Page 5 - FATTORINI2011
P. 5

Williams et al. (2009), residuals were examined with Shapiro-       alternative approaches (McGarigal et al., 2000). CCA identifies
Wilk statistics to determine which distributional assumption was    geographical/environmental gradients which mainly influence
the most appropriate for modelling the dataset. These tests indi-   species composition of sites, whereas CANCOR correlates
cated a marginally significant lack-of-fit for the log-transformed  dominant gradients in species assemblages with geographical/
model (W = 0.916, P = 0.04), but not for the power function         environmental gradients. Thus, CCA is particularly useful for
fitted using non-linear regression (W = 0.970, P = 0.65). Thus,     investigating the influence of geographical/environmental gradi-
preference was given to the untransformed model.                    ents on beta diversity (second level of abstraction sensu
                                                                    Tuomisto & Ruokolainen, 2006) whereas CANCOR investi-
  The resulting z-value (z | 0.25) matched values typically         gates the importance of geographical/environmental gradients in
found for island systems (Rosenzweig, 1995; Drakare et al.,         regulating variation in beta diversity, i.e. “variation in variation
2006). The SAR is a consistent phenomenon in insular ecosys-        in community composition data” — third level of abstraction
tems, and the best way to consider other sources of variation in    sensu Tuomisto & Ruokolainen (2006).
species numbers is through the analysis of residuals from
species-area regressions (Crowell, 1986; Rosenzweig, 1995;          RESULTS
Price, 2004). Thus, residuals from the SAR were correlated with
other geographical variables and environmental heterogeneity        Species richness
indices using Spearman correlation tests, which simply assume
monotonic relationships without any reference to particular           Area was an important correlate of species richness and
functions.                                                          the species-area relationship (SAR) was well modelled by
                                                                    a power function (S = 11.5 A0.23; R2 = 0.84). When
Analysis of variation in species composition                        residuals of the SAR were plotted against other geo-
                                                                    graphical variables, no relationship was found (Table 3).
  Variation in species composition between the islands was          Species richness was also tightly correlated with all meas-
analysed with CCA using the CANOCO program, version 4.5A            ures of landscape heterogeneity (Table 3). When residuals
(Ter Braak & Šmilauer, 2002). A Detrended Correspondence            of the SAR were correlated with landscape heterogeneity
Analysis with the option “detrending-by-segments” (Hill &           indices, significant correlations were found for Pielou
Gauch, 1980) produced a first axis of 9.239 SD, which is more       equitability and, possibly, for Simpson dominance and
than 2 SD units and hence indicates that CCA is suitable for this   Berger-Parker dominance (Table 3). These results suggest
data set (Ter Braak & Prentice, 1988).                              that relationships between species richness and landscape
                                                                    heterogeneity were mainly through area. When the
  Significance of individual environmental parameters (geo-         stronger effect of area was removed, the influence of
graphical and environmental variables) was tested using a for-      landscape heterogeneity was less evident, although there
ward selection with 999 Monte Carlo permutations (see Fatto-        is an indication that richness tends to increase with land-
rini, 2011 for details). The influence of geographical and land-    scape diversity and equitability and decrease with land-
scape variables were tested separately. Both the extent of dif-     scape homogeneity (Table 3).
ferent landscape categories and their proportions can be impor-
tant characteristics of the landscape of a given island. Thus,        TABLE 3. Values of Spearman correlation coefficients of spe-
separate CCAs were performed using alternatively the raw and        cies richness and residuals from the species-area relationship
proportional extent of landscape categories.                        (SAR) with geographical and landscape parameters. Abbrevia-
                                                                    tions are the same as those in Tables 1 and 2. Residuals from
Inter-island biogeographical similarity                             SAR were calculated using the power function. Values in bold
                                                                    are significant at P < 0.05.
  Canonical Correlation Analysis (CANCOR) was used to ana-
lyse the influence of geographical and landscape variables on       Correlation coefficients Correlation coefficients
inter-island biogeographical similarity. The original presence/
absence matrix was then subject to a non-metric multidimen-         between number of spe- between residuals from SAR
sional scaling (NMDS) using the Kulczynski 2 coefficient to
construct a dissimilarity matrix (for a discussion of the use of    cies and environmental and environmental parame-
this coefficient in biogeographical analyses, see Hausdorf &
Hennig, 2005). This technique is designed to construct a “map”      parameters   ters
showing the relationships between a number of objects, given
only a table of distances or similarity between them, and is often  A 0.851      0.082
best at capturing patterns in community data when similarity
coefficients are used (Legendre & Legendre, 1998). The good-        Ds –0.040    –0.194
ness of results obtained by NMDS was measured as stress val-
ues. On the basis of the increase in stress values when the         Da –0.250    0.101
number of dimensions was decreased (Shi, 1993) the retention
of three dimensions was considered to be sufficiently represen-     Di 0.409     0.188
tative. These three dimensions were used as dependent variables
in CANCORs. Separate CANCORs were performed for geo-                SDs –0.024   0.281
graphical and environmental variables to meet “rule C” of
McGarigal et al. (2000). Inter-island faunal similarity was also    SDa –0.114   0.133
investigated by cluster analysis using the Kulczynski 2 coeffi-
cient as a measure of distance and the UPGMA (Unweighted            SDi 0.232    0.081
pair-group method, arithmetic average) amalgamation rule.
NMDS and CANCORs were performed using Statistica 6.0 soft-          N 0.873      0.314
ware.
                                                                    C –0.840     –0.403
  Both CCA and CANCOR investigate the effects of
geographical/environmental variables on species composition         H 0.831      0.364
from complementary points of view and cannot be considered
                                                                    eH/N –0.628  –0.090

                                                                    J 0.737      0.569

                                                                    DMg 0.828    0.358
                                                                     d –0.874    –0.504

                                                                                         663
   1   2   3   4   5   6   7   8   9   10