Statistical analysis

We evaluated the association between habitat structure and the natural history traits of B. pandi by multiple correlation analysis, with P < 0.05 as the significance level. The following variables were considered: SVL (mm), weight (g), perch height (mm), leaf litter depth (mm), vegetation layers, vegetation life form, percent vegetation cover and elevation (meters above sea level). Using the habitat variables, we performed a principal component analysis (PCA) to explore which of these variables presented greater variability between plots with regards the detection/non-detection of B. pandi and, therefore, which of these could explain the observed differences between plots. The variable suitability for PCA was tested performing a Kaiser–Meyer–Olkin test (KMO > 0.5, P < 0.05). Afterward, a quadratic discriminant analysis was performed to determine which of the habitat variables had the greatest discrimination capacity between plots where B. pandi was detected/non-detected.

We assessed the variability in salamander abundance observed in the Supatá population through multiple regression analysis. First, we considered the salamander abundance as a dependent variable and the habitat structure variables recorded at each sampling plot as independent variables: leaf litter depth (mm), vegetation layers, vegetation growth forms, percent vegetation cover, elevation (meters above sea level), temperature (°C), and environment relative humidity. All variables were Ln–transformed prior to perform the statistical analysis.

Second, we evaluated assumptions of normality, autocorrelation, and homoscedasticity using Kolmogorov–Smirnov test, Durbin–Watson test and Breusch–Pagan test, respectively. Given that the p–value of the Durbin–Watson test can easily be less than 0.05 when sample size is very large, we used the Durbin–Watson statistic test (DW) as an autocorrelation criterion. According to Durbin & Watson (1950), a DW of less than 1 indicates a strong positive autocorrelation, a DW greater than 4 indicates a strong negative autocorrelation, values between 1 and 3 suggest a moderate autocorrelation, and a value close to 2 means that there is no autocorrelation.

Third, we tested for multicollinearity between the variables using the variance inflation factor (VIF) with a threshold of 10. Fourth, we selected the “best” regression model employing the Akaike Information Criterion (AIC; Akaike, 1973), considering that models with ΔAIC values of less than two are equally plausible (White & Burnham, 1999). Finally, we used the hierarchical partitioning method to evaluate the contribution of all the independent variables of the regression model (Chevan & Sutherland, 1991).