Table S2. Previous studies of PA effect on deforestation
author / region / data source / number of PAs and level of aggregation / observed/response variables / sampling method / variables used to differentiate/ match samples or estimate deforestation / methods / statistical assumptions / conclusionsCurran et al .(5) / West Kalimatan, Indonesian Borneo. / Forest cover maps from LandSat TM/ETM+ (1988, 1994, 1997, 1999, 2001, 2002) and MODIS.
Protected areas database from the Consensus Forest Land Use Plan. / Number of PAs not mentioned. / Deforestation. / Wall- to- wall deforestation in and out PAs. / Infrastructure, logging concessions, plantations and migration, but do not use this information to adjust PA effect. / Analyze percentages of deforestation in and out PA (buffer 10 km) over the years and throughout Kalimatan in various land uses (e.g. logging concessions). / Not applicable. / PAs show no effect at all.
Pfaff and Sanchez-Azofeifa (6) / Costa Rica. / Forest cover obtained from remote sensing for 1963, 1979, 1986, 1997, and 2000. Biological corridors proposed by Costa Rica. Holdridge Life zones and 436 districts. / Number of PAs not mentioned. Comparison is done for each corridor and for all corridors together. / Deforestation rate in each land unit per year. / Each observation for the logistic regression consists of a forest patch within a life zone and district. / Crop return, distance to major markets, “life zones” classified per precipitation and temperature, The same equation applies for the observed years using years as dummy variables. / Logistic regression to determine the probability of deforestation based on spatial variables. The probabilities are estimated for zones that fall inside the proposed biological corridors and outside them. / Parametric, variables must be independent and not correlated. Spatial dependence was tested and confirmed, but attempts to correct it showed no effect. / Deforestation rate modeled only differs by 3%. (0.90% for PAs and 0.93% for outside).
author / region / data source / number of PAs and level of aggregation / observed/response variables / sampling method / variables used to differentiate/ match samples or estimate deforestation / methods / statistical assumptions / conclusions
Defries et al.(7) / Tropical and subtropical forests. / Landsat imagery (30m), Modis VCF, 500m, and NOAA-AVHRR (8km), spatial resolution of RS data degraded to 8km, time-period from 1980-2000
WDPA (World database on PAs). / 198 PAs, binary: protected versus non-protected. / Mean forest area in and out PAs (50 km outside buffer). / PAs comprising IUCN types 1 and 2, covering dry and moist forests with extent larger than 25600 ha, consisting of 4 image cells of 8 km resolution. / Not included. / Analyze percents and use t-test to verify difference in forest covers in PAs and their buffer zones. / Normal distributions and homoscedasticity of samples. The study does not test assumptions. / 68% of PA showed forest loss in their outside, whereas only 25% showed deforestation inside.
author / region / data source / number of PAs and level of aggregation / observed/response variables / sampling method / variables used to differentiate/ match samples or estimate deforestation / methods / statistical assumptions / conclusions
Nepstad et al.(8) / Brazilian Amazon. / PRODES deforestation data 1997-2000 (30 meters), fire hotspots from GOES-8, 4 km spatial resolution. Instituto Socio-Ambiental PA database. / 165 PAs in
four categories (Parks, indigenous land, sustainable use reserves and national forests). / Deforestation rate and fire hot spot density within 20 km buffers in and out PAs. / Only PAs larger than 10 thousand ha for deforestation, and 50 k for fire hot spot analysis. / Not included. / Wilcoxon signed ranks tests for dependent samples for comparing in and out deforestation and fire within each PA category.
Kruskal-Wallis test to compare mean effectiveness between PA categories and Mann-Whitney U test for comparing parks and inhabitable reserves.
/ Non parametric tests (analyzed variables must come from the same continuous distribution and be independent).
The study does not test all assumptions. / PAs inhibit deforestation and fire. Deforestation is up to 20 times larger outside these areas.
author / region / data source / number of PAs and level of aggregation / observed/response variables / sampling method / variables used to differentiate/ match samples or estimate deforestation / methods / statistical assumptions / conclusions
Oliveira et al.(9) / 79% of Peruvian Amazon / Annual forest degradation maps from 30 m remote sensing imagery (1999-2005). PAs and forest concessions from INRENA and Instituto del Bien Comum. / Two categories
of PAs ( indigenous territories and natural protected) and forest concession. Number not mentioned. Analyses are performed for three individual categories. Number of PAs is not mentioned / Percents of forest disturbance and deforestation. / Wall- to- wall forest degradation data / roads / Compare percents in and out reserves (including PAs, indigenous territories and forest concessions) and across road distance buffers / Not applicable / Only 2% of forest disturbances occurred in PAs, while 11% occurred in indigenous territories. Forest concessions may be an effective deterrent against clear-cutting, once deforestation rates inside these areas were up to two orders smaller than disturbance resulted from logging. However, neighboring nonconcession lands had their deforestation rates increased by (304%-468%).
author / region / data source / number of PAs and level of aggregation / observed/response variables / sampling method / variables used to differentiate/ match samples or estimate deforestation / methods / statistical assumptions / conclusions
Joppa et al.(10) / Amazon, Congo, Atlantic coast and West of Africa / Forest cover from Global Land Cover 2000; original forest cover from WWF ecoregions; PAs from WDPA (World database on PAs) and IBAMA / Four categories I/II,
III/IV, V/VI/Misc (IUCN), and Indigenous reserves. Number of PAs in Amazon, Atlantic coast, West Africa and Congo are respectively: 341, 48, 531 and 142. / Forest cover (in the article, remaining natural vegetation) / 2000 Forest cover maps at 1km2 spatial resolution / Not analyzed / Analyze percents of natural vegetation as a function of distance inward (30 km) and outward (30 km) PA (2km increments) / Not applicable / Most PAs in the Atlantic coast and West Africa protect de jure, whereas in the Amazon and Congo, they only protect de facto, i.e. the effects of their designation may be negligible.
author / region / data source / number of PAs and level of aggregation / observed/response variables / sampling method / variables used to differentiate/ match samples or estimate deforestation / methods / statistical assumptions / conclusions
Andam et al. (11) / Costa Rica. / Aerial photography (1960), LandSat (1986, 1997). PA network from SINAC (National System of Conservation Areas). / 42 PAs analyzed all together for two cohorts. / Deforestation within 1960-1997 for PA designated before 1979, and deforestation within 1986-1997 for PA post 1981.
Binary response: < 80% canopy cover = deforested else forest / Random sampling, 15283 observations, each one of 3 ha). / Distance to roads, distance to forest edge, land use capacity, distance to nearest major city, distance to railroads, and river transportation network, district-level, population density, immigrants, % of houses using fuel-wood, district size. / Covariate matching using Mahalanobis distance with and without Caliper;
Test of McNemar for comparing deforestation in and out PAs. Post-matching weighted regression to estimate deforestation in each sampled plot. / Non parametric
covariate matching: "Ignorable treatment assignmet".
Parametric regression:
covariables and observations must be independent, but these assumptions were not tested / 11-12.4% avoided deforestation by PA in the older cohort and 2.7-5.3% in post 1981 cohort after applying covariate matching.
Regression method yielded similar effect of covariate matching (around 1% higher at the pre-1979 cohort and less than 0.5 for the post-1981).
Avoided deforestation is much lower than reported in other studies (27-51%), pointing out the need for covariate matching.
author / region / data source / number of PAs and level of aggregation / observed/response variables / sampling method / variables used to differentiate/ match samples or estimate deforestation / methods / statistical assumptions / conclusions
Gaveau et al. (12) / Sumatra. / Landsat imagery from 1990 and 2000, PA network from Minister of Forestry. / All PAs together compared to their adjacent areas (10 km buffer) and wider landscape (farther away than 10 km distance).
Number of PAs not mentioned / Percents of deforested area in a cell. / Random sampling cells. The method excludes cells if their centers are within 10 km distance from previously sampled cell, 1264 observations, each one of 25 km2). / Elevation, slope, distance to forest edge, distance to roads, and logging roads. / Method uses the output from a logistic regression obtained with covariates to determine scores for the matching method (propensity score matching). Regression validated with ROC and R2 measures.
Matching pairs are formed only within the same political province. Kolmogorov-Smirnov test used to check matching performance. Wilcoxon and t tests to compare whether PAs reduced deforestation. / Logistic regression:
covariables and observation must be independent. Spatial independence and multi-collinearity tested. / In average PAs reduced deforestation by 27.2 % from their adjacent 10 km buffers and 58.6 % from wider landscape before applying matching method. Applying propensity score matching these effects reduced, respectively, to 7.4% and 24%.
author / region / data source / number of PAs and level of aggregation / observed/response variables / sampling method / variables used to differentiate/ match samples or estimate deforestation / methods / statistical assumptions / conclusions
Pfaff et al. (13) / Costa Rica. / LandSat at 28m spatial resolution (1986, 1997). PA network from Instituto Tecnológico de Costa Rica). / Two categories: parks and biological reserves / deforestation within 1986-1997. / 4229 random sampled cells throughout Costa Rica / Distance to roads, distance to logging milling centers, distance to nearest schools, distance to forest edge, slope, distance to rivers, and ecosystem's characteristics / Propensity score matching using logistic regression. / Parametric: Logistic Regression:
covariables and observation must be independent, but these assumptions were not tested.
R2 (Mcfadden R2) of the regression used for the propensity score was 0.38. / PAs reduces deforestation only 1 to 2% after applying the propensity score matching.