4. Quantitative Methods for Modeling Species Habitat
45
friendly web interface of some versions of GARP is contrasted by the lack of an
easily interpretable model, a complex and largely inaccessible background procedure, and few published applications. ANUCLIM and DOMAIN are based on
simple concepts, and ANUCLIM has more documentation and published applications than DOMAIN. Nevertheless, it appears that none of the nonstatistical
methods are accompanied by a comprehensive text, and they may require more
effort or communication with the authors if they are to be thoroughly understood.
Assumptions
Austin and co-workers (1994b), McCullagh and Nelder (1989), and Hastie and
Tibshirani (1990) discuss assumptions of each of the methods. All five methods
assume that the species data represent an unbiased sampling of the full range of
environmental conditions in the modeling and prediction regions and that the
environmental variables are measured without error. Models and predictions
based on data that do not conform to these assumptions will have unspecified
errors that could only be investigated by some additional analyses (e.g., sensitivity
analyses).
Models for Seven Plant Species
As an example of the use of these methods, predictions for seven Australian plant
species are presented here from a larger study reported by Elith and associates
(1998) and Elith and Burgman (in review) (see Table 4.2). The study area was the
Central Highlands Regional Forest Agreement (RFA) area in Victoria, Australia
(Fig. 4.1). The study aimed to investigate the suitability of existing species and
environmental data for modeling plant distribution at a scale applicable to conservation reserve planning. Species data were presence-absence records from quadrat surveys completed by government agencies in the previous 20 years. There
were 3,522 such quadrats in the RFA, with some spatial and environmental biases
in their distribution (Elith et al. 1998). Quadrats tended to be located close to roads
and away from ecotonal vegetation. They were clustered close to major population centers and in higher rainfall areas and occurred relatively frequently in
particular vegetation types (e.g., in rainforest and alpine areas). The location of
quadrats was estimated to be accurate to within 250 m of their true position for
approximately 95% of the records (Elith et al. 1998). The full suite of 35 environmental variables described climate, topography, parent rock type, and vegetation
class at a 9-second scale.
Modeling Approaches
Models were developed with five of the methods (ANUCLIM, DOMAIN, GAMs,
GLMs, and GARP). Modeling was implemented for each method in a way that
could be adapted to large numbers of species over several different regions.
45
friendly web interface of some versions of GARP is contrasted by the lack of an
easily interpretable model, a complex and largely inaccessible background procedure, and few published applications. ANUCLIM and DOMAIN are based on
simple concepts, and ANUCLIM has more documentation and published applications than DOMAIN. Nevertheless, it appears that none of the nonstatistical
methods are accompanied by a comprehensive text, and they may require more
effort or communication with the authors if they are to be thoroughly understood.
Assumptions
Austin and co-workers (1994b), McCullagh and Nelder (1989), and Hastie and
Tibshirani (1990) discuss assumptions of each of the methods. All five methods
assume that the species data represent an unbiased sampling of the full range of
environmental conditions in the modeling and prediction regions and that the
environmental variables are measured without error. Models and predictions
based on data that do not conform to these assumptions will have unspecified
errors that could only be investigated by some additional analyses (e.g., sensitivity
analyses).
Models for Seven Plant Species
As an example of the use of these methods, predictions for seven Australian plant
species are presented here from a larger study reported by Elith and associates
(1998) and Elith and Burgman (in review) (see Table 4.2). The study area was the
Central Highlands Regional Forest Agreement (RFA) area in Victoria, Australia
(Fig. 4.1). The study aimed to investigate the suitability of existing species and
environmental data for modeling plant distribution at a scale applicable to conservation reserve planning. Species data were presence-absence records from quadrat surveys completed by government agencies in the previous 20 years. There
were 3,522 such quadrats in the RFA, with some spatial and environmental biases
in their distribution (Elith et al. 1998). Quadrats tended to be located close to roads
and away from ecotonal vegetation. They were clustered close to major population centers and in higher rainfall areas and occurred relatively frequently in
particular vegetation types (e.g., in rainforest and alpine areas). The location of
quadrats was estimated to be accurate to within 250 m of their true position for
approximately 95% of the records (Elith et al. 1998). The full suite of 35 environmental variables described climate, topography, parent rock type, and vegetation
class at a 9-second scale.
Modeling Approaches
Models were developed with five of the methods (ANUCLIM, DOMAIN, GAMs,
GLMs, and GARP). Modeling was implemented for each method in a way that
could be adapted to large numbers of species over several different regions.