Conservation performance of different conservation governance regimes in the Peruvian Amazon

State-controlled protected areas (PAs) have dominated conservation strategies globally, yet their performance relative to other governance regimes is rarely assessed comprehensively. Furthermore, performance indicators of forest PAs are typically restricted to deforestation, although the extent of forest degradation is greater. We address these shortfalls through an empirical impact evaluation of state PAs, Indigenous Territories (ITs), and civil society and private Conservation Concessions (CCs) on deforestation and degradation throughout the Peruvian Amazon. We integrated remote-sensing data with environmental and socio-economic datasets, and used propensity-score matching to assess: (i) how deforestation and degradation varied across governance regimes between 2006–2011; (ii) their proximate drivers; and (iii) whether state PAs, CCs and ITs avoided deforestation and degradation compared with logging and mining concessions, and the unprotected landscape. CCs, state PAs, and ITs all avoided deforestation and degradation compared to analogous areas in the unprotected landscape. CCs and ITs were on average more effective in this respect than state PAs, showing that local governance can be equally or more effective than centralized state regimes. However, there were no consistent differences between conservation governance regimes when matched to logging and mining concessions. Future impact assessments would therefore benefit from further disentangling governance regimes across unprotected land.

cloud-free image of the area from 2011; and (iv) it is the area where the previous validation survey had taken place 6 . For the field validation, we randomly selected 'forest', 'degraded' and 'deforested' sites from the analysis among those that met the following criteria: (i) affected an area of at least 2.25 ha, calculated based on the root mean square error (RMSE) of the georeferencing accuracy of the Landsat images used in the analysis 7 to ensure that field sites selected for validation would coincide with those on the map generated through the analysis; (ii) located within 1 km of a paved or unpaved road, 500 m within a navigable river, and/or 6 km of the indigenous community Naranjal, south of Turnavista, in order to ensure that the site could be reached within one day walking or driving distance from Pucallpa or Naranjal; and (iii) sites within the same class were at least 1.5 km apart from each other. To maximize the sample size, we visited the field validation sites in clusters, first selecting a 'degraded' site and then a 'forest' and a 'deforested' site within up to 2 km of the 'degraded' site, if possible. A total of 69 field sites were reachable that were classed into 'forest',  Supplementary Table S5).
For the validation with high-resolution satellite images, we used 90 RapidEye images (5 m resolution, 25 by 25 km) from 2011 from across the study area ( Supplementary Fig. S2), made available by MINAM. The 90 images were selected based on being nearly cloud free and being taken in the same year as the corresponding Landsat image used in the analysis. We selected a stratified random sample of 588 pixels with at least 100 pixels detected as deforested, following the equation of Tortora 8 and recommendation of Olofsson et al. 9 .
The accuracy assessment yielded a 98.1% overall accuracy based on the number of sample plots in the high-resolution satellite images (n=588) and an 85.5% overall accuracy based on the sample plots evaluated during the field survey (n=69), further details of which are given in Supplementary Table S6. Unsurprisingly, the accuracy was therefore lower in areas relatively easily accessible to humans (Supplementary Table S6c) than that of a stratified random sample taken across the study area (Supplementary Table S6a). The user's accuracy for both deforestation and forest degradation was at least 90.8%, while the producer's accuracy was 85.2% or above, based on the number of sample plots. Given that the highest proportion of the study area remains covered in forest, the overall accuracy based on the extent of each class was higher (99.8%; Supplementary Table S6b) than that based on the number of sample plots (Supplementary Table S6a).
It was not possible to separately validate the 2006 forest cover map due to the lack of availability of high resolution satellite image from the required time period. Given that the same methods were applied as for the validated maps and previous assessments, we are confident that the 2006 forest cover map has a similar level of accuracy.

Matching analysis
Matching allows for a counterfactual approach to assess treatment effects, in this case national state PAs, CCs or ITs 10,11 . Through matching, deforestation and degradation rates inside treatment areas can therefore be compared to the rates inside artificial control groups matched according to socio-economic and biophysical factors likely to affect both the location bias of the treatment areas, and deforestation or forest degradation rates [11][12][13] . We matched the treatment areas to three types of controls, namely logging concessions, mining concessions and the wider unprotected landscape beyond the main official land use designations and mainly under the jurisdiction of the state.
We matched with a calliper of 0.25 standard deviations of the propensity score 14 . If no matching control pixel could be found within this caliper, the treatment pixel was excluded and treatment areas with less than 50 successfully matched pixels were excluded from the analysis. As a result, sample sizes varied between analyses for state PAs and ITs (see Supplementary Table S3), but not CCs (n=13). The order of finding matches (i.e. random, smallest to largest, or largest to smallest propensity score) 15 varied between the individual matching runs, depending on which yielded the best balance.
The size of the buffer areas excluded from the analysis was set to 5 km around state PAs, and 1 km around CCs and ITs as these areas are on average much smaller than state PAs. These buffer sizes were judged meaningful in the national context. We also excluded the official buffer areas, designated by the National Service of Natural Areas Protected by the State (SERNANP) around most of the national state PAs. From the unprotected landscape, we excluded state PAs and CCs designated between 2007 and 2012. We further excluded other types of conservation governance regimes that are region-specific, such as Brazil nut concessions, or those that have small sample sizes such as Indigenous Reserves, regional PAs, and Private Conservation Areas. For mining and logging concessions, we only included those areas that were considered as active during the study period (Supplementary Table S1). We further excluded any areas of overlap between mining and logging concessions, and between mining or logging concessions and the treatment areas. In cases, where there were overlaps between treatment categories, we assigned them to the land use category with the stricter resource use restrictions. It was not possible to account for the presence of agricultural land titles, as these have not been mapped across the country. We could also not include hydrocarbon concessions as (i) they occupy a large proportion of the national territory, leaving few potential areas for matched controls and (ii) the location of the considerably smaller areas where exploration and exploitation activities take place are not disclosed.
Prior to the matching analysis, we performed a power analysis to determine whether the sample size of CCs (n=13) and national state PAs (n=30) would be large enough to detect any potential effects. The power analysis was based on data published in Vuohelainen et al. 16 using GPower 3.1 and confirmed that the sample sizes were sufficiently large (n≥11) to detect an effect size of at least 0.98 at a 0.05 significance level and a power of 0.8.

Supplementary Figures and Tables.
Supplementary  Distances (i) and (ii) were correlated (>0.65); (i) was included as it explained a larger proportion of the residual deviance.

MTC and MINAM
Distance to rivers (km) Euclidean distance to main navigable rivers. Layers obtained from MINAM and updated manually by digitizing them in ArcMap 10 to match the rivers in the Landsat images used in the analysis. Given the larger number of zeros in the resulting data layers (62 to 79%), data were transformed into a binomial variable (presence/absence of settlements) to be modelled adequately.
Distances (i), (ii), and (iii) were correlated (>0.65) with distance to settlements, the latter was retained in the models ¶.

MINAM Elevation (m)
Elevation was based on the Shuttle Radar Topographic Mission (SRTM) 90m digital elevation data, processed by Jarvis and colleagues 19 for missing data. Layer was resampled to 30m.
Elevation was correlated (>0.65) with slope. Elevation was retained in the degradation model and slope in the deforestation model ¶. 19 Slope (°) Determined the slope using ArcMap's Slope tool, based on the SRTM 90m digital elevation data processed by Jarvis and colleagues 19 for missing data. Layer was resampled to 30m.
See 'elevation'. 19 Number of wet months Calculated number of wet months per year, with >100mm monthly rainfall following 20 , using ArcMap's Raster Calculator tool, based on the WorldClim Global Climate data (~1950-2000) provided at 30 arc-seconds resolution and resampled to 30m resolution.
This variables was correlated (>0.65) with rainfall. Number of wet months was retained in the degradation model and rainfall in the deforestation model ¶. 21 Rainfall (mm) Mean annual precipitation data were obtained from the WorldClim Global Climate data (~1950-2000) provided at 30 arc-seconds resolution and resampled to 30m resolution.
See 'number of wet months'. 21 Ecoregion Ecoregions included were Amazonico, Puna and Yungas; excluded Seco Ecuatorial as only a small number of data points fell within it. Included Sabana de palmeras within Amazonico as it could not be modelled independently due to its small size and being restricted to MDD. -MINAM

Administrative region
The Peruvian Amazon comprises 14 administrative regions, some of which cover only a small part of the study area and were therefore grouped together into a total of 7 regions. Significance levels: * significant at p<0.05, ** significant at p<0.01, *** significant at p<0.001. Note: The pair of predictor variables distance to small and large town, slope and elevation, and rainfall and number of wet months were highly intercorrelated (>0.65). Therefore for each pair only the predictor explaining more of the deviance of the null model was retained in the model. Distance to rivers was dropped from the degradation model as it was not significant. Supplementary