3. Identifying the Ecological Correlates of Extinction-Prone Species
31
the species to extinction pressure. In other words, if major efforts had not been
made to save these species, they would have been either extinct or functionally
extinct by the mid- or late 1980s.
The explanatory variables included the species’ body mass, trophic level, flying
capability, and taxonomic status. In addition, species endemic to the Chatham
Islands were labeled in the analysis. The Chatham Islands have a peculiar representation of birds, and several common mainland species are absent (Atkinson
and Millener 1991). Explicit recognition was necessary in the model to offset the
peculiarity of this distribution. This was the only instance in which subspecies
were included in the data set.
The selection of these variables was based either on hypotheses from the
literature on correlates of extinction risk or from past empirical studies. For
instance, body mass is correlated with reproductive rates, range requirements, and
possibly population density. All these factors have been shown to be correlated
with extinction risk (Pimm et al. 1988). The level of endemism was proposed by
Mills and Williams (1984) as a trait linked to extinction risk. Small islands
frequently have a higher turnover of species (Pimm et al. 1988), implying that this
type of isolation may be a risk factor. Trophic level has been linked to range
requirements, and as many reserves are believed to be too small to maintain viable
populations of high trophic-level carnivores, trophic level may also be an extinction risk correlate. Flightlessness and poor flying ability have also been proposed
as extinction correlates for New Zealand birds (Gill 1991; Mills and Williams
1984). Hence in the first instance, variable selection was determined by biological
realism.
Model specification generates a range of problems.The first is that some explanatory variables may manifest multicollinearity. This may place restrictions on
how a particular biological factor is represented. For instance, the first iterations
of the model used several different measures of endemism. These variables were
highly correlated amoung themselves and with other variables. New Zealand had
many flightless birds, so some measures of endemism were correlated with “flight
ability.” The variable representing endemism that was least correlated with the
others was used.
The particular approach to parameter selection in this example was based on the
Hendry method (Gilbert 1986). The starting point for the regression is a model
with as many parameters as possible that are likely to explain the phenomenon of
interest. Then, the variable with the least significant coefficient is eliminated.
Another regression is performed, and the process is repeated until the simplest
possible model is found that includes all important explanatory variables. In this
example, variables that were not significant at a 95% level of confidence for at
least one of the regressions (i.e., whose coefficient was not significantly different
from zero) were eliminated.
The advantage of this approach is that usually it results in a model with good
explanatory power, without retaining variables that do not contribute to explaining the response variable. Multicollinearity can sometimes be detected informally
with this approach, as regression coefficients are usually sensitive to specification
31
the species to extinction pressure. In other words, if major efforts had not been
made to save these species, they would have been either extinct or functionally
extinct by the mid- or late 1980s.
The explanatory variables included the species’ body mass, trophic level, flying
capability, and taxonomic status. In addition, species endemic to the Chatham
Islands were labeled in the analysis. The Chatham Islands have a peculiar representation of birds, and several common mainland species are absent (Atkinson
and Millener 1991). Explicit recognition was necessary in the model to offset the
peculiarity of this distribution. This was the only instance in which subspecies
were included in the data set.
The selection of these variables was based either on hypotheses from the
literature on correlates of extinction risk or from past empirical studies. For
instance, body mass is correlated with reproductive rates, range requirements, and
possibly population density. All these factors have been shown to be correlated
with extinction risk (Pimm et al. 1988). The level of endemism was proposed by
Mills and Williams (1984) as a trait linked to extinction risk. Small islands
frequently have a higher turnover of species (Pimm et al. 1988), implying that this
type of isolation may be a risk factor. Trophic level has been linked to range
requirements, and as many reserves are believed to be too small to maintain viable
populations of high trophic-level carnivores, trophic level may also be an extinction risk correlate. Flightlessness and poor flying ability have also been proposed
as extinction correlates for New Zealand birds (Gill 1991; Mills and Williams
1984). Hence in the first instance, variable selection was determined by biological
realism.
Model specification generates a range of problems.The first is that some explanatory variables may manifest multicollinearity. This may place restrictions on
how a particular biological factor is represented. For instance, the first iterations
of the model used several different measures of endemism. These variables were
highly correlated amoung themselves and with other variables. New Zealand had
many flightless birds, so some measures of endemism were correlated with “flight
ability.” The variable representing endemism that was least correlated with the
others was used.
The particular approach to parameter selection in this example was based on the
Hendry method (Gilbert 1986). The starting point for the regression is a model
with as many parameters as possible that are likely to explain the phenomenon of
interest. Then, the variable with the least significant coefficient is eliminated.
Another regression is performed, and the process is repeated until the simplest
possible model is found that includes all important explanatory variables. In this
example, variables that were not significant at a 95% level of confidence for at
least one of the regressions (i.e., whose coefficient was not significantly different
from zero) were eliminated.
The advantage of this approach is that usually it results in a model with good
explanatory power, without retaining variables that do not contribute to explaining the response variable. Multicollinearity can sometimes be detected informally
with this approach, as regression coefficients are usually sensitive to specification