useful insights about the determinants of forest clearing. The results are generally consistent with an economic model in which the conversion of forested land varies with potential profitability. Nelson and Chomitz (2009) and Rudel, et al. (2009) have studied this relationship across countries over multi-year intervals. Within countries, numerous econometric studies have estimated the impact of economic, social and geographic drivers on deforestation during multi-year intervals. Some studies have used aggregate data for states, provinces or sub-provinces (e.g. studies for Brazilian municipios by Pfaff (1997) and Igliori (2006), and Mexican states by Barbier and Burgess (1996)). Many studies have also used GIS-based techniques to obtain multi-year estimates at a higher level of spatial disaggregation (e.g., Cropper, et. al. (1999, 2001) for Thailand; Agwaral, et al. (2002) for Madagascar; Deininger and Minton (1999, 2002), Chowdhury (2006) and Vance and Geoghegan (2002) for Mexico; Kaimowitz, et al. (2002) for Bolivia; and De Pinto and Nelson (2009) for Panama). In rarer cases, studies have used annual national or regional aggregate time series for extended periods (e.g. Zikri (2009) for Indonesia; Ewers, et al. (2008) for Brazil). 3 While econometric work on long-run deforestation drivers is well-advanced, previous data problems limited treatments of economic dynamics to theoretical work and simulation. Arcanda et al. (2008) and others studied the theoretical relationships between macroeconomic drivers and forest clearing. Notable simulation exercises include Cattaneo (2001) for Brazil and San et al. (2000) for Indonesia. Recently, the advent of monthly and annual remote sensing databases has led to the first spatial estimation exercises that explicitly incorporate economic dynamics (Wheeler et al., 2011; Dasgupta et al., 2014). Direct impact studies in Latin America using the new databases have included high-resolution work on new road construction and deforestation in Brazil, where satellite monitoring has been available for a longer period (Laurance, et al., 2009) and Bolivia, Panama, Paraguay and Peru (Reymondin et al., 2013). To our knowledge, the present paper is the first extension of such work to Sub-Saharan Africa. Li et al. (2015) have investigated the potential impact of decreased travel time from improvement of road links in the DRC, but their cross-sectional analysis employs land use information (JRC 2003) that predates the new high-resolution satellite data. 3. Modeling the Economics of Road Improvement and Deforestation As we noted in the introduction, this research aims to develop an analysis of road upgrading and forest clearing that incorporates both economic and conservation concerns. Our objective is a methodology for prospective assessment that can inform infrastructure planning at the outset. To motivate the exercise, we consider the potentially-adverse impact of traditional road improvement planning, in which decision-making is sequential: Decisions on road improvement projects in an area are made first, followed by an EIA (Environmental Impact Assessment) that seeks to mitigate forest clearing by strengthening environmental management rather than affecting the selection of projects. Our modeling exercise shows why coordinated 4 infrastructure planning in such a sequential decision regime, while otherwise desirable for its direct economic contribution, may actually reduce welfare because it increases deforestation and the associated ecological impacts. It also identifies cases where decentralized decision making generates greater (lesser) deforestation when there are diminishing returns to forest clearance. Consider two regions, labeled i and j, which sell their produce at a market at given distances di and dj respectively. Each region is endowed with a given amount of land that can either be left forested, or converted to some alternative use such as agriculture whose outputs are transported to the market and sold at a given price. Let = + be the total endowment of land in region i, where is forested land and is agricultural land. The payoffs in region i to each activity are given by: 2 1. Π = � − (1 − )� − + ( − ) where is the exogenous market price and 0 ≤ < 1 is an index of road quality, with q = 0 representing an unimproved forest track. Improvements in road quality (q) lower transport costs and increase the profitability of agricultural production. With convex costs there are assumed to 2 be diminishing returns to forest conversion ( ). Finally, forests left in their natural state generate a return of v, which could include livelihood and other unmarketed benefits obtained from the forest. Maximizing (1), taking road quality as given, yields the optimal level of clearing (output) in region i: − (1− ) 2. = 2 , where = + . The corresponding indirect profit function from the land use decision is: ( − (1− ))2 3. Π = + 4 5 Region j is symmetric and its specification is suppressed for brevity. It is clear from equation (3) that higher market prices or improvements in road quality lead to greater land 1 conversion (i.e. = 2 > 0 and = 2 > 0). This straightforward result reflects the fact that higher profitability of agriculture or increased market access renders deforestation and land conversion more profitable. The less well known question is how these incentives might vary under different decision making regimes. In what follows, we compare deforestation levels in each region under two contrasting forms of management. In the first the decision on road quality improvements is made autonomously in each jurisdiction. In the alternative the budget allocation is coordinated to maximize the joint welfare of the two regions. In what follows we assume that environmental impacts are not considered in the decision making process. 1 Consider first the case of autonomous decision making. For simplicity assume that there is a fixed budget Bi available for road quality improvements in region i. The cost of improving road quality is given by: Bi=rdiqi. If the budget constraint binds then road quality is given by: 4. = Thus road quality declines with distance (di) and the costs of road construction (r). Substituting in (2) the amount of land conversion is: − (1− ) 5. = 2 In contrast. under coordinated management the budget is allocated to maximize the joint welfare of the two regions: 1 This describes a not uncommon situation where decision making is sequential, with decisions on road location made first, followed by an EIA process that typically seeks to mitigate impacts by strengthening environmental management rather than altering road routing. 6 2 ( − (1− ))2 � − (1− )� 6. = + + + 4 4 subject to − − Which yields solutions for the optimal improvements in road quality: +( − ) +( − ) 7. = and = (+ ) (+ ) where = − , for (k = i,j) ( + � + �) For later use it is instructive to observe that = 2 > 0. �+ � As land conversion costs rise in a region, road quality is improved. Intuitively, road investments are made to equate the marginal payoffs from agricultural sales from each region. With diminishing returns to land conversion, it eventually pays to invest in the higher cost region. The following Lemmas compare road investments and deforestation rates under the different regimes: Lemma 1 If the cost of land conversion between regions differs sufficiently, then road improvements in the high cost region will be greater under coordinated management than under ( +2 ) autonomous management (i.e. < if > ( +2 ). ) Intuitively, under autonomous management, budgets available to each region are fixed, whereas under coordinated management the region with the higher costs of land conversion could receive a higher allocation to equalize marginal payoffs. An implication of this result is that total deforestation rates may differ across management regimes. This result is summarized in Lemma 2 and illustrated in Figure 1 below. 7 Lemma 2 Total deforestation rates will differ under the management regimes and will be higher under coordinated management if the cost of land conversion is sufficiently large in one of the (+2 ) regions (i.e. + < + if > �+2 � ). When differences between regions are large, under coordinated management total payoffs are maximized by ensuring that the region with the comparative advantage in land clearing receives greater support on the margin, thus expanding the total volume of land cleared. This result follows from the convex cost of land conversion, which implies that the incremental returns to further deforestation will be lower in the region with a sufficiently high level of land conversion. By implication, coordinated regional infrastructure planning, while promoted for its economic benefits, may actually increase deforestation in a sequential decision regime that selects road projects first and introduces environmental impact analysis ex-post. As an alternative, the approach developed in this paper seeks to improve the decision process by enabling simultaneous consideration of potential road projects and deforestation impacts. Figure 1: Road planning regimes, road quality and deforestation Planning regimes and deforestation rates 8 Marginal net benefits and road quality 4. Data 4.1 Congo Basin Roads Figure 2 and Table 1 summarize information for seven Congo Basin countries in digital road maps (shapefiles) provided by the Africa Infrastructure Country Diagnostic (AICD) of the African Development Bank Group, augmented by information for the Democratic Republic of the Congo provided by Delorme, and for Equatorial Guinea by DIVA-GIS. 2 As Table 1 shows, the AICD database includes quality information on 1,710 segments of 380 roads, with data on both road surfaces and road conditions for 941 segments. All seven 2 Data available online at http://www.diva-gis.org/gdata. 9 countries are amply represented, with the exception of joint information on surfaces and conditions for roads in Gabon. Table 1: Congo Basin roads data from the AICD Segment Data on Road Road Surface and Country Roads Segments Condition Burundi 39 158 37 Cameroon 103 725 397 Central African Republic 30 103 103 Democratic Republic of the Congo 121 347 180 Gabon 51 122 7 Republic of Congo 25 155 144 Rwanda 11 100 73 Total 380 1,710 941 Table 2 summarizes the information on road surfaces and conditions provided by the AICD database. Among rated road segments, 32.6% are asphalt, 26.7% gravel, and 40.7% earth. Asphalt is undoubtedly over-represented in the sample, but this provides useful information for the econometric analysis. Road conditions vary widely across segments, with 30.6% rated good, 25.6% fair, 35.7% poor and 8.1% very poor. Higher quality appears to be over-represented, but again this provides useful information for the econometric analysis. Table 2: Road surface and condition, AICD database Road Very Good Fair Poor Total Surface Poor % Asphalt 102 104 62 39 307 32.6 Gravel 69 41 111 30 251 26.7 Earth 117 96 163 7 383 40.7 Total 288 241 336 76 941 % 30.6 25.6 35.7 8.1 10 Figure 2: Congo Basin roads 11 4.2 Forest Clearing Hansen et al. (2013) have recently published the first high-resolution, consistently-derived estimates of global forest clearing. The data are available as tiff panels at 30 m spatial resolution for 2000-2012. We have converted the tiff panels to annual files in which cleared pixels are assigned the value 1 in the year when most clearing occurred. Uncleared pixels are assigned the value 0. For the exercise in this paper, we have created a tractable database by summarizing the Hansen estimates in 2.7 km grid cells. Each cell contains 8,100 30 m Hansen pixels, so total counts of Hansen cleared pixels within cells are equivalent to deforestation rates. Figure 3 displays gridded Hansen estimates for the border region of Cameroon, Gabon, Republic of Congo and Central Africa Republic. Each grid cell is color-coded for cumulative percent deforested in 2000-2012. Figure 3 reveals a striking pattern of deforestation along some of the roads in the AICD database. In other cases, however, deforestation is much less pronounced. Figure 3: Forest clearing and road networks, 2000-2012 12 To illustrate characteristic patterns of deforestation in road corridors, Figure 4 plots average cleared pixels in grid cells at varying distances from six roads in Cameroon, Central African Republic and Republic of Congo. Although factors other than distance introduce significant variation, each graph displays a pattern of decline from steep to nearly flat over the interval 0 to 15 km, with the steepest portion in the range 0-5 km. The basic shape can be captured by an exponential function that we employ for the econometric analyses in this paper: = 0 ℎ1 where β1 < 0 p = Hansen pixels cleared per grid square h = Distance from the road For econometric work, we translate the function to its linear form in the logarithms of h and p. 5. Estimation of the Forest Clearing Model The estimation exercises in this paper draw on the insights of previous research and the theoretical model in Section 3 to incorporate seven critical determinants of forest clearing in road corridors: road quality, distance from the road, transport cost to the nearest market center, the agricultural opportunity value of the land, terrain elevation, legal protection status and the incidence of violent conflict. Our regressions for the combined sample of Congo Basin countries employ a measure of travel time as a proxy for transport cost. To support a detailed impact analysis, we also estimate separate regressions for the DRC that incorporate an explicit measure of transport cost developed by Ali et al. (2015). We avoid spatial autocorrelation problems by using mean Hansen pixels cleared within grid cells whose distance from a road segment is aggregated to 2.7 km intervals on both sides of the segment. 13 Figure 4: Average pixels cleared per grid cell vs. distance from road, 2000-2012 14 We specify the basic estimating equation in log-log form: log hijt = α0 + α1 log dij + α2 log pidij + α3 log qi + α4 log mi + α5 log ci + α6 log ei + α7 log vi + Σβtyt + Ɛit where hit = Hansen pixels cleared, grid cell i, road segment j, year t dij = Distance of cell i from road segment j pi = Legal protection status of cell i qi = Condition of road segment j mi = Travel time from cell i to the nearest urban center (transport cost for DRC). ci = Agricultural opportunity value of cell i ei = Elevation of cell i vi = Conflict incidence in cell i yt = Dummy variable with value 1 in year t and 0 otherwise Ɛit = Random error term 5.1 Variable Measures We stack data sets for individual years in the period 2000-2012, allowing for variable mean clearing in grid cells by year. All variables are calculated as centroid values for 2.7 km Hansen grid cells. Data and sources are as follows. Hansen pixels cleared: per 2.7 km grid cell, the number of 30 m pixels identified by Hansen, et al. (2013) as cleared, by year, for 2000-2012. Distance from road segment: distance from the centroid of each grid cell to the nearest road, calculated in ArcGIS 10. Legal protection status: 1 if the parcel is in a protected area identified by the World Database on Protected Areas (WDPA); 0 otherwise. The WDPA shapefile has been downloaded from http://www.protectedplanet.net/. Condition of the road segment: road surface type and road condition, obtained from the Africa Infrastructure Country Diagnostic (AICD). Surface types are earth, gravel and asphalt. Road conditions are identified as good, fair, poor and very poor. We have translated each variable into ordered cardinal values for estimation: Surface: earth (1), gravel (2), asphalt (3); Condition: very poor (1), poor (2), fair (3), good (4). Travel time: time from a road segment’s mean point to the nearest urban center with a population of 50,000 or greater, as estimated by Nelson and Uchida (2009). Raster resolution: .0083 decimal degrees. 15 Transport cost (DRC): total transport cost from each Hansen grid centroid to the nearest urban center with a population of 50,000 or greater, developed from a network analysis. Agricultural opportunity value: mean value for a grid cell, calculated from the high-resolution global grid developed by Deveny, et al. (2009). Raster resolution: .0025 decimal degrees. Elevation: Average elevation for a grid cell, calculated from the CGIAR-SRTM data set (3 seconds resolution), aggregated to 30 seconds resolution by DIVA-GIS (http://www.diva- gis.org/gdata). Conflict incidence: Armed conflict fatalities per unit area, 1997-2007, calculated by Ali, et al. (2015) at 0.017 decimal degrees resolution from data in the Armed Conflict Location Events Dataset (ACLED) (Raleigh, 2010). 5.2 Results Table 3 presents final estimates for the Congo Basin countries using several estimation techniques. Simultaneity bias may be significant in this context, since forest clearing and road placement are jointly determined in a properly-specified spatial economic model. Our measure of travel time to the nearest urban center along existing roads should therefore be treated as endogenous in our estimating equation. To address estimation bias, we introduce Euclidean distance to the nearest urban center as an instrumental variable. In Table 3, the first two columns present OLS results for travel time and Euclidean distance. Columns (3-5) present three results that employ Euclidean distance as an instrument for travel time: standard 2SLS, GLS (IV) with covariance matrix adjustments for road-specific error variances, and robust regression (IV). Our results are distilled from numerous experiments that tested the interactions of distance to road with road quality, travel time, agricultural opportunity value, elevation and protection status. The experiments were critical for determining whether these variables affect the slope of the relationship between forest clearing and distance from the road. Only protection status has a strong, consistent effect on the slope, so we have retained the interaction of protection and distance in Table 3. We have included country dummies to control for systematic differences in 16 forest-clearing intensity. Prior experimentation also revealed that road surface type has no significance for forest clearing, controlling for road condition, so we have excluded this variable from the final regressions. The estimated coefficients in Table 3 have generally high significance, and their signs are consistent with prior expectations. Among the IV regressions, the GLS corrections for non- uniform error variances across roads result in substantially higher standard errors (and lower t- statistics) but, with the exception of conflict intensity, all variables retain significance levels of 95% or higher. We use the GLS (IV) estimates for our discussion of results. The most critical variable for this exercise is distance from the road. In Table 3, the results for this variable are strong and consistent across estimators: Ceteris paribus (in GLS (IV)), forest clearing intensity declines 0.67% with each 1% increase in distance. The result for protection status indicates that the relationship between clearing intensity and distance steepens significantly in protected areas. Our results for road quality are also consistent across specifications, and for numerous experiments that test interactions with distance from the road. Road surface (earth, gravel, asphalt) never has a significant effect, 3 but the impact of road condition (1=very poor; 2=poor; 3=fair; 4=good) is large and highly significant. Neither road surface nor road condition interacts significantly with distance from the road. Among other regression variables, forest clearing is negatively related to travel time to the nearest market center (elasticity -.63 in GLS (IV)) and elevation (elasticity -.26), and positively related to land opportunity value (.07) and conflict intensity (.03). As expected, our IV results for travel time to the nearest urban center differ markedly from the OLS results. 3 We have tested the effect of road surface using categorical variables as well as the previously-described cardinal measure. 17 Table 3: Regression results (all non-dummy variables in log form) Dependent Variable: Cumulative Hansen pixels cleared (1) (2) (3) (4) (5) OLS OLS 2SLS GLS (IV) Robust (IV) Distance from road -0.535 -0.672 -0.672 -0.672 -0.658 (68.68)** (102.73)** (102.73)** (24.18)** (102.11)** Protected area x Distance from road -0.132 -0.164 -0.164 -0.164 -0.162 (22.59)** (28.32)** (28.32)** (4.28)** (28.48)** Road condition 0.209 0.205 0.205 0.205 0.219 (14.80)** (14.48)** (14.48)** (2.15)* (15.71)** Travel time to nearest urban center -0.471 -0.628 -0.628 -0.583 (37.35)** (30.57)** (4.84)** (28.81)** Euclidian distance to market center -0.236 (30.57)** Land opportunity value 0.031 0.069 0.069 0.069 0.067 (5.95)** (14.01)** (14.01)** (2.04)* (13.71)** Elevation -0.185 -0.258 -0.258 -0.258 -0.244 (15.79)** (22.41)** (22.41)** (3.23)** (21.55)** Conflict intensity (1997 - 2007) 0.025 0.031 0.031 0.031 0.018 (8.52)** (10.65)** (10.65)** (1.64) (6.45)** D2002 0.795 0.794 0.794 0.794 0.791 (27.88)** (27.68)** (27.68)** (30.68)** (28.01)** D2003 1.079 1.078 1.078 1.078 1.077 (37.91)** (37.67)** (37.67)** (37.51)** (38.21)** D2004 1.308 1.306 1.306 1.306 1.307 (45.98)** (45.71)** (45.71)** (42.66)** (46.41)** D2005 1.547 1.546 1.546 1.546 1.543 (54.48)** (54.17)** (54.17)** (48.81)** (54.87)** D2006 1.749 1.748 1.748 1.748 1.748 (61.69)** (61.35)** (61.35)** (53.31)** (62.29)** D2007 1.952 1.951 1.951 1.951 1.952 (68.91)** (68.52)** (68.52)** (58.06)** (69.61)** D2008 2.082 2.081 2.081 2.081 2.079 (73.54)** (73.12)** (73.12)** (61.97)** (74.15)** D2009 2.262 2.261 2.261 2.261 2.257 (79.94)** (79.50)** (79.50)** (65.67)** (80.55)** D2010 2.454 2.453 2.453 2.453 2.444 (86.78)** (86.30)** (86.30)** (70.15)** (87.27)** D2011 2.549 2.548 2.548 2.548 2.539 (90.16)** (89.67)** (89.67)** (72.52)** (90.69)** D2012 2.641 2.640 2.640 2.640 2.630 (93.45)** (92.93)** (92.93)** (74.81)** (93.96)** D[Dem. Rep. of the Congo] 2.043 2.095 2.095 2.095 2.132 (31.26)** (31.84)** (31.84)** (6.67)** (32.89)** D[Cameroon] 0.336 0.322 0.322 0.322 0.279 (5.05)** (4.81)** (4.81)** (0.98) (4.24)** D[Gabon] 0.594 0.304 0.304 0.304 0.415 (5.35)** (2.73)** (2.73)** (0.60) (3.79)** D[Republic of Congo] 0.547 0.387 0.387 0.387 0.411 (8.01)** (5.67)** (5.67)** (1.16) (6.11)** D[Central African Republic] 0.164 0.114 0.114 0.114 0.062 (2.43)* (1.70) (1.70) (0.35) (0.92) D[Rwanda] 0.791 0.805 0.805 0.805 0.805 (9.38)** (9.50)** (9.50)** (2.21)* (9.63)** Constant 5.339 4.252 6.985 6.985 6.668 (46.48)** (39.65)** (45.28)** (7.06)** (43.88)** Observations 44743 44743 44743 44743 44743 R-squared 0.56 0.56 0.56 0.56 0.56 Absolute value of t statistics in parentheses * significant at 5%; ** significant at 1% 18 The yearly dummy variables increase steadily in size, reflecting the increase in cumulative clearing, but the increments decrease markedly during the period 2000-2012. All country dummies except CAR’s are highly significant (and highly-varied) in the 2SLS and robust (IV) results, although the GLS (IV) results suggest that the large positive estimate for the DRC is the most robust in the set. 4 6. Implications for Forest Clearing We explore the implications of our results with the GLS (IV) estimates in Table 3 and the Appendix table for DRC, employing a three-step approach: (1) We use the regression variables � to predict forest clearing ( 0 ) for 520,732 2.7 km pixels in the eight Congo Basin countries. Where road condition estimates are missing (most often for rural tertiary roads), we assume very poor conditions (value 1). (2) We simulate the effect of upgrading by progressively resetting all � road conditions to higher values (2,3,4) and re-predicting forest clearing ( 1 ) for all pixels. (3) � To ensure consistency with the initial pattern of deforestation, we form the ratio [ρ = ( � 1 )/( 0 )] ̂ = ρf]. and multiply by actual clearing (f) to obtain the final prediction result [ 6.1 Micro-Results: Kasese, DRC Using the GLS (IV) results in Table 3, we illustrate our approach and the micro- implications with a road segment in the Kasese area of Maniema Province, DRC (Figure 5). We simulate upgrading with two adjustments to the data. Starting with the initial condition (1 - very poor), we improve the condition in unit increments (2, 3, 4). At the same time, we adjust the data on travel time to the nearest urban center to reflect the improvements. For this adjustment, 4 As previously noted, we have also tested the model separately for the DRC. We include the DRC results in Appendix 1 and explore their implications in Section 6.2. In comparison with Table 3, we obtain identical signs and simlar significance levels for distance from the road, its interaction with protected-area status, road condition, Euclidian distance from the nearest urban center, and elevation. The DRC parameter estimates for road condition are higher than the Congo Basin estimates in Table 3. Some variation in estimated parameter values is unsurprising, given the huge difference in degrees of freedom (44,743 observations for the Congo Basin vs. 13,422 for the DRC). 19 we estimate average road speed by road condition from estimates for 12,322 road segments in the DRC produced by Ali, et al. (2015). Figure 5: Kasese, Maniema Province, DRC Table 4 presents our results, along with derived travel time multipliers for a road segment whose initial condition is very poor. In this condition, average travel speed on the segment is 24 km/hr. Progressive improvements in the segment’s condition increase average speed to 36, 55 and 73 km/hr, respectively. With increased speed, travel time to the nearest urban center falls to 75%, 49% and 33% of the initial time. 5 Table 4: Road condition and travel time to nearest urban center Average Travel Road Speed Time Condition (km/hr) Multiplier 1 24 1.00 2 36 0.75 3 55 0.49 4 73 0.33 5 A reduction in travel time should also produce an increase in the opportunity value of land in the road corridor. The opportunity value measure employed for this exercise is based on physical measures (e.g., soil quality) and does not incorporate the effect of distance to market. We are therefore unable to estimate the effect of road quality improvement on this variable. By implication, our results should be interpreted as conservative estimates of road improvement impacts on forest clearing. 20 Figure 6 presents our forest clearing results for the Kasese road segment as its condition improves from very poor (1) to good (4). With the road in very poor condition (value 1), 34% of previously-forested land is cleared within 200 meters of the road. Further away, clearing declines to 11% at 1 km, 7% at 2 km and 2% at 10 km. When the segment’s condition is improved to fair (value 3), predicted clearing percent at 200 m, 1 km, 2 km and 10 km is 66%, 22%, 14% and 5%, respectively. With further upgrading to good condition (value 4), clearing at the four distances increases to 91%, 30%, 19% and 6%, respectively. Two patterns are noteworthy in these results. First, upgrading from very poor to good produces near-complete deforestation within a narrow corridor straddling the road. Second, deforestation intensity falls rapidly as distance from the road increases. Despite this decline, the overall spatial magnitude of road improvement’s impact is striking. To illustrate, suppose that we adopt 5% clearing as the criterion for identifying pixels at the outer margin of corridor deforestation. For the road in poor condition, this yields an affected corridor 6 km wide (3 km on either side of the road). Upgrading to good condition widens the affected corridor to 26 km. Figure 6: Effect of road quality on forest clearing intensity 21 6.2 Regional Results -- Eastern DRC We explore the regional implications for eastern DRC, using the DRC-specific regression results reported in the Appendix because they incorporate a more precise measure of transport cost developed by Ali, et al. (2015). 6 Figures 7, 8 and 9 illustrate the predicted impact of road upgrading for adjoining parts of Kasai Oriental, Equateur, Orientale and Kivu Provinces of eastern DRC. Figure 7 displays cumulative forest clearing from 2000 to 2012 in our Hansen 2.7 km pixels. Colors range from dark green for clearing below 5.3% to dark brown for clearing between 59.2% and 99.9%. Intensive clearing is almost entirely within relatively narrow corridors along roads between urban areas, with clusters of clearing in settled areas where road networks are relatively dense. Clearing is generally moderate: typically 5-16% in rural areas, 5-33% in areas with denser road networks, and, in rare cases, higher than 33% near major settlements. For an instructive contrast, Figure 8 uses the same color coding to display the predicted result of generalized road upgrading to good condition (value 4). Changes are particularly striking in Kasai Oriental and Kivu: Clearing near many rural roads moves into the range 33%- 59% and large areas near dense road networks move into the range 59% -99.9%. The differences between Figures 7 and 8 are displayed in Figure 9, which provides a striking picture of the widespread “browning” that accompanies generalized road upgrading. 6.3 National Results -- DRC 6 From Ali, et al. (2015), we incorporate the impact of road upgrading on transport cost using an approach analogous to the method illustrated in Table 4. We calculate unit transport cost (transport cost to the nearest market center divided by distance) for road condition 4 (good). For each pixel, we multiply this value by distance to the nearest market center to estimate transport cost after upgrading. 22 Figure 10 extends the regional results to the national level for DRC, color-coding by predicted change in the deforestation rate along road corridors 2 km wide. Our results are presented at the sub-district level. For each sub-district, we compute mean deforestation rates Figure 7: Eastern DRC: Percent clearing along roads before upgrading Figure 8: Eastern DRC: Percent clearing along roads after upgrading 23 Figure 9: Eastern DRC: Change in percent clearing along roads with upgrading 24 before and after upgrading, for pixels whose centroids lie within 2 km corridors straddling all roads in the sub-district. Figure 10 displays changes in mean deforestation rates produced by road upgrading to level 4 (good). The heaviest impacts (increases of 12.6% to 22.3%) are evident in west Orientale, east Equateur, central Kasai-Occidental, northeast Kasai-Oriental and central Maniema. Adjacent areas in all five provinces also have significant impacts (7.7% to 12.6%). The heaviest impacts are concentrated in relatively isolated rain forest areas with poor roads, since market access for these areas would be most improved by upgrading. Overall, our results indicate that 10-20% increases in deforestation would be common after upgrading in rain forest road corridors. Figure 10: Changes in road corridor deforestation with generalized upgrading \\\ 25 7. Incorporating Biodiversity Deforestation may reduce or eliminate critical habitat for endangered animal and plant species. Effective protection of biodiversity requires a variety of approaches that are tailored for different political, economic and geographic conditions. In the Congo Basin countries, three broad approaches have emerged: (1) Traditional strategies that focus on monitoring, enforcement and local community outreach for the protected areas identified in Figure 11. Figure 11: Protected areas in the Congo Basin countries (2) Strategies that identify ecologically-representative areas where scarce administrative and financial resources can be focused on the conservation components of integrated development plans. To illustrate, Figure 12 displays eleven “Priority Landscapes” in the Congo Basin countries that have been identified by a consortium of national and international conservation planners (WWF, 2000). 26 Figure 12: Priority landscapes in the Congo Basin countries Source: WWF (2000) (3) “Ecological gradient” strategies that use high-resolution biodiversity measures to steer infrastructure development toward sites where ecological damage will be minimized. An illustrative planning problem is displayed in Figure 13, which overlays roads and an ecological gradient that measures mammal concentrations in the Cameroon portion of the Sangha Tri- National Priority Landscape. Figure 13: Sangha Tri-National Priority Landscape, Cameroon Source: IUCN (2010), Figure 3, p. 90 27 As Figure 13 shows, some road corridors are in areas of modest mammal concentration, while others pass through areas of high concentration. With such information, an ecological gradient strategy attempts to minimize ecological damage by favoring road improvements in areas of modest concentration. We adopt the ecological gradient approach for this paper, because it incorporates both economic and ecological concerns into infrastructure planning. More generally, ecological gradient strategies require high-resolution data on concentrations of diverse species and their vulnerability to encroachment and habitat loss. This paper mobilizes such information to develop ecological gradients (henceforth ecogradients) for the Congo Basin countries. We develop ecogradients for animal species using digital range maps (shapefiles) provided by the International Union for the Conservation of Nature (IUCN) and BirdLife International. Table 5 summarizes the results of range map overlays for the Congo Basin and global tropical forests. Among species classes, the Congo Basin is most heavily represented among birds (1,141/8,370 (13.6%)) and mammals (484/3,547 (13.6%)), followed by amphibians (264/4,467 (5.9%)) and reptiles (37/1,744 (2.1%)). Overall, the Basin accounts for 10.6% of global tropical forest species. Table 5: Animal species counts for tropical forest areas Congo Congo Basin Global Basin Tropical Tropical Percent Forest Forest of Areasa Areasa Total Amphibiansb 264 4,467 5.9 Birdsc 1,141 8,370 13.6 Mammalsb 484 3,547 13.6 Reptilesb 37 1,744 2.1 Total 1,926 18,128 10.6 a Based on a Vegetation Continuous Field (VCF) threshold value of 25 in 2000. See Hansen, et al. (2003). b Shapefiles provided by IUCN at http://www.iucnredlist.org. c Shapefiles provided by BirdLife International on request. 28 7.1 Ecological Gradients for Animal Species 7.1.1 Species Density A useful ecogradient database should have tractable size, while incorporating sufficient spatial resolution for realistic local applications. This exercise employs a 5 km spatial grid that covers the Congo Basin with 133,782 square cells. We measure species density by overlaying the grid on 1,926 range maps for Basin species and registering the presence or absence of each species in each grid cell. 7.1.2 Species Vulnerability Species density provides critical information for developing ecogradients, but at least two other elements are needed: (1) Geographic vulnerability, which can be proxied by endemicity: the proportion of each species’ range that lies within each grid cell. Species that reside in very few grid cells may be particularly vulnerable to habitat encroachment. Endemicity can be computed from the database constructed for this paper. (2) A measure of extinction risk that adds the insights of the international scientific community. Recent work by Mooers, et al. (2008) has explicitly modeled the relationship between extinction probability and the risk indicator that is provided for each species in the IUCN and BirdLife databases. 7 Endemicity Our ecogradient database includes endemicity -- the percentage of each species’ range that is found in each grid cell. Endemicity treats all species equally at the global level, since each species has a total count of 1. Total endemicity for each grid cell -- the sum of its species 7 The IUCN's current classification categories are Critically Endangered, Endangered, Vulnerable, Near Threatened and Least Concern. 29 endemicity measures -- assigns higher values to cells inhabited by species whose ranges are relatively limited. By implication, forest clearing in higher-value cells may be particularly destructive for remaining critical habitat. Extinction Risk Species differ in vulnerability for many reasons that are not captured by our endemicity measure. To incorporate these factors, we use the threat status code assigned to each species by the IUCN’s Red List. An appropriate measure of vulnerability in this context is extinction risk, so we convert Red List status codes to extinction probabilities using the methodology of Mooers, et al. (2008). For species indicator construction, we normalize these probabilities so that a weight of 1.0 is assigned to species in the highest category (Critically Endangered). Table 6 tabulates conversions from Red List codes to normalized species weights, using four probability assignments. Three employ IUCN estimates to derive measures of extinction probability over the next 50, 100 and 500 years. The fourth draws on recent work by Isaac, et al. (2007), who combine a direct extinction risk measure with a measure of each species' isolation on a phylogenetic tree. 8 8 A phylogenetic tree is a branching tree diagram that traces the evolutionary descent of different species from a common ancestor. Species in sparse (isolated) branches of a phylogenetic tree are relatively unique, since they share common descent patterns with fewer other species. 30 Table 6: Normalized species aggregation weightsa Normalized Extinction Probabilities IUCN: Future Years IUCN Isaacb 50 100 500 Code Status CR Critically Endangered 1.00000 1.00000 1.00000 1.00000 EN Endangered 0.50000 0.43299 0.66770 0.99600 VU Vulnerable 0.25000 0.05155 0.10010 0.39000 NT Near Threatened 0.12500 0.00412 0.01000 0.02000 LC Least Concern 0.06250 0.00005 0.00010 0.00050 Rounded Weight Ratios CR:EN 2 2 1 1 CR:VU 4 19 10 3 CR:NT 8 243 100 50 CR:LC 16 20,000 10,000 2,000 a Data source: Mooers, et al. (2008). b From calculations by Mooers, et al., based on Isaac, et al. (2007). Table 6 shows that Isaac’s inclusion of the phylogenetic isolation factor changes the weight ratios substantially, particularly for species in the lowest threat category (Least Concern). We explore the implications for hypothetical areas A and B in Table 7. A is populated by only 2 species, both rated as Critically Endangered. B is populated by 20,000 species, but all are rated as of Least Concern. Our extinction risk indicator for each area is the sum of normalized extinction probabilities for resident species. Assignment of weights for Mooers' IUCN-derived 50-year extinction probabilities yields a total risk indicator of 2 for A -- twice the total for B, because each Critically Endangered species is weight-equivalent to 20,000 Least Concern species. In contrast, assignment of the Isaac weights yields an overall risk rating for B (1,250) that is 625 times greater than the rating for A (2), because each Critically Endangered Species is weight-equivalent to 16 species of Least Concern. The other two cases are intermediate, but far closer to the 50-year IUCN case. 31 Table 7: Implications of alternative weighting schemes Total Scores IUCN Extinction Probabilities: Future Years Status Area Species (Uniform Count Within Areas) Isaac 50 100 500 A 2 CR 2 2 2 2 B 20,000 LC 1,250 1 2 10 7.1.3 Composite Ecogradients for Animal Species With such potentially-huge differences in metrics, it is important for an ecogradient methodology to accommodate different risk-weighting schemes in a consistent and plausible way. In addition, it seems highly unlikely that the weights assigned to species classes (amphibians, birds, mammals, reptiles) by the conservation community would be proportional to their grossly unequal representation in overall species counts (Table 5). Finally, conservation specialists are not likely to agree on the relative importance that should be assigned to endemicity and extinction risk in composite ecogradients. To accommodate these diverse concerns, we adopt a conservative strategy for ecogradient construction. First, we divide all measures of endemicity and extinction risk by their maximum values and multiply by 100 to create indexes in the range 0-100. This ensures comparability in measurement. Then, for each grid cell, we select the maximum index value as our ecogradient measure for the cell. This approach ensures parity treatment for all animal species and alternative vulnerability indicators in determining composite ecogradients. 32 7.2 Incorporating Biomes An ecogradient measure based on animal species alone provides an incomplete accounting of biodiversity. A more complete measure would incorporate plants and insects, using indices similar to those we have developed for animals. Although no such indices exist at the requisite geographic scale, WWF has provided a first approximation by segmenting the world into 825 terrestrial ecoregions (Figure 14). WWF defines an ecoregion as “a large unit of land or water containing a geographically distinct assemblage of species, natural communities, and environmental conditions.” 9 Accordingly, we adopt the ecoregion as a general proxy for distinctive plant and inspect species, as well as animal species that are not represented in the range maps provided by IUCN and BirdLife International. Our method for incorporating WWF ecoregions resembles our treatment of species endemicity. For this exercise, we identify all ecoregions in the Congo Basin countries (Figure 15). Then we extend geographic coverage to surrounding countries that include parts of those ecoregions. Figure 16 identifies these countries, which extend from Senegal to South Sudan in the north, and from Angola to Malawi in the south. For the group of selected countries, we compute the percent of total area accounted for by each ecoregion. Then we compute its vulnerability index as the inverse of its area share and assign the appropriate index value to each pixel in the Congo Basin countries. This accounting assigns high values to pixels in smaller ecoregions, where clearing single pixels may pose more significant threats to biome integrity. 9 Complete information about the WWF terrestrial ecoregions is available online at http://wwf.panda.org/about_our_earth/ecoregions/ 33 Figure 14: WWF Terrestrial Ecoregions 34 Figure 15: WWF Terrestrial Ecoregions in the Congo Basin countries 35 Figure 16: WWF Ecoregions in the Greater Congo Basin region 36 7.3 A Composite Biodiversity Indicator No consensus exists on relative weights for our animal species and ecoregion indices in a composite index. As before, we could provide a conservative accommodation by normalizing both indices to range 0-100 for each pixel and choosing the larger index value as our measure. However, the ecoregion index is far more skewed than the species index. As a consequence, choosing maximum values for each normalized index pair will select the species index in the great majority of cases. For a more robust and balanced measure, we substitute ranks -- measured as percentiles -- and choose the greater percentile value for each pixel. Figures 17, 18 and 19 map the two component indexes and the composite index as percentiles. We have included national and provincial boundaries for ease of reference. Comparison provides a clear rationale for the composite index, since the two component maps look very different. The species map (Figure 17) reflects counts, extinction threats and endemicity measures for 1,976 species in the Congo Basin countries. The 90-percentile areas are mostly in southern and western Cameroon, eastern DRC, Rwanda and Burundi. Intermediate 70- and 80-percentile areas separate these from the lower-percentile areas. The ecoregion map (Figure 18) has some overlaps with the species map, particularly for 90-percentile areas in western Cameroon, eastern DRC, Rwanda and western Burundi. The two maps also overlap in northern and southern low-percentile areas. However, they have completely different patterns in the central area, which is very low-percentile in the species map but relatively high-percentile in the ecoregion map. Figure 19 combines both sets of information, displaying the highest percentile value for each pixel. 10 This approach ensures some “leveling up”, visible in the near-absence of values in 10 All indices are measured as centroid values for our 2.7 km Hansen pixels. Where necessary, we have resampled with the Stata routine geonear for pixel-level matching. 37 Figure 17: Animal species index, Congo Basin countries 38 Figure 18: Ecoregion index, Congo Basin countries 39 Figure 19: Composite species-ecoregion index, Congo Basin countries 40 the range 0-10. Nevertheless, even maximum scoring produces a map with a wide distribution of pixel values. One striking feature is the blue/green (0-50) band that arcs from northern Cameroon to eastern DRC and back to southern DRC. Another is the prominent clustering of very high values in western Cameroon, along the border between Congo and the DRC, and along the eastern margin of the Basin. For the purposes of this paper, the most important message in our results is the striking non-uniformity of ecological vulnerability across forested areas in the Congo Basin countries. By implication, a full assessment of the benefits and costs of road upgrading should go beyond simple measurement of forest loss to an assessment of the potential impact of that loss on biological diversity. 8. Ethno-Linguistic Diversity Within the Congo Basin countries, indigenous territories are occupied by over 500 ethno- linguistic groups. In many cases, their cultures and livelihoods reflect centuries of adaptation to environments that may be altered by forest clearing. Although we focus on biodiversity impacts in this paper, we believe that information about ethno-linguistic domains may also be of interest to infrastructure planners. Accordingly, we construct an ethno-linguistic index that resembles our index for ecoregions. Our data are drawn from Felix and Meur (2001), who have constructed and digitized a map of African ethno-linguistic domains. 11 For the Congo Basin countries, we compute each domain’s share of the total area for identified domains. Then we compute a relative sensitivity index for each ethno-linguistic domain as the inverse of its area share. We assign an appropriate index value to each pixel for the Congo Basin countries. This approach assigns higher values to smaller domains, which may be disproportionately affected by forest 11 The digitized map is available as a shapefile at http://worldmap.harvard.edu/data/geonode:etnicity_felix. 41 Figure 20: Ethno-linguistic index, Congo Basin countries 42 clearing. The resulting ethno-linguistic map (Figure 20) is a patchwork pattern of potential sensitivity that only overlaps significantly with the species and ecoregion maps in narrow strips of western Cameroon, eastern DRC, Rwanda and Burundi. 9. Road Improvement Revisited - the Stakes for Vulnerable Areas In this section, we join the two strands of our research by combining predicted deforestation from road upgrading with our composite index of ecological vulnerability. As we show in Figures 7-9, the multiple determinants of forest clearing can produce highly-varied patterns of road corridor deforestation within the same region. Even more variety is to be expected when we incorporate the vulnerability index, since the correlation between our measures of clearing and ecological vulnerability is close to zero in the database (ρ = -0.0356). Our primary expositional task is simplification, since combining the two indicators produces results for thousands of road segments. In this paper, we communicate the general tenor of our results by focusing on the four-province area of DRC that we have featured in Figures 7-9. Full results for the Congo Basin countries are available from the authors on request. Figure 21 sets the stage by overlaying the road network on a composite vulnerability map for the four-province area in eastern DRC (Kivu, Orientale, Equateur, Kasai Oriental). This provides an immediate sense of potential vulnerability from road upgrading at different points in the network. Some roads pass through high-vulnerability areas for long distances, while others remain within low-vulnerability zones. Figure 9 displays projected forest clearing from road improvement, showing extensive deforestation along many primary arteries and around denser parts of the network. Figure 22 displays the combined impact indicator, which is the product of predicted percent deforestation from upgrading (Figure 9) and the vulnerability index (Figure 21). We have 43 standardized the indicator so that its range is 0 - 100. To discern the effect of combining the two indicators, compare Figures 9 and 22 for the dense road cluster south of the Kasai Oriental label. Figure 9 has consistently dark brown areas throughout the cluster, indicating intense deforestation in the road corridors. In Figure 22, however, the southern and western parts of the cluster are dominated by indicator scores in the range 0-10 (yellow), while the central and eastern parts have large clusters in the range 30-100 (red). In part this reflects variation in biodiversity vulnerability, which is relatively low (green) in the southern part of the cluster and significantly higher (yellow) in the northern part. 10. Summary and Conclusions In this paper, we have used high-resolution spatial data for eight Congo Basin countries to develop and estimate an econometric model of deforestation that incorporates the economics of road improvement. We find large, highly-significant effects of upgrading on the intensity and extent of forest clearing in road corridors. In addition, our results highlight the powerful roles played in proprietors’ forest-clearing decisions by market access, land opportunity values, official protection status and topography. They also provide the first estimate of the impact of violent conflict on deforestation in Sub-Saharan Africa. Using our econometric estimates, we predict the impact of generalized road upgrading on forest clearing along road corridors in the Congo Basin. We illustrate the results with a detailed assessment of impacts in the DRC. Predicted effects in road corridors vary widely with prior road conditions and locational economics, but increases of 10-20% are typical. In addition, many corridors have significant extensions in the outer margin of forest clearing. 44 Figure 21: Eastern DRC - Road networks and ecological vulnerability 45 Figure 22: Eastern DRC - Ecological risk from road network improvement After investigating the impact of road improvement on forest clearing, we extend the analysis to potential ecological impacts. Using spatially-formatted databases from IUCN, BirdLife International and WWF, we construct a pixel-level measure of biodiversity risk for thousands of animal species and plant biomes. The resulting high-resolution map reveals a complex geographic pattern of ecological vulnerability. We overlay the Basin-wide road network on this map to provide a first-order guide to risk assessment for proposed road corridor improvements. In the concluding stage of the analysis, we combine our ecological risk indicator with pixel-level predictions of forest clearing produced by road upgrading. The result is a high- resolution map of expected risks for road upgrading in road segments, corridors and regional 46 networks. Predicted deforestation is uncorrelated with our vulnerability indicator, so the variation in the two indicators separately is compounded in the combined measure. We illustrate the implications in a detailed assessment for the eastern DRC. Overall, our results cast doubt on the utility of generalizations about the impact of road upgrading on deforestation and vulnerable biomes. In our high-resolution spatial assessment, we find impacts as varied as the economic, social and ecological conditions that prevail in different road corridors. However, our results are also cautionary, because we find very large impacts in some road corridors and significant impacts in many more. By implication, road improvement planning in tropical forest regions is unlikely to maximize welfare unless it anticipates and incorporates such impacts. 47 References Ali, Rubaba, A. Federico Barra, Claudia N. Berg, Richard Damania, John D. Nash and Jason Russ. 2015. Infrastructure in post-conflict and fragile environments: Evidence from Democratic Republic of Congo. Africa Region, World Bank (forthcoming). Arcanda, Jean-Louis, Patrick Guillaumont, Sylviane Guillaumont Jeanneney. 2008. Deforestation and the real exchange rate. Journal of Development Economics, 86(2): 242-262. Barbier, Edward and Joanne C. Burgess. 1996. Economic analysis of deforestation in Mexico. Environment and Development Economics, 1:203-239. Buys, Piet, Uwe Deichmann and David Wheeler. 2010. Road Network Upgrading and Overland Trade Expansion in Sub-Saharan Africa. Journal of African Economies 19(3): 399-432. June. Cattaneo, Andrea. 2001. Deforestation in the Brazilian Amazon: Comparing the Impacts of Macroeconomic Shocks, Land Tenure, and Technological Change. Land Economics 77(2): 219- 240. Chowdhury, Rinku Roy. 2006. Driving forces of tropical deforestation: The role of remote sensing and spatial models. Singapore Journal of Tropical Geography, 27:82–101. Cropper, Maureen, Jyotsna Puri, and Charles Griffiths. 2001. Predicting the Location of Deforestation: The Role of Roads and Protected Areas in North Thailand. Land Economics, 77(2):172-186. Cropper, Maureen, Charles Griffiths and Muthukumara Mani. 1999. Roads, Population Pressures, and Deforestation in Thailand, 1976-1989. Land Economics, 75(1): 58-73. Dasgupta, Susmita, Daniel Hammer, Robin Kraft and David Wheeler. 2014. Vyaghranomics in space and time: Estimating habitat threats for Bengal, Indochinese, Malayan and Sumatran tigers. Journal of Policy Modeling (forthcoming). Deininger, Klaus and Bart Minten. 1999. Poverty, Policies, and Deforestation: The Case of Mexico. Economic Development and Cultural Change, 47:313–344. Deininger, Klaus and Bart Minten. 2002. Determinants of Deforestation and the Economics of Protection: An Application to Mexico. American Journal of Agricultural Economics, 84(4): 943- 960. De Pinto, Alessandro and Gerald C. Nelson. 2009. Land Use Change with Spatially Explicit Data: A Dynamic Approach. Environmental and Resource Economics, 43(2): 209-229. 48 Deveny, A., J. Nackoney, N. Purvis, R. Kopp, M. Gusti, G. Kindermann,, M. Macauley, E. Myers Madeira, M. Obersteiner and A. Stevenson. 2009. Forest Carbon Index: The geography of forests in climate solutions. Joint report by Resources for the Future and Climate Advisers. December Ewers, Robert, William Laurance and Carlos Souza. 2008. Temporal fluctuations in Amazonian deforestation rates. Environmental Conservation, 35:303-310. Felix, M. L., and C. Meur, C. 2001. Peoples of Africa: Enthnolinguistic map. Brussels: Congo Basin Art History Research Center. Hansen, M.C., P. V. Potapov, R. Moore, M. Hancher, S. A. Turubanova, A. Tyukavina, D. Thau, S. V. Stehman, S. J. Goetz, T. R. Loveland, A. Kommareddy, A. Egorov, L. Chini, C. O. Justice and J. R. G. Townshend. 2013. High-Resolution Global Maps of 21st-Century Forest Cover Change. Science 342 (6160): 850-853. Hansen M, et al. 2003. Global percent tree cover at a spatial resolution of 500 meters: First results of the MODIS vegetation continuous fields algorithm. Earth Interact 7:1–15. Available at http://ams.allenpress.com/archive/1087–3562/7/10/pdf/i1087–3562-7–10-1.pdf. Igliori, Danilo Camargo. 2006. Deforestation, Growth and Agglomeration Effects: Evidence from Agriculture in the Brazilian Amazon. Department of Land Economy, University of Cambridge. Isaac N., S. Turvey, B. Collen, C. Waterman and J. Baillie. 2007. Mammals on the EDGE: conservation priorities based on threat and phylogeny. PLoS One 2: e29. IUCN. 2010. David Yanggen, Kenneth Angu and Nicodème Tchamou (eds). Landscape-Scale Conservation in the Congo Basin: Lessons Learned from the Central African Regional Program for the Environment (CARPE). Gland, Switzerland: IUCN. Joint Research Center (JRC). 2003. European Commission Global Land Cover 2000 Project. http://bioval.jrc.ec.europa.eu/products/glc2000/glc2000.php. Kaimowitz, D., P. Mendez, A. Puntodewo and J. Vanclay, 2002. Spatial regression analysis of deforestation in Santa Cruz, Bolivia. In C.H. Wood and R. Porro (eds) Land Use and Deforestation in the Amazon. University Press of Florida, p. 41-65. Laurance, William F. 2012. As Roads Spread in Rainforests,The Environmental Toll Grows. Yale Environment 360. 19 January. http://e360.yale.edu/feature/as_roads_spread_in_tropical_rain_forests_environmental_toll_grows /2485/ 49 Laurance, William F., Miriam Goosem and Susan G.W. Laurance. 2009. Impacts of roads and linear clearings on tropical forests. Trends in Ecology & Evolution 24(12): 659-669. 11 September Laurance, William F., Mark A. Cochrane, Scott Bergen, Philip M. Fearnside, Patricia Delamônica, Christopher Barber, Sammya D'Angelo and Tito Fernandes. 2001. The Future of the Brazilian Amazon. Science 291(5503): 438-439. 19 January. Li, Man, Alessandro De Pinto, John M. Ulimwengu, Liangzhi You and Richard D. Robertson. 2015. Impacts of Road Expansion on Deforestation and Biological Carbon Loss in the Democratic Republic of Congo. Environmental and Resource Economics 60: 433-469 Mooers, A., D. Faith and W. Maddison. 2008. Converting Endangered Species Categories to Probabilities of Extinction for Phylogenetic Conservation Prioritization. PLoS ONE 3(11): e3700. doi:10.1371/journal.pone.0003700. Nelson, Andrew and Kenneth M. Chomitz. 2009. Protected Area Effectiveness in Reducing Tropical Deforestation: A Global Analysis of the Impact of Protection Status. Evaluation Brief 7, Independent Evaluation Group, World Bank, October. Pfaff, Alexander. 1997. What Drives Deforestation in the Brazilian Amazon? Evidence from Satellite and Socioeconomic Data. Policy Research Working Paper No. 1772, World Bank. May. Raleigh, Clionadh, et al. "Introducing ACLED: An armed conflict location and event dataset special data feature." Journal of peace Research 47.5 (2010): 651-660. Reymondin, Louis, Karolina Argote, Andy Jarvis, Carolina Navarrete. Alejandro Coca, Denny Grossman, Alberto Villalba, Paul Suding. 2013. Road Impact Assessment Using Remote Sensing Methodology for Monitoring Land-Use Change in Latin America: Results of Five Case Studies. Inter-American Development Bank Environmental Safeguards Unit, Technical Note No. IDB - TN - 561. July. RFF (Resources for the Future and Climate Advisors). 2011. The Geography of Forests in Climate Solutions: Price Geography. http://www.forestcarbonindex.org/maps.html Rudel, Thomas, Laura Schneider, Maria Uriarte, B. L. Turner II, Ruth DeFries, Deborah Lawrence, Jacqueline Geoghegan, Susanna Hecht, Amy Ickowitz, Eric F. Lambin, Trevor Birkenholtz, Sandra Baptista, and Ricardo Grau. 2009. Agricultural intensification and changes in cultivated areas, 1970–2005. Proceedings of the National Academy of Science. 106 (49): 20675–20680. San, Nu Nu, Hans Löfgren and Sherman Robinson. 2000. Structural Adjustment, Agriculture, and Deforestation in the Sumatera Regional Economy. TMD Discussion Paper No. 52, International Food Policy Research Institute, March. 50 Stifel, D., and B. Minten. 2008. Isolation and Agricultural Productivity. Agricultural Economics 39 (1): 1–15. Stifel, David, Bart Minten and Bethlehem Koro. 2012. Economic Benefits and Returns to Rural Feeder Roads: Evidence from a Quasi-Experimental Setting in Ethiopia. International Food Policy Research Institute and Ethiopian Development Research Institute. ESSP II Working Paper No. 40. May Uchida, H. and A. Nelson. 2009. Agglomeration Index: Towards a New Measure of Urban Concentration. Background paper for World Development Report 2009. Washington, DC: World Bank. University of Maryland. 2014. Global Forest Change 2000–2012: Data Download. Department of Geographical Sciences. Available online at http://www.earthenginepartners.appspot.com/science-2013-global-forest/download.html Vance, Colin and Jacqueline Geoghegan. 2002. Temporal and spatial modelling of tropical deforestation: a survival analysis linking satellite and household survey data. Agricultural Economics, 27(3): 317-332. Wheeler, David, Daniel Hammer, Robin Kraft, Susmita Dasgupta, Brian Blankespoor. 2011. Economic Dynamics and Forest Clearing: A Spatial Econometric Analysis for Indonesia. Center for Global Development Working Paper No. 280. http://www.cgdev.org/content/publications/detail/1425820 WWF. 2000. Workshop: Priority Landscapes in the Congo River Basin. Libreville, Gabon http://wwf.panda.org/what_we_do/where_we_work/congo_basin_forests/wwf_solutions/congo_l andscapes_management/ Zikri, Muhammad. 2009. An Econometric Model for Deforestation in Indonesia. Working Paper in Economics and Development Studies, No. 200903. Center for Economics and Development Studies, Department of Economics, Padjadjaran University. July. 51 Appendix Table: Regression results for DRC (all non-dummy variables in log form) (1) (2) (3) (4) (5) OLS OLS 2SLS GLS (IV) Robust (IV) Distance from Road -0.298 -0.309 -0.309 -0.309 -0.296 (29.97)** (30.43)** (30.43)** (8.80)** (32.77)** Protected area x Distance from road -0.165 -0.204 -0.204 -0.204 -0.152 (18.56)** (22.79)** (22.79)** (2.91)** (19.18)** Road condition 0.544 0.455 0.455 0.455 0.513 (17.56)** (13.76)** (13.76)** (1.57) (17.49)** Transport cost to nearest urban center -0.481 -0.952 -0.952 -0.876 (34.15)** (23.54)** (3.28)** (24.39)** Euclidian distance to nearest -0.279 urban center (23.54)** Land opportunity value -0.091 -0.023 -0.023 -0.023 0.004 (10.32)** (2.63)** (2.63)** (0.37) (0.46) Elevation -0.066 -0.130 -0.130 -0.130 -0.238 (2.06)* (3.97)** (3.97)** (0.59) (8.18)** Conflict intensity (1997 - 2007) 0.016 0.029 0.029 0.029 -0.000 (5.52)** (9.96)** (9.96)** (1.23) (0.10) D2002 0.935 0.934 0.934 0.934 0.925 (21.93)** (21.46)** (21.46)** (19.61)** (23.92)** D2003 1.269 1.269 1.269 1.269 1.246 (29.78)** (29.15)** (29.15)** (22.76)** (32.24)** D2004 1.559 1.558 1.558 1.558 1.542 (36.58)** (35.81)** (35.81)** (27.75)** (39.91)** D2005 1.827 1.826 1.826 1.826 1.810 (42.87)** (41.97)** (41.97)** (32.71)** (46.86)** D2006 1.983 1.982 1.982 1.982 1.972 (46.56)** (45.58)** (45.58)** (36.81)** (51.07)** D2007 2.155 2.154 2.154 2.154 2.145 (50.59)** (49.52)** (49.52)** (38.55)** (55.55)** D2008 2.278 2.278 2.278 2.278 2.271 (53.50)** (52.37)** (52.37)** (40.65)** (58.81)** D2009 2.462 2.462 2.462 2.462 2.456 (57.81)** (56.60)** (56.60)** (44.19)** (63.62)** D2010 2.651 2.650 2.650 2.650 2.647 (62.25)** (60.94)** (60.94)** (48.34)** (68.56)** D2011 2.747 2.746 2.746 2.746 2.738 (64.50)** (63.14)** (63.14)** (48.95)** (70.92)** D2012 2.832 2.831 2.831 2.831 2.821 (66.48)** (65.09)** (65.09)** (49.73)** (73.06)** Constant 4.810 4.806 6.747 6.747 7.255 (23.39)** (22.07)** (25.94)** (4.07)** (31.42)** Observations 13758 13758 13758 13758 13758 R-squared 0.49 0.47 0.47 0.47 0.51 Absolute value of t statistics in parentheses * significant at 5%; ** significant at 1% 52