Page 8 - Pretto_Celesti_2012

Non-native plant species in Mediterranean islands 2565

established non-native species richness with the tenfold cross validation was used to obtain estimates of
explanatory variables. Regression trees represent a relative errors. The 1-Standard Error rule was
flexible statistical procedure that explains the variation employed to select the tree with the best number of
of a single numeric response variable by repeatedly splits, whereas the best tree is the minimum cost tree.
splitting the data into more homogeneous groups The relative contribution of each group of explanatory
characterized by the average value of the response variables to the non-native plant species composition on
variable, the number of observations and the range of each island was assessed by means of the variation
the values of the explanatory variables involved in the partitioning approach (Borcard et al. 1992). We repeated
split (De’ath and Fabricius 2000). As it is a non- the analyses for the datasets of total non-native species
parametric procedure, it is not necessary to test and established non-native species separately. In both
normality or other assumptions regarding the statisti- cases, we used data on presence/absence and we
cal distribution of data in advance. Results are preliminarily tested the unimodal or the linear response
graphically presented in the form of inverted tree of the species along the environmental gradient by means
diagrams which only include independent variables of a Detrended Correspondence Analysis (DCA). There-
that appear to be predictive of the dependent variable fore, we conducted a series of Canonical Correspondence
and are easy to interpret. In recent years, classification Analyses (CCA) and partial Canonical Correspondence
and regression trees (CART) have become widely Analyses (pCCA). The total explained variation was thus
used techniques for the analysis of large and complex decomposed in seven components: the pure effect of
environmental data even in the field of biological geographical, environmental and human-related vari-
invasions (Pys ˇek et al. 2009; Pys ˇek et al. 2010a, b). ables, three first-order shared components (geographi-
In order to account for species-area relationships, the cal ? environmental, geographical ? human-related,
number of total non-native and established non-native environmental ? human-related) and the component
species on each island were regressed on the log-island shared by all three groups of variables. The statistical
area, while Pearson’s standardized residuals were used significance of the unique contributions was tested using
as the response variable. To reduce the multicollinearity 4,999 unrestricted permutations under the null model.
between the 14 explanatory variables and to select only The explanatory variables used in the variation
the most appropriate ones, we included the variables in a partitioning analysis were the same as those employed
bivariate correlation matrix. When a pair was highly in the regression tree analysis plus island area. Prior to
correlated (r C 0.70), we retained only one variable for performing the canonical ordination, we reduced the
further analysis. After this procedure, the following number of descriptors by means of a forward selection
predictor variables were entered in the regression tree procedure associated with Monte Carlo permutation
analysis: latitude, longitude, distance from the closest tests in order to avoid an overestimation of the
island, altitude, percentage of volcanic lithology, per- explained variation in the data (Økland and Eilertsen
centage of island area covered by artificial surfaces, 1994). We thus also obtained an approximately equal
percentage of island area covered by agricultural number of descriptors in each component and reduced
surfaces, percentage of island area subjected to envi- the risk of groups with higher numbers of descriptors
ronmental protection measures and tourist pressure. being comparatively overvalued in the partial analysis
Distance from the mainland, perimeter and percentage (Borcard et al. 1992). All the variables included in the
of sedimentary lithology were correlated, respectively, final model were considered to be significant at
with the distance from the nearest island, area and p \ 0.05.
percentage of volcanic lithology, whereas population All the multivariate analyses were implemented in
density was eliminated because of its strong correlation the program CANOCO version 4 (ter Braak and
ˇ
with the percentage of artificial surfaces. Tourism Smilauer 1998).
pressure and the extension of artificial surfaces are
positively correlated, though not significantly (r =
0.61). Regression trees were performed by means of Results
the Salford Predictive Miningsuite (SPM; http://salford-
systems.com/) using the default settings: the least- The 37 islands host 154 non-native vascular plant
squares was chosen as the splitting criterion and species, which account for 16.7 % of the 920 non-

123