1. Land-use change threatens biodiversity in tropical landscapes, but its impact on forest regeneration remains poorly known. In fact, the landscape-scale patterns driving the diversity of regenerating plants within forest fragments have been rarely explored, and we are uncertain whether such drivers vary across regions with different land-use change patterns. 

2. We assessed the effect of landscape composition (forest cover and matrix openness) and configuration (forest patch density) on species diversity of sapling assemblages (trees ≥30 cm height and <1 cm diameter) in old-growth forest fragments from three Mexican rainforest regions with different disturbance levels (n = 20 landscapes per region). We separately assessed old-growth forest specialists (OGS) and forest generalist (FG) species to test the hypothesis that: (i) OGS species shows recruitment limitation (‘loser species’), and can therefore be negatively impacted by landscape changes, especially by forest loss and matrix openness in more deforested regions; and (ii) FG species can regenerate and even proliferate in more disturbed landscapes (‘winner’ species). 

3. We recorded ~24,000 plants from 415 species. Landscape composition showed stronger effects than landscape configuration. The diversity of OGS species generally decreased in more deforested landscapes dominated by open matrices, and FG species followed the opposite response, especially in regions with high-to-intermediate degree of disturbance. Overall, forest fragmentation (patch density) showed weak or no effects on species diversity, especially after controlling for forest cover effects (i.e., fragmentation per se). In contrast to the fragmentation threshold hypothesis, the effect of fragmentation was independent of the regional context. Moreover, OGS species were affected by landscape attributes operating at larger scales than FG species. 

4. Synthesis. Our findings support our hypotheses, and indicate that forest loss and matrix openness, not fragmentation per se, can cause the recruitment failure of tree assemblages in highly deforested rainforests. This can be related to source and dispersal limitation in more deforested landscapes with treeless matrices. Therefore, to promote the regenerative potential (resilience) of forest patches in human-modified tropical landscapes, conservation programs should focus on preventing forest loss (even the smallest forest patches) and improving matrix quality with treed elements, particularly in highly deforested tropical regions.

We worked in three rainforest regions from southeastern Mexico with different patterns and history of land-use change: (1) Marqués de Comillas region (in Selva Lacandona rainforest, Chiapas) and labeled as low-deforestation region (LDR) in the database; (2) Los Tuxtlas rainforest (Veracruz), labeled as intermediate-deforestation region (IDR); and (3) Northern Chiapas, labeled as high-deforestation region (HDR). In each region, we systematically selected 20 old-growth forest patches (that is, 60 old-growth forest patches in total). At each forest patch, we established 25 circular plots (1.60-m radius, 8 m² each; total sampling area per site = 200 m²) at the center of each forest patch, in a grid of 5 x 5 plots with 30 m separation between plots. In each plot, all tree saplings (excluding palms and lianas) with ≥ 30 cm height and < 1 cm of diameter at breast height (DBH) were counted and identified. We classified sapling species into three groups (but we worked with only two): old-growth forest specialists ('OG' in the database) and forest generalists ('Gen' in the database). Basically, old-growth forest specialists (or late successional species) are species that mainly inhabits old-growth forests and forest generalists are species that can be found both in old-growth forests and at intermediate successional stages due to their greater adaptability to both environments. 

We estimated species diversity (i.e., total number of species and number of common species) of OGS and FG saplings for each study site in each region. We summed the data of the 25 plots from each forest patch to avoid pseudoreplication. Then, we measured species diversity with Hill numbers using the entropart package. These diversity indexes measure the effective number of species in a community, which facilitates the comparison of diversities between different ecological communities. Specifically, we calculated Hill numbers of order 0 (⁰D = species richness), 1 (¹D = exponential of Shannon index) and 2 (²D = inverse of Simpson index). As sample coverage varied among forest sites, ranging from 0.81 to 0.98, we estimated species diversity with both observed data and expected one with Chao-Shen extrapolation (CS in the database).

Regarding the landscape variables, we estimated four metrics of landscape structure in 13 different spatial extents: i.e., concentric landscapes or buffers of 100- to 1300-m radius (at 100-m intervals) from the geographic center of each focal forest site. We used this multi-scale approach because we did not know the scale of effect, that is, the spatial scale that yields the strongest responses for each landscape predictor. We selected four landscape metrics that can have strong influence on sapling diversity: two landscape composition variables (i.e., FC = forest cover, MO = matrix openness), and two landscape configuration variables (i.e., PD = patch density, ED = forest edge density). We estimated forest cover as the percentage of landscape area covered by old-growth forest. We calculated matrix openness as the percentage of the anthropogenic matrix covered by open areas (i.e., human settlements, pastures, annual crops, and water bodies). Patch density is the number of old-growth forest patches within the landscape divided by landscape area (n/ha). Finally, edge density is the perimeter length of all old-growth forest patches within the landscape divided by landscape area (m/ha).