Even the most intact forest fragments remaining across this vast study region had lost some of their midsized and large-bodied mammal fauna, and none of the sites surveyed had a full complement of species. Of the maximum number of 3,528 populations (18 species at 196 sites) that once occupied our entire forest region, only 767 (21.7%) were still extant, regardless of their level of demographic and genetic viability. These results are in contrast with far more optimistic predictions from classic species-area relationships (SARs) [15], which would forecast an overall persistence rate of 46.8% to 82.7% of all mammal populations for the fragments surveyed on the basis of reasonable bounds of z-values ranging from 0.4 to 0.1 [16], and the total proportional forest area remaining across the four BSRs that we surveyed (15.02%, Ref. 9). However, applying SARs derived from these 196 surveyed fragments to all 738,887 fragments remaining across the study region (Fig. 2) would yield an even lower overall population persistence rate of 15.1% (predicted total of 2,008,295 from a maximum of 13,299,966 populations).

Both small and large forest remnants varied widely in the number of species they retained. On average, only 3.3 (±2.1 SD) of the 18 species sampled persisted in patches smaller than 50 ha, which vastly dominated the remaining forest cover in the study region (Fig. 2). This mean patch-scale persistence rate increased only moderately to 5.0 (±3.3) and 4.0 (±2.1) species in larger forest patches of 500–1,000 ha and 1,000–5,000 ha, respectively (Table S2). Even fragments larger than 5,000 ha on average harbored only 7.2 species (±4.1), further confirming that high rates of local extinctions were widespread even in the largest forest patches located within the former ranges of the 18 mammal species. Contrary to most empirical SARs for tropical vertebrates, forest patch area alone explained only a small fraction of the variation in mammal species richness (R²_adj = 9.4%; Fig. 3), and none of the eight SAR models in the literature (Text S1) provided an adequate fit to the transformed or untransformed data. Likewise, patch area alone failed to predict most of the variation in a measure of aggregate mammal biomass (R²_adj = 4.4%; Fig. 3) and aggregate mammal vulnerability of the species persisting in each forest patch (R²_adj = 10.2%; Fig. 3). On a positive note, however, the five protected areas established throughout the study region substantially elevated overall species persistence: 5 of the 9 sites retaining 10 or more species were strictly protected forest reserves. Protected areas on average retained over three times as many species, and those species were on average 1.5-fold larger and shared a species vulnerability index 3.4-fold higher than those in unprotected areas (Walsh's approximation unequal variance t-tests, p<0.02 in all cases; Fig. 3).

On average, forest patches contained only 3.9±2.5 species [range = 0–15], but most of these were small to midsized, disturbance-tolerant species, such as small primates (marmosets and titi monkeys) and nine-banded armadillos, that were still ubiquitous throughout the survey region (Table 1; Fig. 4). Of the 18 species sampled, only the smallest species (marmosets) persisted in over half of all 196 forest patches. Three species occupied less than 1% of the combined area of all surveyed patches (Table 1; Fig. 4). Eleven of the 18 species sampled are larger than 5 kg and these fared far worse, with only 230 (10.7%) from a maximum of 2,156 populations (11 species at 196 sites) still persisting. White-lipped peccaries had been completely extirpated throughout the entire region, and four of the other seven large-bodied species (>9.5 kg) (lowland tapir, woolly-spider monkey, jaguar and giant anteater) nearly so, occurring in only 1 to 6 of the 196 patches (Table 1; Fig. 4). On the basis of the minimum observed forest patch size where each species were known to occur, we conservatively scaled up these local extinctions to all 738,887 forest fragments remaining across the entire ∼25.3×10⁶ ha study region. This translates into dramatic contractions in the remaining effective range size of all species, with only the smallest-bodied species occurring in ∼14% of their former geographic range (Table 1). Eleven of 12 species larger than 5 kg were restricted to only <3% of the entire region. Moreover, these extinctions were often reported to pre-date most of the historical deforestation and land-use change across the region; for instance, interviews clearly indicated that there was no record in recent memory of these otherwise widespread species in all but a few forest fragments surveyed. Only white-lipped peccary and woolly-spider monkey were reported to have been locally extirpated in the last decade prior to interviews (but in only 2 and 3 patches, respectively) and the largest five species were reported by 98% of all interviewees to have disappeared from local folklore, and were no longer pursued by subsistence hunters as viable game species.

Forest protection status and patch area were the only predictors consistently appearing in all best candidate models explaining contemporary species richness, but the importance of protection was consistently greater than patch size in terms of both species persistence and key traits that predispose species to local extinctions such as body mass, hunter preference and per capita fecundity (Tables S3 and S4). Most other patch and landscape variables were relatively unimportant, with the exception of topographic slope, whereby steeply dissected terrains — that are less physically accessible to both present and past hunters — boosted persistence of extant populations, particularly large-bodied, vulnerable taxa. Higher-elevation sites were also more likely to retain large-bodied species. Surprisingly, contemporary household density had only a modest effect on species persistence, but this can be attributed to considerable changes in the size and spatial distribution of the human population in this region since most local extinctions occurred. Likewise, measures of landscape connectivity including inter-patch proximity and the proportion of forest cover remaining within matrix areas had a marginal effect on levels of species persistence, likely because source populations in the neighborhood of fragments were no longer available regardless of the extent to which the intervening matrix was permeable.

All field workers in this study are Brazilian citizens who are both affiliated with Brazilian research institutions and completely aware of the ethics regulations in Brazilian federal and state-level legislation in relation to local interviews in rural parts of northeastern Brazil, where the study was conducted. Prior to the onset of the project, we had obtained the appropriate agreement from the Ministry of Environment (MMA), which funded part of this work, to deploy our field campaigns. All field activities, including wildlife population surveys and local interviews, were therefore conducted under the full rigor of Brazilian legislation (Law No. 5,197 of 3rd January 1967 (see http://www.planalto.gov.br/ccivil_03/leis/L5197.htm). Verbal permission was obtained from landowners/managers for wildlife population surveys, and we made every effort to ascertain that all face-to-face interviewees were willing informants for whom special ethics procedures in terms of property rights of “traditional indigenous knowledge” do not apply. In order to conduct interviews in small households and national parks we were only required to have a verbal agreement by landowners or managers of protected areas, all of whom had been contacted by phone/email or met personally prior to the onset of interviews. We are therefore satisfied that all field activities conducted as part of this study were legitimate, even considering the most rigorous interpretation of Brazilian law regarding the acquisition and use of field data that does not involve scientific specimens.

In Brazilian law, local non-indigenous communities and/or households can be approached without prior ethics approval provided that none of the procedures and/or methods involve any form of coercion, which in this study was never the case. Interviewees were not financially compensated for those interviews, which often only lasted 20–30 min. Furthermore, there were no signs of resentment following interviews from any of our informants. We therefore conclude that these interviews were conducted lawfully both within and outside the protected areas we visited.

All interviewers were sensitive of the often harsh local realities faced by smallholders in this part of Brazil, and our interviews were conducted in a socially benign context without the use of translators. Many of our informants were illiterate or semi-literate, yet our verbal face-to-face communication and exchanges with them were clearly unambiguous, and there were no misunderstandings as to the purpose of the interviews or the destination of any data gathered therein. Moreover, in all cases we guaranteed full anonymity of our interviewees to protect their privacy or any legal (mis)interpretation of wrongdoing on their part, particularly in relation to previous hunting activities in either protected or unprotected areas. We are therefore content with the way in which these interviews were conducted, and this is vindicated by the fact that we remain completely unaware of any (formal or informal) complaints or other forms of discontent in relation to the way in which local informants were approached.

We used 30-m resolution satellite images (Landsat 7 ETM, 2001, UTM, Datum Córrego Alegre) to pinpoint and select 196 forest fragments across a study region of ∼252,669 km² encompassing portions of the Brazilian states of Bahia, Sergipe and Minas Gerais. This region included four of the eight recognized BSRs of the Brazilian Atlantic Forest biome [29]: Bahia (N = 141 sites), Diamantina (N = 37), São Francisco (N = 17), and Pernambuco (N = 1). Subsequently, we travelled a total distance of >205,000 km on paved and unpaved roads between January 2003 and January 2005 to survey individual fragments and interview a select set of 408 local residents to obtain patch-scale occupancy data on 18 medium to large-bodied mammal species within these fragments (Text S1). Because all mammal species addressed in this study are ubiquitous in forest habitat [29], they were almost certainly once widespread across the original forest cover throughout the study region. Recorded absences can therefore be interpreted as recent or historical local extinctions. Because these fragments were selected to maximize their size-spectrum and degree of forest structure integrity, they were considerably larger and in better structural condition than the total pool of 738,887 forest remnants (>0.1 ha) persisting across the entire region (Fig. 1), ∼85% of which are smaller than 10 ha (Fig. 2).

All 18 terrestrial and arboreal species considered in this study are coarse-scale forest habitat generalists, in that they occur within the Brazilian Atlantic Forest right across all forest types within the elevation range that we considered here, wherever extant populations persist today. These species were partly selected for the interviews and surveys because of their low level of forest microhabitat specificity. Furthermore, historical records of their geographic occurrence also show that they were widely distributed throughout the study region, and several species still occur well outside the typical forest domain of the Atlantic Forest biome, into more seasonally-dry transitional areas in the neighboring Cerrado and Caatinga domains. These species were therefore widespread right across the study region, and this is particularly the case of large-bodied, wide-ranging ‘landscape’ species that integrate large areas consisting of slightly different forest types. Our collective knowledge was combined with both IUCN maps and expert information obtained from other mammalogists working in the region to further ascertain that all forest fragments targeted in this study were indeed well within the known historical distributional range of the 18 species.

Interviewees were only selected during visits to each forest fragment if they were very familiar with the large mammal fauna and interviews were aided by color plates and high-quality photos, and in some cases playbacks of vocalizations, of each focal species to maximize accuracy in patch occupancy records. Additionally, color plates of five Amazonian and Mesoamerican mammal species (within the same spectrum of body mass as those surveyed here) known not to occur in the study region were also presented to each interviewee to quantify the prevalence of type II errors in the interview data. Further methodological details on forest patch selection, semi-structured interviews, field surveys, variable definition and data analysis are presented in Text S1.

Patch occupancy records of any given species were defined as patch-level incidence when interviewees had no doubt as to whether the species was locally present at the time of interviews or within 12 months prior to interviews, whether the species was thought to be a full-time resident or an occasional transient within that patch. Local extinction events were conservatively defined as unambiguous absences when a species had been reported to have once occurred within a patch but had not been sighted, heard or otherwise detected through tracks, scats, scrapes, burrows, vocalizations, or any other signs of their presence for at least three years. When compiling conflicting information from more than one interviewee referring to the same forest patch, we recorded a species as present if it had been unambiguously reported by at least one independent interviewee even if it had not been reported or otherwise omitted by another. We also excluded from the analysis other forest patches visited during the field campaigns (not shown in Figures S1, S2, S3) for which any of the 18 target species considered here was not reliably known to be either present or absent. Finally, this interview-based presence/absence sampling protocol has been successfully validated at 145 forest fragments elsewhere by four independent field sampling techniques, including diurnal line-transect censuses, camera trapping, intensive monitoring of sand track-plates, and counts of armadillo burrows [30]–[34].

We used 23 Landsat 7 ETM scenes (2001) to quantify key patch metrics (Area, Core Area, Perimeter, Gyrate and Shape indices) and landscape metrics (e.g. patch proximity and proportion of forest within 1-km external buffers) for all 196 forest patches surveyed, using Fragstats v. 3.3 coupled with ArcGIS 9.0. In addition, we calculated the percentage of total old-growth (primary) and successional forest cover (secondary forest older than ∼6 years) within a 1-km buffer area immediately outside the boundaries of the each forest patch. A total of 127 high-resolution digital maps (1∶100,000), which were obtained from the Brazilian Institute of Geography and Statistics (IBGE), were overlaid with the Landsat 7 ETM images to quantify topographic relief and the number of 5-m elevation contour lines (m above sea level) defining the steepness of the terrain of each patch. To verify patch-scale gradients in Δ-elevation we also extracted the minimum and maximum elevation for each forest fragment from four SRTM radar images (tiles 28/13, 28/14, 29/13, 29/14) downloaded from srtm.csi.cgiar.org/. We built elevation maps from SRTM images creating TIN (triangular irregular network) images with 30-m contour lines using the 3D Analyst tool from Arctool. We also measured the straight-line distance from each patch to both the nearest river or major perennial stream and paved/unpaved road used by motor vehicles. Based on local interviews and the high-resolution digital maps, we also recorded the total number of active rural households inside each patch and within a 2-km outside buffer zone from their perimeter to estimate the mean internal and external household density. Topographic slope and distance to roads and rivers were considered to be good proxies of the degree to which forest patches were physically accessible to extractive activities, including subsistence and commercial hunting; household density was considered as a good proxy of the degree to which this potential access could be realized; and forest protection status was a good proxy of the degree to which patches were legally accessible to hunters. All of these data were initially compiled using the best available databases, images, digital/hardcopy maps, and the aggregate information obtained from local informants, all of which were subsequently ground-truthed and verified during site visits (Text S1).

We considered a wide spectrum of explanatory variables describing the history of forest disturbance, physical accessibility, human population density, and patch and landscape variables associated with each forest isolate sampled. Based on generalized linear models (GLMs) followed by an AIC-based model averaging approach, we assess the relative importance of predictors of three assemblage-wide response variables: total species richness, aggregate body mass; and a measure of vulnerability of the mammal species retained in each patch (Text S1 for a description of these response variables).