2. -1
A comparison of predictive methods in extinction risk
studies: contrasts and decision trees
MATTHEW S. SULLIVAN1,
*, MARTIN J. JONES1
, DAVID C. LEE2
,
STUART J. MARSDEN2
, ALAN H. FIELDING1
and
EMILY V. YOUNG3
1
Department of Biological Sciences, Manchester Metropolitan University, Chester Street,
Manchester M1 5GD, UK; 2
Department of Environmental and Geographical Sciences, Manchester
Metropolitan University, Chester Street, Manchester M1 5GD, UK; 3
4/111 Rushall Crescent, Fitzroy
North, VIC 3068, Australia; *Address for correspondence (e-mail: m.sullivan@mmu.ac.uk; phone:
+44-0-161-247-1664; fax: +44-0-161247-6325)
Received 8 March 2005; accepted in revised form 5 July 2005
Key words: Comparative analysis, Decision trees, Extinction risk, Galliformes
Abstract. Over the last two decades an increasing emphasis has been placed on the importance of
controlling for phylogeny when examining cross-species data; so-called comparative methods.
These methods are appropriate for testing hypotheses about correlations between evolutionary
events in the history of a clade and adaptive responses to those changes. When this approach is
applied to extinction risk, possible correlations between evolutionary changes in, for example, body
size or habitat specialisation and some measure(s) of current threat status are examined. However,
there may be a mismatch here between the results of such studies, and the real, pragmatic needs of
species conservation. This kind of approach certainly adds to our knowledge of some fundamental
processes, but it is more difficult to see how this can be applied to conservation decision-making.
For more practical purposes a decision-tree approach can be extremely useful. This paper illustrates
the use of a contrasts based analysis of extinction risk compared with a decision-tree analysis for
Galliformes (Aves). While the contrasts analyses concur with some general macroecological trends
found in other studies, the decision-tree models provide lists of species predicted to be more at risk
than current assessments would suggest. We argue that in practical terms, decision tree models
might be more useful than a macroecological linear model-based approach.
Introduction
As we move into a new period of widespread and rapid biological extinctions,
analyses of correlates of extinction risk and prediction of extinction risk are
increasing. Much of this effort uses cross-species correlative analysis (e.g. see
review in Fisher and Owens 2004). However, this kind of analysis is not
straightforward because of phylogenetic non-independence among species (e.g.
Harvey and Pagel 1991). A problem arises because species within a clade may
have similar values for the independent and dependent variables, not because
selection has operated separately and similarly in each case but simply because
the species share common ancestry. The unit of analysis therefore cannot be the
species but should instead be the independent evolutionary events within
the clade (e.g. Garland et al. 1992, and Fisher and Owens 2004 give a recent
Biodiversity and Conservation (2006) 15:1977–1991 Ó Springer 2006
DOI 10.1007/s10531-005-4315-2
3. illustration). To carry out such analyses requires some estimate of phylogeny
and as these have become more readily available so the application of methods
which control for phylogeny (e.g. Felsenstein (1985), Grafen (1989), Purvis and
Rambaut (1995) and Pagel (1999)) has increased. These methods generate and
analyse contrasts, changes in the causal and response variables of interest at
nodes in the phylogeny. Indeed, Fisher and Owens (2004) identify some 20
studies which explicitly control for phylogeny while examining correlates of
extinction risk. However, it is important to remember that these methods were
developed to test evolutionary hypotheses, such as concerning perhaps life
history evolution or more generally adaptive responses to changes in ecological
pressures.
In moving from the species to the evolutionary event as the unit of analysis,
so the application of the results beyond a coevolutionary hypothesis becomes
less clear. This is a particular issue for the applied discipline of extinction-risk
studies. By looking for correlates of extinction risk one is implicitly looking for
predictors of extinction risk, and when we talk about predicting extinction risk
we are usually talking about predicting risk for individual species or clusters of
species. So, perhaps sometimes there is a mis-match between what comparative
methods do and what legislators, decision-makers, managers and donor
organisations actually want, which is a species-by-species analysis. To give an
example from our own work, Sullivan et al. (2000) used a comparative method
to identify correlates of extinction risk in European hoverflies. We found that a
decrease in flight period and an increase in body size in the history of a clade
were correlated with increased present day extinction risk. This does not mean
that large species with short flight periods were at risk (although many were),
but that larger changes in those variables in their evolutionary history placed
species at relatively greater risk.
Comparative methods, then, can provide information on evolutionary pro-
cesses which affect present-day extinction risk. The question posed here is
whether this is a wholly appropriate approach for the applied discipline of
predicting extinction risk. We suggest that a method is also required which will
provide predictions of extinction risk for species which can be used in setting
priorities and other conservation decision-making. A decision tree is a logical
model represented as a binary splitting tree that shows how the value of a
target variable can be predicted by using the values of a set of predictor
variables. A regression tree is generated for continuous target variables (see
Jones et al. this issue), and a classification tree for binary targets. Ultimately,
the tree produces a list of species which it predicts to be at risk according to a
given definition, such as Red List classification. Some of these will be ‘cor-
rectly’ classified while others will be misclassified. Of the misclassifications,
some may be ‘false positives’ i.e. species which are not currently classified as at
risk but share characteristics of those which are at risk. We suggest that the
identification of this group of species is a key objective in predicting extinction
risk. In arriving at these classifications the decision tree method has a number
of features which make it particularly suited to biological data: categorical
1978
4. predictor variables can be included alongside continuous variables; missing
values can be substituted by combinations of other predictors and variables
can be used in the model more than once at different stages. It is particularly
useful when there is strong conditional dependence amongst predictors, and
responses to variables do not have to be linear or monotonic (Bell 1999), all of
which make it particularly appropriate for our species data and objectives.
Jones et al. (this issue) give a thorough description of the advantages of
decision trees in the extinction-risk context.
We present analyses of extinction risk using a data set of 285 species of
Galliformes (Aves). A contrasts analysis is compared with a decision tree
analysis to illustrate the kinds of information each produces. We suggest that
the species lists produced by the decision trees are more easily interpretable for
end users of such data.
Methods
Taxonomy
This work considers those species belonging to the avian order Galliformes,
which contains the families Megapodidae (the megapodes), Cracidae (cracids),
Odontophoridae (New World Quails), Tetraonidae (grouse), Meleagrididae
(turkeys), and Phasianidae (partridges, Old World Quails, and pheasants). We
follow the taxonomy utilised by the Action Plans of the four relevant Species
Specialist Groups (Brooks and Strahl 2000; Dekker et al. 2000; Fuller et al.
2000; Fuller and Garson 2000; Storch 2000). Therefore, species taxonomy
followed: Sibley and Monroe (1990, 1993) for the Phasianidae; Jones et al.
(1995) (Megapodidae); Sibley and Monroe (1990, 1993), del Hoyo et al. (1994)
and American Ornithologists’ Union (1998) (Tetraonidae), with the latest
publication given priority in cases of disagreement (Storch 2000); and Sibley
and Ahlquist (1990) (Cracidae), with some modifications resulting from
International Cracid Symposia (Brooks and Strahl 2000). A recently described
species of grouse, Centrocercus minimus (Young et al. 2000), subsequent to
those references followed in Storch (2000), is included here. The extinct Co-
turnix novaezelandiae (Holdaway 1999; Birdlife International 2000) was also
included. The number of subspecies for each species recognised by this study
follow the aforementioned taxonomic references, except for cracids, New
World quails, partridges, guineafowl, snowcocks, and grouse where del Hoyo
et al. (1994) is adhered to.
Phylogeny
Sibley and Ahlquist (1990) was the starting point for a phylogeny estimate of
this group and subsequent studies are used to complete a tapestry estimate. We
place Meleagris, Tympanachus, Perdix, Alectoris, Coturnix, Bambusicola, and
1979
5. Francolinus within the pheasants (Kimball et al. 1999). Remaining species of
Perdicini and Phasianini not included by Kimball et al. (1999) are placed
according to Johnsgard’s (1999) dendogram of the pheasant genera, except for
the genus Polyplectron which follows the phylogeny constructed by Kimball
et al. (2001). The intrafamilial phylogenetic relationship in grouse is taken
from Lucchini et al. (2001), with Centrocercus minimus positioned according to
the species status it now occupies (Young et al. 2000). The evidence reviewed
by Brom and Dekker (1992) resulted in the megapodes being currently treated
as the sister group of all other galliforms. This taxonomic arrangement is
followed here with phylogenetic branching taken from Jones et al. (1995)
(adapted from Brom and Dekker 1992).
Data coding
Red List coding was the dependent variable in all analyses. Species conserva-
tion status was taken from the 2000 Red List (Birdlife International 2000).
Each IUCN category was coded according to level of extinction threat: ‘LC’
(0); ‘NT’ (1), ‘VU’ (2), ‘EN’ (3), ‘CR’ (4), ‘EW’ and ‘EX’ (5).
A range of predictor variables was used which can be divided into a number
of groups. Variables reflecting evolutionary radiation at the genus and
species level were number of congeners, number of subspecies and species/
genus ratio.
Morphology was represented by body mass and wing length. Species body
mass values were taken primarily from Dunning (1993) and supplemented
by: del Hoyo et al. (1994) (all Galliformes); Madge and McGowan (2002)
(pheasants, partridges and grouse); Jones et al. (1995) (megapodes); and
Johnsgard (1999) (pheasants). Where a body mass range was given with no
mean, the midpoint was taken. For polytypic species, several body mass values
were taken from each subspecies, if these were available, and the mean value of
these subspecific means taken. Missing body mass values were estimated by
regressing known body masses (or log-transformed body masses) against wing
length, independently for each sex and within a genus, clade or subfamily to
reduce potential scale differences across a wider taxonomic range (Lee and
Marsden this issue). We derived from these the ratio of male to female mass.
Species wing length data were taken from Madge and McGowan (2002)
(pheasants, partridges and grouse), Jones et al. (1995) (megapodes), Johnsgard
(1999) (pheasants), Johnsgard (1988) (partridges, quails, snowcocks and
francolins), Johnsgard (1983) (grouse), Urban et al. (1986) (African species),
and Cramp and Simmons (1980). We supplemented these data with wing
length measurements taken from specimens housed at The Walter Rothschild
Zoological Museum, Tring, UK. Again, from these we derived the male to
female wing length ratio.
1980
6. Other life history and ecological variables
Clutch size and egg mass. The midpoint, or mean if given, of the typical or
usual clutch size range in the wild was considered. If clutch size data were
available only from a captive situation, then this was taken from Delacour
(1977) and Howman (1979). Megapode clutch size was taken as ‘1’ with eggs
being regularly laid throughout the breeding season, which follows the defi-
nition of Jones et al. (1995). For mating system, species were classified by their
exclusive or dominant mating system: monogamous (coded 0) or polygamous
(coded 1). Migratory status of a species was considered as an indication of
some reliance upon more than one seasonal environment and the level of
adaptation to a particular environment. Three seasonal migration variables
were considered: geographical; altitudinal; and habitat migration. Species were
classified as whether they did or did not undergo each of these migrations.
Species were also classified as ‘sedentary’ (no migration pattern) or vagile
(undergoing one or more migration types). Megapodes that undergo small
migrations between small islands to communal nesting sites were scored under
habitat migration. Majority and regular migratory patterns were considered as
positive migrations. This excluded irruptive or nomadic patterns determined by
rainfall or food resources.
Altitudinal ranges were classified as occurrence or not within 500 m bands,
from sea level to 500, 500 to 1000 m and so on until >2000 m. Altitudinal
range was defined as the absolute altitudinal range a species was recorded from
to allow for local differences in natural and anthropogenic factors. Altitudinal
records for introduced populations were excluded from the analyses: some
conflicted with or increased the overall native range, e.g. the native altitude
range of Lady Amherst’s pheasant Chrysolophus amherstiae is >1800 m,
whereas the introduced population in Britain is situated in the lowlands.
Latititudinal range midpoint for each species was calculated from the species
distribution maps in del Hoyo et al. (1994).
Species were classified as ‘mainland’ species (occurring exclusively on
mainland land masses), ‘island’ species (occurring exclusively on islands), or
‘mainland and island’ species (occurring on both land mass types). We clas-
sified any land mass equal to or smaller in area than Madagascar as an island.
The number of natural habitat types in which each species is recorded was
used to give a measure of habitat specialisation. Habitat categories followed
those of the IUCN Habitats Authority File (http://www.iucn.org/themes/ssc/
sis/authority.htm) at the highest hierarchal level: ‘forest’; ‘savanna’; ‘shrub-
land’; ‘grassland’; ‘wetlands (inland)’; and we combined the remaining cate-
gories into ‘other’ habitats (e.g. ‘rocky barren areas’ and ‘desert’). ‘Forest edge
and clearings’ was added as an important ecotone for forest and more open
habitat species. The number of anthropogenic habitat types in which each
species is recorded was used to give a measure of adaptability to anthropogenic
habitats. Anthropogenic habitats were divided into ‘man-altered landscapes’
(including logged forest, regenerating vegetation, plantations), ‘agriculture’ (all
1981
7. monocultures, excluding plantations), and ‘other’ (urban habitats). Species
were recorded as either present (scores 1) or absent (0) in each habitat type.
Analytical methods
To examine correlates of IUCN status while controlling for phylogeny CAIC
v2.6.9 (Purvis and Rambaut 1995) was used. For continuous variables, we used
the algorithm Crunch, and for binary variables we used their Brunch algo-
rithm. CAIC generates contrasts at nodes throughout the tree, which are
changes in the independent and dependent variables at these nodes, and which
are themselves statistically independent and amenable to traditional analysis
methods. IUCN status was always the dependent variable. For continuous
predictors, univariate linear regression through the origin was used. For binary
predictors, contrasts are produced at each node where the predictor changes
from state 0 to state 1. A sign test is then appropriate to examine any bias in the
signs of the changes in IUCN status at these nodes.
Decision tree analyses
In this paper we are less interested in specific IUCN categories, which are also
lost in the contrasts analysis, and so the IUCN status of species was here
dichotomised either side of vulnerable (VU) or above, giving 75 of the 285
species as threatened. The QUEST algorithm (Loh and Shih 1997) as imple-
mented by SPSS AnswerTree v3.0, which is the appropriate algorithm for
dichotomous dependent variables, was used to classify species as either threa-
tened or not using all the variables defined earlier. (Jones et al. (this issue) give
an example of using the CHAID decision tree algorithm when the dependent
variable has multiple categories). No training data were available so instead a
10-fold cross-validation was used. A maximum of five levels below the root was
specified and the minimum number of cases for the parent and child nodes were
set at 10 and 5, respectively. The algorithm produces a tree in which in the
terminal nodes all cases are classified as either threatened or not in such a way so
as to minimise the costs of misclassification, which can be either false positives
or false negatives. Of particular interest to us are the false positives – species
which share all the characteristics of threatened species but which are currently
classified as unthreatened (see also Reed and Shine 2002). One advantage of
decision tree analysis is the opportunity to specify different relative costs for the
two types of misclassification. When costs are equal this is equivalent to max-
imising the proportion of cases correctly classified using the available data. It is
important to recognise that by not manipulating these costs one is making an
implicit assumption that these costs are equal. Within conservation it is rea-
sonable that we should place a greater relative cost on wrongly classifying a
species as unthreatened when it is actually threatened (a false negative), and so
1982
8. we investigate the effect of making the cost of a false negative equal to or to be 2-
fold, 3-fold or 6-fold greater than that of a false positive.
Results
Contrasts analyses
One of the main confounding traits in analyses like this can be body size
(Fisher and Owens 2004), so firstly we examined the relationship between
contrasts in body size and extinction risk category. Two measures are available,
contrasts in wing length and in mass. For both males and females neither of
these measures was significantly related to extinction risk (regressions through
the origin, all d.f. = 1, 148, F values from 0.01 to 2.96, all n.s.). Body size was
therefore not a confounding variable and has not been controlled for in sub-
sequent analyses.
Table 1a shows the results of individual regressions of contrasts in IUCN
status against continuous predictor variables. Increase in number of subspecies
and increase in the sum of natural habitat types occupied are associated with a
decrease in extinction risk. The results of the sign tests on binary predictors of
extinction risk are shown in Table 1b. Here, nodes at which movement, alti-
tudinal migration and occurrence in forest edge habitats arise are associated
with decreases in extinction risk.
Decision tree analyses
The tree produced by QUEST when misclassification costs are equal (Tree 1) is
shown in Figure 1. Thirty-eight false positive species are identified in terminal
node 8 and a further 3 in terminal node 5 giving a total of 41. The tree was
robust to the effect of increasing misclassification costs. At 2- and 3-fold in-
creases in the cost of a false negative a further 3 species were classified as false
positives in terminal node 10, with no change in the use of variables to split the
tree. When the relative costs were increased to 6-fold, there was some change in
the variables used to split the tree (Tree 2, Figure 2), but the core group of 38
species in terminal node 8 remained constant. At this level of costs, the rules on
the right hand side of the tree change slightly. The use of species occurrence on
an island drops out and is replaced by relative egg mass (node 6) and mass ratio
(node 10). These results are summarised in Table 2.
Discussion
We have approached the analysis of extinction risk in two ways. The contrasts
analysis has identified a small number of correlates of change in extinction risk.
An increase in the number of subspecies was negatively correlated with change
in risk. Subspeciation may be associated with an increase in the dietary and
1983
9. habitat range of the species, but it may also reflect the age of the taxon. Under
the taxon cycle hypothesis (e.g. Ricklefs and Cox 1978) subspeciation occurs in
the middle age of a taxon, followed by subsequent further full speciation and
hence a decline in the number of subspecies. Under this model, then, there is a
non-linear correlation between taxon age and extinction risk. We assumed
Table 1.
Variable d.f. F Coefficient Significance
a. Outcomes of linear regressions through the origin of contrasts in continuous variables against
contrasts in Red List status
Spp/genus 1, 150 < 0.001 0 n.s.
No. ssp 1, 150 11.82 À0.06 0.001
Mass ratio 1, 148 0.17 0 n.s.
Wing length ratio 1, 148 0.56 0 n.s.
Clutch size 1, 140 3.29 À0.05 n.s.
Relative egg mass 1, 114 1.50 2.35 n.s.
Relative egg length 1, 141 0.54 2.49 n.s.
Sum natural habitats 1, 149 6.55 À0.16 0.011
Sum man-altered
Habitats 1, 149 1.90 À0.13 n.s.
Latitudinal mid-point 1, 148 1.09 0 n.s.
Variable N (non-zero) contrasts Bias Significance
b. Sign tests of contrasts in Red List status for binary predictors
Mainland sp 22 n.s.
Island sp 11 n.s.
Mating system 16 n.s.
Movement 15 negative 0.0074
Sedentary 17 n.s.
Habitat migration 10 n.s.
Altitudinal migration 17 negative 0.013
Geographical migration too few contrasts
Forest 12 n.s.
Forest edge 35 negative <0.001
Savannah 15 n.s.
Grassland 20 n.s.
Scrub 29 n.s.
Wetlands 9 n.s.
Other natural habitats 13 n.s.
Man altered habitat 35 n.s.
Agricultural 29 n.s.
Other man-made 10 n.s.
0–500 m 26 n.s.
500–1000 m 30 negative 0.043
1000–1500 m 30 n.s.
1500–2000 m 34 n.s.
Over 2000 m 28 negative 0.036
1984
13. equal branch lengths in our contrasts analysis, and branch lengths were not
estimated for subspecies, but molecular data would be needed to confirm this.
We also found a correlation between change in number of natural habitats
occupied and extinction risk. Habitat specialisation has been found to correlate
with extinction risk in some studies (e.g. Harcourt et al. 2002; Brashares 2003;
Norris and Harper 2004). In the binary predictors, movement, and specifically
altitudinal migration, was associated with a decrease in risk. This, again, might be
correlated with an increase in habitat breadth. Two of the altitudinal categories
were negatively associated with risk. One of these was the very high altitude band
(>2000 m), areas where presumably species may have rather small ranges but yet
are not under particular anthropogenic threat. Why species occupying the alti-
tudinal band 500–1000 m should be little prone to extinction is not readily
explicable but bodes well for galliforms in this foothill zone. A movement to-
wards occupancy of forest edge habitats, will presumably be reflected in a species’
tolerance of habitat alterations and fragmentations, and will also serve taxa well
in the face of anthropogenic habitat alteration (e.g. Watson et al. 2004).
So, some correlates of change in risk have been identified, however it remains
the case that this analysis is of contrasts. The data points are changes in the
variables at nodes in the phylogeny (which is itself an estimate). How is this
correlative information to be used by managers, legislators and fund raisers?
For these users it may be that a species list is preferable to the results of a test
of an evolutionary hypothesis. In our contrasts analysis there was no effect of
Table 2. (continued).
Species list by Tree node and classification rule Relative cost of false negative to false
positive
Equal 2-Fold 3-Fold 6-Fold
Tree 1 node 10
Rule: N natural habitats > 2 AND Island = 0 AND
M:F mass ratio > 1.69
Francolinus castaneicollis • •
Meleagris gallopavo • •
Tetrao parvirostris • •
Tree 2 node 6
Rule: N natural habitats > 2 AND Relative egg
mass > 0.36
Coturnix chinensis •
Megapodius bernsteinii •
Megapodius reinwardt •
Tree 2 node 10
Rule: N natural habitats > 2 AND Relative egg
mass < 0.36 AND M:F mass ratio > 1.74
Francolinus castaneicollis •
Meleagris gallopavo •
Tetrao parvirostris •
1988
14. body size, which, as Fisher and Owens (2004) point out is a fairly common
finding in these studies. It may be that body size is related to other factors
which make a decline longer to detect (Pimm 1988, cited in Fisher and Owens
2004) or perhaps body size would be important within certain contexts, which
this kind of linear analysis cannot detect without subdividing the data. Using a
more restricted phylogeny, Keane et al. (2005) identify some correlates of
extinction risk in a data set for the Phasianidae, a family within the galliformes,
but it is clear that, particularly when attempting to include human impact,
interactions become much more important. These are just the kinds of context-
dependence that a decision tree can pick out.
The decision tree analysis found that two of the variables which were sig-
nificant in the contrasts analysis, number of habitats occupied and number of
subspecies, were also informative in classification of species into extinction risk.
This does lend support to the value of these variables in predicting extinction
risk, but the decision tree shows how the importance of number of subspecies
depends on the value of the sum of natural habitats occupied. At all costs of
misclassification, number of natural habitats occupied, number of subspecies,
and primary occupancy of forest habitat were characteristic of false positives.
In addition to these variables, island occurrence, but only when number of
natural habitats occupied was greater than two, picks up a further three species.
Further illustration of the non-linear and context-dependent relationships
that decision trees can detect is shown by the way that classificatory variables
change as the costs of false negatives increases. Actually, this tree is quite robust
to the effect of increasing these costs, the core group of 38 species remains
throughout. But, at higher relative costs when number of natural habitats
occupied was greater than two, species with mass dimorphism greater than 1.69
are picked up but only if they do not occur on islands. As costs increase still
further relative egg mass becomes important and mass dimorphism re-enters the
selection rules below a certain threshold of relative egg mass.
The decision tree thus produces lists of species which share the characteristics
of threatened species but which are currently classified as unthreatened. The
particular combinations of traits which lead to this classification are also
identified, allowing some elucidation of underlying processes, but not restricted
to those produced from linear models. Species lists have been derived from
contrasts analyses (Jones et al. 2003). This is done by determining the model
which best explains the variation in contrasts in extinction risk, thus giving the
relationship between contrasts in dependent and independent variables while
controlling for phylogeny. This model is then re-applied to the raw species data
to find those outlier species whose predicted IUCN status is very different to
observed. However, this is several steps removed from the original data and is
still restricted to linear relationships, which may not be the best descriptors.
Overall, although traits analyses may provide some macroecological insights,
we suggest that given the urgent need to redress threats of extinction, a decision
tree approach with it’s freedom from assumptions of linearity and ease of
interpretation, may provide information of more immediate management value.
1989
15. References
American Ornithologists’ Union 1998. Check list of North American birds. American Ornitholo-
gists’ Union, Washington, DC.
Bell J.F. 1999. Tree-based methods. In: Fielding A.H. (ed.), Machine Learning Methods for
Ecological Applications. Kluwer Academic, Boston, MA and London, pp. 89–106.
Birdlife international 2000. Threatened Birds of the World. Lynx Ediciones and Birdlife Interna-
tional, Barcelona and Cambridge, UK.
Brashares J. 2003. Ecological, behavioural and life-history correlates of mammal extinctions in
West africa. Conserv. Biol. 17: 733–743.
Brom T.G. and Dekker R.W.R.J. 1992. Current studies on megapode phylogeny. In: Dekker
R.W.R.J and Jones D.N. (eds), Proceedings of the First International Megapode Symposium,
Christchurch, New Zealand, December 1990. Zoologische Verhandelingen 278: 7–17.
Brooks D.M. and Strahl S.D. 2000. Currasows, Guans and Chachalacas: Status, survey and
conservation Action Plan for Cracids 2000–2004. IUCN/SSC Cracid Specialist Group. IUCN,
Gland, Switzerland and Cambridge, UK.
Cramp S. and simmons K.E.L. 1980. Handbook of the Birds of Europe, the Middle East and North
Africa Vol. 2. Oxford University Press, Oxford, UK.
Dekker R.W.R.J., Fuller R.A. and Baker G.C. 2000. Megapodes: Status, Survey and Conservation
Action Plan 2000–2004. WPA/Birdlife/SSC Megapode Specialist Group. IUCN and World
Pheasant Association, Gland, Switzerland and Cambridge, UK and Reading, UK.
Delacour J. 1977. The pheasants of the world. 2nd edn. Spur Publications, Hindhead, UK.
del Hoyo J., Elliott A. and Sargatal J. 1994. Handbook of the Birds of the World. Barcelona, Lynx
Ediciones.
Dunning J.B.Jr. 1993. CRC Handbook of Avian Body Masses. CRC Press, London.
Felsenstein J. 1985. Phylogenies and the comparative method. Am. Nat. 125: 1–15.
Fisher D.O. and Owens I.P.F. 2004. The comparative method in conservation biology. Trend. Ecol.
Evol. 19(7): 391–398.
Fuller R.A., Carroll J.P. and McGowan P.J.K. 2000. Partridges, Quails, Francolins, Snowcocks
and Guineafowl: Status, Survey and Conservation Action Plan 2000–2004. WPA/Birdlife/SSC
Partridge, Quail and Francolin Specialist Group. IUCN and the World Pheasant Association,
Gland, Switzerland and Cambridge, UK, and Reading, UK.
Fuller R.A. and Garson P.J. 2000. Pheasants: Status, Survey and Conservation Action Plan 2000–
2004. WPA/Birdlife/SSC Pheasant Specialist Group. IUCN and the World Pheasant Associa-
tion, Gland, Switzerland and Cambridge, UK, and Reading, UK.
Garland T., Harvey P.H. and Ives A.R. 1992. Procedures for the analysis of comparative data
using phylogenetically independent contrasts. Systematic Biology 41 (1): 18–32.
Grafen A. 1989. The phylogenetic regression. Philos. Trans. R. Soc. Ser. B 326: 119–157.
Harcourt A.H., Coppeto S.A. and Parks S.A. 2002. Rarity,specialization and extinction in pri-
mates. J. Biogeogr. 29: 445–456.
Harvey P.H. and Pagel M.D. 1991. The Comparative Method in Evolutionary Biology. Oxford
University Press, Oxford.
Holdaway R.N. 1999. Introduced predators and avifaunal extinction in New Zealand. In: MacPhee
R.D.E. (eds), Extinctions in Near Time: Causes, Contexts and Consequences. Plenum Press, New
York, pp. 189–238.
HowmanK.C.R.1979.Pheasants:TheirBreedingandManagement.K.andR.Books,Edlington,UK.
Johnsgard P.A. 1983. The Grouse of the World. Croom Helm Ltd.
Johnsgard P.A. 1988. The Quails, Partridges and Francolins of the World. Oxford University Press,
Oxford, UK.
Johnsgard P.A. 1999. Pheasants of the World: Biology and Natural History, 2nd edn. Swan Hill
Press, Shrewsbury, UK.
Jones D.N., Dekker R.W.R.J. and Roselaar C.S. 1995. The Megapodes. Oxford University Press,
Oxford, UK.
1990
16. Jones K.E., Purvis A. and Gittleman J.L. 2003. Biological correlates of extinction risk in bats. Am.
Nat. 161: 601–614.
Jones M.J., Fielding A.H. and Sullivan M.S. 2005 (this issue). Analysing extinction risk in parrots
using decision trees. In: Sullivan M.S. and Marsden S.J. (eds), 2005. Extinction Risk: Predicting,
Assessing, Prioritising and Redressing. Biodiversity and Conservation.
Keane A., de L., Brooke M.D. and Mcgowan P.J.K. (in press). Correlates of extinction risk and
hunting pressure in gamebirds (Galliformes). Biol. Conserv. 126 (2): 216–233.
Kimball R., Braun E.L., Zwartjes P.W., Crowe T.M. and Ligon J.D. 1999. A molecular phylogeny
of the pheasants and partridges suggests that these lineages are not monophyletic. Mol. Phylo-
genet. Evol. 11(1): 38–54.
Lee D.C. and Marsden S.J. 2005 (this issue). Accumulation of knowledge and changes in Red List
classifications within the family Galliformes. In: Sullivan M.S. and Marsden S.J. (eds), 2005.
Extinction Risk: Predicting, Assessing, Prioritising and Redressing Biodiversity and Conservation.
Loh W.-Y. and Shih Y.-S. 1997. Split selection methods for classification trees. Stat. Sinica 7:
815–840.
Lucchini V., Hoglund J., Klaus S., Swenson J. and Randi E. 2001. Historical biogeography and a
mitochondrial DNA phylogeny of grouse and ptarmigan. Mol. Phylogenet. Evol. 20(1): 149–162.
Madge S. and Mcgowan P.J. 2002. Pheasants, Partridges and Grouse: A guide to the Pheasants,
Partridges, Quails, Grouse, Guineafowl, Buttonquails and Sandgrouse of the World. Christo-
pher Helm, London.
Norris K. and Harper N. 2004. Extinction processes in hot spots of avian biodiversity and the
targeting of pre-emptive conservation action. Proc. R. Soc, Lond. Ser. B 271: 123–130.
Pagel M. 1999. Inferring the historical patterns of biological evolution. Nature 401: 877–884.
Purvis A. and Rambaut A. 1995. Comparative analysis by independent contrasts (CAIC): an Apple
Macintosh application for analysing comparative data. CABIOS 11(3): 247–251.
Reed R.N. and Shine R. 2002. Lying in wait for extinction? Ecological correlates of conservation
status among Australian elapid snakes Conserv. Biol. 16: 451–461.
Ricklefs R.E. and Cox G.W. 1978. Stage of taxon cycle, habitat distribution, and population
density in the avifauna of the West Indies. Am. Nat. 112: 875–895.
Sibley C.L. and Ahlquist J.E. 1990. Phylogeny and Classification of Birds: A Study in Molecular
Evolution. Yale University Press, New Haven, USA.
Sibley C.L. and Monroe B.L.Jr. 1990. Distribution and Taxonomy of Birds of the World. Yale
University Press, New Have, USA.
Sibley C.L. and Monroe B.L.Jr. 1993. A Supplement to Distribution and Taxonomy of Birds of the
World. Yale University Press, New Haven, USA.
Storch I. 2000. Grouse: Status, Survey and Conservation Action Plan 2000–2004. WPA/Birdlife/
SSC Grouse Specialist Group. IUCN and the World Pheasant Association, Gland, Switzerland
and Cambridge, UK, and Reading, UK.
Sullivan M.S., Gilbert F., Rotheray G., Croasdale S. and Jones M.J. 2000. Comparative analyses of
correlates of red data book status: a case study using European hoverflies (Diptera: Syrphidae).
Anim. Conserv. 3: 91–95.
Urban E.K., Fry C.H. and Keith S. 1986. The birds of Africa. Vol. 2. Academic Press, London.
Watson J.E.M., Whittaker R.J. and Dawson T.P. 2004. Habitat structure and proximity to forest
edge affect the abundance and distribution of forest-dependent birds in tropical coastal forests of
southeastern Madagascar. Biol. Conserv. 120(3): 311–327.
Young J.R., Braun C.E., Oyler-McCance S.J., Hupp J.W. and Quinn T.W. 2000. A new species of
Sage-Grouse (Phasianidae: Centrocercus) from south-western Colorado. Wilson Bull. 112(4):
445–453.
1991