wPS IsS% POLICY RESEARCH WORKING PAPER 25 83 How the Location of Roads Establishing protected areas (national parks together with and Protected Areas wildlife sanctuaries) in North Affects Deforestation Thailand did not reduce the likelihood of forest clearing, in N orth Thailand but wildlife sanctuaries may have reduced the probability of deforestation. Where new Maureen Cropper roads are located affects how Jyotsna Puri much of a threat they are to Charles Griffiths protected areas. The World Bank Development Research Group Infrastructure and Environment U April 2001 I POLICY RESEARCH WORKING PAPER 2583 Summary findings Using plot-level data, Cropper, Puri, and Griffiths Road building, by reducing the impedance-weighted estimate a bivariate probit model to explain land clearing distance to market, has promoted clearing, especially and the siting of protected areas in North Thailand in near the forest fringe. 1986. The authors simulate the impact of further road Their model suggests that protected areas (national building to show where road building is likely to have parks together with wildlife sanctuaries) did not reduce the greatest impact on forest clearing and where it is the likelihood of forest clearing, but wildlife sanctuaries likely to threaten protected areas. may have reduced the probability of deforestation. This paper-a product of Infrastructure and Environment, Development Research Group-is part of a larger effort in the group to examine factors affecting deforestation in developing countries. The study was funded by the Bank's Research Support Budget under the research project "Spatial Models of Environmental Processes: A Study of Deforestation in Thailand" (RPO 683-17). Copies of this paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Viktor Soukhanov, mail stop MC2-205, telephone 202-473-5721, fax 202-522-3230, email address vsoukhanov@worldbank.org. Policy Research Working Papers are also posted on the Web at http:// econ.worldbank.org. The authors may be contacted at mcropper@worldbank.org, jpuri@worldbank.org, or griffiths.charles@epa.gov. April 2001. (38 pages) 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 view of the World Bank, its Executive Directors, or the countries they represent. Produced by the Policy Research Dissemination Center HOW THE LOCATION OF ROADS AND PROTECTED AREAS AFFECTS DEFORESTATION IN NORTH THAILAND Maureen Cropper World Bank, MC2-521, 1818 H Street NW, Washington, DC, 20433. Email mcropper@worldbank.org, tel. 202-473-1277, fax 202-522-3230 Jyotsna Puri World Bank, MC2-524, 1818 H Street NW, Washington, DC, 20433. Email ipuri@worldbank.org, tel. 202-458-4799, fax 202-522-3230 Charles Griffiths U.S. EPA, Ariel Rios Building, MC 2177, 1200 Pennsylvania Avenue NW, Washington, DC 20460. Email griffiths.charles@epa.gov, tel. 202-260-6711, fax 202-260-5732 The authors are, respectively, Lead Economist, World Bank, Professor of Economics, University of Maryland, College Park, and University Fellow, Resources for the Future; Consultant, World Bank, and Ph.D. student, University of Maryland; and Economist, U.S. EPA. The authors would like to thank Ken Chomitz, Louise Scura, Tom Tomich, seminar participants at the Kennedy School of Government, and participants in a joint seminar hosted by CIFOR and ICRAF, Indonesia, for comments and helpful discussions. The authors also thank the World Bank for providing financial support. Cropper, Puri, Griffiths I. INTRODUCTION Concern over the rate at which forests are being converted to agriculture has given rise to a literature that quantifies the impact of forces that drive deforestation. The literature has focused on two questions: (1) What factors affect the location of deforestation? and (2) What factors affect the rate of deforestation? Each question has policy significance. It is clearly important to know where deforestation is likely to occur, especially if it is in environmentally sensitive areas, and it is also important to know how fast the process is taking place. This paper focuses on the first question. We estimate an equilibrium model of land use in North Thailand in the mid-1980's, using coarse-resolution (1:1,000,000) plot-level data. The purpose of the model is to explain where deforestation is likely to occur and to examine the impact of two government policies that can affect the location of deforestation: the establishment of protected areas, and road building. Protected areas are often suggested as a means of conserving tropical ecosystems and have, at least on paper, been created in many tropical developing countries. In 1985, Thailand declared that 15% of its land area should be set aside for conservation or protected forests. By 1986, 10% of the country's land lay within protected areas. Fifty-two percent of the land in protected areas was devoted to national parks and 42 percent to wildlife sanctuaries.' Whether such areas can in fact protect biodiversity depends on their size and location, and on how they are managed. Protected areas are less likely to experience encroachment if they have the political support of surrounding communities, and if these communities can produce sufficient income without encroaching upon the protected area. This suggests that understanding the reasons for the success or failure of protected areas requires on-the-ground knowledge, and is best evaluated using a case study approach. 3 Cropper, Puri, Griffiths The contribution we make to the topic is to evaluate statistically whether protected areas have reduced the probability of deforestation in national parks and wildlife sanctuaries in Thailand. Other authors who have tackled this issue (Chomitz and Gray 1996; Deininger and Minten 1996) have estimated a land use model that predicts the probability that land located in protected areas is cleared. The fraction of land predicted to be cleared is then compared with the fraction of land actually cleared to determine the impact of protected areas on clearing. This approach does not, however, allow one to determine whether the impact of protected areas on clearing is statistically significant, or to test hypotheses about its magnitude. We estimate a bivariate probit model to explain the probability that a plot of land is cleared and the probability that it lies within a protected area. Protected area status enters the clearing equation, and variables that affect the designation of an area as protected (but not the clearing decision) are used to identify the coefficient of protected area status. This allows us to control for the selectivity problem inherent in single-equation models of land use: In a single-equation model of clearing, the coefficient of protected area status is likely to overstate (in absolute value) the impact of protected areas on clearing. This is because protected areas are likely to be located in places that have not yet been cleared. The second topic on which we focus is the impact of roads on the land clearing decision. Qualitatively, the impact of roads on land clearing is well understood: Road building facilitates access to markets, and thus raises the probability that forests will be cleared for agriculture. Understanding the quantitative impact of road building on clearing is, however, crucial for policy. Suppose a government wishes to build a road to a proposed national park. Where should the road be located to reduce the likelihood of development en route to the park? As Chomitz and Gray (1996) emphasize in their study of the impact of roads on agricultural development in 4 Cropper, Puri, Griffiths Belize, the impact of roads depends on the topography of the area, and on soil quality. One goal of our study is to show where road building in North Thailand is likely to have the greatest impact on the probability that forests are cleared, and to identify the impact of further road building on protected areas. To investigate the issues discussed above, we have assembled a GIS database on land use, roads, physiographic variables (slope, elevation and soil quality), populated places and population density for the 17 provinces of North Thailand. The data also include protected area boundaries, and provincial and district boundaries. The model of land clearing and protected area status estimated with these data is described in section II. Section III contains a more detailed description of the data and our sampling strategy. Econometric results are presented in section IV. We conclude the paper by showing how our model can be used to estimate the threat of encroachment in protected areas. II. A MODEL OF LAND CLEARING AND PROTECTED AREA STATUS Economic theory predicts that forested land will be cleared if the profits from clearing land exceed the profits from leaving land under forest cover. We follow Chomitz and Gray (1996) (see also Nelson and Hellerstein 1996) in assuming that the profit from land use k on plot i, Ri,k may be defined as the difference between the value of outputs and inputs Qik and Xik at their respective location-specific prices Pik and C,k, Rik = PikQik - CikX,k. [1] Chomitz and Gray (1996) demonstrate that when output is a Cobb-Douglas function of X,k and plot characteristics, s1,js2j .... Q,k = S,kX,kk with 0 < pk < 1 [2] 5 Cropper, Puri, Griffiths Sik 1, li 2i ... [3] R,k may be written Rik = (6 )Ck '8' (PzkSjk8k) )4 Ik By taking logs and collecting coefficients, this can be transformed into an expression of the form ln R,k =k +akklnPk + Ok lnC,k + ,u,nk lnsm, [5] n Empirically we distinguish between two forms of land use, agriculture (k = 1) and forestry (k = 0) and note that plot i will be devoted to agriculture if ln R,, > ln R,o. In practice, data on input and output prices are unavailable at the plot level. We assume that both Pik and C,k vary with the impedance-weighted distance of the plot from the nearest market (Cost;), and, also, with the population density of the district in which the plot is located (Population densityi). District population affects Pj by shifting the demand for agricultural output, and C,, by shifting the demand and supply curves of labor. We use district population density, rather than population, to control for the fact that districts vary in area. Plot characteristics {s,n } that affect the profitability of clearing include slope, elevation, measures of soil quality and the plot's protected area status. Since the government has the right to evict persons living in parks or wildlife sanctuaries, there is at least some threat of expropriation if output is grown in these areas. The province in which the plot is located is also likely to affect the profitability of agriculture. Provincial dummy variables capture differences in rainfall and may proxy differences in tenure security. Representing protected area status by Y2, = 1, if a plot lies in a protected area (and = 0 otherwise), and all other factors that influence 6 Cropper, Puri, Griffiths the profitability of conversion (including distance to markets and population density) by vector Z,,a plot will be cleared if Z,B1 + yY2, > 0.2 In our empirical model, Zi includes the slope of the plot, its elevation, population density in the district in which the plot is located, the natural logarithm of impedance-weighted to market, provincial dummy variables and dummy variables for soil categories. The two factors that have not been included in this model are logging and subsistence agriculture. Land clearing has been blamed on the extensive commercial logging that took place over the three decades prior to Thailand's 1989 logging ban, but it should be remembered that a logged forest is not a permanent condition. While logging does reduce the cost of clearing, we assume that the land will not be permanently cleared unless it is economically profitable to do so. Subsistence farming, on the other hand, does produce permanently cleared land. Although we have not presented a model of subsistence farming, we note that the factors included in Zi are also likely to be important in explaining whether subsistence farmers will clear land (Angelsen 1996). There is no well-developed theory to explain which plots of land are designated protected areas; however, political and economic considerations suggest that land where the opportunity costs of protection are low (land of low agricultural value) would be more likely to be selected than land of high agricultural value. This suggests that the factors, Z,, that affect the profits of clearing land (the opportunity cost of protection) should enter the equation to explain protected area status. The benefits of protecting a plot, which accrue more broadly, should, however, depend on different factors. Areas that serve as habitat to endangered species or that contain fragile ecosystems clearly yield higher benefits from preservation than areas that are ecologically 7 Cropper, Puri, Griffiths unremarkable. Riverine forests constitute fragile ecosystems that are often home to diverse species. We posit that location near rivers increases the chance that a plot is protected. The econometric model that we estimate is thus given by Yli =ZiBI +YY2i +eli Yl =1 if Y1, >0;=0 otherwise [61 Y2i = ZiB2 + aWi + e2i Y2i =1 if Y2i > 0; = O otherwise [7] where the plot is cleared (Yt, = 1) if the net profits from clearing plot i (Y,*) are positive, and the plot lies in a protected area (Y2, = 1) if the net benefits from protecting plot i (Y2,) are positive. Wi indicates that the plot is located near a river (watershed dummy). We estimate this structural model as a bivariate probit model, assuming that eli and e2, are jointly normally distributed.3 This allows us to estimate the impact of protected area status on the probability that a plot is cleared. The model is estimated for two definitions of protected area: national parks and wildlife sanctuaries (hereafter referred to as "protected areas"), and wildlife sanctuaries only. The focus on wildlife sanctuaries is prompted by anecdotal evidence that the Thai government has made stronger efforts to prevent encroachment in wildlife sanctuaries than in national parks. III. STUDY AREA AND DATA The area we have chosen for this study-the 17 provinces that constitute North Thailand-remains heavily forested, especially the Upper North portion of the region.4 Protected areas constituted 11 percent of the region in 1986, the year of our study (see Figure 1), and continue to be established. North Thailand is one of the poorest regions of Thailand, and road building is part of the government's strategy to reduce rural poverty.5 Between 1973 and 8 Cropper, Puri, Griffiths 1985, extensive road building increased road density in North Thailand by 57% (Cropper, Griffiths and Mani 1999). The policy issues raised in the introduction are, therefore, relevant to North Thailand. Data We model the land clearing decision in North Thailand in 1986 using coarse resolution data. Land use information comes from a 1: 1,000,000 Land Development Department map that originally contained 15 land use categories. Urban areas and water were omitted from the study area and the remaining land uses classified as "forest" or "non forest."6 The term "clearing," as used in section II, is thus synonymous with "non forest." Physiographic factors that should influence the profitability of clearing include the soil characteristics of the plot, its slope and elevation. All soils in North Thailand are classified by the FAO Soil Map of the World as falling in one of 12 soil categories, defined on the basis of soil texture and slope class (Acrisol, Fluvisol, Gleysol, etc.). We represent these soil categories using a series of dummy variables.7 Elevation (in meters) was obtained at a resolution of 30 arc seconds, and the slope of each plot was calculated as the maximum difference between the elevation of the plot and the elevation of each of neighboring plot. (The sources of our data are described in the Appendix.) To compute ease of access to markets, we digitized a 1982 road map of Thailand (1: 1,000,000 scale), distinguishing between paved and unpaved roads. The locations of market towns were obtained from the Digital Chart of the World. To calculate the impedance-weighted distance from each plot to the nearest market town, travel along a paved road was assigned an impedance factor of 1, travel along an unpaved road an impedance factor of 2, and travel from a 9 Cropper, Puri, Griffiths plot to a road a factor of [100 + (Slope of Plot)2]. An algorithm was used to compute the shortest distance from each point to the nearest market town.8 River distances were computed in a similar fashion. Population, a proxy for the demand for agricultural products and for labor supply, is measured at the district level using 1990 census data.9 Population density is calculated using 1990 district boundaries. Because each district is large relative to the size of a plot, we treat district population density as exogenous to the pixel. Protected area boundaries, obtained from the IUCN, indicate that 14.4 percent of our sample points lie within protected areas (parks and wildlife sanctuaries), while 9.1 percent lie within wildlife sanctuaries. The percent of protected areas remaining under forest cover is 87% whereas it is 70% for all sample points. Sampling Strategy All layers of the GIS database were converted to a resolution of 100 square meters, which resulted in over 28,000,000 data points. We sampled points systematically, at 5-km intervals, which yielded 6,550 observations. The 3 provinces that contained no protected areas were dropped from estimation of the protected area equations, while the 5 provinces that contained no wildlife sanctuaries were dropped from those equations (see Table 1). Exact collinearity between protected areas and four soil categories (and between wildlife sanctuaries and the same soil categories) necessitated that observations in these soil categories also be dropped (see Table 1). The means and standards deviations of variables for each of the protected area and wildlife sanctuary samples are presented in Table 1. 10 Cropper, Puri, Griffiths IV. ECONOMETRIC RESULTS Determinants of Land Clearing in North Thailand We begin by examining how well our model explains land clearing in North Thailand (see Tables 2 and 3). North Thailand is a mountainous area, characterized by parallel hills and valleys that run north to south (see Figure 2). Steep slopes and high elevations have helped to protect much of the area from clearing. Indeed, 70 percent of the study area was classified as forested in 1986. The model of Table 2 correctly predicts land use within sample for 91 percent of the sample points under forest cover (Yli = 0). The model predicts clearing within sample less accurately--only 57 percent of cleared plots are correctly predicted to be cleared. When the model does predict clearing, however, it is correct 75 percent of the time (see Table 3). The quantitative impacts of factors that affect the probability of clearing are shown in columns (4) and (5) of Table 2. Phyisographic factors have a significant impact on clearing: Calculated at the means of explanatory variables, the elasticity of probability of clearing with respect to the slope of the plot is -0.48, and the elasticity with respect to elevation is -0.61. 1 Soil quality also matters. Sixty percent of the observations in our sample lie in FAO soil category Ao90-2/3c, which is the omitted soil category in our models. This soil type is distinguished by shallow soils, found on steep slopes, with low potassium content. The few pockets of better soil in North Thailand have a higher probability of being cultivated. For example, the marginal effect of moving from FAO soil unit Ao9O-2/3c to FAO soil unit LcI00-c is to increase the probability of cultivation by 36 percentage points. The latter soil is distinguished by finely textured soils that drain well, have good chemical properties, and are well-suited to growing sugarcane and rice. In general, the soil categories that significantly 11 Cropper, Puri, Griffiths increase the probability of clearing are loamy, occur at greater depth than soils in the reference category, and are found on flat or moderately undulating plains Deininger and Minten, in their study of deforestation in Mexico, note that physiographic factors alone explain land clearing almost as well as a model to which socioeconomic variables-specifically, population density and market access-are added. The same is true of North Thailand. If we exclude population density and impedance-weighted distance from the model, the percent of within sample observations correctly predicted by the model hardly changes: the percent of within sample observations correctly predicted by the clearing equation falls from 81.1 percent to 80.7 percent. Nonetheless, population density and market access do have a statistically significant impact on clearing. Figures 3 and 4 show the impact of changes in these variables on the probability of clearing, when all other variables are held at their mean values for plots in forest areas. In forest areas mean population density is approximately 40 persons per square kilometer. Doubling this density (and holding all other variables at their means in forest plots) increases the probability that a plot is cleared from about 0.15 to 0.18 (see Figure 3). This relatively modest effect can be explained by the fact that higher population density has two opposing effects on clearing-increases in population density may imply higher agricultural prices, which should encourage clearing, but may also reflect higher wages, which should discourage clearing." The impact of roads is much larger, especially at the forest fringe. Consider a forest plot that is 2.5 km from the nearest paved road and 6 km along the road to the nearest market (i.e., with an impedance-weighted distance of 256). As Figures 4 shows, bringing this plot 1.5 km closer to a paved road (i.e., reducing impedance-weighted distance by 150) increases the 12 Cropper, Puri, Griffiths probability of clearing from 0.18 to 0.23, i.e., by 5 percentage points. The impact of changes in the road network is further explored in section V. below. Determinants of the Location of Protected Areas and Wildlife Sanctuaries As one would expect, the variables that increase the probability that a pixel is cleared in general reduce the probability that it lies within the boundary of a protected area (see Table 2) or wildlife sanctuary (see Table 4). Steeper slopes, higher elevations and locations farther from market centers increase the chance that land is designated a protected area. The same is true for wildlife sanctuaries, although slope and elevation have a smaller quantitative impact on the siting of wildlife sanctuaries than they do on all protected areas. Higher population density in a district increases the probability that a pixel within the district lies in a protected area, although the effect is quantitatively small. This may reflect a desire to locate national parks near population centers. By contrast, higher population density reduces the probability of siting a wildlife sanctuary in a district. Our results in Tables 2 and 4 support Dixon and Sherman's (1990) observation that, in developing countries, areas of low agricultural value are more likely to be designated protected areas in order to avoid political conflict. This point is brought home by estimating univariate probit versions of equation (6) (without either protected area or wildlife sanctuary dummy variables) and using them to predict the probability that plots in protected areas and wildlife sanctuaries are cleared. The average predicted probability of clearing is 0.165 for protected areas and 0.125 for wildlife sanctuaries. These numbers are much lower than the average predicted probability of clearing for all sample points, which are 0.308 for the protected area sample and 0.26 for the wildlife refuge sample. 13 Cropper, Puri, Griffiths Impacts of Protected Areas and Wildlife Sanctuaries on Land Clearing We turn now to the impact of protected areas on the probability that land is cleared. The coefficient of the protected area dummy in the clearing equation in Table 2 is insignificant, suggesting that protected areas had no statistically significant impact on forest clearing in North Thailand.'2 A much different impression is obtained from a univariate probit model with the same variables as equation (6). In the univariate probit model (not shown) the coefficient of the protected area variable = -0.199, with a standard error of .076. The impact of switching Y2i = 1 from Y2i = 0 is to reduce the probability of clearing by 6 percentage points. This erroneous conclusion occurs because areas designated as protected are less likely to be cleared in the first place. Measuring the impact of protected areas using the Chomitz and Gray/Deininger and Minten approach also leads to a different conclusion than Table 2. Their approach is to estimate a single equation probit model for clearing and then use this to predict the probability that pixels in protected areas are cleared. If we estimate a single equation model for clearing (without the protected area or watershed dummies) the average probability that protected areas are cleared equals 0. 165. This is higher than the fraction of protected areas actually cleared (0.132). The analysis of Table 2 however indicates that this difference is not statistically significant. The story is somewhat different for wildlife sanctuaries. In the single-equation version of equation (7) in Table 4, wildlife sanctuaries have a much larger impact on clearing (coefficient = -0.303 with standard error - 0.104) than do all protected areas. In Table 4, the coefficient of wildlife sanctuaries is approximately the same as in the single equation model (-0.334), but has a larger standard error (0.257). Had we been able to identify a better instrument for wildlife sanctuaries than the watershed dummy, we would very likely have estimated the impact of 14 Cropper, Puri, Griffiths wildlife sanctuaries with greater precision. We therefore conclude that there is weak evidence to suggest that wildlife sanctuaries may have deterred deforestation in North Thailand. These results are consistent with anecdotal evidence (Albers 1999). National parks in Thailand are designed without formal buffer zones to separate parks from adjacent land uses. Park boundaries often become de facto buffer zones, a result supported by our analysis. By contrast, anecdotal evidence suggests a deliberate policy to prevent encroachment in wildlife sanctuaries. V. POLICY IMPLICATIONS OF THE MODELS In this section we use the model to answer two questions of policy relevance for North Thailand: (1) Which protected areas are under the greatest threat of encroachment? and (2) What is the likely impact on protected areas of increased road building? We define the areas of North Thailand under greatest threat of deforestation as those areas under forest cover in 1986 (Y1 = 0) for which the predicted probability of clearing exceeds one-half. Two hundred ninety-three sample points are so threatened, and are plotted on Figure 5. Most of these points are clustered in the low-lying portions of the lower half of the region. This is not surprising given the importance of slope and elevation in explaining clearing. Although only 8 of the 293 points lie strictly within the boundaries of protected areas, most of the points are clustered near protected areas. The national parks of Nam Nao and Thung Salaeng Luang, near the southeastern border of North Thailand are surrounded by areas under high threat of conversion, as are the Khao Sanam Phriang wildlife sanctuary and the Ramkamhaeng national park, located to the west. We note, in the case of Thung Salaeng Luang, that three-quarters of the area of the park under forest cover in 1986 had a probability of clearing greater than or equal to 0.4. This is not to suggest that protected areas are an improper policy response. We simply 15 Cropper, Puri, Griffiths note that the "paper boundary" of the official protected area, which is an admittedly weak proxy of forest protection and management, has not been sufficient to deter land clearing. To show how further road building might affect deforestation, we use equation (6) (Table 2) to compute the impact of a 1 00-unit reduction in impedance-weighted distance to market on the probability of clearing for all our sample points. This is equivalent to bringing a paved road one kilometer closer to each point. We then identify the areas where such an improvement in access raises the probability of clearing above 0.5. There are 207 such points. These points (along with the points predicted to be cleared in Figure 5) are plotted in Figure 6. Not surprisingly, the plots that we predict will be cleared as a result of road building are often clustered near the plots predicted to be cleared in Figure 5. In some cases we predict that road- building will result in clearing within protected area boundaries. In other cases, road building will lead to development around a park or wildlife sanctuary, suggesting the likelihood of eventual encroachment. This is especially true for the national parks labeled in Figure 6. What are the policy implications of these exercises? We emphasize that our predictions of forest areas under threat are subject to the limitations of our data, and have not been verified by comparing our predictions with actual changes in forest cover that have occurred since 1986. That said, analyses such as ours can suggest where effort should be placed if the goal of protected area management is to prevent deforestation within park boundaries. While our work says little about what tools are likely to be effective in preventing encroachment, it suggests where these tools should be applied. Our models also suggest where road building is likely to increase the threat of encroachment in protected areas, but also where it will not. There are, for example, areas in Figure 6 where improved access to markets is likely to encourage land clearing 16 Cropper, Puri, Griffiths (and may thereby achieve other objectives, such as reducing poverty), but where protected areas are not threatened. 17 Cropper, Puri, Griffiths APPENDIX SOURCES AND LAYERS COMPRISING THE GIS DATABASE Data Layer Source Year Attribute Categories Land Use Land Development Department 1986 15 land use categories Bangkok, Thailand Political University of New Hampshire 1990 17 Provinces and 168 districts Boundaries Elevation Digital Elevation Model (EROS web site) NA I meter intervals http://edcwww.cr.usgs.gov Rivers Digital Chart of the World Unknown Perennial and non-perennial waterways Roads Digitized from paper maps provided 1982 Paved and Unpaved roads by the Land Development Department, Thailand Soil FAO 1972 12 FAO soil categories Population Housing and Population Census, Thailand 1990 Population at the district level Populated Places Digital Chart of the World Unknown 620 populated places in study area Slope Derived from the Elevation Map Derived using 'slope' module in IDRISI Protected Areas IUCN (World Conservation Union) 1991 National Parks (IUCN category No. II ) providing World Conservation Monitoring & Wildlife Sanctuaries (IUCN category Centre data / The World Bank No. IV) 18 Cropper, Puri, Griffiths Properties of Soils Of Thailand (%) FAO Soil Too Infertile Sandy Loamy Clayey Slope Slope Slope Depth> Category Wet 1-8% 8-30% >30% lOOcms Af60-1/2ab 10 20 30 70 0 25 75 0 100 Agl6-2a 70 30 0 100 0 70 30 0 100 Agl7-1/2ab 55 20 15 85 0 35 65 0 100 AolO7-2bc 0 20 0 100 0 0 75 25 90 Ao90-2/3c* 0 10 0 65 35 0 25 75 20 I-Lc-Bk-c 0 0 0 100 0 16 50 34 66 Je72-2a 40 0 0 100 0 100 0 0 100 LclOO-c 0 0 0 100 0 0 25 75 10 Lg39-3ab 70 0 0 60 40 60 40 0 100 AolO8-2ab 10 60 0 90 10 30 70 0 100 Nd65-3ab 0 0 0 50 50 30 65 5 90 Vp64-3a 10 10 0 40 60 75 15 10 40 * Is the comparison Soil Category Source: FAO/UNESCO Soil Map of The World. Note: These 12 categories of soils are an exhaustive list of soils occurring in North Thailand. The numbers in the table show the percentage of each soil category in all of Thailand with the property shown in the column. 19 Cropper, Puri, Griffiths References Angelsen, Arild. 1996. "Deforestation: Population or Market Driven?" Working Paper. Chr. Michelsen Institute, Bergen, Norway. Albers, Heidi. 1999. Personal communication, Resources for the Future. Chomitz, Kenneth M., and David P. Gray. 1996. " Roads, Land Markets and Deforestation: A Spatial Model of Land Use in Belize." The World Bank Economic Review 10: 487-512. Cropper, Maureen L., Muthukumara Mani and Charles Griffiths. 1999. "Roads, Population Pressures and Deforestation in Thailand, 1976-1989." Land Economics 75: 58-73. Deininger, Klaus, and Bart Minten. 1996. "Determinants of Forest Cover and the Economics of Protection: An Application to Mexico." Working Paper Number 10. The Poverty, Environment and Growth Working Paper Series. The World Bank. Dixon, John A., and Paul B. Sherman. 1990. Economics of Protected Areas: A New Look at Benefits and Costs. Honolulu: East-West Center Island Press. International Labor Organization. N.D. "Macroeconomic Policies and Poverty: The Thai Experience." http://www.ilo.org/public/english/region/asro/bangkok/paper/thaiexp.htm Nelson, Gerald C., and Daniel Hellerstein. 1997. "Do Roads Cause Deforestation? Using Satellite Images in Econometric Analysis of Land Use." American Journal of Agricultural Economics 79:80-88. 20 Cropper, Puri, Griffiths FOOTNOTES The remaining 8% of protected areas included arboretums, botanical gardens and reserved areas. 2If P1k and C,k are exponential functions of population density and distance to market, then these variables will enter Z, linearly. Likewise if {s11i } are an exponential function of plot characteristics they will enter Z, linearly. 3To reduce the problem of spatial autocorrelation, we sample plots at intervals of 5 km. We have also estimated the model including average values of slope, elevation and distance to market within a 1 0-km radius of plot i. The coefficients of these variables measured for plot i are robust to the inclusion of the average values of the variables on surrounding plots. 4The Upper North consists of the provinces of Chiang Mai, Chiang Rai, Nan, Lampang, Lamphun, Mae Hong Son, Uthai Thani, Tak and Phrae. 5 In 1986 Northeast Thailand had the lowest per capita income of any area in Thailand. The North had the second lowest per capita income (International Labor Organization N.D., Table 3.3). 6Since this data comes from a paper map rather than a satellite image, "forest" means land under the control of the Royal Forest Department, and does include land without significant tree cover. Therefore, we are technically modeling a proxy of forest cover. 7The distribution of more familiar soil properties (nitrogen or phosphorous context) is known for all plots in a soil category; however, it is not known at the level of an individual plot. 8 Costdistance is a module in Arc/InfoTM that calculates for each cell the least accumulative cost of travel from a set of source cells, over a cost surface. 21 Cropper, Puri, Griffiths 9 This population data may understate the true population due to the large number of ethnic minorities and illegal aliens in the upland portions of Northern Thailand. 0 If we calculate the elasticity at means of forested plots, the elasticities with respect to slope and elevation are much higher: -0.66 and -0.84, respectively. Our discussion here focuses on the models reported in Table 2. Results for the clearing equations in Table 4 are qualitatively and quantitatively similar to those in Table 2. "As a referee noted, increases in rural population density may reduce agricultural wages through the factor proportions effect. 12Following the suggestion of a referee, we also used the length of time a pixel had been designated protected to explain the probability of clearing. This variable was, however, insignificant. 22 Cropper, Puri, Griffiths TABLE 1 Summary Statistics North Thailand Sample Protected Area Sample Wildlife Sanctuary Sample Variable Mean or proportion Mean or proportion Mean or proportion (S.D) (S.D) (S.D) Total no. of observations 6550 4946 4355 Cleared Land 0.425 0.307 0.263 Slope of plot (degrees) 3.54 (3.87) 4.24 (3.94) 4.46 (3.94) Elevation (meters) 472.54 (352.13) 546.32 (645.06) 578.93 (341.15) Population density (1990) 63.44 (67.14) 45.64 (53.78) 42.56 (55.63) (people/kM2) Cost82 (impedance-weighted 546.92 (621.68) 636.45 (676.85) 652.96 (700.85) distance to nearest market) Watershed dummy 0.600 0.569 0.562 Protected area dummy 0.108 0.144 0.263 Wildlife sanctuary dummy 0.069 0.091 0.151 Province dummy (Chiang Rai) 0.062 Province Omitted Province Omitted Province dummy (Chiang Mai) 0.134 0.164 0.186 Province dummy (Mae Hong Son) 0.077 0.102 0.116 Province dummy (Phayao) 0.037 0.029 0.033 Province dummy (Nan) 0.069 0.091 0.104 Province dummy (Lampang) 0.075 0.095 0.108 Province dummy (Phrae) 0.040 0.047 0.054 Province dummy (Lamphun) 0.026 0.025 0.029 Province dummy (Uttaradit) 0.046 0.054 0.061 Province dummy (Tak) 0.103 0.136 0.155 23 Cropper, Puri, Griffiths Province dummy (Sukhothai) 0.040 0.035 Province Omitted Province dummy (Phitsanulok) 0.062 0.059 0.067 Province dummy (Phetchaboon) 0.072 0.085 Province Omitted Province dummy (Khamphaeng 0.047 0.034 0.039 Phet) Province dummy (Phichit) 0.026 Province Omitted Province Omitted Province dummy (Nakhon Sawan) 0.045 Province Omitted Province Omitted Province dummy (Uthai Thani) 0.039 0.044 0.050 Soil dummy (Af6O-1/2ab) 0.119 0.147 0.136 Soil dummy (Ag16-2a) 0.007 0.009 0.010 Soil dummy (Agl 7-2ab) 0.086 Category Omitted Category Omitted Soil dummy (AolO7-2bc) 0.062 0.056 0.049 Soil dummy (Ao9O-2/3c) 0.479 0.598 0.634 Soil dummy (I-Lc-Bk-c) 0.029 0.038 0.038 Soil dummy (Je72-2a) 0.045 Category Omitted Category Omitted Soil dummy (LcI00-c) 0.012 0.016 0.018 Soil dummy (Lg39-3ab) 0.046 Category Omitted Category Omitted Soil dummy (AolO8-2ab) 0.068 0.090 0.079 Soil dummy (Nd65-3ab) 0.043 0.047 0.036 Soil dummy (Vp64-3a) 0.005 Category Omitted Category Omitted 24 Cropper, Puri, Griffiths TABLE 2 Bivariate Probit Model Estimated Using Protected Area Sample Dependent variable Cleared Land (Yl = 1) Equation [6] Equation [7] Independent variable Coefficient Z Marginal Elasticity Coefficient Z Marginal Elasticity Effect Effect Slope (degrees) -0.088 -10.652 -0.027 -0.475 0.034 5.297 0.005 0.272 Elevation (ms.) -0.001 -8.095 -0.0003 -0.614 0.001 9.058 0.0001 0.917 Population densityl990 0.003 4.532 0.001 0.154 0.001 2.297 0.0002 0.09 (people/km2)* * * Log(cost) (1982)** -0.191 -9.729 -0.059 -0.24 0.192 7.477 0.028 0.362 Provincial dummy -0.12 -1.085 -0.039 0.363 1.422 0.063 (Chiang Mai) Provincial dummy -0.725 -5.573 -0.179 1.052 4.163 0.253 (Mae Hong Son) Provincial dummy (Phayao) -0.341 -2.249 -0.094 1.042 3.748 0.265 Provincial dummy (Nan) -0.278 -2.453 -0.082 -0.574 -1.737 -0.059 Provincial dummy (Lampang) -0.493 -4.4 -0.133 0.501 1.865 0.096 Provincial dummy (Phrae) -0.394 -3.099 -0.105 1.381 5.217 0.388 Provincial dummy (Lamphun) -0.491 -2.769 -0.123 1.851 6.395 0.579 Provincial dummy (Tak) -0.343 -3.061 -0.095 1.197 4.755 0.292 Provincial dummy (Sukhothai) -0.288 -2.101 -0.079 1.331 4.678 0.373 Provincial dummy 0.384 2.916 0.141 1.446 5.454 0.407 (Phitsanulok) Provincial dummy 0.746 6.46 0.274 0.855 3.216 0.194 (Phetchaboon) Provincial dummy 0.03 0.213 0.014 1.334 4.752 0.374 25 Cropper, Puri, Griffiths (Kamphaeng Phet) Provincial dummy -0.107 -0.687 -0.02 2.176 8.253 0.68 (Uthai Thani) Soil dummy (Af60-1/2ab) 0.326 4.773 0.108 -0.452 -4.29 -0.052 Soil dummy (Agl6-2a) 0.563 2.263 0.224 1.397 6.103 0.406 Soil dummy (Ao l O7-2bc) -0.17 -1.677 -0.05 0.175 1.473 0.028 Soil dummy (I-Lc-Bk-c) 0.101 0.761 0.038 0.573 5.193 0.116 Soil dummy (LclO0-c) 0.947 5.75 0.361 0.309 1.921 0.054 Soil dummy (AolO8-2ab) 0.215 2.536 0.071 -0.52 -3.326 -0.055 Soil dummy (Nd65-3ab) -0.062 -0.599 -0.018 -0.06 -0.387 -0.007 Watershed dummy*** 0.188 3.543 0.026 Protected Area dummy -0.077 -0.332 -0.059 (1986)*** Constant 1.295 8.87 -4.098 -14.01 Rho 1.295 8.87 -0.068 0.309 Log Likelihood -3714.743 No. of observations 4946 Marginal Effects calculated from univariate reduced-form equations. ** Cost is measured as units of primary road traveled, in km. *** Watershed dummy =1 if the impedance-weighted distance to the nearest river is less 3 km, assuming no primary roads. Protected area dummy =1 if pixel lay in a Protected Area in 1986. Population density is measured at the district level. Population density is measured at the district level. 26 Cropper, Puri, Griffiths Table 3 Within-Sample Accuracy of Bivariate Probit Model in Explaining Clearing (Protected Area Sample) Actual - Cleared Forested Percentage of modeled Model Prediction 1- predictions correct Cleared 872 296 75% Forested 657 3133 83% Percentage of sample points correctly 57% 91% predicted by the Bivariate Probit Model Note: Diagonal (Bold) figures show correct predictions. 27 Cropper, Puri, Griffiths TABLE 4 Bivariate Probit Model Estimated Using Wildlife Sanctuary Sample Dependent variable Cleared Land (Yl = 1) Equation [6] Equation [7] Independent variable Coefficient Z Marginal Elasticity Coefficient Z Marginal Elasticity Effect Effect Slope (degrees) -0.09 -10.111 -0.026 -0.561 0.019 2.514 0.001 0.191 Elevation (ms.) -0.001 -6.861 -0.0002 -0.62 0 2.997 0.00003 0.459 Population density 1990 0.003 4.509 0.001 0.163 -0.008 -4.401 -0.0005 -0.753 (people/km2)* * * Log(Cost)(1982)** -0.179 -8.4 -0.051 -0.25 0.292 9.198 0.017 0.638 Provincial dummy -0.199 -1.781 -0.053 0.321 1.251 0.061 (Chiang Mai) Provincial dummy -0.734 -5.602 -0.159 0.868 3.485 0.17 (Mae Hong Son) Provincial dummy (Phayao) -0.341 -2.218 -0.084 1.438 5.116 0.388 Provincial dummy (Nan) -0.335 -2.942 -0.084 -0.472 -1.456 -0.006 Provincial dummy -0.571 -5.046 -0.131 0.047 0.161 0.032 (Lampang) Provincial dummy (Phrae) -0.405 -3.264 -0.097 0.55 1.793 0.113 Provincial dummy -0.66 -3.73 -0.138 1.398 4.562 0.378 (Lamphun) Provincial dummy (Tak) -0.342 -3.06 -0.088 0.94 3.786 0.179 Provincial dummy 0.379 2.913 0.12 0.835 3.003 0.175 (Phitsanulok) Provincial dummy 0.048 0.341 0.013 0.778 2.439 0.17 28 Cropper, Puri, Griffiths (Kamphaeng Phet) Provincial dummy -0.021 -0.135 -0.009 1.972 7.59 0.578 (Uthai Thani) Soil dummy (Af60-1/2ab) 0.277 3.748 0.085 -0.572 -3.881 -0.022 Soil dummy (Agl6-2a) 0.619 2.703 0.212 0.057 0.221 0.003 Soil dummy (Ao l O7-2bc) -0.202 -1.686 -0.052 -0.343 -1.553 -0.015 Soil dummy (I-Lc-Bk-c) 0.012 0.076 -0.002 0.729 6.158 0.077 Soil dummy (Lcl00-c) 0.949 5.647 0.34 0.377 2.271 0.03 Soil dummy (AolO8-2ab) 0.41 4.312 0.131 -0.15 -0.701 -0.008 Soil dummy (Nd65-3ab) -0.104 -0.848 -0.029 0.266 1.427 0.017 Watershed dummy*** 0.133 2.052 0.007 Wildlife sanctuary dummy -0.334 -1.296 -0.077 (1986)*** Constant 1.211 8.055 -4.037 -12.746 Rho 0.018 0.017 Log Likelihood -2890.715 No. of observations 4355 * Marginal Effects calculated from Univariate Reduced form equations. ** Cost is measured as units of primary road traveled in km. *** Watershed dummy =1 if the impedance-weighted distance to the nearest river is less 3 km, assuming no primary roads. Protected area dummy =1 if pixel lay in a Protected Area in 1986. Population density is measured at the district level. 29 Cropper, Puri, Griffiths Figure Titles Figure 1. Forest and Protected Areas Map of North Thailand, 1986 Figure 2. Elevation Map of North Thailand Figure 3. Impact of Population Density on Probability of Clearing, Evaluated at Forest Means Figure 4. Impact of Impedance-Weighted Distance on Probability of Clearing, Evaluated at Forest Means Figure 5. Areas Predicted to be Cleared Figure 6. Areas Predicted to be Cleared after a 100 unit Reduction in Impedance-Weighted Distance 30 Cropper, Puri, Griffiths DForest Non-Forest Protected Areas (National Parks & Wildlife Sanctuaries) 3 1 Cropper, Puri, Griffiths 0 - 250 ms ;i 250 -500ms 0 0 ; 0. 08 N 500 -775 ms 3 ¢0 0000000000:000000;00000<4000 * 775-1200 ms * 1200 -2517 ms 32 Cropper, Puri, Griffiths Impact of Population Density on Probability of Clearing at Forest Means 0.25 0.2 0.15 -_ 8.86 9.36 15.01 28.35 43.71 82.14 104.97 Population Density (p_eoplesquare km) 33 Cropper, Puri, Griffiths Impact of Impedance Weighted Distance to Market on Probability of Clearing at Forest Means 0.3 . .. .. .. 0.25 0.2 0.1 5 0.1 0.05 0 63.31 106.91 256.21 568.50 1011.31 1587.48 2034.49 Impedance Weighted D)stnceto larket 34 Cropper, Puri, Griffiths Thmg Salaeng Luang Ramkamhaeng Nam Nao Khao Sanam Phriang 35 ON CD DO~~~~~~~~~~~~~~~~- z~~~~~~~~~~ z~~~~~~~~~~ Et (0 LX- 0 e pz r:M C7 " 0 I0 Cropper, Puri, Griffiths 1 The remaining 8% of protected areas included arboretums, botanical gardens and reserved areas. L 2If Pik and Cik are exponential functions of population density and distance to market, then these variables will enter Z, linearly. Likewise if {s,1 } are an exponential function of plot characteristics they will enter Z, linearly. 3To reduce the problem of spatial autocorrelation, we sample plots at intervals of 5 km. We have also estimated the model including average values of slope, elevation and distance to market within a 1 0-km radius of plot i. The coefficients of these variables measured for plot i are robust to the inclusion of the average values of the variables on surrounding plots. The Upper North consists of the provinces of Chiang Mai, Chiang Rai, Nan, Lampang, Lamphun, Mae Hong Son, Uthai Thani, Tak and Phrae. 5 In 1986 Northeast Thailand had the lowest per capita income of any area in Thailand. The North had the second lowest per capita income (International Labor Organization 1 99X, Table 3.3). 6 distribution of more familiar soil properties (nitrogen or phosphorous context) is known for all plots in a soil category; however, it is not known at the level of an individual plot. 7 Costdistance is a module in Arc/InfoTM that calculates for each cell the least accurnulative cost of travel from a set of source cells, over a cost surface. 8 This population data may understate the true population due to the large number of ethnic minorities and illegal aliens in the upland portions of Northern Thailand. 37 Cropper, Puri, Griffiths 9 If we calculate the elasticity at means of forested plots, the elasticities with respect to slope and elevation are much higher: -0.66 and -0.84, respectively. Our discussion here focuses on the models reported in Table 2. Results for the clearing equations in Table 4 are qualitatively and quantitatively similar to those in Table 10 As a referee noted, increases in rural population density may reduce agricultural wages through the factor proportions effect. Following the suggestion of a referee, we also used the length of time a pixel had been designated protected to explain the probability of clearing. This variable was, however, insignificant. 38 Policy Research Working Paper Series Contact Title Author Date for paper WPS2566 Convexity and Sheepskin Effects Norbert R. Schady March 2001 T. Gomez in the Human Capital Earnings 32127 Function: Recent Evidence for Filipino Men WPS2567 Policy Reform, Economic Growth, Susmita Dasgupta March 2001 Y. d'Souza and the Digital Divide: An Somik Lall 31449 Econometric Analysis David Wheeler WPS2568 Ownership Structure and the Robert Cull March 2001 Z. Kranzer Temptation to Loot: Evidence from Jana Matesova 38526 Privatized Firms in the Czech Mary Shirley Republic WPS2569 From Users to Custodians: Changing Liz Alden Wily March 2001 M. Yue Relations between People and the Peter A. Dewees 34102 State in Forest Management in Tanzania WPS2570 Asymmetries in Union Relative Wage Niels-Hugo Blunch March 2001 M. Clarke Effects in Ghanaian Manufacturing: Dorte Verner 31752 An Analysis Applying Quantile Regressions WPS2571 Stock Market Responses to Bank Daniela Klingebiel March 2001 R. Vo Restructuring Policies during the Randy Kroszner 33722 East Asian Crisis Luc Laeven Pieter van Oijen WPS2572 Nonfarm Income, Inequality, and Richard H. Adams, Jr. March 2001 N. Obias Poverty in Rural Egypt and Jordan 31986 WPS2573 The Gender Implications of Public Martin Rama March 2001 H. Sladovich Sector Downsizing: The Reform 37698 Program of Vietnam WPS2574 How Adverse Selection Affects the Paolo Belli March 2001 H. Sladovich Health Insurance Market 37698 WPS2575 The Quality of Bureaucracy and Chong-En Bai March 2001 H. Sladovich Capital Account Policies Shang-Jin Wei 37698 WPS2576 Trade Policy, Standards, and Gary Clyde Hufbauer March 2001 L. Tabada Development in Central America Barbara Kotschwar 36896 John S. Wilson Policy Research Working Paper Series Contact Title Author Date for paper WPS2577 Developing Rainfall-Based Index Jerry Skees April 2001 P. Kokila Insurance in Morocco Stephanie Gober 33716 Panos Varangis Rodney Lester Vijay Kalavakonda WPS2578 How Accession to the European Bartlomiej Kaminski April 2001 L. Tabada Union Has Affected External Trade 36896 and Foreign Direct Investment in Central European Economies WPS2579 Public Health and Education Sudharshan Canagarajah April 2001 M. Clarke Spending in Ghana in 1992-98: Xiao Ye 31752 Issues of Equity and Efficiency WPS2580 On "Indirect" Trade-Related Research Olivier Lumenga-Neso April 2001 L. Tabada and Development Spillovers Marcelo Olarreaga 36896 Maurice Schiff WPS2581 Securities Clearance and Mario Guadamillas April 2001 H. Issa Settlement Systems: A Guide to Robert Keppler 30154 Best Practices WPS2582 Development Financing during a Suhas Ketkar April 2001 S. Crow Crisis: Securitization of Future Dilip Ratha 30763 Receivables