**2.5. Laboratory determination of soil properties and soil classification**

Selected soil physical properties were determined in the field such as soil depth (cm). Chemical and soil texture were determined in the laboratory using methods by Page and Keeney [33] and Klute [34] respectively. Micronutrients (iron, manganese, copper zinc) were determined using Diethylenetriaminepenta-acetic acid (DTPA) according to Moberg [35]. The field and laboratory data were used to classified soils to level-2 of the FAO World Reference Base [36]. Although chemical soil properties were not used in modelling, it was used for soil classification. For modelling only topsoil depth and texture were used as input data.

#### **2.6. Statistical analysis**

The data was organised for multiple regression analysis. There were two dependent variables (plant cover (%) and total rodent burrows. The independent variable examined were 25, which were landform types, slope gradient (degrees), slope length (m), slope form (concave, convex, straight, compound), elevation (m a.s.l.), drainage, erosion type, rock outcrops and surface stones (number), slope aspect, hillshade (radians), slope curvature types (radians), soil depth (cm), soil texture (textural class), atmospheric temperature (degrees Celsius), topsoil (10 and 30 cm) temperature and topsoil (10 and 30 cm depth) relative humidity (%) were model input data. There was a total of 487 data entries collected. Categorical data such as textural class were given dummy number.

Abiotic factors explaining spatial distribution of plants and animals species were established by inputting 25 factors in a Generalised Linear Model, distribution family 'Gaussian' which is a multinomial for multiple dependent variables [37] applying a formula:

$$\mathbf{Y}\_{\text{i}} = \boldsymbol{\beta}\_{0} + \boldsymbol{\beta}\_{1}\mathbf{X}\_{\text{ii}} + \boldsymbol{\beta}\_{2}\mathbf{X}\_{\text{2i}} + \boldsymbol{\beta}\_{3}\mathbf{X}\_{\text{3i}} + \boldsymbol{\varepsilon}\_{\text{i}} \tag{1}$$

Where: Yi = respondent (dependent) variables (plant cover, trapped animals/rodent burrows as a proxy); β<sup>o</sup> = Intercept; β<sup>1</sup> X1i+….β<sup>3</sup> X3i = predictors or independent variables; ε<sup>i</sup> = error term.

Using R software the GUI rattle [38]. Model validation was addressed by portioning the data. The 70% of the data was allocated for training while 30% was used to develop the model. Different runs were made first using all predictors then reduced or added examining the model goodness of fit by looking the null and residual deviance and Akaike information criteria (AIC), whereby a model with a smallest AIC and a narrower gap between null and residual deviance was opted as model explaining the factors influencing species distribution along the landscape. Multicollinearity, was tackled by keying or deleting weakly correlated variables serially in the model.
