Environmental Engineering Reference
In-Depth Information
Fig. 13.1 Overview of the species distribution modelling workflow. The three phases contain
various tasks, for which typical examples are given in the
right
column
13.2.1 Pre-processing and Visualization
The Response Variable
When the data are presence-absence (i.e. binary) no further preparation is needed.
When data are counts or continuous, we have to make sure that assumptions of the
modelling approach are met. For parametric modelling approaches (regressions by
means of GLM or GAM), count data are usually assumed to be Poisson distributed
but all too often are not. Continuous responses are generally assumed to be
normally distributed. These assumptions can be checked only after modelling,
because we need to look at the residuals or compare log-likelihoods of different
distributions. Generally, if too many zeros have been observed, the data are over-
dispersed and we have to resort to one of three alternative approaches: a quasi-
Poisson distribution (where over-dispersion is explicitly modelled); a negative
binomial distribution (where a clumping parameter is fitted); or a separate analysis
of zeros and non-zeros (as in zero-inflated or other mixed distribution models:
