2.3 Data analysis
Non-metric multidimensional scaling (NMDS) ordination analysis based on the Bray-Curtis similarity matrix (Clarke 1993), were used to classify the spatial (stream order) and temporal (season) variations of fish assemblage structure. Relative abundance data for each species was log (x+1) transformed to normalize the data and avoid skew. Bray-Curtis similarity coefficient was calculated based on the relative abundance matrices. Rare species that occurred in less than 3 sites were excluded from the analysis. Next, one-way analysis of similarity (ANOSIM) was carried out to determine whether fish assemblages changed significantly among stream orders or seasons. Then, a similarity of percentage analysis (SIMPER) was used to identify species that contributed most to the spatial or temporal dissimilarities of fish assemblages. All these analyses were performed with the PRIMER 5 software package (Clarke and Warwick 2001), including modules “NMDS”, “SIMPER” and “ANOSIM”.
Difference in environmental factors between seasons were tested by one-way analysis of variance (ANOVAs) using SPSS statistical programs (Version 20.0). Relationships between fish assemblage structure and environmental factors for each season were examined by constrained canonical ordinations. Detrended Correspondence Analysis (DCA) were used to determine the appropriate model for direct gradient analysis (Leps and Smilauer 2003). As the length of gradients for the first axis was estimated at 1.908 (<3) for wet season and 1.746 (<3) for dry season, Redundancy Analysis (RDA) was applied in further analyses. Fish relative abundance data were log10(x+1) transformed and the downweighting option was used to reduce the influence of rare species. Environmental variables that did not meet the normality assumption (Shapiro-Wilk test, p<0.05) were transformed using natural logarithms, and collinear environmental viable with high variation inflation factors (VIF>20) were eliminated from further analyses (McCune and Grace 2002). Stepwise forward selection with Monte Carlo permutation tests (999 permutations, p<0.05) was used to select a parsimonious set of explanatory variables. These analyses were performed using the software CANOCO for Windows 4.5 version (Ter Braak and Smilauer 2002).