53370 Protected Area Effectiveness in Reducing Tropical Deforestation A Global Analysis of the Impact of Protection Status 7 on ce iti en Ed fer n Co Protected Area Effectiveness in Reducing Tropical Deforestation A Global Analysis of the Impact of Protection Status Andrew Nelson and Kenneth M. Chomitz Evaluation Brief 7 October 2009 The World Bank http://www.worldbank.org/ieg Washington, D.C. ©2009 Independent Evaluation Group Communications, Learning, and Strategy The World Bank 1818 H Street, NW Washington, DC 20433 E-mail: ieg@worldbank.org Telephone: 202-458-4487 Fax: 202-522-3125 http://www.worldbank.org/ieg All rights reserved This Evaluation Brief is a product of the staff of the Independent Evaluation Group (IEG) of the World Bank. The findings, in- terpretations, and conclusions expressed here do not necessarily reflect the views of the Executive Directors of the World Bank or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this work. The boundaries, colors, denomina- tions, and other information shown on any map in this work do not imply any judgment on the part of the World Bank or IEG concerning the legal status of any territory or the endorsement or acceptance of such boundaries. Rights and Permissions The material in this publication is copyrighted. Copying and/or transmitting portions or all of this work without permission may be a violation of applicable law. IEG encourages the dissemination of its work and will normally grant permission to re- produce portions of the work promptly. For permission to photocopy or reprint any part of this work, please send a request with complete information to ieg@worldbank.org. ISBN-13: 978-1-60244-123-1 ISBN-10: 1-60244-123-5 Contents v Acknowledgments vii Executive Summary 1 1 Introduction 5 2 Matching Methods 7 3 Data and Sampling 21 4 Results 27 5 Conclusions 29 References Acknowledgments The authors thank Lucas Joppa and Alex Pfaff for Salvemini, Cheikh M’Backe Falle, and Janice useful discussions on the spatial data and the Joshi. Heather Dittbrenner edited the paper. matching methods employed here and for access This work was funded in major part by the to unpublished manuscripts. They thank Andrew Evaluation Department of the Norwegian Agency Warner for helpful comments and Lucy Fish at for Development Cooperation (Norad). UNEP-WCMC for advice on the World Database on Protected Areas. The authors are grateful for Comments on this working paper are welcome: the assistance of Dinara Akhmetova, Diana kchomitz@worldbank.org. v Executive Summary The REDD agenda (Reducing Emissions from evaluations. One area of dispute is the relative Deforestation and Degradation) seeks to effectiveness in deforestation reduction of mobilize positive incentives for countries to strictly protected areas versus areas that allow reduce deforestation, the source of 20 percent of some degree of sustainable use by local people. anthropogenic greenhouse emissions. To be successful, this agenda requires not only financ- This study assesses the impact of tropical ing and international agreement on procedures, protected areas on deforestation fires, which are but it also needs practical guidance on how to the best available globally consistent proxy for accomplish such reductions in ways that also deforestation at a fine spatial scale. The paper promote local environmental and development covers the entire tropical forest biome to goals. estimate the avoided deforestation afforded by several thousand protected areas. Building on Such guidance may come from existing efforts in recent advances, the authors use matching the establishment of protected areas and indige- methods to compare protected area points with nous areas. Motivated by biodiversity, environ- similar unprotected points, controlling for slope, mental, social, and land rights concerns, these rainfall, road proximity, and other factors affect- interventions encourage forest conservation and ing both deforestation and protected area sustainable use and would often be expected to placement. Unlike previous studies, this work reduce deforestation. Protected areas have provides a continuous measure of the effective- expanded in recent years and now cover 27 ness of protection as a function of varying percent of the tropical forest biome. Forests degrees of deforestation pressure, as well as for controlled by local and indigenous communities different classes of protection (strict, multi-use have also expanded. An assessment of the and indigenous). effectiveness of these areas in reducing de- forestation could inform the design of interven- Across the biome, the paper finds that protected tions to promote REDD: reduced carbon areas generally have significantly lower fire rates emissions from deforestation and degradation. than comparable nonprotected areas, but this Yet there is considerable uncertainty and contro- differential declines as remoteness increases. versy over the impacts and effectiveness of Multi-use protected areas generally provide protected areas and very few well-designed greater deforestation reduction (in absolute vii E V A L U AT I O N B R I E F 7 terms) than strict protected areas (see Figures deforestation pressure. Indigenous areas have an ES.1 and ES.2). This protective effect may be even higher protective impact. Estimates for Africa obscured because the multi-use protected areas indicate modest impact of strict protected areas, tend to be established in zones of higher but results are not robust for multi-use areas. Figure ES.1: Forest Fire Rates in Figure ES.2: Forest Fire Rates in Latin America and the Caribbean, Latin America and the Caribbean, Strict Protection Multi-Use Protection viii CHAPTER 1 Introduction Tropical deforestation accounts for between one been the most prominent and best funded, by fifth and one quarter of the total human contri- the World Bank, other donors, and host bution to greenhouse gases (Gullison and others countries. The Global Environment Facility says 2007; Kindermann and others 2008), and a larger that its investments in protected areas include proportion of emissions from developing $1.6 billion of its own resources and $4.2 billion countries. Reduction of deforestation therefore in cofinancing; much of this has been im- contributes to climate change mitigation and plemented through the World Bank. Protected may also provide development benefits areas have expanded rapidly in recent years (Canadell and Raupauch 2008; Miles and Kapos (UNEP/IUCN 2009; UNEP-WCMC 2008) and now 2008; Chomitz 2007). The REDD (Reduced cover around 27.1 percent1 of the tropical forest Emissions from Deforestation and Degradation) estate. In many ways they provide a model for agenda seeks to integrate deforestation re- broader classes of intervention, since most duction into the global climate regime under the efforts to reduce deforestation will involve some United Nations Framework Convention on kinds of restrictions on land use practices Climate Change, rewarding countries that (Chomitz 2007). reduce forest emissions (Canadell and Raupauch 2008; FAO, UNDP , and UNEP 2008). But there is a Yet there is considerable uncertainty and contro- dearth of rigorous evaluations of the impact of versy over the impacts and effectiveness of specific interventions on deforestation (Hansen protected areas, and very few well designed and others 2008; Chomitz 2007). evaluations (Andam and others 2008; Ferraro and Pattanayak 2006). On one hand, protected Although the REDD agenda is new, the forest areas are sometimes characterized as ineffective protection agenda is not. Conservation and “paper parks.” On the other, there is increasing sustainable management of forests have been evidence that deforestation rates are lower in motivated by biodiversity and livelihood protected areas (see, for example, Nepstad and concerns for decades. Where deforestation is a others 2006). However, this observed impact threat to biodiversity, successful conservation or may be partially illusory, because protected areas sustainable management efforts will have a side tend to be established in areas that are unattrac- benefit of reducing forest carbon emissions. So tive to agricultural conversion. A small but an evaluation of the effectiveness of past conser- growing literature has applied increasingly vation efforts can inform the design of interven- sophisticated statistical procedures to control for tions to promote REDD. This is especially salient this source of bias (Andam and others 2008; in the humid tropical forests, where deforesta- Ferraro and Pattanayak 2006; Chomitz and Gray tion rates and carbon densities are both high. 1996; Cropper, Puri, and Griffiths 2001; Joppa Their loss is the major source of forest carbon and Pfaff 2009c; Pfaff 2009a; Pfaff and others emissions. 2009; Ferraro 2008). Among conservation interventions in tropical Building on and extending some recent method- forests, the establishment of protected areas has ological advances (Andam and others 2008; Pfaff 1 E V A L U AT I O N B R I E F 7 2009a; Pfaff and others 2009; Joppa and Pfaff control for the location of the protected area and 2009b), this study is an impact evaluation of the its characteristics to ensure that any comparison of effect of tropical protected areas on deforesta- land cover change—particularly deforestation— tion fires, which are the best available consistent between protected and unprotected lands is proxy for deforestation at a fine spatial scale. It unbiased. Location must be accounted for, uses the spatial analysis of remote sensing data because protected areas may be disproportion- to characterize the tropical forest biome and ately located in areas characterized by higher matching methods (Morgan and Harding 2006) slopes, greater distance to cities, and lower suitabil- to control for the effects of location and ity for agriculture (Joppa and Pfaff 2009b). These landscape characteristics to ensure unbiased factors, which are strongly associated with lower estimates of the avoided deforestation fires pressures for deforestation, presumably reduce provided by different classes of protection. Thus, the political and economic costs of imposing land location-specific estimates are generated based use restrictions (Chomitz 2007). If a protected area on almost 3,000 protected areas covering 2 is remote, has poor-quality soil or difficult terrain, million square kilometers (km) of the tropical or is subject to extremely high rainfall, then it may forest biome. Unlike previous work, this work well benefit from de facto protection. Comparing provides a continuous measure of effectiveness these “low-pressure” areas to unprotected lands in as a function of varying degrees of deforestation general might show that legislated protection has pressure (proxied by travel time to the nearest significant benefits for avoiding deforestation. But city). if unprotected lands with similarly unappealing characteristics also exhibited little or no change in This study does not evaluate the impact of forest cover, then such legal protection would be protected areas on local welfare or livelihoods— minimal. a controversial subject on which there is very little rigorous evaluation (Ferraro and Pattanayak Conversely, protected areas in “high-pressure” 2006; Ferraro 2008). However, it does address zones, with good access to roads and markets the relative impacts on deforestation of strict and containing agriculturally suitable environ- protected areas versus areas in which local ments, may exhibit greater levels of degradation people have greater rights of use and than unprotected lands in general. Again, if only management. such high-pressure protected areas were compared with unprotected areas facing similar 1.1. Assessing protected area pressures, the result would likely be that such effectiveness protected areas, although possibly degraded, do provide a degree of protection (Joppa and Pfaff Assessing the impact of protection in terms of 2009b, 2009c; Adeney, Christensen, and Pimm land cover change is challenging, whether it is 2009; Joppa, Loarie, and Pimm 2008). assessed as part of a detailed park study, region- ally (Nepstad and others 2006), or globally Joppa and Pfaff (2009c) provide a recent review (Bruner and others). Earth observation data of the empirical literature on the impact of provide ever more detailed and more frequent protected areas on deforestation. Most conven- pictures of land cover, climate, and events such tional studies have not fully controlled for the as fires and as a result have become a key source bias in location. However, a number of recent of information for such studies, along with other studies have introduced controls for attractive- spatial information on population density, ness of conversion (Chomitz and Gray 1996; transport networks, protected area boundaries, Deninger and Minten 2002), econometrics with and the like. controls for endogeneity of protected area placement (Cropper, Puri, and Griffiths 2001), As recent studies have demonstrated (Joppa and and, more recently, matching methods that are Pfaff 2009c; Pfaff and others 2009), it is vital to thought to be less sensitive to specification error 2 P R O T E C T E D A R E A E F F E C T I V E N E S S I N R E D U C I N G T R O P I C A L D E F O R E S TAT I O N (Andam and others 2008; Pfaff 2009a; Pfaff and This study builds on the matching approach used others 2009; Joppa and Pfaff 2009a, 2009b). in Andam and others (2008) and Joppa and Pfaff These methods seek to pair protected forest (2009a). It differs from the latter in several plots with unprotected but otherwise similar important respects. First, it focuses on the “control” plots. tropical forest biome, where deforestation rates and carbon emission rates are highest. Second, Andam and coauthors (2008) used matching as a result of this focus, it can use what is for the methods to assess the deforestation-reducing moment the most consistent and up-to-date impact of Costa Rica’s system of protected areas. high-resolution proxy for deforestation: forest They found that protected areas on average did fires. Third, it presents results by continent modestly reduce deforestation, but by substan- rather than country but disaggregates protected tially less than a naïve comparison of mean area impacts by distance to a city (a proxy for deforestation rates in protected versus un- deforestation pressure). Finally, it breaks out protected areas. Pfaff and others (2009) qualify results for multi-use protected areas (Interna- this result, showing that Costa Rican parks had a tional Union for Conservation of Nature [IUCN] greater protective effect in areas facing greater categories V and VI, and indigenous areas) to pressure, such as those close to the capital. inform the debate about the advantages and Joppa and Pfaff (2009a) extend the approach of disadvantages of strict protection. Andam and others to the global set of protected areas, assessing impacts by country on forest The next section briefly describes the most cover in 2000 and 2005, and for the 2000–05 commonly used matching estimators and is change in cover. (Because the two land cover followed by a description of the study area and datasets used different methods, the change an assessment of the suitability of available measure is acknowledged to be “noisy.”) They spatial data for a global scale analysis of avoided found, again, that deforestation reduction was deforestation in the tropical forests. The final generally less than a simple comparison would section presents the results of the two analyses. indicate. In short, there have been several well-defined Note studies of the effectiveness of protection for 1. Boundary and area data are not available for a avoiding deforestation, but none has simultane- small percentage of protected areas, and so this ously addressed differences in pressure, location may be a conservative estimate. bias, and protection status on a global scale. 3 CHAPTER 2 Matching Methods Recent reviews and evaluations of matching have all covariables. It can, however, be combined presented its benefits in providing robust es- with other matching approaches to force exact timates of causal effects (Morgan and Harding matching on a subset of the covariates, for 2006) and as a nonparametric preprocessing tool example, to force matching pairs to selected (Ho and others 2007). This chapter provides a from the same country. Such selective exact brief background to the methods used in this matching is an important requirement for this paper. analysis, because it ensures that average results across the biome can be disaggregated by Matching has become a popular method of country. causal inference, particularly in econometrics, but also in fields as diverse as law, medicine, and Nearest neighbor matching identifies the most conservation policy (Morgan and Harding 2006; similar treatment cases to each control case by Ho and others 2007; Joppa and Pfaff 2009c; means of a distance measure derived from the Sekhon 2007). Matching works by identifying a difference across all the covariables. One control group that is “very similar” to the common measure of similarity is the Mahala- treatment group with only one key difference: nobis distance metric, a scale invariant measure the control group did not participate in the of the multidimensional distance between two program of interest. In this case, the program of points. Typically, the algorithm randomly orders interest is designated protection: was a patch of the treatment cases and for each one in turn land protected (treatment group) or not (control selects the control case with the smallest group)? Defining “very similar” based on the distance. Poor matches are avoided by assigning covariates of each case is one challenge facing a tolerance to judge the quality of the match. This the researcher when applying matching to data. distance tolerance is termed a caliper and simply The aim here is to identify a matching control determines the acceptable similarity for a match. case for each treatment case to produce a balanced dataset, where a perfectly balanced The authors use remotely sensed data on forest dataset would consist of pairs of cases with fire activity between 2000 and 2008 as a measure identical covariables in the treatment and control of deforestation and used tropical forest cover in groups. year 2000 and information on protected areas to characterize the tropical forest biome in year Matching algorithms take different approaches 2000. The matching aims to provide unbiased to defining “very similar.” A variety of approaches estimates of avoided deforestation fires in exists; this paper relies on exact matching and protected areas in the tropical forests for differ- nearest neighbor matching. ent classes of protection and for different levels of pressure. The data, the study area, the Exact matching simply identifies pairs of identi- disaggregation by protection type, and the cally matching cases in the two groups, based on definition of pressure are described in chapter 3. 5 CHAPTER 3 Data and Sampling All spatial data were projected to equal area separately. Papua New Guinea and Micronesia sinusoidal projection, with a WGS84 datum and are considered part of Asia for this analysis. spheroid. Unless otherwise stated, raster resolu- tion is 1 km. The relevant data from each data 3.2. Tropical forest area layer were extracted at 1-km spacing and stored in a PostgreSQL database (version 8.3). The Within this area, the extent of the remaining following sections describe these data layers. tropical forest in 2000 was extracted from two land cover data sources: Global Land Cover for 3.1. Study area the year 2000 (Bartholome and Belward 2005) derived from ~1 km resolution SPOT data and The study is limited to developing countries Percentage Forest Canopy Cover for 2000 (recipient countries of World Bank loans) and (Hansen and others 2003) derived from ~500 m the extent of the tropical forest biome. The MODIS data. biome—derived from the World Wildlife Fund’s Terrestrial Ecoregions of the World (Olson and All 11 land cover classes from GLC2000 that others 2001)—contains the maximum spatial contain forest or forest mosaics were extracted, extent of the world’s tropical and subtropical along with all ~1-km pixels where the average moist broadleaf forests. percent forest cover was greater than 25 percent (Hansen and others 2008). This is a higher Figure 1 shows the spatial intersection of these threshold than the 10 percent used in the FAO countries and the biome. The area in green is the Forest Resource Assessment (FAO 2006) and in a maximum extent of the study area covering 19.73 recent assessment of global forest protection million km2. The biome is clearly split across (Schmitt and others 2009). One justification for three continents; each will be analyzed using the 10 percent threshold in those global Figure 1: Extent of the Tropical Forest Biome 7 E V A L U AT I O N B R I E F 7 analyses was to capture woodland areas in Africa; Figure 2: Fire Activity and Forest Extent however, these are not part of the tropical forest biome. Twenty-five percent was chosen to minimize the risk of including tropical woodlands/savannas and other land that was already largely cleared of forest, that was predominantly used for agriculture, and that could exhibit high fire activity that was not necessarily related to deforestation events. This delineation of tropical forest extent is a conservative estimate based on the common area of both these forest layers within the boundaries of the biome, covering 13.15 million km2 of tropical forest area in 2000. For reference, a tropical forest extent based on the MODIS data alone or on GLC2000 alone would amount to 15.13 million km2 or 14.51 million km2, respec- biome and 0.70 million of these occurred in tively. Agreement between the two across the forested areas (Table 1). Of the 13.15 million biome is 83.1 percent. 1-km tropical forest pixels, 5.31 percent had at least one fire event in that time frame. 3.3. Outcome variable: Fire activity on forests The outcome variable is a binary measure of forest fire activity per square km: was there ever Fire activity (Figure 2)—overlaid on forest a fire event in that pixel during 2000–08? This extent—was used as a proxy for tropical time period is reflected in the choice of covari- deforestation fire events. (The overlay screens ables and the definition of the control/treatment out fires used for land management on groups below. The lack of coverage until October previously cleared areas such as pastures.) Fire 2000 and then partial coverage until July 2002 activity was estimated from spatially referenced implies that the binary measure here is slightly remote sensing data on forest fires from the conservative as an estimate of fire-affected area. MODIS Active Fires dataset (Justice and others 2002). MODIS Active Fire data are provided on Another dataset was considered as a proxy for two satellite platforms, Terra from October 2000 tropical deforestation events: the recently and Aqua from July 2002, both to present day. released MODIS Collection 5 Burned Area Thus, there is partial coverage from October Product, which includes global, monthly 500- 2000 (two passes per day) and complete coverage from July 2002 (four passes per day), Table 1: 1-km Forest and Fire Pixel including both day and night passes. Statistics in the Tropical Forest Biome (2000–08) Following Morton and others (2008) in their study of fire activity in the Amazon, this paper Forest Fire Fire extracted only the high-confidence fires—all fires Region pixels pixels rate occurring at night and daytime fires with > 330K Biome 13,154,816 698,514 0.0531 brightness temperature in the 4 μm channel— LAC 6,989,019 365,074 0.0522 from more than 1 million MODIS fires scenes between 2000 and 2008.1 Some 1.21 million 1-km Africa 2,529,918 142,913 0.0565 pixels recorded at least one fire between October Asia 3,635,879 190,527 0.0524 2000 and January 2009 in the tropical forest Note: LAC = Latin America and the Caribbean. 8 P R O T E C T E D A R E A E F F E C T I V E N E S S I N R E D U C I N G T R O P I C A L D E F O R E S TAT I O N meter (m) resolution maps of burn dates.2 A burned area data as a proxy for tropical provisional version of this data was made deforestation events. available for evaluation but is currently offline. A direct comparison between the active fire and The presence of one or more fires in a 1-km pixel burned area data for July 2001–June 2002 made cannot be directly translated into an estimate of the following pertinent finding for burned area deforested area. A fire event may represent and fire detections by land cover class: anything from a small clearing of a single hectare to complete deforestation of the 1-km pixel. Savannas, woody savannas, grasslands and However, it can be assessed whether this fire shrublands account alone for 85% of the presence/absence data can be used as a plausible MODIS burned areas (over 3.1×106 km2), proxy for deforestation activity in the tropical a figure consistent but greater than with forest biome. The authors compared the binary the active fires detections, which account measure of forest fire activity to a recently for 73.7% (over 2.38×106 km2). Conversely, published, Landsat-calibrated, biome-wide dataset the five forest classes (evergreen needle- with a spatial resolution of 18.5 km that quantifies leaf, evergreen broadleaf, deciduous forest cover loss from 2000 to 2005 (Hansen and needleleaf, deciduous broadleaf and mixed others 2008). They plotted the area of fire activity forest) account for only 5.5% of the global for 2000–05 as a proportion of forest area against MODIS burned areas (0.20×106 km2) but percent forest cover loss for 2000–05 per 18.5-km for 11.6% of the active fire detections (over pixel (Figure 3). The analysis was repeated for 5 0.37×106 km2), highlighting the fact that percent (left of figure) and 1 percent (right of many forest fires are detected by the active figure) bins of forest cover loss. fire product but not by the burned area product (Roy and others 2008). There is a strong trend of increasing fire activity with increased loss of forest cover across the This higher detection rate, albeit including both biome, from 0 to 30 percent forest cover loss. medium- and high-confidence fires, and the fact The trend continues for higher forest cover loss that the burned area data are still provisional, led percentages, but there are very few 18.5-km to a preference for the active fire data over the pixels (<0.2 percent of the tropical forest biome Figure 3: Forest Fire Rate (fire area/forest area) against Forest Cover Loss for 2000–05, with Linear Trend Lines Fire area as a proportion of tropical forest area against tropical forest cover loss (2000-2005) Fire area as a proportion of tropical forest area against tropical forest cover loss (2000-2005) 0.45 0.45 Biome LAC Africa Asia Biome LAC Africa Asia 0.40 Linear (Biome) Linear (LAC) Linear (Africa) Linear (Asia) 0.40 Linear (Biome) Linear (LAC) Linear (Africa) Linear (Asia) R2 = 0.9681 R 2 = 0.9667 R 2 = 0.2489 R 2 = 0.916 0.35 R2 = 0.9551 R2 = 0.9243 R2 = 0.0234 R2 = 0.4261 0.35 Fire area as a proportion of forest area Fire area as a proportion of forest area 0.30 0.30 0.25 0.25 0.20 0.20 0.15 0.15 0.10 0.10 0.05 0.05 0.00 0.00 0 5 10 15 20 25 30 0 5 10 15 20 25 30 Tropical forest cover loss % (2000-2005) Tropical forest cover loss % (2000-2005) 9 E V A L U AT I O N B R I E F 7 area) in these areas. Latin America and the scientific research and/or environmental Caribbean and Asia show the same clear trend as monitoring. the whole biome, but the case is less clear for Africa. It should be noted that the remote CATEGORY Ib: Wilderness Area: Protected area sensing estimate of African deforestation differed managed mainly for wilderness protection. drastically from the Forest Resources Assessment Definition: Large area of unmodified or (2005) by the FAO (Hansen and others 2008; FAO slightly modified land and/or sea retaining its 2006), so the deviation between the fire natural character and influence, without measures and the remote sensing measures may permanent or significant habitation, which is not be solely due to misclassification of the fire protected and managed so as to preserve its data. natural condition. From this it is reasonably sure that the chosen CATEGORY II: National Park: Protected area subset of active fires is a plausible proxy for managed mainly for ecosystem protection deforestation events, especially in Latin America and recreation. and Asia. The case is less convincing for Africa Definition: Natural area of land and/or sea but is still plausible. designated to (a) protect the ecological integrity of one or more ecosystems for 3.4. Protected areas and IUCN present and future generations; (b) exclude management classes exploitation or occupation inimical to the purposes of designation of the area; and (c) The World Database on Protected Areas (WDPA) provide a foundation for spiritual, scientific, (UPED/IUCN 2009) is the source for protected area educational, recreational, and visitor opportu- information. Protected area information, including nities, all of which must be environmentally park boundaries (and park center coordinates and and culturally compatible. area for areas with unknown boundaries), designa- tion date, IUCN protected area management classi- CATEGORY III: Natural Monument: Protected fication, and status were extracted from the WDPA area managed mainly for conservation of database for all protected areas that were inside or specific natural features. that intersected the tropical forest biome. Definition: Area containing one or more specific natural or natural/cultural feature that This list of protected areas includes all nationally is of outstanding or unique value because of (IUCN protected area management classes I its inherent rarity, representative or aesthetic through VI as well as unknown) and internation- qualities, or cultural significance. ally (UNESCO MAB reserves, Ramsar sites, and World Heritage sites) recognized parks and CATEGORY IV: Habitat/Species Management amounts to 4.13 million km2 of protected area Area: Protected area managed mainly for within the biome, of which 3.62 million km2 is conservation through management interven- forested. tion. Definition: Area of land and/or sea subject to The six management classes as described by active intervention for management purposes IUCN are— to ensure the maintenance of habitats and/or to meet the requirements of specific species. CATEGORY Ia: Strict Nature Reserve: Protected area managed mainly for science. CATEGORY V: Protected Landscape/Seascape: Definition: Area of land and/or sea possess- Protected area managed mainly for landscape/ ing some outstanding or representative seascape conservation and recreation. ecosystems, geological or physiological fea- Definition: Area of land, possibly with coast tures and/or species, available primarily for and sea, where the interaction of people and 10 P R O T E C T E D A R E A E F F E C T I V E N E S S I N R E D U C I N G T R O P I C A L D E F O R E S TAT I O N nature over time has produced an area of • Nonstrict or multi-use protection—IUCN distinct character with significant aesthetic, classes V and VI ecological, and/or cultural value, and often • Unknown protection—Nationally recognized with high biological diversity. Safeguarding but with no IUCN class the integrity of this traditional interaction is • Indigenous—A subset of the unknown class, vital to the protection, maintenance, and but under indigenous stewardship. evolution of such an area. Strict protection means areas that are designed CATEGORY VI: Managed Resource Protected specifically for nature protection. Multi-use Area: Protected area managed mainly for the protection means that the areas allow some form sustainable use of natural ecosystems. of sustainable use. The indigenous group of Definition: Area containing predominantly protected areas occurs in Latin America, predom- unmodified natural systems, managed to inantly in Brazil, with a few areas in Panama and ensure long-term protection and mainte- Colombia. Figure 4 shows the IUCN classified nance of biological diversity, while providing a protected areas that were designated before sustainable flow of natural products and 2000; the dominance of the protected tropical services to meet community needs. forest area in Latin America and the Caribbean is clear. There were 2,974 IUCN classified (IUCN Two treatment groups were considered, based classes I through VI, plus unknown) protected on protected areas with boundary information. areas designated before 2000 in the tropical The first group consists of all protected areas that forest biome that contained at least 1 km2 of were designated pre-2000. The second group is tropical forest. restricted to protected areas that were designated between 1990 and 2000. Use of the The control groups are based on areas that have restricted group allows us to examine the impact never been protected, up through 2008. Thus, of more recently created protected areas and any tropical forest area that has ever been provides a check against the possibility of protected based on the entire tropical forest endogeneity in the matching variables. coverage of the World Database of Protected Areas (WDPA) park boundaries is excluded. Each protected area has been assigned an IUCN Where boundary data were missing, protected management class. These two groups are areas were represented by circles around center disaggregated further based on the IUCN coordinates. management classes: Summary statistics for tropical forest area and • Strict protection—IUCN classes I though IV protected tropical forest area are shown in Tables Figure 4: Protected Areas in the Tropical Forest Biome with an IUCN Management Classification Designated Before 2000 11 E V A L U AT I O N B R I E F 7 Table 2: Total Tropical Forest Protected (km2 and %) by Protection Class and Region Area Biome Latin America and the Caribbean Africa Asia Forest Area 13,154,816 6,989,019 2,529,918 3,635,879 Protected Area 3,619,941 (27.5) 2,719,301 (38.9) 411,761 (16.3) 488,879 (13.4) Ia 166,892 (1.3) 152,650 (2.2) 1,425 (0.1) 12,817 (0.4) Ib 21,207 (0.2) 10,415 (0.1) 1,097 (0.0) 9,695 (0.3) II 740,910 (5.6) 482,193 (6.9) 127,902 (5.1) 130,815 (3.6) III 57,837 (0.4) 47,140 (0.7) 483 (0.0) 10,214 (0.3) IV 142,896 (1.1) 21,211 (0.3) 20,447 (0.8) 101,238 (2.8) Strict (I–IV) 1,129,742 (8.6) 713,609 (10.2) 151,354 (6.0) 264,779 (7.3) V 239,072 (1.8) 190,400 (2.7) 52 (0.0) 48,620 (1.3) VI 799,854 (6.1) 716,626 (10.3) 26,069 (1.0) 57,159 (1.6) Multi-use (V–VI) 1,038,926 (7.9) 907,026 (13.0) 26,121 (1.0) 105,779 (2.9) Unknown 544,336 (4.1) 215,721 (3.1) 216,377 (8.6) 112,238 (3.1) Indigenous 850,394 (6.5) 850,394 (12.2) 0 (0.0) 0 (0.0) Other 56,543 (0.4) 32,551 (0.5) 17,909 (0.7) 6,083 (0.2) Note: Numbers in parentheses are percentages of the region’s total forest area. 2–4. Comparing Tables 2 and 3, it is clear that protected area almost doubled in size, with there has been a massive expansion of the notable expansions in types Ia, II, V, VI, and protected area in the biome between 2000 and indigenous areas in Latin America and the 2008, from almost 2 million km2 to 3.6 million Caribbean. The major gains in Africa and Asia km2, or 15–27 percent of the biome, well above come from the unknown classification. the Convention on Biological Diversity target Unknown areas—which may signify an inclusion area of 10 percent (Schmitt and others 2009). For of new but incomplete protected area data to the Latin America and the Caribbean and Africa, the WDPA database—increased in all regions. Table 3: Pre-2000 Tropical Forest Protected (km2 and %) by Protection Class and Region Area Biome Latin America and the Caribbean Africa Asia Protected Area 1,972,474 (15.0) 1,418,225 (20.3) 224,362 (8.9) 329,887 (9.1) Ia 75,391 (0.6) 62,157 (0.9) 671 (0.0) 12,563 (0.3) Ib 10,785 (0.1) 10,411 (0.1) 257 (0.0) 117 (0.0) II 588,005 (4.5) 348,957 (5.0) 122,201 (4.8) 116,847 (3.2) III 40,709 (0.3) 35,557 (0.5) 91 (0.0) 5,061 (0.1) IV 116,814 (0.9) 15,594 (0.2) 18,949 (0.7) 82,271 (2.3) Strict (I–IV) 831,704 (6.3) 472,676 (6.8) 142,169 (5.6) 216,859 (6.0) V 144,595 (1.1) 113,150 (1.6) 52 (0.0) 31,393 (0.9) VI 487,342 (3.7) 420,399 (6.0) 21,653 (0.9) 45,290 (1.2) Multi-use (V–VI) 631,937 (4.8) 533,549 (7.6) 21,705 (0.9) 76,683 (2.1) Unknown 119,808 (0.9) 30,405 (0.4) 54,088 (2.1) 35,315 (1.0) Indigenous 359,914 (2.7) 359,914 (5.1) 0 (0.0) 0 (0.0) Other 29,111 (0.2) 21,681 (0.3) 6,400 (0.3) 1,030 (0.0) Note: Numbers in parentheses are percentages of the region’s total forest area. 12 P R O T E C T E D A R E A E F F E C T I V E N E S S I N R E D U C I N G T R O P I C A L D E F O R E S TAT I O N Table 4: Tropical Forest Protected (km2 and %) by Protection Class and Region, 1990–2000 Area Biome Latin America and the Caribbean Africa Asia Protected Area 807,704 (6.1) 631,591 (9.0) 46,574 (1.8) 129,539 (3.6) Ia 19,222 (0.1) 17,892 (0.3) 0 (0.0) 1,330 (0.0) Ib 10,525 (0.1) 10,411 (0.1) 0 (0.0) 114 (0.0) II 200,036 (1.5) 102,365 (1.5) 34,617 (1.4) 63,054 (1.7) III 16,144 (0.1) 14,315 (0.2) 0 (0.0) 1,829 (0.1) IV 24,512 (0.2) 9,175 (0.1) 0 (0.0) 15,337 (0.4) Strict (I–IV) 270,439 (2.1) 154,158 (2.2) 34,617 (1.4) 81,664 (2.2) V 60,229 (0.5) 57,231 (0.8) 0 (0.0) 2,998 (0.1) VI 195,355 (1.5) 170,830 (2.4) 4,042 (0.2) 20,483 (0.6) Multi-use (V–VI) 255,584 (1.9) 228,061 (3.3) 4,042 (0.2) 23,481 (0.6) Unknown 42,100 (0.3) 14,836 (0.2) 3,889 (0.2) 23,375 (0.6) Indigenous 219,258 (1.7) 219,258 (3.1) 0 (0.0) 0 (0.0) Other 20,323 (0.2) 15,278 (0.2) 4,026 (0.2) 1,019 (0.0) Note: Numbers in parentheses are percentages of the region’s total forest area. Table 5. Forest and Fire Area (km2) and Crude Fire Rates per Region/Protection Group Fire rate Expected Avoided relative to fire pixels fire pixels mean at mean un- at mean un- Protection class Forest pixels Fire pixels Fire rate* unprotected* protected rate* protected rate * Latin America and the Caribbean Never 4,269,718 317,608 0.0744 Strict (I–IV) 472,676 7,597 0.0161 Ϫ0.0583 35,161 27,564 Multi-use (V–VI) 533,549 16,245 0.0304 Ϫ0.0439 39,689 23,444 Unknown 30,405 646 0.0212 Ϫ0.0531 2,262 1,616 Indigenous 359,914 5,414 0.0150 Ϫ0.0593 26,773 21,359 Africa Never 2,118,157 128,499 0.0607 Strict (I–IV) 142,169 2,538 0.0179 Ϫ0.0428 8,625 6,087 Multi-use (V–VI) 21,705 654 0.0301 Ϫ0.0305 1,317 663 Unknown 54,088 3,393 0.0627 0.0021 3,281 Ϫ112 Asia Never 3,147,000 172,212 0.0547 Strict (I–IV) 216,859 9,801 0.0452 Ϫ0.0095 11,867 2,066 Multi-use (V–VI) 76,683 2,810 0.0366 Ϫ0.0181 4,196 1,386 Unknown 35,315 495 0.0140 Ϫ0.0407 1,933 1,438 *This table compares aggregate mean fire rates between protected and unprotected areas and does not control for differences in deforestation pressure between protected and unpro- tected areas. 13 E V A L U AT I O N B R I E F 7 Comparing Tables 3 and 4, almost half (45 naïve estimate of impact is modified when other percent) of the pre-2000 protected area factors affected deforestation are controlled (see expansion in Latin America and the Caribbean Figure 4). happened between 1990 and 2000, though most of this is associated with multi-use and indige- 3.5. Pressure on protected areas nous areas. In Africa there was little expansion (21 percent) of the protected area network Some protected areas may be naturally protected during 1990–2000, and that expansion was because of their remoteness and inaccessibility, limited to IUCN classes II, VI, and unknown. This regardless of the level and effectiveness of small area will have implications in the interpre- designated protection. Examples of this de facto tation of the following matching analyses for the protection can be observed in the Amazon and African 1990–2000 treatment groups. For Asia, Congo basins. Conversely, protected forest areas the 1990–2000 expansion accounts for almost 40 in densely populated and easily accessible percent of the pre-2000 protected area network. regions—such as those in Ghana that are clearly visible as islands of intact forest—remain The number of observed tropical forest fire forested because of their designated status and pixels and the tropical forest area for each region enforced or de jure protection (Pfaff and others and protection group (pre-2000 areas only) are 2009; Joppa, Loarie, and Pimm 2008; Joppa and shown in Table 5. The last three columns show Pfaff 2009c). The application of the matching crude measures of the amount of avoided fire approach estimates the average effect of protec- activity without accounting for the nonrandom tion across each continent, but there is strong location of the protected areas or the character- evidence to suggest that the effect will vary istics of the protected and non-protected areas. depending on the ease of access to the protected In all cases (except unknown protection in area (Pfaff and others 2009). Africa), these tabulations show lower fire activity in protected versus unprotected areas, with To assess this, a recent model of travel time to differences as high as 5.9 percent for indigenous major cities in 2000 (Nelson 2008) was used as a areas. Strict protection has lower fire rates than measure of accessibility. (In the Latin American multi-use protection in Latin America and the example in Figure 5, deep red areas are remote, Caribbean and Africa by 1.2–1.4 percent, whereas and yellow areas are near cities.) Major cities are the converse is true in Asia, where strict protec- defined as having a population of 50,000 or more tion appears quite ineffective compared to multi- in 2000. Protected pixels that are easily accessible use and unknown. Nonprotected rates are are assumed to face higher pressure for land higher in Latin America and the Caribbean than cover conversion and require de jure protection; in Africa and Asia (7.4, 6.1, and 5.5 percent, conversely, those that are remote face a much respectively), but protected versus nonprotected lower pressure of land cover change are assumed differentials in Latin America and the Caribbean to have a degree of de facto protection. exceed those in Africa and Asia (differences between protected and nonprotected are –5.3, As a first, crude estimate of the relationship –3.1, and –1.5 percent, respectively). between tropical forest fire activity and pressure, the fire activity (fire area/forest area) for the When this percentage reduction is related to the tropical forest biome is plotted against travel time protected forest area, the result is an for protected (for protected areas designated pre- uncorrected measure of the number of avoided 2000 with any type of protection) and unprotected forest fire pixels due to protection (remember, forest areas and the difference between the two, fire activity cannot be directly translated to an for the biome (Figure 6) and each continent estimate of area deforested), amounting to some (Figure 7). Again, these (unmatched) estimates are 85,500 1-km pixels or 4.4 percent of the naïve: they do not correct for other determinants protected areas in the tropical forest biome. This of deforestation pressure and make no correction 14 P R O T E C T E D A R E A E F F E C T I V E N E S S I N R E D U C I N G T R O P I C A L D E F O R E S TAT I O N Figure 5: Accessibility to Cities Figure 6: Crude Forest Fire Rate (fire area/forest area) against Travel Time for the Tropical Forest Biome (no controls for other variables) for the bias in location of the protected areas or The greatest difference is in the 0–12-hour range, their environmental similarity or lack of it with peaking at 6–7 hours. unprotected areas. Rates across Africa (Figure 7, middle) vary too, The lines in Figures 6 and 7 show the 95 percent but the difference between protected and confidence limits around a best fitting loess unprotected is only more than 2–3 percent in curve3 through the points. The best fit was very accessible regions. The average rates across determined via cross validation assessed by the Africa are 0.0607 and 0.0302 for unprotected and Akaike Information Criterion. The confidence protected areas, respectively (with a difference limits were derived from bootstrapping the loess of 0.0304, 2 percent lower than Latin America fit with 1,000 replications. and the Caribbean). The rates are above average for travel times of less than 6–9 hours, and the The average fire rates across the biome are difference between fire rates in protected and 0.0649 and 0.0255 for unprotected and protected unprotected areas becomes negligible at around areas, respectively (with a difference of 0.0393), 24 hours travel time. Fire activity peaks in very but this varies considerably in both protected accessible areas (0–3 hours travel time). and unprotected areas, with more accessible regions having the expected higher fire rate and The plot for Asia (Figure 7, bottom) shares the greater difference between protected and characteristics with both Latin America and the unprotected. The rates are above average for Caribbean and Africa. The average rates across travel times less than 12–15 hours, and the differ- Asia are 0.05548 and 0.0399 for unprotected and ence between fire rates in protected and protected areas, respectively (with a difference unprotected areas becomes negligible at around of 0.0149). The rates are above average for travel 48 hours travel time. times less than 12–15 hours. Fire activity peaks in accessible areas (0–9 hours travel time). The same trend, although much more pronounced, is visible in Latin America and the These plots suggest that deforestation pressure, Caribbean (Figure 7, top); the average rates and the protective effect of protected areas, across the region are 0.0744 and 0.0214 for might differ systematically with remoteness from unprotected and protected areas, respectively cities. Hence exact matching on travel time is (with a difference of 0.0530). The rates are above used as a covariable in the model. This allows average for travel times less than 18–21 hours. computation of treatment effects per travel time 15 E V A L U AT I O N B R I E F 7 Figure 7: Crude Forest Fire Rate (fire area/forest area) zone, as well as the average treatment effect. against Travel Time for Latin America and the Aggregation is to 15-minute zones, which allows Caribbean, Africa, and Asia (no controls for other for further aggregation, to compute fire rates per variables) 1-hour zone, for example. When the four protection classes (strict, multi- use, unknown, and indigenous) are combined across three continents, the result is 10 cohorts of control/treatment samples (ignoring the combinations for indigenous protection in Asia and Africa). Each cohort is used as input to the matching procedures described in chapter 2 (nearest neighbor matching based on Mahala- nobis distance with/without calipers)—thus there are two analyses per cohort. A description of the other covariables that will be considered for their role as controlling factors, and the sampling procedure used to create the cohorts for matching, follows. 3.6. Environmental characteristics In addition to the proxy of pressure for conver- sion (described in section 3.5), a suite of spatial data layers was collected to characterize the different environments within the biome. Distance to road network. Roads provide quick and easy access to areas. In this case, they make forest areas accessible to small- and large-scale deforestation agents (Chomitz and Gray 1996). A similar distance measure—distance to roads (Figure 8)—was created based on a vector road network extracted from the fifth edition of the Vector Smart Map Level 0 dataset (NIMA 2000). The primary source for the database is the 1:1 million scale Operational Navigation Chart series. The reference period is 1979–99 (Nelson, de Sherbinin, and Pozzi 2006).4 Here red indicates proximity to roads; green indicates extreme remoteness. Distance to major cities. The proximity of a Note: In all four cases the fire activity varies with accessibility, with peaks of activity in highly ac- patch of land to a potential market is a key cessible regions and much lower rates in more distant forest areas. Rates in protected areas are con- explanatory variable for land use change (Barbier sistently lower than in unprotected areas, with average differences ranging from 5.3 percent in Latin America and the Caribbean to 1.5 percent in Asia, but these differences can be as large as 9 per- and Burgess 2001). The major problem is the cent in high-pressure areas of Latin America and the Caribbean (c.f. Figure 7, top, at the seven-hour identification of such markets from a dataset of mark). populated places. 16 P R O T E C T E D A R E A E F F E C T I V E N E S S I N R E D U C I N G T R O P I C A L D E F O R E S TAT I O N Figure 8: Distance to Roads Figure 9: Distance to Major City A similar distance measure—distance to major unlikely to be converted to agriculture, and the cities (Figure 9)—was created based on a point associated cloud cover and humidity preclude dataset of city centroids (CIESIN 2004). Follow- the use of fire activity as a reliable measure of ing the definition of major cities in the travel time deforestation. layer, the distance to the nearest city with an estimated population of at least 50,000 in 2000 Rainfall estimates (Figure 12; lighter areas was measured. indicate low rainfall and darker areas high rainfall) were extracted from data provided by Terrain. Terrain is a factor for land use suitability. the Tropical Rainfall Monitoring Mission, specifi- Mild slopes and lower elevations are likely to be cally from the 3B42-TRMM-Adjusted Merged- more accessible, more productive, more val- Infrared Precipitation product (Huffman and uable, and thus more attractive for conversion to others 1997). This dataset provides monthly agriculture. As well as having a direct relation to estimates of rainfall rates at a ¼-degree resolu- suitability, slope and elevation are proxies for tion. These rates were converted to millimeters physical soil properties, and elevation is a proxy (mm) per month, then aggregated into annual for temperature. Figure 10: Elevation Elevation (Figure 10) and slope (Figure 11) were derived from the Consortium for Spatial Informa- tion (CGIAR-CSI) version (Reuter, Nelson, and Jarvis 2007) of the 90-m resolution SRTM digital elevation model from NASA (Farr and Kobrick 2000). The CGIAR-CSI version of the data has filled in the data void areas with auxiliary digital elevation model data and topographically correct interpolation algorithms. The mean and variance of both slope and elevation were extracted at 1- km resolution, so each 1-km estimate is based on 100 or so elevation or slope values. Rainfall. Rainfall is another factor for land use suitability. Areas of extremely high rainfall are 17 E V A L U AT I O N B R I E F 7 Figure 11: Slope Figure 12: Rainfall rainfall estimates and finally into an estimate of • Were designated as protected pre-2000 based the average annual rainfall in mm for 2000–08. on protected area boundary information from the WDPA Country. Detailed country boundaries were • Classified as forest cover in 2000, based on extracted from the Global Administrative Areas the 11 land cover classes in GLC2000 that are database (Hijmans and others 2008). This forest or forest mosaic information is used for exact matching to ensure • Met the 25 percent forest cover threshold from that each control/treatment pair belongs to the MODIS forest cover for 2000 same country. • Fell into the relevant protection group (strict, multi-use, unknown, indigenous) for the cohort. Summary statistics for all the above variables in the tropical forest and protected tropical forest areas The two forest criteria reflect the conservative are shown in Table 6. In general, protected tropical estimate of tropical forest area in 2000. forest areas are more remote, have lower fire incidence rates, and have higher elevation/slope The corresponding control group was based on than the tropical forests as a whole. another random sample that was five times as large. The control points had to meet the follow- 3.7. Sampling strategy and software ing criteria: All data layers were stored as a table in a • Had never been protected up to the end of 2008 PostgreSQL database (version 8.3), amounting to • Classified as forest cover in 2000, based on some 19 million records, one record per 1-km the 11 land cover classes in GLC2000 that are pixel. The matching analysis was split into three forest or forest mosaic geographic regions: Latin America and the • Met the 25 percent forest cover threshold from Caribbean, Africa, and Asia. A list of points that the MODIS forest cover for 2000. would be used to form the control and treatment groups was extracted from the database for each The never protected area takes into account any region. The list of points for the treatment group form of recognized protection from the WDPA was based on a 10 percent random sample of through the end of 2008 and including protected points.5 The treatment points had to meet the areas with information on their designation date. following criteria: Those protected areas with boundary information 18 P R O T E C T E D A R E A E F F E C T I V E N E S S I N R E D U C I N G T R O P I C A L D E F O R E S TAT I O N Table 6: Summary Statistics for Variables in Tropical Forest Areas Forest Area Protected forest area Region/Variable Mean St. Dev. Median Mean St. Dev. Median Biome Travel time (minutes) 1,353 1,401 817 1,678 1,528 1,181 Rainfall (mm) 2,135 712 2,051 2,102 621 2,026 Dist. to cities (km) 185 142 149 207 139 180 Dist. to roads (km) 47 73 14 72 94 28 Fire pixels (proportion) 0.053 0.224 0 0.026 0.158 0 Elevation (meters) 410 483 245 449 510 281 Slope (degree) 6.4 6.9 3 6.9 7.2 4 Latin America and the Caribbean Travel time (minutes) 1,772 1,564 1,323 1,913 1,596 1,481 Rainfall (mm) 2,197 571 2,186 2,099 499 2,060 Dist to cities (km) 226 150 200 235 141 208 Dist to roads (km) 76 87 44 94 101 54 Fire pixels (proportion) 0.052 0.223 0 0.022 0.145 0 Elevation (meters) 314 439 181 361 449 229 Slope (degree) 4.8 5.8 2 5.5 6.3 3 Africa Travel time (minutes) 646 563 486 889 652 736 Rainfall (mm) 1,569 408 1,533 1,632 482 1,587 Dist to cities (km) 145 92 131 166 97 160 Dist to roads (km) 9 11 5 13 12 9 Fire pixels (proportion) 0.057 0.231 0 0.030 0.170 0 Elevation (meters) 493 362 441 581 533 446 Slope (degree) 4.2 3.9 3 5.2 4.6 4 Asia Travel time (minutes) 1,039 1,180 558 1,201 1,354 685 Rainfall (mm) 2,410 885 2,365 2,436 905 2,438 Dist to cities (km) 132 129 85 117 107 87 Dist to roads (km) 18 28 7 19 29 9 Fire pixels (proportion) 0.052 0.223 0 0.040 0.195 0 Elevation (meters) 540 584 348 741 605 629 Slope (degree) 11.3 8.1 11 14.0 7.9 14 are simply masked out. Protected areas with no presence/absence from 2000–08 as a proxy for boundary information but with latitude/longitude deforestation events. The treatment variable is point location and area information are treated as protected/nonprotected. The covariates represent circles centered on their latitude/longitude factors that affect deforestation and the location of coordinate, and those areas are also masked out. protected areas. The covariates are: The analysis is on 1-km resolution data. The 1. Average elevation outcome variable is a binary measure of fire 2. Average slope 19 E V A L U AT I O N B R I E F 7 3. Average rainfall (2000–08) Notes 4. Distance to roads 1. http://modis-fire.umd.edu/MOD14.asp. 5. Distance to cities 2. http://modis-fire.umd.edu/MCD45A1.asp. 6. Country 3. Loess is a form of local polynomial regression 7. Travel time to nearest city in 15-minute incre- fitting that acts something like a moving average; the bandwidth is analogous to the width of the ments. window used for the moving average. 4. The start date is debatable; the third edition of The slope, rainfall, and distance covariates are VMAP0, published 1997, also has a 20-year similar to those used in the Andam and coauthors reference period—1974–94! The fifth edition was (2008). The last two covariates were used as exact published in 2000, but given the minor changes matches, to ensure that each control/treatment throughout editions (1st in 1992, 2nd in 1995, pair belonged to the same country and faced 3rd/4th in 1997, and 5th in 2000), it is unlikely to comparable pressure for land conversion as well have much post-1990 data. as having similar environmental characteristics. 5. Ten percent was chosen to comply with the memory and time limits that arise from matching Several matching software libraries, for use in on large datasets, based on personal communica- common statistical packages, are available (for tion with Lucas Joppa, Duke University. A control/treatment group of around 100,000 points example, Ho and others 2007; Sekhon 2007; and 7 covariables requires around 20 hours on a Abadie and others 2004). The matching package fast PC running Windows XP . Fortunately, most of (Sekhon 2007) (version 4.7-6) running in the the samples in the following analyses are smaller open source statistical program R (version 2.8.1) than this. Both samples were saved into a on MS Windows XP SP3 was used. temporary table in PostgreSQL, and this table was then read directly into R (via an ODBC connec- Matching was performed on all cohorts of tion) for analysis in the matching package. This protected areas defined by geographic region was repeated for each analysis, with results saved and protection type, using the Mahalanobis as text files. distance metric, both with and without a 0.5SD caliper. Matching was performed with replace- ment and bias adjustment. 20 CHAPTER 4 Results points against a mean loss of 5.8 percent (Table 5) 4.1 Average estimate of avoided forest over 2000–08. Multi-use protected appears to be fires area due to protection more effective than strict by approximately 2 percentage points, and this also translates into a Table 7 shows the results of the matching larger area. “Unknown” is less effective, but the analyses—the estimated avoided fire activity as a area is quite small. Indigenous areas are shown to proportion of all pre-2000 protected areas— reduce forest fire incidence by 16.3–16.5 percent- alongside the crude estimates from Table 5.1 age points, more than two and a half times as Table 8 repeats but uses the 1990–2000 protected much as the crude estimates (5.9 percent) and areas as the treatment group. twice as effective as any other group in the matched results, with a greater estimated avoided Looking at the results for pre-2000 against never fire pixel area than strict, multi-use, and unknown protected (Table 7) in the Latin America and the combined. Strictly protected areas in Africa are Caribbean region, the matched results for strict only one-quarter as effective (about a 1 percent- protection suggest a much lower level of avoided age point impact), as the uncorrected estimates fire activity than the crude (uncorrected) would suggest. The estimated impacts for multi- estimates. Nonetheless, protected areas reduced use areas are not robust: a significant 3 percent the incidence of forest fires by 2.7–4.3 percentage for the Mahalonobis, but 0 percent (with wide Table 7: Estimated Impact on Fire Incidence (cumulative over 2000–08) Comparing All Pre-2000 Protected Areas against Never Protected Mahalanobis Mahalanobis with calipers Region/Protection Crude Estimate [SE] Pairs Estimate [SE] Pairs Latin America and the Caribbean Strict Ϫ0.058 Ϫ0.027 [0.002] 46,015 Ϫ0.043 [0.001] 28,039 Multi-use Ϫ0.044 Ϫ0.048 [0.003] 52,505 Ϫ0.064 [0.002] 29,993 Unknown Ϫ0.053 Ϫ0.038 [0.010] 2,232 Ϫ0.023 [0.004] 511 Indigenous Ϫ0.059 Ϫ0.165 [0.003] 36,166 Ϫ0.163 [0.003] 28,482 Africa Strict Ϫ0.043 Ϫ0.010 [0.002] 13,507 Ϫ0.013 [0.001] 7,582 Multi-use Ϫ0.031 Ϫ0.030 [0.008] 1,592 § Ϫ0.001 [0.004] 715 Unknown 0.002 § Ϫ0.010 [0.007] 4,980 § 0.000 [0.004] 2,306 Asia Strict Ϫ0.010 Ϫ0.017 [0.003] 20,683 Ϫ0.020 [0.002] 12,101 Multi-use Ϫ0.018 Ϫ0.049 [0.006] 7,408 Ϫ0.043 [0.004] 4,319 Unknown Ϫ0.041 § Ϫ0.010 [0.005] 3,528 Ϫ0.044 [0.003] 1,072 § All estimates significant at p < 0.001 except those marked with §. 21 E V A L U AT I O N B R I E F 7 Table 8: Estimated Impact on Fire Incidence (cumulative over 2000–08, not annualized) Comparing 1990–2000 Protected Areas against Never Protected Mahalanobis Mahalanobis with calipers Region/Protection Crude Estimate [SE] Pairs Estimate [SE] Pairs Latin America and the Caribbean Strict Ϫ0.065 Ϫ0.038 [0.003] 14,409 Ϫ0.077 [0.002] 5,749 Multi-use Ϫ0.030 Ϫ0.062 [0.004] 21,972 Ϫ0.075 [0.003] 15,032 Unknown Ϫ0.063 Ϫ0.026 [0.006] 889 too few points 80 Indigenous Ϫ0.061 Ϫ0.128 [0.004] 21,813 Ϫ0.127 [0.003] 15,276 Africa Strict Ϫ0.047 Ϫ0.022 [0.004] 2,730 Ϫ0.045 [0.004] 1,056 Multi-use Ϫ0.060 too few points 153 too few points 12 Unknown Ϫ0.059 Ϫ0.066 [0.008] 203 too few points 18 Asia Strict Ϫ0.022 Ϫ0.029 [0.005] 7,355 Ϫ0.031 [0.002] 2,536 Multi-use 0.031 Ϫ0.067 [0.020] 1,832 Ϫ0.051 [0.008] 559 Unknown Ϫ0.049 Ϫ0.023 [0.006] 2,349 Ϫ0.070 [0.004] 569 Note: The full set of balance metrics and other outputs from these matching analyses are available on request. error bands) for the estimate with calipers. In broad scope of protected areas, each with Asia, strictly protected areas perform better than advantages and disadvantages. The conclusion that in the crude estimates, but multi-use is twice as protected areas are effective is seen to be robust. effective as strict. At first glance, it may seem paradoxical that in Table 8 estimates suggest that, with the some cases the mean reduction in fire incidence exception of indigenous areas, protected areas is greater than the mean incidence of fires—for designated between 1990 and 2000 offer better instance, in the case of Latin American indige- protection than pre-2000 protected areas as a nous areas. This implies that the protected areas whole, with improvements ranging from 1 to 3.5 are located in regions of higher-than-average percentage points, disregarding results with few deforestation pressure. For further insight, the matched pairs. In Latin America and the next section disaggregates impacts by remote- Caribbean, multi-use protected areas appear to ness—a strong correlate of pressure. be as effective or more effective than strict, but indigenous areas are almost twice as effective as 4.2. Disaggregated estimates any form of protection. In Asia, strictly protected areas perform better than in the crude estimates, To assess the importance of location when but multi-use is twice as effective. In Africa, these estimating the effectiveness of protection, the recently established protected areas appear fire rate in the matched treatment and control much more effective than the larger set consid- groups is disaggregated by travel time. This is ered in Table 7, with a robustly estimated impact done only for the pre-2000 treatment group, as of about 4.5 percentage points. There are too few the 1990–2000 group often has too few points to points to estimate an impact for multi-use areas. allow disaggregation. Table 9 summarizes the results. The range of The fire rate per travel time band was plotted and estimates represents a robustness test—use of two a loess curve was fitted through them using cross kinds of matching procedures and a more or less validation and Akaike’s information criterion to 22 P R O T E C T E D A R E A E F F E C T I V E N E S S I N R E D U C I N G T R O P I C A L D E F O R E S TAT I O N Table 9: Summary of Estimate Protected Area Impacts on Fire Incidence (%) Mean reduction Mean reduction Mean reduction Mean fire due to strict due to multi-use due to indigenous Area incidence protected areas protected areas areas Latin America and Caribbean 7.4 2.7–4.3 4.8–6.4 16.3–16.5 3.8–7.7 6.2–7.5 12.7–12.8 Africa 6.1 1.0–1.3 (0.1)–3.0 Not applicable 4.4–4.5 Not calculated Asia 5.5 1.7–2.0 4.3–5.9 Not applicable 2.9–3.1 6.7–5.1 Note: Italics indicate estimates for protected areas established between 1990 and 2000. determine the best fitting smoothing factor or Second, except for strict protection in Africa, bandwidth. Furthermore, the loess estimator protected areas generally have significantly lower (1,000 repetitions) was bootstrapped to fire rates than comparable nonprotected areas. determine 95 percent confidence intervals However, this differential declines as remoteness around the curve. This was done for the fire rates increases. Natural protection is often as effective from the matched control (never protected, as strict protection in remote areas—at least for red), and treatment data (protected pre-2000, the moment. Third, in both Latin America and green) and for the difference between the two the Caribbean and Asia, nonremote multi-use (gray). This difference is essentially a disaggre- areas are located in areas of higher deforestation gated version of the estimates in Table 7 and pressure than strict areas. For instance, at 1 hour provides an unbiased estimate of the avoided from cities in Latin America and the Caribbean, deforestation fires due to protection for different the control for multi-use areas experience fire degrees of remoteness. The following figures rates of about 16 percent whereas the controls (13, 14, and 15) show these confidence intervals for strict areas had fire rates of about 6 percent. around the loess curve as shaded polygons, as Fourth, in Latin America and the Caribbean, fire well as the points that they are fitted though. The rates are generally higher in multi-use than in results are reported for strict, multi-use, and strict protected areas, controlling for remote- indigenous areas for Latin America and the ness. Yet the absolute impact of multi-use areas is Caribbean, strict for Africa (there are insufficient greater than that of strict areas. At 1–12 hours pairs for multi-use to permit disaggregation), and from cities, for instance, multi-use protected strict and multi-use for Asia, although the areas reduce fire rates by about 6–12 percentage number of pairs for multi-use in Asia is just points, and strict protected areas reduce rates by acceptable. These estimates provide an unbiased only about 5 or 8 percentage points. Indigenous and more realistic view than the naïve estimates areas also have a very high absolute impact. in Figures 6 and 7. In Asia, the pattern is different. Controlling for Some strong regularities emerge. First, in almost distance, deforestation rates are higher in strict all cases, fire activity inside protected areas than in multi-use protected areas. Strict declines with increasing remoteness. Although protected areas appear to be ineffective at the same is generally true for areas outside deterring fires in nonremote areas. Their protected areas, in some cases (strict and multi- effectiveness increases with remoteness, peaking use in Latin America and the Caribbean and strict at about 12 hours distance from the city and in Asia) the outside rate and hence effectiveness declining thereafter. In contrast, multi-use of protection increases with remoteness protected areas are most effective in regions reaching a maximum at around 9–12 hours. proximate to population centers. 23 E V A L U AT I O N B R I E F 7 In Africa, strict protected areas appear to have a use areas are limited by a small sample and are modest impact. Estimates of the impact of multi- not robust. Figure 13: Unbiased Estimated Fire Rates (red, never protected; green, protected; and grey, difference) for Tropical Forests in Latin America (with matching) Note: Top – Strict protection in Latin America and the Caribbean, with Mahalanobis matching (left) and Mahalanobis matching with calipers (right). Bottom – Multi-use protection in Africa, with Mahalanobis matching (left) and Mahalanobis matching with calipers (right). 24 P R O T E C T E D A R E A E F F E C T I V E N E S S I N R E D U C I N G T R O P I C A L D E F O R E S TAT I O N Figure 14: Unbiased Estimated Fire Rates (red, never protected; green, protected; and grey, difference) for Tropical Forests in Latin America and Africa (with matching) Note: Top – Indigenous protection in Latin America and the Caribbean, with Mahalanobis matching (left) and Mahalanobis matching with calipers (right). Note the change in scale on y axes. Bottom – Strict protection in Africa, with Mahalanobis matching (left) and Mahalanobis matching with calipers (right). 25 E V A L U AT I O N B R I E F 7 Figure 15: Unbiased Estimated Fire Rates (red, never protected; green, protected; and grey, difference) for tropical forests in Asia (with matching) Top – Strict protection, with Mahalanobis matching (left) and Mahalanobis matching with calipers (right) Bottom – Multi-use protection, with Mahalanobis matching (left) and Mahalanobis matching with calipers (right) 1. In all cases the crude (comparing all protected pixels against all never protected pixels) and prematch rates (comparing an unmatched 10 percent sample of protected pixels against a Note similar proportion of never protected pixels) were 26 CHAPTER 5 Conclusions This paper uses forest fires as a proxy for than designation of strict protection in areas of deforestation and associated carbon release. higher population density and less remoteness. Using global data for the tropical forest biome, it is apparent that protected areas have a substan- This analysis does not however attempt to tially and statistically significantly lower incidence measure “leakage”—the degree to which protec- of forest fires than nonprotected areas, even after tion of one forest plot merely displaces conver- controlling for terrain, climate, and remoteness. sion to another, unprotected plot. This is a more The protective effect is greatest in nonremote significant issue for carbon emissions than for areas (for Latin America and Africa) and areas of biodiversity conservation, because the latter intermediate remoteness (Asia). Very remote might be preferentially concerned with certain areas have low deforestation rates even if unique biodiversity locations whereas the former unprotected—at least for the moment. cares only about the density of carbon. Chomitz (2002) reviews theoretical and empirical studies Importantly, it is clear that mixed-use protected of leakage and concludes that on both grounds areas—where some degree of productive use is leakage is far less than the 100 percent feared by allowed—are generally as effective or more critics. He points out that complementary effective than strict protected areas, especially in policies (such as sponsoring crop intensification) less remote areas with greater pressure for could neutralize any leakage thought to arise agricultural conversion and timber extraction. In from forest protection. Latin America, where indigenous areas can be identified, they are found to have extremely large In addition, this analysis is unable to detect some impacts on reducing deforestation—much larger kinds of forest degradation. Surreptitious than a naïve, uncontrolled comparison would removal of timber can result in biodiversity suggest. These results suggest that mixed-use damage and lower carbon densities, but may not and indigenous areas are disproportionately be detected through fire data. located in areas of higher deforestation pressure. This is noteworthy, given increasing attention to Extension of this line of evaluation will be facili- indigenous land rights. tated as better data become available. Improve- ments in remote sensing techniques and From a policy viewpoint, these findings suggest interpretation offer the prospect of more direct that some kinds of land use restrictions— and precise measurement of deforestation and variations of protection—can be effective of forest carbon emissions. There is also a need contributors to biodiversity conservation and to assemble, harmonize, and make public assess- climate change mitigation goals. The results ments of protected area management resources suggest that indigenous areas and multi-use and practices in order to better understand the protected areas can help accomplish these goals, specific interventions that can contribute to also suggesting some compatibility between reduced carbon emissions. Finally, there is a environmental goals (carbon storage and great need to complement land cover and land biodiversity conservation) and support for local management measures with monitoring of livelihoods. Zoning for sustainable use may be human welfare and conditions in protected and more politically feasible and socially acceptable unprotected forest areas. 27 E V A L U AT I O N B R I E F 7 It is important to stress that protected areas important, it is easier to reach consensus on the may be effective along other dimensions, even necessity and approach to protecting a forest where there is little impact on current before there are large economic pressures for deforestation rates. This is especially true for conversion, often by people from outside the protected areas established in remote regions forest itself. A well-established protection with little current pressure for agricultural regime may be better able to withstand conversion. Such areas may already be effective pressures for unsustainable exploitation when in mitigating other threats, such as poaching of the frontier arrives, as it eventually will in many mammals and selective logging. Equally currently remote places. 28 References Abadie, A., and G. Imbens. 2006. “Large Sample and Protected Areas in North Thailand.” Land Properties of Matching Estimators for Average Economics 77: 172–86. Treatment Effects.” Econometrica 74: 235–67. Deninger, K., and B. Minten. 2002. “Determinants of Abadie, A., D. Drukker, J.L. Herr, and G. W. Imbens. Deforestation and the Economics of Protection: 2004. “Implementing Matching Estimators for An Application to Mexico.” American Journal of Average Treatment Effects in Stata.” Stata Journal Agricultural Economics 84: 943–60. 4: 290–311. FAO (Food and Agriculture Organization of the Adeney, J.M., N.L. Christensen, and S.L Pimm. 2009. United Nations). 2006. “Global Forest Resources “Reserves Protect against Deforestation Fires in Assessment 2005.” FAO Forestry Paper 147. Amazon.” PLoS ONE 4, e5014. FAO, UNDP, and UNEP. 2008. UN Collaborative Andam, K., P . Ferraro, A. Pfaff, A. Sanchez-Azofeifa, and Programme on Reducing Emissions from J. Robalino. 2008. “Measuring the Effectiveness of Deforestation and Forest Degradation in Develop- Protected Area Networks in Reducing Deforesta- ing Countries (UN-REDD). tion.” Proceedings of the National Academy of Farr, T.G., and M. Kobrick, M. 2000. “Shuttle Radar Sciences 105: 16089–94. Topography Mission Produces a Wealth of Data.” Barbier, E.B., and J.C. Burgess. 2001. “The Economics American Geophysical Union. Earth—Oceans— of Tropical Deforestation.” Journal of Economic Atmosphere 81: 583–5. Surveys 15: 413–32. Ferraro, P.J. 2008. “Protected Areas and Human Well- Bartholome, E., and A.S. Belward. 2005. “GLC2000: A Being.” Paper presented at conference, Economics New Approach to Global Land Cover Mapping and Conservation in the Tropics: A Strategic from Earth Observation Data.” International Dialogue, Conservation Strategy Fund, January Journal of Remote Sensing 26: 1959–77. 31–February 1. Bruner, A.G., R.E. Gullison, R.E. Rice, and G.A.B. Ferraro, P.J., and S.K. Pattanayak. 2006. “Money for Fonseca. 2001. “Effectiveness of Parks in Protect- Nothing? A Call for Empirical Evaluation of ing Tropical Biodiversity.” Science 291: 125–28. Biodiversity Cnservation Ivestments.” PLoS Biol 4: Canadell, J.G., and M.R. Raupach. 2008. “Managing e105. Forests for Climate Change Mitigation.” Science Gullison, R.E., P.C. Frumhoff, J.G. Canadell, C.B. Field, 320: 1456–57. D.C. Nepstad, K. Hayhoe, R. Avissar, L.M. Curran, P . Chomitz, K. 2007. “At Loggerheads? Agricultural Friedlingstein, C.D. Jones, and C. Nobre. 2007. Expansion, Poverty Reduction, and Environment “Environment: Tropical Forests and Climate in the Tropical Forests.” World Bank Policy Policy.” Science 316: 985–6. Research Report 308. Hansen, M., R. DeFries, J.R. Townshend, M. Carroll, C. Chomitz, K. 2002. “Baseline, Leakage, and Measure- Dimiceli, and R.A. Sohlberg. 2003. “Global Percent ment Issues: How Do Forestry and Energy Projects Tree Cover at a Spatial Resolution of 500 Meters: Compare?” Climate Policy 2(1): 35–49. First Results of the MODIS Vegetation Continuous Chomitz, K., and D. Gray. 1996. “Roads, Land Use and Fields Algorithm.” Earth Interactions 7: 1–15. Deforestation: A Spatial Model Applied to Belize.” Hansen, M.C., S.V. Stehman, P.V. Potapov, T.R. World Bank Economic Review 10: 487–512, Loveland, J.R.G. Townshend, R.S. DeFries, K.W. CIESIN. 2004. “Global Rural-Urban Mapping Project Pittman, B. Arunarwati, F. Stolle, M.K. Steininger, (GRUMP), Alpha Version: Population Density M. Carroll, and C. DiMiceli. 2008. “Humid Tropical Grids.” Socioeconomic Data and Applications Forest Clearing from 2000 to 2005 Quantified by Center (SEDAC), Columbia University, New York. Using Multitemporal and Multiresolution Re- Cropper, M., J. Puri, and C. Griffiths. 2001. “Predicting motely Sensed Data.” Proceedings of the National the Location of Deforestation: The Role of Roads Academy of Sciences 105: 9439–44. 29 E V A L U AT I O N B R I E F 7 Hijmans, R., N. Garcia, J. Kapoor, A. Rala, A. Nelson, A., A. de Sherbinin, and F. Pozzi. 2006. Maunahan, and J. Wieczorek. 2008. Global “Towards Devlopment of a High Quality Public Administrative Areas (version 0.9). Domain Global Roads Database.” Data Science Ho, D., K. Imai, G. King, and E. Stuart. 2007. Journal 5: 223–65. “Matching as Nonparametric Preprocessing for Nepstad, D., S. Schwartzman, B. Bamberger, M. Reducing Model Dependence in Parametric Causal Santilli, D. Ray, P. Schlesinger, P. Lefebvre, A. Inference.” Political Analysis 15: 199–236. Alencar, E. Prinz, G. Fiske, and A. Rolla. 2006. Huffman, G.J., R.F. Adler, P. Arkin, A. Chang, R. “Inhibition of Amazon Deforestation and Fire by Ferraro, A. Gruber, J. Janowiak, A. McNab, B. Parks and Indigenous Lands.” Conservation Rudolph, and U. Schneider. 1997. “The Global Biology 20: 65–73. Precipitation Climatology Project (GPCP) NIMA. 2000. Vector Map Level 0 (Digital Chart of the Combined Precipitation Dataset.” Bulletin of the World), Edition 5, National Imagery and Mapping American Meteorological Society 78: 5–20. Agency. Joppa, L., and A. Pfaff, A. 2009a. “Global Park Impacts: Olson, D.M., E. Dinerstein, E. Wikramanayake, N. How Much Deforestation Has Protection Avoided?” Burgess, G. Powell, E.C. Underwood, J. D’Amico, Unpublished manuscript, Duke University. I. Itoua, H. Strand, J. Morrison, C. Loucks, T. Joppa, L.N., and A. Pfaff. 2009b. “How the World’s Allnutt, T.H. Ricketts, Y. Kura, W. Wettengel, and K. Protected Areas Have Avoided Threat. Mimeo, Kassem. 2001. “Terrestrial Ecoregions of the Duke University. World: A New Map of Life on Earth.” BioScience Joppa, L.N., and A. Pfaff. 2009c. ”Re-Assessing the 51: 933–8. Forest Impacts of Protection: The Challenge of Pfaff, A. 2009a. “Evaluating Deforestation/Carbon Non-Random Location and a Corrective Method.” Impacts of Protected Areas: Challenge, Approach, Ecological Economics Reviews. Forthcoming. Costa Rican Case and Two Applications to the Joppa, L.N., S.R. Loarie, and S.L. Pimm. 2008. “On the Brazilian Amazon.” WWF / Moore / Linden Protection of ‘Protected Areas.’” Proceedings of workshop. the National Academy of Sciences 105: 6673–8. Pfaff, A. 2009b. “Evaluating Deforestation Impacts of Justice, C.O., L. Giglio, S. Korontzi, J. Owens, J.T. Protected Areas.” Presented at conference, Morisette, D.P. Roy, J. Descloitres, S. Alleaume, F. Connecting Amazon Protected Areas and Indige- Petitcolin, and Y. Kaufman. 2002. “The MODIS Fire nous Lands to REDD Frameworks, World Wildlife Products.” Remote Sensing of Environment 83: Fund, February 11–12. 244–62. Pfaff, A., J. Robalino, A. Sanchez-Azofeifa, K. Andam, Kindermann, G., M. Obersteiner, B. Sohngen, J. and P . Ferraro. 2009. “Location Affects Protection: Sathaye, K. Andrasko, E. Rametsteiner, B. Observable Characteristics Drive Park Impacts in Schlamadinger, S. Wunder, and R. Beach. 2008. Costa Rica.” The B.E. Journal of Economic “Global Cost Estimates of Reducing Carbon Analysis & Policy. Forthcoming. Emissions through Avoided Deforestation.” Reuter H.I, A. Nelson, and A. Jarvis. 2007. “An Evalua- Proceedings of the National Academy of Sciences tion of Void Filling Interpolation Methods for 105: 10302–7. SRTM Data.” International Journal of Geographic Miles, L., and V. Kapos. 2008. “Reducing Greenhouse Information Science 21: 983–1008. Gas Emissions from Deforestation and Forest Roy, D.P., L. Boschetti, C.O. Justice, and J. Ju. 2008. Degradation: Global Land-Use Implications.” “The Collection 5 MODIS Burned Area Product— Science 320: 1454–5. Global Evaluation by Comparison with the MODIS Morgan, S.L., and D.J. Harding. 2006. “Matching Active Fire Product.” Remote Sensing of Environ- Estimators of Causal Effects: Prospects and Pitfalls ment 112: 3690–3707. in Theory and Practice.” Sociological Methods Schmitt, C.B., N.D. Burgess, L. Coad, A. Belokurov, C. Research 35: 3–60. Besançon, L. Boisrobert, A. Campbell, L. Fish, D. Morton, D.C., R.S. Defries, J.T. Randerson, L. Giglio, Gliddon, K. Humphries, V. Kapos, C. Loucks, I. W. Schroeder, and G.R. Van Der Werf. 2008. Lysenko, L. Miles, C. Mills, S. Minnemeyer, T. “Agricultural Intensification Increases Deforesta- Pistorius, C. Ravilious, M. Steininger, and G. tion Fire Activity in Amazonia.” Global Change Winkel. 2009. “Global Analysis of the Protection Biology 14: 2262–75. Status of the World’s Forests.” Biological Nelson, A. 2008. “Travel Time to Major Cities: A Global Conservation. Map of Accessibility.” European Commission. Sekhon, J.S. 2007. “Multivariate and Propensity Score 30 P R O T E C T E D A R E A E F F E C T I V E N E S S I N R E D U C I N G T R O P I C A L D E F O R E S TAT I O N Matching Software with Automated Balance York: World Conservation Union and UNEP-World Optimization: The Matching Package for R.” Conservation Monitoring Center. Journal of Statistical Software. UNEP-WCMC. 2008. State of the World’s Protected UNEP/IUCN. 2009. Protected Areas Extracted from Areas 2007—An Annual Review of Global Conser- the 2009 World Database on Protected Areas. New vation Progress. New York: UNEP . 31