The Third Spanish National Forest Inventory (3SNFI, Ministerio de Medio Ambiente 1997–2007) was used as the data source for native tree species richness, type of forest or functional group, harvest intensity, successional stage, basal area and forest canopy cover. The 3SNFI had a systematic sampling design, with plots located at the intersections of a 1 × 1 km UTM grid that fall within forests with a minimum canopy cover of 5%. A total of 24 498 inventory plots were surveyed in the study area from 2000 to 2004. For each plot, the total number of tree species, including saplings, was recorded within a 25-m radius circle using a predefined list of 177 tree species. For this study, we excluded exotic species and plots corresponding to plantations, dehesas, riparian and burnt forests, leaving a total of 78 native species in 14 306 inventory plots (referred to hereafter as ‘forest plots’). However, all 24 498 plots were used to assess land use/cover composition in the landscape surrounding these forest plots (see below).
A functional group, or forest type, variable (FUNC) was defined based on the functional identity of the dominant tree species of each plot. The following categories were distinguished: broad-leaved deciduous (BD), sclerophyllous evergreen (SE) and coniferous forests (CO). Mixed-forest plots, those with two co-dominant species of different functional groups (<1%), were not considered in the analysis of tree species richness.
The 3SNFI surveys also characterized the type of harvesting – clear-cutting, shelterwood cutting or selection cutting – applied in each forest stand, based on either direct knowledge or indirect evidence such as stumps or slash indicating management activity within approximately 10 years since the previous SNFI survey. Because a one-time field assessment only gives evidence of the degree of canopy removal, rather than the whole set of forestry practices during the rotation period, we renamed harvesting categories as the following: complete harvesting (clearcutting), partial harvesting (selection cutting) and no treatment. Species richness was not assessed in forest plots with shelterwood cutting due to the small sample size (only 79 plots). The numbers of plots and dominant tree species for each combination of environmental zone (EnZ), functional group (FUNC) and harvesting intensity category are presented in Table 1.
Table 1. Sample size (number of plots, n) and proportion of dominant tree species within the stands for each combination of environmental zone (EnZ), functional group (FUNC) and harvesting intensity
|EnZ||FUNC||Complete harvesting||Partial harvesting||No treatment|
| n ||Composition|| n ||Composition|| n ||Composition|
|MDM||BD||21||90% Quercus pyrenaica||49|| ||1321|| |
|SE||0||–||26||100% Quercus ilex||793||96% Quercus ilex|
|CO||179||90% Pinus sylvestris||523|| ||2649|| |
|MDN||BD||0||–||37|| ||846|| |
|SE||0||–||64||97% Quercus ilex||2260|| |
|CO||427||90% Pinus pinaster||772|| ||2974|| |
|SE||0||–||42||98% Quercus ilex||939||96% Quercus ilex|
|CO||0||–||74|| ||214|| |
|Total||627|| ||1587|| ||12065|| |
We also obtained data for each plot from the 3SNFI on four additional variables that may affect species richness: basal area, forest canopy cover (FCC), slope and successional stage. Basal area and FCC accounted for coarse differences in forest types across the study area. Basal area (m2 ha−1) was estimated in the 3SNFI from a sample of trees selected depending on the stem diameter at breast height (DBH), and its distance to the plot centre, ranging from 5-m radius for trees with DBH from 7.5 cm to 12.5 cm up to a maximum radius of 25 m for trees with DBH of at least 42.5 cm. Slope and FCC were measured in the field in the 3SNFI. The number and composition of tree species naturally varies across successional stages. Successional stage was therefore considered and estimated as an ordinal variable based on the average development stage of the three dominant tree species inventoried in each plot: (i) recently regenerated (up to canopy closure); (ii) thicket (up to natural pruning); (iii) trees with DBH ≤ 20 cm and (iv) trees with DBH > 20 cm.
Three climatic variables important for tree species richness in this region were also considered: total precipitation from September to May (representing the Mediterranean growing season), maximum temperature in July (as a proxy for maximum annual temperature) and minimum temperature in January (representing the lowest annual temperature). Climate variables were obtained at a resolution of 200 m from the Climatic Atlas of the Iberian Peninsula (Ninyerola, Pons & Roure 2005) and matched to each of our plots. We also considered plot elevation, which was calculated from the official Spanish Digital Elevation Model at a resolution of 25 m (Ministerio de Fomento 1999). The selection of these variables was based on exploratory analyses and on the literature about factors affecting plant species diversity. Temperature and precipitation are critical determinants of productivity and stress (e.g. Currie & Paquin 1987; Mittelbach et al. 2001). Elevation was intended as a proxy for altitudinal gradients in additional factors difficult to quantify, such as total altitudinal bands area, net primary productivity or those derived from evolutionary geometric constraints or historical human land-use intensity (see Nogués-Bravo et al. 2008).
The landscape context surrounding each forest plot was characterized using all 3SNFI plots, including those outside the boundary of the study area to avoid edge effects. For this landscape context analysis, plots from the 3SNFI were classified into the following categories: forests, plantations, dehesas, riparian forests and ‘others’. Dehesas were defined in the 3SNFI as woodlands of scattered trees with a minimum FCC of 5% (and normally with a FCC lower than 20%) and croplands or pastures in the understorey. Additionally, plots within the forest category were subclassified into three harvesting categories according to the observed harvesting intensity: completely harvested forest (including both clear-cut and shelterwood plots), partially harvested forest and unmanaged forest. All missing plots (i.e. the intersections of the 1 × 1 km grid falling outside of forest cover) were assigned to an ‘others’ category, which encompassed agricultural, urban and other nonforest land uses. The percentage cover of each of these seven land-use categories [forests (partially or completely harvested or unmanaged), plantations, dehesas, riparian forests and others] was calculated within two radii of each forest plot, 3·5 and 5·5 km. Selection of these radii was based on a trade-off between exploratory analyses of spatial residuals, ecological criteria on long distance dispersal (e.g. Jordano et al. 2007) and assuring a minimum number of plots to represent landscapes.
Tree species richness was defined as the number of native species of trees surveyed in a particular 3SNFI forest plot. We constructed a hierarchical, or multilevel, model to account for the different scales, and potential interactions, represented in the data (see Fig. 2 for a conceptual diagram of the hierarchical structure). We followed a Bayesian approach to estimate the model parameters, as is suitable for multilevel models (Gelman & Hill 2007). For our final model (see Table S1 for models considered), tree species richness in plot i, Si, was estimated as a function of the climatic, topographic, forest structure and management variables presented in the previous section using a Poisson likelihood:
with the plot level mean richness modelled as:
where αhfe is a nested effect of harvest intensity, h, for each functional group, f, and environmental zone, e (nested structure discussed further below). The parameter δ describes the effect of successional stage on species richness, and κe(q) is a vector of coefficients describing the nested effects of q Environmental Factors (precipitation, temperature, elevation, slope, basal area and FCC) in each environmental region, e (for each q variable, κe ~ Normal(0,10 000)). We also included quadratic terms for some of the environmental variables, according to exploratory analyses. The vector of parameters μ(c) represents the regression coefficients associated to each Landscape Cover variable (c = percentage of plantation, dehesa and riparian forest). The vector θh(k) represents effects of each Landscape Management variable (k = percentage in the landscape of complete harvesting, partial harvesting or no management), depending on each plot's harvesting practice, . Finally, spatially explicit random effects, ρ(j)(i), were included to reduce autocorrelation in the residuals (Legendre & Legendre 1998). For reasons of computational limitation, plots, i, were grouped in j = 192 cells of 30 × 30 km to estimate these spatial random effects. Spatial effects were assigned to a multivariate Gaussian distribution with covariance expressed as a negative exponential function of the distance between cell centroids (e.g. Diggle, Tawn & Moyeed 1998). Further detail can be found in Fig. S1 in the online Supporting Information.
Figure 2. Conceptual diagram of the hierarchical structure of the model. Boxes represent levels of ecological organization and arrows represent parameter organization (i.e. prior probability distributions for model parameters and hyperparameters, defined by other hyperparameters).
Download figure to PowerPoint
As mentioned above, the effects of local harvest intensities, h, were nested within functional groups, f, and environmental zones, e, as follows:
Note that γe represents an estimate of the mean species richness in plots within each environmental zone, controlling for differences in functional groups and management among stands and is therefore a stand level, rather than regional level, indicator of biodiversity.
To assess whether managed plots had higher or lower species richness with regard to unmanaged plots, for each combination of environmental zone, e, and forest functional type, f, differences between intercepts for complete harvesting, or partial harvesting, and unmanaged plots were calculated within the model:
The resulting posterior distributions of these differences were used to calculate whether treatments had significant effects on species richness. The probability, Pr, that a particular harvest intensity resulted in the same species richness than no treatment was then calculated as Pr = Maximum(Pr(D > 0),Pr(D ≤ 0)), where Pr(D ≤ 0) is the cumulative distribution of D up to zero and Pr(D > 0) = 1−Pr(D ≤ 0). The maximum of these values yields a probability that two treatments are different, a Bayesian ‘test statistic’ from which the significance of a treatment may be assessed. A value greater than 0·95 could be interpreted as analogous to a frequentist P-value of 0·05, but these probabilities can be interpreted as more continuous probability measures than frequentist P-values. Because the hierarchical modelling approach used in this study represents the data complexity and inherent variability better than more conventional multivariate analyses, we have considered here as statistically significant those values of Pr ≥ 0·90.
We assigned noninformative priors to all parameters. All variance terms were assigned noninformative Gamma priors, 1/σ*2 ~ Gamma(0·01,0·01). Similarly, regression coefficients and hyperparameters were assigned noninformative Normal distributions:
Models were fit using OpenBUGS 3.1.0 (Thomas et al. 2006). The final model was run for 405 000 iterations on three independent chains, and convergence was assessed, after discarding pre-convergence burn-in interactions, via visual inspection and using the Gelman–Rubin statistic (Gelman & Rubin 1992). Although we structured these models to reflect our understanding of these ecological processes and to answer the main questions of the study, we also used model selection to help determine the best model. We calculated the Deviance Information Criteria (DIC) for models of increasing complexity (see Table S1), to evaluate their fit to the data while penalizing for increased numbers of parameters (Spiegelhalter & Best 2000). Also, to quantify the explanatory power of the model, explained variance was calculated. Finally, as a tool to explore the results, we generated predictions of species richness at the environmental zone level using the γe parameters, as a function of each explanatory variable (conditional on mean values of the others).