Habitat loss, fragmentation and degradation effects on small mammals: Analysis with conditional inference tree statistical modelling

Biological Conservation 176 (2014) 80–98

Contents lists available at ScienceDirect

Biological Conservation journal homepage: www.elsevier.com/locate/biocon

Habitat loss, fragmentation and degradation effects on small mammals: Analysis with conditional inference tree statistical modelling Christopher P. Johnstone ⇑, Alan Lill, Richard D. Reina School of Biological Sciences, Monash University, Clayton Campus, Victoria 3800, Australia

a r t i c l e

i n f o

Article history: Received 25 October 2013 Received in revised form 24 April 2014 Accepted 27 April 2014

Keywords: Antechinus Agilis Swainsonii Fuscipes Habitat loss Relative abundance Performance indices Conditional inference tree Random forests

a b s t r a c t Anthropogenic habitat loss, fragmentation and degradation often co-occur in a landscape and their relative influence on a native animals’ health and survival can be difficult to determine. We examined the influence of these environmental variables on the estimated relative abundance of some small mammal species in a large area (2500 km2) of southeastern Australia. Using the agile antechinus (Antechinus agilis) as a model, we also examined the association between these variables and three population performance indices, mass-size residuals (MSR; indexing fat reserves), the neutrophil/lymphocyte ratio (N:L; indexing physiological stress) and red blood cell counts (RBC; indexing regenerative anaemia). Study sites were in either highly disturbed and fragmented, or relatively undisturbed, continuous Eucalyptus forest. We generated conditional inference tree statistical models to identify the relative importance of up to 49 ecological variables in explaining variation in small mammal abundance and performance indices. Habitat loss was important in explaining small mammal abundance, as were the abundances of the same species in neighbouring study sites. The models also suggested that the habitat area required to support a ‘healthy’ population was greater in the larger species examined. Autocovariates of neighbouring site same-species abundances and habitat fragmentation were the next most important influences on small mammal relative abundance, implying that metapopulations may be important for population persistence, especially in bush rats (Rattus fuscipes). Habitat degradation, reflected in structural and floristic features, was less important, but explained some variance in relative abundances. For agile antechinus populations, time of year, degree of forest fragmentation and extent of native tree cover were important in explaining performance indices. Results indicated that habitat reduction per se was a significant threatening process for small mammals. Habitat loss requires at least the same research attention as that currently devoted to anthropogenic habitat fragmentation and degradation. Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction Studies of vertebrate conservation are largely concerned with the effects of anthropogenic habitat loss, fragmentation and degradation on native biota. Substantial evidence indicates that anthropogenic habitat fragmentation negatively affects terrestrial vertebrate assemblages and populations (Andrén, 1994). Fewer studies have examined the effects of anthropogenic habitat degradation, and habitat loss per se has received the least research attention (Fazey et al., 2005), despite a general consensus that it is probably the world’s leading cause of native species’ decline (Fahrig, 1997; Foley et al., 2005). One underlying difficulty is that habitat loss, fragmentation and degradation often co-occur in a

⇑ Corresponding author. Tel.: +61 0 401288692. E-mail addresses: [email protected] (C.P. Johnstone), alan. [email protected] (A. Lill), [email protected] (R.D. Reina). http://dx.doi.org/10.1016/j.biocon.2014.04.025 0006-3207/Ó 2014 Elsevier Ltd. All rights reserved.

landscape, and thus their independent effects can be difficult to isolate (Fischer and Lindenmayer, 2007). Vertebrate conservation studies in anthropogenically-disturbed landscapes are typically concerned with comparing a population response variable (e.g. site occupancy or abundance) or performance indices (such as brood size or level of physiological stress) (Fletcher et al., 2007) with multiple environmental variables in order to identify possible relationships. Survivorship and reproduction are products of complex interactions of an animal’s genome, behaviour, physiology and autecology. Demographic studies can be informative about whether a population is declining or at risk, but to understand why a population is declining functional studies must be undertaken and they must involve a whole-organism consideration of behaviour, physiology and genetics. This functional approach has been advocated and discussed by several authors (Homyack, 2010; Janin et al., 2011). Such performance measurements can be useful as conservation tools. For example, because

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

physiological stress can, in some situations, deterministically reduce reproductive output and survivorship, elevated physiological stress must always be considered at least a potential early warning sign, even if populations appear to be stable. Stress measurements can sometimes be used to index features of autecology that might not be easy to measure, such as parasite loads (Martínez-de La Puente et al., 2011) or food availability (Herring et al., 2011). However, these performance-ecological relationships may well be species-specific (Johnstone et al., 2012b) and for them to be useful more work is needed to ascertain if there are generalisable relationships or if each endangered vertebrate species would need to be studied independently. The tendency in small mammal research has been to examine metrics of population distribution or demography, whilst not considering other features of the animal’s relationship with its environment. Behavioural studies of small mammals are sometimes conducted (e.g. Banks and Dickman, 2000; Cockburn and Lazenby-Cohen, 1992; Dickman, 1986) that are informative about a species’ use of environment and its relationship with other species. However, physiological ecology is perhaps less well studied in mammals than is behaviour, and certainly much less focus has been given to mammal ecophysiology than bird ecophysiology. Here, for one of the small mammal species studied, we examined (a) indices of physiological stress, which may be both indicative of, and contributing to, decline of free-living vertebrates in degraded or fragmented habitats (Johnstone et al., 2012a; Martínez-Mota et al., 2007; Suorsa et al., 2004); and (b) body condition indices, which can be informative about metabolic reserves in individual animals (Peig and Green, 2009; Schulte-Hostedde et al., 2005). Experimental field manipulations are useful for testing potential animal-environment relationships (Mac Nally and Horrocks, 2002), but are not always feasible because of cost or ethical considerations (studying habitat loss through large-scale experimental habitat removal would be contentious, to say the least; (Diamond, 1983). Consequently, most large-scale studies use a naturally-occurring experimental design (a natural experiment sensu Diamond, 1986). However, there may be multiple, correlated, environmental factors that influence a population, and factors may interact, have synergistic effects or partially negate one another (Laurance and Cochrane, 2001). The variables measured or indexed may be continuous, ordinal or nominal (or a mixture of these) and data may be non-linear or non-normally distributed. Linear regression approaches must be modified using non-Gaussian distributions and/or non-linear equations in order to characterise these sorts of data (Zuur et al., 2007), whereas conditional inference tree models can represent non-linear relationships with relative ease. For example, a conditional inference tree will model a U-shaped or J-shaped curve if such a relationship exists, whereas such curves are impossible to model using standard general linear or additive models and require careful use of non-linear link functions for generalised linear or additive models. A conditional inference tree could furthermore easily represent a complex wave-like relationship with multiple peaks that would not easily be represented by a mathematical equation. Such relationships are either absolutely outside the realm of general or generalised linear/additive models or so complex as to be effectively so. A further advantage of conditional inference tree model approaches is that they are the first step towards a random forests analysis. Random forests analysis is a powerful, predictive, model-averaging approach, where random bootstraped samples of predictor variables are used to general a ‘forest’ of models, and from this forest the relative importance of predictor variables can be calculated. Random forests analyses are being increasingly applied when exploring complex relationships in ecology (Cutler et al., 2007; Prasad et al., 2006) and genetics (Bureau et al., 2005). Random forests approaches tend to outperform other modelling techniques for predicting known

81

relationships. Lawler et al. (2006) found that a random forests approach consistently outperformed generalised linear models, generalised additive models, artificial neural networks and genetic algorithms for rule-set prediction for predicting actual species presence or absence. The authors further reported that a random forests approach allowed for better prediction of species presence or absence than just relying on the single best conditional inference tree model. Conditional inference tree modelling is an intuitive, easily implemented and interpreted statistical method that copes well with complex data, but it is underused in ecology. It is a tool for examining the relationship between a single response variable and multiple potentially explanatory variables (Quinn and Keough, 2002; Zuur et al., 2007). Such models are popular in medical and genetic research, probably because they tend to be better at predicting known relationships from data than more commonly used methods, such as logistic regression (Nagy et al., 2010). The models produced are predictive and robust to non-linearity, nonnormality, multicolinearity and multiple interactions among explanatory variables (Quinn and Keough, 2002; Zuur et al., 2007). From a conservation management perspective, conditional inference tree models are useful because they generate decision trees. By using conditional inference tree models, clear decision paths can be used to determine, for example, how much habitat in a given area would correlate with a given mean response variable, be this species richness, occupancy, abundance or performance metrics. We used conditional inference tree models to investigate the relative roles of habitat loss, fragmentation and degradation and other environmental variables in determining the relative abundance or performance indices of three native Australian small mammal species common to modified forests in south-eastern Australia.

2. Materials and methods 2.1. Defining habitat fragmentation, degradation and loss Here we define habitat loss as the complete removal of native canopy cover. In the area studied, habitat loss was usually the result of agriculture, and in particular the creation of open fields for grazing domestic stock. The native small mammals studied are not thought to make extensive use of open fields, and there are no reports of these species being caught at high numbers in open fields. It remains possible of course that individuals may make exploratory incursions into fields or move across fields between habitat patches, but once tree-cover is removed the landscape no longer has the foraging patches or nest sites that would be required for persistance of a population. Habitat fragmentation was defined as any separation of habitat (contiguous native tree-cover) by matrix where the separation distance was >20 m from the edge of the canopy (not tree-trunks) to the nearest canopy. A researcher walked around the perimeter of all fragments used in the study (i.e. those that were <300 ha in size) and measured canopy gaps. Potential corridors were also noted and recorded, but they were discarded from the analysis, as indices derived from number of corridors per fragment and apparent corridor quality did not help to explain variation in any of the small mammal population metrics measured. Habitat degradation is harder to define than habitat loss or fragmentation. The reason is that in order to know precisely what constitutes degradation of habitat for a given species, all aspects of the species’ ecology must be known. For the species in this study, we considered degradation to comprise windfall at forest edges, livestock grazing in forest, firewood collection, recreational trail-bike riding and invasion of a forest by non-native plants and animals.

82

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

For the latter, we considered populations of the introduced European red fox (Vulpes vulpes) and feral domestic cat (Felis catus) to be potential factors rendering habitat less suitable for the small mammals studied. To measure degradation of vegetation and habitat structure we counted and identified plant species in two 20  20 m quadrats per site. These floristic and habitat quadrats were randomly placed within trapping grids at sites and there was one quadrat per trapping grid (see Section 2.4) In each quadrat we counted logs and measured stumps and trees and recorded the number of living or dead trees. Non-indigenous shrubs or trees were found in only a few sites (<5) and never at high densities. Measures of weediness were therefore excluded. We did not average values for the two quadrats by site, because they were arranged so as to capture the conditions at the edge and interior of fragments; rather than use a less refined factor (fragment edge, fragment interior, continuous forest), we considered it better to try to identify whether differences in vegetation and habitat complexity could be identified and then examine whether they related to positions near or far away from forest edges. Invasive predator activity was more difficult to establish, and with hindsight we would now advise the use of camera traps to monitor predator activity. However, we recorded the number of live-traps that had been disturbed by a predator overnight (signs of canine tooth marks and moved >5 m from original location) and used this as an index of predator ‘activity’, although we acknowledge this approach is not ideal.

and Dickman, 1998) that forage in leaf litter and on vertical tree trunks (Dickman, 1988). They nest communally in tree-hollows in groups of up to 20 individuals (Cockburn and Lazenby-Cohen, 1992). Prior to 1998, the species was included in the brown antechinus (A. stuartii) species-complex (Dickman et al., 1998). Insectivorous dusky antechinus are largely terrestrial and tend to dig for prey in leaf litter and nest individually or in small family units in burrows (Cockburn and Lazenby-Cohen, 1992; Dickman, 1988). Bush rats (50–225 g) nest socially in burrows and typically are herbivorous, but they will opportunistically eat invertebrates. The native broad-toothed rat (Mastacomys fuscus) and two exotic rodents, the black rat (Rattus rattus) and domestic mouse (Mus musculus), were also captured, but their abundance estimates are not analysed here because they were low in forest fragments and zero in continuous forest.

2.2. Study area and species

2.4. Live-trapping protocol

In 2007 and 2008, we undertook a large-scale (study area 2500 km2, approximate centre at 38°370 S, 146°100 E) (Fig. 1), comparative study using live-trapping to examine effects of habitat fragmentation on populations of the agile antechinus (Antechinus agilis), a small, insectivorous marsupial native to south-eastern Australia. Recently, authors have argued strongly that population density and distribution measures alone do not provide a full understanding of population wellbeing (Davis et al., 2008; Fletcher et al., 2007; Homyack, 2010), and so we also measured performance metrics (estimates of metabolic reserves, physiological stress and regenerative anaemia). Concurrent to the study on agile antechinus, we recorded the numbers of other native, small mammals trapped as by-catch and some local environmental variables. The study area has one of the longest histories of native vegetation clearing for agriculture and mining in mainland Australia. The Gippsland Company was founded in 1840 and gold was discovered in the area in 1853. Extensive forest clearing was triggered by an 1865 Lands Act, which made settlement easier. The European population of Gippsland in 1854 was 1965 and 2 years later in 1856 it was 3634, and continued to grow (Daley, 1962). Consequently, the study landscape can provide insights into how comparable, but more recently cleared, areas are likely to change with respect to their mammal fauna. The dominant canopy trees at all 60 study sites were Eucalyptus species, primarily E. obliqua, E. radiata and E. regnans, and habitat similarity was achieved for comparative purposes by restricting study sites to forest composed of the three Ecological Vegetation Classes (EVC) ‘Lowland Forest’, ‘Wet or Damp Forest (Wet)’ and ‘Wet or Damp Forest (Damp)’ (Davies et al., 2002). Live-trapping was conducted from April to August, 2007 and from March to August, 2008. In these years, mean monthly rainfall was 73 mm and mean monthly maximum and minimum mean ambient temperatures were 17.7 °C and 7.8 °C, respectively (Australian Bureau of Meteorology, 2009). The two small marsupials (family Dasyuridae) studied were the agile (16–44 g) and the dusky antechinus (Antechinus swainsonii) (38–170 g). Agile antechinus are scansorial insectivores (Sumner

Trapping was conducted using Elliott live-traps (Elliott ScientificTM, Australia) insulated with two plastic bags fitted over the non-door end of the trap and baited with rolled oats, peanut butter, water and vanillin and containing insulating bedding. It was timed to occur between two major A. agilis life-history annual events, male-biased dispersal (January–February) and the synchronised breeding rut (mid-August–September). In 2007 and 2008, we conducted 3600 trapping-nights (nights  number of traps set) in 60 study sites (Fig. 1), trapping at 30 sites in each year. Traps were set <3 h before dusk and animals released <3 h after dawn. Trapping was conducted on three nights at each site and two trapping grids were used per site. In forest fragments, one grid was placed <60 m from the edge and one >80 m from the edge. In fragments <10 ha the grids were sometimes only 20 m apart, but fragments ranged from 5 to 300 ha, so that in large fragments the grids were as much as 500 m apart. The spatial pattern was replicated as closely as possible in the paired pseudofragment using GPS. The arrangement of traps was designed to examine edge effects, but these have already been reported elsewhere (Johnstone et al., 2011). Here, we are more interested in whether the floristic and habitat variables measured for each trapping grid show relationships with the small mammal population condition indices being examined. Although trap grids in the same fragment are not independent, we think that there was enough floristic variation among grids that averaging them would result in a loss of information. Thus, we have used a statistical approach to address this non-independence (see Section 2.11). Live-trapping was conducted for three consecutive nights at each site to reduce the risk of trapping deaths. It is generally thought that when small mammals are recaptured over successive nights the capture stress increases risk of trapping death with each additional consecutive night of capture. Trapping alternated between sites that were deemed highly disturbed/fragmented and relatively undisturbed/continuous (i.e. trapping of paired disturbed and undisturbed sites was distributed evenly from March to August in each study year). In disturbed sites, trapping grids comprised three rows of seven traps spaced

2.3. Study sites Forest fragments studied were <30 ha, 30–60 ha and 60–300 ha in area (n = 10 in each case). All fragments had at least 100 m of cleared agricultural matrix separating them from other forest fragments, and all fragment sites were >5 km from the nearest area of continuous forest. Continuous forest was defined as areas of >1000 ha with total, continuous tree cover. Forest fragments were paired with ‘pseudofragments’ i.e. areas of similar size and shape (Mac Nally and Bennett, 1997) in continuous forest for comparison.

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

83

Fig. 1. Study region in south-east Australia. White = cleared, agricultural land. Darker shaded areas = tree cover (includes native re-growth, old growth forest and native plantations). Approximate locations of study sites are indicated by white boxes (h). Map based on Victorian Department of Sustainability and Environment (DSE) interactive ‘Forest-Explorer Online’ maps (http://www.dse.vic.gov.au/). Bottom image = satellite image of study area from Google Earth (http://www.google.com/earth/index.html).

evenly in a 40  120 m area. Capture rates in continuous/undisturbed sites were unexpectedly high and trapping effort was therefore reduced for welfare reasons, so that grids comprised three rows of three traps spaced evenly in a 40  120 m area. 2.5. Relative abundance index We estimated relative abundance from only the first night of trapping to avoid possible confounding effects of learning over a three-day period. After adjusting the data to account for sprung traps (Beauvais and Buskirk, 1999), we employed a density-frequency transformation of the trap success data (Caughley, 1977) to construct estimated relative abundances. Each sprung trap, including traps that contained by-catch, was considered the equivalent of a half-night’s trapping effort, as we did not know when in the night it was sprung (Beauvais and Buskirk, 1999). The densityfrequency transformation was used because of different trap numbers in the fragmented and continuous habitats, as this can cause biased results if only a simple ratio of captures per trap is used (Atchley et al., 1976). As traps are progressively filled they cannot capture additional animals, and the Caughley density-frequency transformation attempts to estimate how many animals would have been caught per trap if traps could capture multiple animals. Thus we deemed estimated relative abundance (RA) to be density of captures per night’s trapping effort, which corresponded to

1  LN (1  (captures/full trap nights + half trap nights)). The resultant values were square root transformed to achieve normality. The assumptions of the Caughley density-frequency transformation (equal catchability of individuals, no temporal or environmental effects of catchability) are not likely to be strictly met for any small mammal species (i.e. some individuals will always be more likely to be trapped than others because of differences in boldness), so this index must be considered only a rough estimate of actual abundances. For this reason we have used the term ‘estimated relative abundance’ throughout to make it clear that this is an estimate of actual abundance only. Near trap saturation in some continuous forest sites, possibly as a result of the reduced trapping effort, may have led to underestimation of RA, the effect of which would have been to increase Type II error with respect to the influence of habitat loss, fragmentation and degradation on RA. This was unavoidable, but we considered that this underestimation was acceptable, although obviously not ideal. 2.6. Agile antechinus’ body condition and stress indices Captured A. agilis were visually sexed and a <1 mm disc of ear pinna tissue was removed to facilitate identification on recapture to avoid resampling. We took a blood sample (approx 100 lL) and measured body mass (±0.5 g) and morphometrics (±0.1 mm)

84

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

to gauge hypothalamus–pituitary–adrenal axis (HPA)-mediated stress and body condition, respectively (Davis et al., 2008; Schulte-Hostedde et al., 2005). Morphometrics measured were distances from each eye to the nose tip, nose-vent length, vent-tail length, both front- and both hind-paw lengths. Only nose-vent length was used to construct BCIs. Typically, the first two A. agilis captured on a grid in a trapping night that were not recaptures were sampled (i.e. we obtained 4 samples per day and 12 over the three days of trapping at each site). Blood sampling was conducted within 15 min of removal of the animal from the trap. Blood was withdrawn by capillarity into a heparinised microhaematocrit tube after puncturing one of the two lateral veins at the base of the tail with a 27 gauge needle. The blood withdrawn from each animal was used for measuring erythrocyte and leukocyte values. About 5–10 lL was used to prepare a blood smear (by the pullwedge method) for estimating the neutrophil-to-lymphocyte ratio (N:L), after staining with May–Grünwald–Geimsa stain (Lewis et al., 2006). N:L is considered an accurate index of HPA axis-mediated chronic stress in vertebrates (Davis et al., 2008). Cell counts were all conducted at 400 magnification by the same researcher, who counted >200 cells by visually sweeping the smear from end to end P5 times, but avoiding the preparation’s edges (Lewis et al., 2006). Leukocyte trafficking due to handling stress could conceivably have confounded the estimation of baseline N:L (Dhabhar et al., 1994), but a validation trial suggested that this did not happen (Johnstone et al., 2012a). Trapping stress was more difficult to assess, because establishing baseline N:L requires instant-killing traps (Fletcher and Boonstra, 2006) and killing agile antechinus in the numbers needed would have been logistically difficult and ethically contentious. We present elsewhere a detailed discussion of trapping stress (Johnstone et al., 2012a) and we think that the evidence in our study supports a best interpretation in which N:L was a positive index of environmental stress. Animals in all sites would have been in traps for approximately the same time on average (up to 15 h), so that observed differences in the stress index should be due to environmental differences rather than variation in trapping stress. However, we acknowledge that trapping stress effects on small mammals are poorly understood (Fletcher and Boonstra, 2006) and thus our N:L interpretation must be viewed as tentative. Indirect evidence for elevated HPA-axis mediated stress can be obtained from erythrocyte metrics through the use of regenerative anaemia profiles (Johnstone et al., 2012b). Regenerative anaemia should not alter quickly enough to be confounded by trapping (Johnstone et al., 2012b), and in marsupials there is convincing evidence that a red blood cell count (RBC) may be a straightforward method for identifying it (see Barnett et al., 1979; Johnstone et al., 2012b). Thus we also used part of each blood sample to conduct standard RBC counts (400 magnification with light microscopy). We calculated a body condition index (BCI) using mass-size residuals (MSR) (Schulte-Hostedde et al., 2005) derived from body mass and the nose-to-vent length for each animal. This index accurately estimates fat stores in some small mammals (SchulteHostedde et al., 2005). There is some dispute about the best method of generating residuals of mass from skeletal measurements, with some authors preferring regression methods other than ordinary least squares (Green, 2001; Peig and Green, 2009). However, we used an ordinary least squares approach because it was simple to calculate and provided residuals in an easy-to-comprehend unit (i.e. g above or below expected mass), although we acknowledge that some loss of accuracy is probably associated with this method. 2.7. Potential explanatory variables Recently conservation biologists have favoured landscape configuration and evaluation tools that are more sophisticated than

those available to most conservation managers or concerned landowners. For this reason, we opted to characterise large-scale landscape variables (e.g. forest area or fragment configuration) using methods that required only access to freely available online maps and free image manipulation and spreadsheet software. This means that the statistical conditional inference tree models presented here should be reasonably easy to use as decision trees for conservation management of a site within the study area. After reviewing previously published work on the current study species or closely-related species (Bennett, 1990a,b, 1993; Dunstan and Fox, 1996; Knight and Fox, 2000), 36 environmental variables that might influence the captured mammals’ relative abundance, condition and chronic stress levels were chosen and measured (Table 1). Local environmental variables were measured in a 20  20 m quadrat positioned randomly in each trapping grid (i.e. there were two measurements per site, unless otherwise stated). We used Simpson’s dominance (Shrub dom.) and Wilson’s evenness (Shrub even.) indices derived from species counts of shrubs to characterise differences in shrub assemblages (Smith and Wilson, 1996). We also used a single-parameter shape descriptor of the distribution of tree diameters at breast height (DBH) in each site. This was derived from a Weibull distribution (Baker et al., 2005; Horner et al., 2010) and can be used to estimate whether a stand of trees represents recent, medium or old-growth forest (Baker et al., 2005). The extent of native tree cover and degree of forest fragmentation were estimated from online native vegetation maps (1:75,000) (Victorian Department of Sustainability and Environment: ‘Forest Explorer Online’ www.dse.vic.gov.au). Using ImageJ, a known area was measured in pixels (using the ‘Lasso’ tool, the ‘Analyse’ menu and ‘Measure’), so that the hectare:pixel ratio could be calculated for maps at 1:75,000 scale (http://rsbweb.nih.gov/ij/). Then the area of native vegetation was measured in pixels cumulatively at 0.5, 1, 2, 3, 4 and 5 km along a radius from each site. Distances were established by using the scale provided in the ‘Forest Explorer Online’ maps. Areas in pixels were then converted to areas in hectares. A similar approach was used to count fragments. Any area of forest separated from other forest by >20 m of cleared matrix was considered a fragment. The cumulative number of fragments was then counted at the same kilometre distances as given above (i.e. in completely continuous forest, the values would be 1 at all distances for ‘fragmentation’). In a configuration of three fragments where the central fragment only occupied the 0.5 km radius circle, a second fragment spanned the 2–4 km circles and the third fragment spanned the 4–5 km circle, the cumulative fragments would be scored for <0.5 km = 1; <1 km = 1; <2 km = 1; <3 km = 2; <4 km = 2; and <5 km = 3. 2.8. Data analysis Statistical analysis was conducted in R (2.15.1, R Core Development Team 2010), using the package ‘party’. Response variables were averaged by trapping grid and, for A. agilis, by sex. Averaging ratios should be avoided (Atchley et al., 1976) and so N:L site means were derived from differential neutrophil and lymphocyte count means. The distributions of these values were checked for normality and homoscedasticity, and relative abundance (RA) was square root and N:L log10 transformed to achieve the former distribution. The object of constructing a tree-branched model is to recursively split the data and produce a ‘tree’ of sub-populations and their associated ‘risk factors’ (Quinn and Keough, 2002; Zuur et al., 2007) (here we use ‘tree-branched’ model to denote the whole family of models including conditional inference trees, regression trees, etc.). The model produced resembles an inverted tree, with the first node being the tree’s root. Each observation

85

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

Table 1 Explanatory variables used in the study. The most likely terrestrial predator to have investigated traps was the introduced European red fox (Vulpes vulpes) (Menkhorst and Knight, 2004). Shrubs deemed likely to be unpalatable to browsers were Acacia verticillata and Coprosma quadrifida. Large trees (circumference at breast height P60) are an important resource for tree hollow-nesting species in the study area (Cockburn and Lazenby-Cohen, 1992). The abbreviations for analyses are Fragmentation (F), Physiological Indices (P [only for agile antechinus’ N:L, RBC and BCI]) and Relative abundance (RA). Parameter

Definition

Measurement

Analyses

% Dead ABH

Percentage of tree area at breast height that was dead

F/P/RA

% Non-Eucalyptus ABH Altitude Autocov

Percentage of tree area at breast height that was not genus Eucalyptus

400 m2 quadrats 400 m2 quadrats Trap grid/GPS Covariate 400 m2 quadrats 400 m2 quadrats Vegetation map

F/P/RA

Trap grid Vegetation map 400 m2 quadrats Trap grid/GPS 400 m2 quadrats 400 m2 quadrats Trap grid/GPS Covariate Trap grid

P/RA P/RA

Trap grid Trap grid Trap grid Trap grid Trap grid Trap grid Trap grid 400 m2 quadrats 400 m2 quadrats 400 m2 quadrats 400 m2 quadrats 400 m2 quadrats 400 m2 quadrats Vegetation map

P P P P P P/RA P F/P/RA

Covariate Covariate

F/P/RA F/P/RA

Browse index

Height above sea level of trap grid Autocovariate of the response variable (to test for spatial autocorrelation) Calculated at 1, 2, 3, 4, 5, 10, 20 and 30 km radius of sites. Each radius is a separate variable in this analysis. Percentage of shrubs not palatable to browsers (thorned/prickly)

DBH median

Median tree diameter at breast height

Frag

Cumulative count of native forest fragments (any patch with >20 m cleared matrix separation)

Gully Habitat

Calculated at 0.5, 1, 2, 3, 4 and 5 km radius of sites. Each radius is a separate variable in this analysis. Was trap grid in a gully? (Y/N) Was site classified as fragment or continuous at outset? (F < 300 ha/C>1000 ha contiguous)

Large tree dens.

Density of trees with circumference >300 cm

Latitude Leaf litter

Latitude of site (digital) Leaf litter depth and extent (scored 0–5, 0 = none/5 = 100% cover and >5 cm thick)

Log dens.

Density of woody debris with circumference >10 cm

Longitude Month Predator index RA Agilis RA Females RA Fuscipes RA Males RA Swainsonii Ridge Sex Shrub dens.

Longitude of trap grid Month of year (March = 3, April = 4, May = 5, etc.) Percentage of traps showing signs of mammal predator disturbance (tooth marks on plastic bag and moved >5 m overnight) Relative abundance of agile antechinus Relative abundance of female agile antechinus Relative abundance of bush rats Relative abundance of male agile antechinus Relative abundance of dusky antechinus Was trap grid on a ridge? (Y/N) Sex of agile antechinus Density of shrubs (shrubs were <10 cm circumference and 60–200 cm height)

Shrub dom.

Simpson’s species dominance index for shrubs

Shrub even.

Wilson’s species evenness index for shrubs

Shrub rich.

Species richness of shrubs

Tree dens.

Density of trees (trees were >10 cm circumference or >2 m height)

Tree rich.

Species richness of trees

Tree cover

Cumulative extent of native tree cover (ha)

Weibull scale Weibull shape

Calculated at 0.5, 1, 2, 3, 4 and 5 km radius of sites. Each radius is a separate variable in this analysis Descriptor of scale of Weibull distribution of tree DBHs Descriptor of shape of Weibull distribution of tree DBHs

included in our models was either averaged by trapping grid (to test for edge effects) and sex (A. agilis) or by trapping grid only. All possible binary splits of the response variable were assessed for each potential explanatory variable. The aim of splitting the data at each step was to establish groups that had a between-variation as large, and within-variations as small, as possible. The advantages of this approach over additive or general linear modelling are that: (1) the number of explanatory variables that can be included is unlimited; (2) interactions among terms can be easily identified and intuitively visualised; (3) the method is not invalidated by multicolinearity of explanatory variables; and (4) there is no requirement for linearity and normality in explanatory variables (Quinn and Keough, 2002; Zuur et al., 2007). However, traditional tree-branched models (classification and regression trees, CARTs) have the serious drawback that splitting is biased

F/P/RA P/RA P/RA

F/P/RA P/RA

F/P/RA P/RA F/P/RA F/P/RA P/RA P/RA P/RA

F/P/RA F/P/RA F/P/RA F/P/RA F/P/RA P/RA

in favour of explanatory variables in which more splitting is possible (Quinn and Keough, 2002; Zuur et al., 2007). Moreover, they can easily be over-fitted and require somewhat subjective ‘pruning’ methods (Quinn and Keough, 2002). We therefore used the more recently developed Conditional Inference Tree (CIT) (‘party’ in ‘R’) (Hothorn et al., 2006a,b). It is similar to a CART in that it is a form of binary recursive partitioning, but it uses a machinelearning algorithm embedded in a conditional inference framework (Hothorn et al., 2006a) i.e. whereas CARTs continue splitting until no further splits are possible, a CIT uses a statistically-determined stopping criterion, an a priori P value, to determine where splitting is no longer valid. Typically, splitting is accepted at P < 0.1 (e.g. Nagy et al., 2010). The P value represents the deviation from the partial null hypothesis, and is generated by allocating weights to nodes (explained in more detail in Hothorn et al.,

86

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

2006a). In CARTs, cross-validation is required to ‘prune’ trees, but this is not required when using CITs because the use of an a priori P value limits the model’s branching and growth (Hothorn et al., 2006b). Conditional inference trees are not affected by over-fitting and are unbiased with regard to the types of explanatory variables used (Hothorn et al., 2006a; Strobl et al., 2007). A further point worth addressing is that although conditional inference trees are robust to co-linearity in that correlated predictors can be included without invalidating the final model, only the best predictor is chosen; thus potentially a variable might be picked that is acting as a proxy or ‘average’ for several other variables that are mechanistically driving the response variable. That said, random forests analyses overcomes this by providing a model that includes all variables that are contributing to explaining variation in the response ranked in order of importance, so that any ‘proxy’ variable and the mechanistic drivers should all appear high in the importance ranks. Experimental manipulations to shed light on actual cause and effect could then be undertaken. Tree-branched models are robust to autocorrelation of explanatory variables (Zuur et al., 2007), but spatial autocorrelation of the response variable may inflate Type I error, at least in CARTs (Segurado et al., 2006). Using Moran’s I (Legendre, 1993) we found that A. agilis and A. swainsonii relative abundances were spatially autocorrelated (P < 0.05). As there is no standard method for handling spatial autocorrelation in tree-branched models, we provide a detailed discussion and solution below (Section 2.10). ‘Random forests’ is a tree-branched modelling method used for establishing the relative importance of multiple explanatory variables with respect to a single response variable. This iterative method selects a random subset of explanatory variables and constructs a tree from them. This step is repeated until a ‘forest’ consisting of a pre-determined number of ‘trees’ is constructed. Using the ‘ctree’ method in ‘party’, the relative importance of each explanatory variable was calculated from the number of times that it was used to construct a tree. Variables are included in trees using an a priori conditional inference framework, so not all variables examined are used in a tree, i.e. those that do not satisfy P < 0.1 are omitted. This is a different approach to that used in a CART random forest method, in which relative importance is derived from the variance explained in the out-of-bag sample for each tree. Using CITs, more variables require more trees to achieve ‘stability’. We used 4000 trees with the ‘ctree’ function ‘replace’ set to ‘false’ to avoid bias (Strobl et al., 2007). We report non-parametric Kendall tau (s) correlations where appropriate to provide an indication of the strength and direction of non-parametric relationships (Arndt et al., 1999). The s correlations were generated using the function ‘cor.test’.

site-level vegetation. Analysis was conducted at the level of floristic quadrats, so that n = 120. Floristic quadrats were chosen as the unit of replication rather than averaging the habitat and vegetation indices and measurements by site, because quadrats were up to 300 m apart and sometimes had different aspects or positions on a slope. This means that sometimes quadrats had quite distinct vegetation species, and averaging the data measured from grids at a site level would have resulted in an unacceptable loss of information.

Although conditional inference tree models are robust to nonnormal distributions or multicolinearity of predictors, they are vulnerable to spatial autocorrelation, which will increase Type I error (Czúcz et al., 2011). We follow the method of (Czúcz et al., 2011) to test for spatial autocorrelation of classification trees, modified here for use with regression tree models. Using trapping grids as ‘sites’ (i.e. n = 120), we calculated residuals of fitted values (from ‘treeresponse’ in ‘party’) and response variables by using the response from the fitted values (fitted values  relative abundances). We then used a Mantel test (‘mantel’ in ‘ecodist’ with nperm = 10,000) to compare a distance matrix of geographic locations (from latitude and longitude) with the residuals of fitted and response variables. Mantel’s r, upper and lower confidence limits, one-tail (Ho: r 6 0 and r P 0), and two tail (Ho: r = 0) P values were tested. If significant autocorrelation is found, a Mantel correlogram (‘mgram’ in ‘ecodist’) can be used to identify the lag distance at which spatial autocorrelation ceases to be significant. Where significant spatial autocorrelation occurs, sites can be discarded at random within the lag distance (i.e. if spatial autocorrelation occurred in sites <1000 m apart, then sites can be discarded randomly so that no sites are closer than 1000 m apart). However, in our dataset patterns of spatial autocorrelation were complex, and autocorrelation could be non-significant at smaller lag distances (e.g. <5 or 10 km) and then become a significant factor, only to disappear again. Where complex spatial autocorrelation occurs, it is more interesting to study the autocorrelation rather than try to correct it out of a biological model. We generated an autocovariate for each response variable (Dormann et al., 2007) using ‘autocov_dist’ in ‘spdep’ in R (using digital longitude and latitude). As conditional inference tree models cannot be over-fitted, this allows for generating multiple spatial autocovariates and testing at what spatial scale autocorrelation is most important. Here we used autocovariates generated at 1, 2, 3, 4, 5, 10, 20 and 30 km distances from study sites, and included these as predictor variables.

2.9. Are fragmented sites also anthropogenically degraded?

2.11. Avoiding pseudoreplication

It is important to examine whether the 30 nominally anthropogenically fragmented and 30 nominally undisturbed study sites differed in their vegetation structure or habitat complexity. The process used to split populations when building tree-branched statistical models makes no assumptions about cause and effect, and to our knowledge there is no inherent reason why treebranched models cannot be used with one explanatory variable and many response variables. To do this we used Fragmentation (nominally fragmented or continuous) in place of the ‘response’ variable and vegetation and habitat complexity variables (Table 1) in place of ‘predictor’ variables. The tree-branched model constructed can be thought of as addressing the question: is it possible to split study sites into sub-populations according to the vegetation and habitat complexity of the environment, and if so, do fragmented and continuous sites group differently? The habitat features (Table 1) included as variables were those that described

We measured vegetation, environmental features and small mammal abundances on two trap grids per study site. If conditions within sites are related, including both grids in analysis is pseudoreplication. The simplest solution is to either average the values for sites or discard one trap grid per site at random. However, we set trap grids in fragments and corresponding pseudofragments at edges and interiors. This means that there is potentially informative variation in environmental features between trap grids at a given site, and we wished to retain this variation in the analysis. To account for this, we have used a conservative adjustment of significance level to reduce Type I error rates. As a rule of thumb, we suggest that the significance level (a = 0.10) in such situations is divided by n2, where n is the average number of sub-samples per sample. In this study there were always two sub-samples per site, so we have used a significance level of 0.025 throughout. Using SITE as an explanatory variable was also tested, but this led to

2.10. Spatial autocorrelation

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

87

over-splitting of the data and regression trees that were difficult to interpret and unlikely to be biologically meaningful.

2.12. Seasonal variation Where effects of time of year were observed, a second CIT was constructed (after correcting for seasonal changes) by taking residuals of the response variable from the month of the year (i.e. using a linear model with month as predictor.) The residuals were then used as response variables to construct a CIT. The seasonally uncorrected and corrected tree models are presented for comparison.

2.13. Limitations on interpretation Reporting splits in explanatory variables at precise values can imply a biological significance that is probably not real. For example, a split in abundance may be reported for native tree cover at 75.2 vs. 75.3 ha, whereas actually there is probably very little difference between these two exact values in terms of environmental conditions. The exact thresholds produced will always depend on the data examined and for this reason the thresholds need to be considered only as approximations of real relationships. Obviously this is of less concern where the explanatory variable is categorical, but at least in some instances, the resulting step-functions may not as accurately represent a relationship as would a model that assumes a continuous relationship (e.g. a linear mixed effects model).

Fig. 2. Conditional inference tree examining the nominal site classifications Fragment (F) and Pseudofragment (in continuous forest) (P) using vegetation and habitat complexity as ‘explanatory’ variables. Fragmented and continuous sites were best classified into four sub-populations. Node 3 is primarily fragmented sites [light grey] (90%), whereas Nodes 4 and 6 are primarily continuous sites [dark grey] (100% and 98% respectively). Node 7 constitutes and equal number of fragmented and continuous sites. LEAF = leaf litter depth and extent score. pcNONEUC = Percentage of trees that were not in genus Eucalyptus. EVENESS.W = Wilson’s evenness index for shrub species.

3.2. Stress and condition metrics for agile antechinus 3. Results In 2007, 967 A. agilis (429 males and 538 females), 51 A. swainsonii and 125 Rattus fuscipes were captured in 1800 trapnights spread among 30 study sites. All trapped A. agilis were adults. Morphometrics and body mass were measured for 183 male and 112 female A. agilis; the corresponding numbers for blood sampling were 182 and 113, respectively. During 2008, we captured 655 A. agilis (379 males and 276 females), 14 A. swainsonii and 191 R. fuscipes in the same number of trap-nights and study sites as in 2007. Morphometric and mass data for A. agilis were taken for 151 males and 105 females, whilst blood sampling was conducted on 150 males and 104 females. Numbers for morphometrics and blood samples differ slightly because some animals escaped from the researchers before all measurements could be taken.

3.1. Differences between fragmented and continuous study sites Fragmented sites were best characterised by: (1) shallow depth and limited extent of leaf litter and a low percentage of tree DBH that comprised native, non-Eucalyptus species (n = 54); and (2) deep, extensive leaf litter and high shrub species evenness (n = 5). Continuous sites were best characterised by: (1) shallow depth and limited extent of leaf litter, but a high percentage DBH that comprised non-Eucalyptus tree species (n = 7); (2) deep leaf litter and low shrub species evenness (n = 42); and (3) deep, extensive leaf litter and high shrub evenness (n = 5) (Fig. 2). Six continuous forest sites apparently had leaf litter and DBH characteristics more typical of fragmented sites, whilst one fragmented site had leaf litter and shrub evenness characteristics more typical of continuous sites (Fig. 2). The key differences appeared to be leaf litter depth (shallow in fragments, deep in continuous forest) and shrub species evenness (high in fragments, low in continuous forest).

N:L ratio increased from March to August. In May–June (after the annual dispersal, but before breeding), there was a forest fragmentation effect on N:L. Where native tree cover was more fragmented within 1 km of sites, N:L was higher (Fig. 3A). When N:L was corrected for seasonal variation, the fragmentation effect remained, except that the split was for fragmentation within 4 km of sites (Fig. 3B). Agile antechinus RBC (cells/L) was higher in sites where native tree cover was less within 0.5 km of the site. At sites where native tree-cover was greater within 0.5 km of a site, there was a second split in RBC, and a seasonal effect was evident, with RBC decreasing from March to August. When RBC was corrected for seasonal changes, the significant effect of native tree cover remained (Fig. 4A and B). Males had higher BCI than did females. Male A. agilis had lower BCI in sites where native tree cover was less within 0.5 km of the site (Fig. 5). 3.3. Relative abundance of the three study species 3.3.1. Agile antechinus Abundances for male and female agile antechinus were analysed separately. Male abundances were higher where native tree cover was greater within 3 km of the site. Female abundances were higher where leaf litter depth and extent were greater (Figs. 6 and 7). 3.3.2. Dusky antechinus The relative abundance of A. swainsonii was greater in sites at higher altitude (Fig. 8). 3.3.3. Bush rat R. fuscipes’ RA was higher in sites that had a greater autocovariate of R. fuscipes’ abundance within 3 km of the site (i.e. RA

88

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

Fig. 3. (A) Conditional inference tree for A. agilis (log)N:L. The number of each node is shown in a small box inset in a larger box bearing the relevant explanatory variable’s name and associated P-value. Categories or numerical ranges for each split are shown immediately below the variable name box. e.g. for Node 1 (MONTH), the split was into populations sampled in March–June vs. those sampled in July–August. The splitting criterion was P < 0.025. Boxplots show medians, ranges and upper and lower quartiles for populations for which no further splitting was possible. e.g. A. agilis sampled in May–June were split by habitat fragmentation in a 1 km radius (frag.1 km); in sites with low fragmentation indices (67), median (log)N:L was lower than in sites with a high index of fragmentation (>7). (B) The same as Figure 3a, except that (log)N:L has been corrected for seasonal changes from March to September before analysis (i.e. the values on y-axes are (log)N:L above or below expected given the Month). Here, A. agilis living in habitats that were more fragmented over a 4 km radius showed the higher indices of stress.

positively associated with neighbouring RA within 3 km of a site). To examine whether other features of the environment were interacting with the autocovariate of R. fuscipes’ abundances, a second tree was produced using a = 0.10. This tree was subject to Type I error, but indicated that where autocovariates of abundance were low (i.e. where neighbouring fragments had fewer bush rats), bush rat abundance was higher where the tree species had a greater non-Eucalyptus DBH (e.g. Acacia, Oleria, Cassinia spp.). Where the DBH of non-Eucalyptus tree species was low, there was a third split, with greater bush rat abundances being associated with a higher Weibull Scale for area at breast height of all trees (i.e. greater area at breast height associated with greater bush rat abundances, but only where the dominant tree genus was Eucalyptus) (Fig. 9A and B).

3.4. Random forests analysis A Random Forest procedure was used to determine the relative importance of all explanatory variables (Table 1). Month, Fragmentation (5 km), density of logs, an autocovariate of N:L (2 km), and density of large trees were important variables that had positive correlations with A. agilis’ N:L (i.e. possibly these features were associated with greater levels of physiological stress) (Fig. 10A). Month, native tree cover (0.5 km radius of study sites), longitude, log density and native tree cover at 1 and 2 km were important for explaining RBC (cells/L), but the correlations were negative (Fig. 10B). Sex was important in explaining A. agilis’ BCI (M > F). Month, extent of native tree cover (0.5 km radius of study sites), female

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

89

Fig. 4. (A) Conditional inference tree for A. agilis’ RBC. RBCs were higher in populations living in sites with 627.97 ha native tree cover in 0.5 km radius of sites. In sites where the tree cover was >27.97 ha in 0.5 km radius of a site, RBC was higher in March–May than in June–September. (B) The same as Figure 4a except that RBC was corrected for seasonal changes from March–September before analysis (i.e. the values on y-axes are RBC above or below expected given the Month). Correcting for Month did not alter the association between RBC and native tree cover extent.

A. agilis abundances, native tree cover at 1 km and shrub species richness were also important explanatory variables (Fig. 11). For agile antechinus, the most important variables for explaining RA were tree cover and altitude (both positively correlated with RA). Autocovariates of abundance were also important, although for females leaf litter depth and extent were more important than the autocovariate of female abundance (Fig. 12). Abundance of A. swainsonii responded most strongly to the autocovariate of A. swainsonii abundance (5 km radius of study sites) and to native tree cover (3 km) (both correlations were positive) (Fig. 13). Abundances of R. fuscipes responded most strongly to the autocovariate of R. fuscipes abundance (3 km radius of study sites), percentage of native tree species that were not in the genus Eucalyptus, and native tree cover extent (within 5 km radius of study sites) (Fig. 13).

4. Discussion Almost half of all mammal species extinctions in the last 200 years have occurred in Australia (Cardillo and Bromham, 2001; Short and Smith, 1994). Many of Australia’s remaining small mammal species are threatened or endangered (terrestrial and flying mammal species that are currently vulnerable = 49, endangered = 32, critically endangered = 4, extinct = 28 of approx. 300 species thought to have been extant at the time of European settlement; Australian Federal Department of Environment, 2014) and the future of species that are currently locally common remains uncertain, as large-scale clearing of native vegetation continues (Gibbons and Lindenmayer, 2007). Previous studies on the same or closely related small mammal species indicate that populations of the genus Antechinus respond positively to habitat complexity (e.g. woody debris; Mac Nally

90

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

Fig. 7. Conditional inference tree for female A. agilis’ relative abundance. Abundance was greater in sites where leaf litter was deeper and more extensive.

Fig. 5. Conditional inference tree for A. agilis BCI. Male BCI was greater than female BCI, but female BCI was not significantly associated with any ecological variable. Male BCI was greater where native tree cover was 620.53 ha within 0.5 km radius of sites.

Fig. 8. Conditional inference tree for A. swainsonii’ relative abundance. Abundance was higher at altitudes >479 m above sea level.

4.1. Anthropogenic degradation in fragmented study sites

Fig. 6. Conditional inference tree for male A. agilis’ relative abundance. Abundance was greater where native tree cover was >2353.05 ha within 3 km of sites.

and Horrocks, 2002). Although previous researchers have reported area effects on population densities, these have been attributed to there being more highly intact habitat complexity in larger forest patches (Knight and Fox, 2000). Population density of native rodent examined, the bush rat R. fuscipes, is thought to respond positively to forest patch area (Dunstan and Fox, 1996), although recent evidence suggests that the occurrence of this species is also affected by vegetation structure (Holland and Bennett, 2007, 2009).

The conditional inference tree model characterising sites in fragmented and continuous habitats in terms of vegetation and habitat structure was consistent with the theory that habitat fragmentation is associated with some degree of habitat degradation (Fischer and Lindenmayer, 2007). The fragmented sites had shallower and less extensive leaf litter and an apparently lower tree species richness than continuous forest sites in the same landscape. There are several reasons why non-Eucalyptus trees may have been less common in forest fragments. Some species have valuable timber and have been selectively logged and milled in historical times (e.g. Acacia melanoxylon), whilst others are palatable to livestock (e.g. Olearia lirata, O. argophylla, Acacia melanoxylon).

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

91

Fig. 9. (A) Conditional inference tree for R. fuscipes’ relative abundance. Abundance was greater in sites where neighbouring sites (within 3 km) had a higher autocovariate of R. fuscipes’ abundance (i.e. emigration and immigration is probably important for maintaining local populations). (B) The same as (a), except that the significance level was changed to 0.10 to explore relationships. In sites with low abundances in neighbouring forest sites, R. fuscipes occurred at higher abundances if the percentage of native tree ABH that was not Eucalyptus was >27.9%. In sites with a low native trees total ABH, R. fuscipes’ abundances were greater in sites where the Weibull Scale of ABH (i.e. overall size of the ABHs) was greater.

Anecdotally at least, some species were also selectively felled for firewood in historical times (e.g. Olearia argophylla). Leaf litter depth has previously been positively associated with native small mammal occurrence (Dunstan and Fox, 1996; Kelly and Bennett, 2008), so its apparent shallowness in fragmented sites is a concern from a conservation viewpoint. Many non-Eucalyptus trees (e.g. Olearia argophylla, Acacia melanoxylon) have fissured bark that shelters the invertebrate prey of A. agilis, so that the absence of these tree species is also potentially concerning (Selwood et al., 2009). 4.2. Relative abundance of small mammals at the landscape level The strongest landscape level influences on estimated RA reveal interesting differences in spatial scale responses by the species to

the environment. Male agile antechinus responded most strongly to native tree cover at 3 km (tau 0.383), male abundance at 5 km (tau 0.310) and forest fragmentation at 2 km (tau 0.253). Females responded to native tree cover extent at 0.5 km (tau 0.232), female abundance at 3 km (tau 0.186) and forest fragmentation at 1 km (tau 0.216). Given our understanding of agile antechinus’ life history, it makes sense that males would respond more strongly to native tree cover, the spatial autocovariate of male abundance and forest fragmentation at a greater spatial scale. Dispersal in this species is male-biased, and females tend to be highly philopatric (Cockburn and Lazenby-Cohen, 1992; Kraaijeveld-Smit et al., 2002a). Our results seem to imply that male agile antechinus are more likely than females to respond negatively to habitat degradation and fragmentation over a larger area (M = 3 km radius cf F = 0.5 km radius). Probably, this is because of reduced male

92

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

Fig. 10. Relative importances of all explanatory variables for agile antechinus’ N:L (A) and RBC (B). The length of the horizontal bar indicates the relative importance (no units) of the variable for explaining (log)N:L or RBC determined using random forests analysis. The numbers to the right of each bar are Kendall tau correlation coefficients. F/C = Fragment/Continuous. RA = Relative abundance. M/F = Male/Female.

immigration into a forest fragment site when the surrounding matrix is highly disturbed and therefore is less suitable for male dispersal. Dusky antechinus responded most strongly to dusky antechinus abundance at 5 km (tau 0.268), native tree cover extent at 3 km (tau 0.214) and forest fragmentation at 4 km (tau 0.061). They were not sexed before release from traps, so we do not know if the effect of habitat loss and fragmentation on the sexes is different (as in agile antechinus). The CIT for dusky antechinus only provided one significant node, where dusky antechinus abundance was greater at higher altitudes. This species is considered to be an alpine wet montane species (Menkhorst and Knight, 2004), so this finding is not surprising. However, this result is also somewhat confounded, because sites at higher elevations appeared to be less anthropogenically disturbed (e.g. firewood collection was never

observed above 400 m elevation). If this interpretation is correct, it could be because sites at higher elevations were more difficult for humans to access. The possibility of relationships between altitude and human disturbance warrants further investigation. The results for the third small mammal species and the only eutherian in this study, the bush rat, showed clear and strong responses to the autocovariate of bush rat abundances, particularly within 3 km of study sites. This is interesting, as previous research has indicated that bush rats respond to forest fragment area (Dunstan and Fox, 1996; Holland and Bennett, 2009) or positively to complex, dense vegetation cover (Holland and Bennett, 2007). Some authors have tested the accuracy of predictive models of bush rat patch occupancy to the results of actual field surveys (e.g. Ball et al., 2003; Lindenmayer and Lacy, 2002; Lindenmayer et al., 2003), but the results of these studies are difficult to apply

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

Fig. 11. Relative importance of all explanatory variables for agile antechinus’ body condition index (BCI). The length of the horizontal bar indicates the relative importance (no units) of the variable for explaining BCI determined with random forests analysis. The numbers to the right of each bar are Kendall tau correlation coefficients. F/C = Fragment/Continuous. RA = Relative abundance. M/F = Male/ Female.

here, as they were conducted in Eucalyptus forest fragments in a matrix of exotic softwood plantation. Plantations and cleared fields would probably present a quite different barrier to dispersal, so we are reluctant to generalise these findings to the populations under study here. However, in a recent field experiment in a fragmented landscape in a grazed matrix, Holland and Bennet (2011) removed bush rats from small forest fragments, simulating local ‘extinctions,’ and then tracked immigration into fragments and undisturbed control fragments via dispersal across a cleared agricultural matrix. Migration of both males and females (required for successful recolonisation) into experimentally expiated fragments was observed at only two of four such fragments 16 months after the removal of bush rats, and the number of immigrants was significantly lower than in fragments where bush rats were not removed. The implication is that bush rats may maintain a metapopulation, but that remigration is relatively slow and they may be ‘reluctant’ to settle in habitat lacking conspecifics. Bush rats can use vegetation corridors in fragmented habitats (Bennett, 1990a) and perhaps such corridors may help maintain

93

metapopulations in this species, but given the strong influence of the autocovariate of local bush rat abundances, more attention should be given to how bush rats maintain metapopulations and whether there is a habitat loss threshold at which metapopulations will collapse. As bush rats were not caught at all sites in continuous forest, presumably they also have a variable population density in continuous forest, and move among suitable habitats in heterogeneous, but contiguous forest. It would be interesting to examine movement of this species in landscapes of heterogeneous, contiguous native forest. The most important native tree cover and fragmentation determinants of bush rat abundances were at 5 km (tau 0.273) and 1 km (tau 0.275), respectively. That habitat loss per se explained the patterns of estimated RA of the mammal species studied better than habitat fragmentation did is intriguing. The effects of habitat loss may not be as well understood as is often assumed; in a review of publications in the conservation biology field, only 2% of studies directly examined native vegetation loss as a threatening process (Fazey et al., 2005), the implication being that most conservation researchers do not consider habitat loss per se to be an important topic to investigate. Here, the estimated RA of agile antechinus, the smallest species investigated, was associated with native tree-cover at 2.5 km (on average), that of dusky antechinus with native tree cover at 3 km and that of bush rats, the largest species, with tree cover at 5 km from a trapping site. The implication of this is that loss of habitat within 5 km of a study site was negatively associated with small mammal RA. This has important implications for faunal reserve design: if a forest reserve is to be functionally protected, the source sites for immigration and emigration have to be considered for protection as well. Arguably, this is already well understood in the scientific literature, but governments responsible for establishing reserves are perhaps unlikely to consider the implications of protecting one area of native vegetation whilst allowing deforestation at sites 4–5 km distant from it. Significant correlations between a species’ body size and the area of habitat required to support a stable population of that species are thought to be widespread (Peters and Raelson, 1984; Robinson and Redford, 1986). However, empirical evidence for such relationships, at least in small mammals, is largely circumstantial e.g. that larger areas tend to have more species of a given taxon (Connor and McCoy, 1979), that larger species tend to occur at lower densities (Peters and Raelson, 1984), that individuals of larger species typically have larger home ranges (Kelt and Van Vuren, 2001) and that larger taxa are often at greater risk of extinction in fragmented habitats (Turner, 1996). The relationship is potentially confounded by trophic level and variation in landscape productivity (Wright, 1983) or distribution of essential habitat structures (nesting sites, water sources, woody debris, etc.) (Tews et al., 2004), so that it is not always easily identifiable. A further point worth addressing is that it is very likely that a key driver causing decline in native small mammals is predation by invasive V. vulpes or F. catus, but we are unconvinced that our ‘predator activity index’ (which was really just an index of overnight trap disturbance by an animal large enough to move a trap >5 m) adequately measured activity by these predators. Very clearly a study is needed that measures predator activity using discrete trap cameras and compares the data obtained to habitat loss and fragmentation and small mammal abundances. Disentangling the causal matrix could then be done using experimental manipulation (e.g. predator exclusion or establishment of small, artificial, predation ‘shelters’ for small mammals throughout a fragment). Also important is that unique marking of individuals for markrecapture analysis was not done here and this means that the indices we are using can only be viewed as a measure of trapping success (sensu Stewart, 1979). We would strongly advise researchers undertaking similar work to consider using PIT tags and possibly

94

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

Fig. 12. Relative importance of all explanatory variables for explaining male (A) and female (B) agile antechinus’ relative abundances. The length of the horizontal bar indicates the relative importance (no units) of the variable for explaining abundance determined using random forests analysis. The numbers to the right of each bar are Kendall tau correlation coefficients, which are suitable for examining non-linear relationships. Definitions of explanatory variables are given in Table 1. F/C = Fragment/ Continuous.

also trap cameras in addition to trapping to establish more accurate estimates of abundance. 4.3. Variables influencing agile antechinus population wellbeing Male agile antechinus were more abundant when tree cover at 3 km was greater, whereas females responded to more extensive and deeper leaf litter. This implies that males were responding to less degraded meta-habitats (or perhaps larger foraging areas), whereas females were responding to less disturbed local habitats (or perhaps smaller foraging areas). The strongest associations for BCI were with sex (males had larger BCI than females) and agile antechinus’ abundance. The finding that male agile antechinus had higher estimated fat reserves where native tree cover was greater within 0.5 km of a site was predictable; we have reported a similar relationship elsewhere (Johnstone et al., 2010) and sex differences in mass are well known for agile antechinus (Cheal et al., 1976). The association between higher male BCI and lower local abundances in sites is more interesting. Elsewhere (Johnstone et al., 2010) we have proposed

that in less degraded habitats (where abundances are higher) a male agile antechinus’ body condition may be food-limited, whereas in the more degraded habitats male agile antechinus’ abundance may be predator-limited (i.e. through predation by invasive V. vulpes or F. catus); consequently at low population densities, male agile antechinus are able to attain a higher body mass prior to breeding, although the habitat may be worse for survival. This makes biological sense, as males with greater fuel reserves have a significant reproductive advantage (Kraaijeveld-Smit et al., 2003, 2002b). Perhaps female BCI might reveal a similar trend during a period when energy requirements are higher for females (e.g. during lactation). An association between the extent and depth of leaf litter and small mammal abundance has been documented previously in Australia (Knight and Fox, 2000), although it is usually assumed to relate to food abundance and ‘bottom-up’ population regulation (i.e. less leaf litter results in a lower arthropod abundance, which in turn results in nutritional stress and lower abundance in omnivorous and insectivorous small mammals). However, bottom-up regulation of small mammal population density is generally not

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

95

Fig. 13. Relative importance of all explanatory variables for bush rat (A) and dusky antechinus (B) relative importances. The length of the horizontal bar indicates the relative importance (no units) of the variable for explaining relative abundance determined using random forests analysis. The numbers to the right of each bar are Kendall tau correlation coefficients. F/C = Fragment/Continuous.

well supported empirically (Krebs, 2009), and in agile antechinus at least, the BCI results are not consistent with this interpretation. Leaf litter extent and depth had a negative relationship with agile antechinus’ BCI (tau 0.158) and associations between BCI and abundances were negative too (male tau 0.063, female tau 0.132) i.e. deeper and more extensive leaf litter was associated with higher abundances, but lower estimated fat reserves. If agile antechinus populations were food-limited, low relative abundance and poor body condition should have been correlated and both should have been observed in food-poor sites (Suorsa et al., 2004; although for a discussion of this also see Caughley et al., 1988). The reason why leaf litter appeared to influence agile antechinus’ abundance is therefore unknown, although here the relationship seems strongest for females, and perhaps is only an indication of a local lack of disturbance. If only an estimator of abundance had been measured in this study, the obvious conclusion would have been that agile antechinus were food limited, as evidenced by the positive correlation with leaf litter. However, once BCI is also taken into account, a food-limitation explanation is clearly not plausible. This illustrates why it is important to

measure performance indicators, such as BCI and N:L, wherever possible, and not rely solely on population density, fecundity or survivorship estimates when addressing ecological questions about possible threatening processes and/or the conservation management of vertebrates. It could be that in our study leaf litter quality was a surrogate for an aspect of forest stand health, canopy density or another unidentified environmental factor that was not measured in this investigation. Experimental manipulation of field sites is difficult, but possible, and could be used to address this issue (Mac Nally and Horrocks, 2002). By manipulating leaf litter depth at sites one could explore whether antechinus are responding directly to leaf litter and why. For example, another explanation of the leaf litter-relative abundance relationship is that antechinus could in theory be attracted to leaf litter primarily because it helps to camouflage their shape and outline rather than because it provides good foraging opportunities. Using native leaf litter and an artificial material that provides similar camouflage but no food, such as fabric leaves, could address this point and help resolve why antechinus numbers are higher where leaf litter is deeper and more extensive.

96

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

The results for N:L and RBC for agile antechinus provide some evidence for elevated HPA-axis mediated stress in more fragmented habitats (higher N:L where forest fragmentation was greater) and where habitat loss is greater (higher RBC where native tree cover was lower). These interpretations assume that N:L really reflected the degree of lymphocyte trafficking out of peripheral blood and into other body compartments in response to long-term physiological stress, and that RBC really reflected regenerative anaemia (where erythrocytes are released prematurely into the blood in response to bleeding, or long-term physiological stress or high blood parasite loads). Trapping stress is a potentially confounding element here, especially for N:L, (Johnstone et al., 2012b), so we think that these interpretations should be considered tentative, although they remain worth exploring further. To interpret the N:L result, some aspects of agile antechinus’ life history need to be considered. The species is a rare example of a largely semelparous mammal (Braithwaite and Lee, 1979). Young are weaned in the late Austral summer (January), soon after which there is male-biased dispersal (Cockburn et al., 1985b). Adults nest socially until the late Austral winter (August in the study area), when a synchronised, competitive, breeding rut occurs (LazenbyCohen and Cockburn, 1988) that has features of lek behaviour (Kraaijeveld-Smit et al., 2002a). Successful mating is mostly achieved by the larger males and those that mate closer to the time of female oestrous (Kraaijeveld-Smit et al., 2003, 2002b). After breeding, there is a complete male die-off (Braithwaite and Lee, 1979), and after weaning their young, only 15% of females live to reproduce in a second year (Cockburn et al., 1985a). Other authors have investigated the relationship between habitat fragmentation and HPA axis-mediated stress in terrestrial vertebrates with equivocal results (Martínez-Mota et al., 2007; Mazerolle and Hobson, 2002; Suorsa et al., 2004). Conceivably the relationship may be environment-, time-of-year- or taxon-specific. We have reported elsewhere evidence of a broad, positive association between habitat fragmentation and N:L in agile antechinus (Johnstone et al., 2012a). Interestingly, in the present study there was strong evidence of this relationship only during a twomonth period between dispersal and the subsequent social reorganisation (Cockburn and Lazenby-Cohen, 1992). Fragmentation was no longer influential in winter (July–August), when psychosocial stress (prior to breeding) or stressors associated with winter conditions could have been more substantial influences on an individual’s wellbeing than stressors associated with forest fragmentation i.e. edge effects, novel barriers to dispersal or foraging and a greater abundance of invasive predators, etc. (Fischer and Lindenmayer, 2007). Although environmental stressors of wild vertebrates sometimes act synergistically (e.g. stress due to predators and food scarcity is multiplicative, not additive; Boonstra et al., 1998a; Zanette et al., 2003), the unusual breeding system of agile antechinus may mean that it is adaptive for males, at least, to reduce the negative and often reproduction-suppressing effects of HPA axis-mediated chronic stress in the breeding season by having a physiological mechanism that down-regulates short-term stress responses at this time of the year (Wingfield and Sapolsky, 2003). Without experimental manipulation of sites, the causal relationship cannot be known, but certainly a comparison of the agile antechinus stress response in fragmented and continuous forest after predator exclusion (Stokes et al., 2004) and/or food supplementation (Dickman, 1989) would be worthwhile. Using an approach similar to that used in the Northern Hemisphere to examine predator and food effects on snowshoe hares (Lepus americanus; Boonstra et al., 1998b; Krebs et al., 1995), would be invaluable, so that a large scale experiment that used treatments of (1) predator exclusion, (2) food supplementation, (3) predator exclusion and food supplementation, and (4) a control, would be a suitable approach. However, as a caveat, Prevedello et al. reported that

a meta-analysis of food supplementation studies suggest that such studies can act as ecological traps by concentrating predators and prey in the same area. The study recommended spreading food over larger areas to reduce the ecological trap effect. 4.4. Random forests analysis of stress metrics Some of the associations between environmental variables and N:L and RBC agree with general expectations about the response of a small mammal to its environment e.g. greater native tree cover associated with a lower N:L (tau 0.071 at 5 km) and a lower RBC (tau 0.214 at 0.5 km) and greater forest fragmentation associated with a higher N:L (tau 0.166 at 5 km) and RBC (tau 0.171 at 0.5 km). The associations, however, were weak and apart from that between N:L and fragmentation, they did not warrant inclusion in the bestsupported model, so conclusions can only be tentative. It is interesting that RBC seemed to respond to tree cover and fragmentation at 0.5 km from a trapping site, whereas N:L was most strongly associated with tree cover and fragmentation at 5 km. As with the results for abundances of small mammals, it was unexpected that a health indicator would be associated with changes in the environment over such a large area (5 km), but possibly generalised degradation of the surrounding landscape was acting as an indicator of local forest stand condition or as a measure of the abundance of invasive predators. Although the RBC association with the density of logs was as expected (tau 0.149), N:L was higher where log density was higher (tau 0.210). Different aspects of an animal’s interior milieu can respond differently to stressors, but it is peculiar that agile antechinus should seemingly show a higher physiological stress level in association with an environmental variable that is thought to be beneficial for this genus’ survival (Kelly and Bennett, 2008; Knight and Fox, 2000; Stokes et al., 2004). The possibility of a spurious correlation cannot be excluded. Windfall at fragment edges might have caused greater woody debris densities in more disturbed sites, but if so RBC should have been higher with a greater log density too. Antechinus abundances are higher in sites with a greater woody debris volume (Mac Nally and Horrocks, 2002), and if this is an effect of both habitat selection (i.e. dispersing individuals are more likely to settle in sites with logs) and survivorship (i.e. logs provide cover from predators), it is plausible that some social stress from crowding could result. Regenerative anaemia tends to be reported in association with injury and parasite stress, whereas N:L can be responsive to frequent, but relatively minor, stressors whose effects add up over time (Johnstone et al., 2012b). The two metrics do not always respond to environmental stressors similarly, as is evident from the agile antechinus’ responses to seasonal change. N:L increased from March to September (tau 0.475), but RBC decreased during the same period (tau 0.231). It is plausible that the higher N:L associated with a greater density of logs was a response to social stress, but that this stress was not sufficient to trigger an emergency state and regenerative anaemia. Experimental manipulation of cover (adding or removing woody debris) and/or manipulation of population density (trapping and moving individuals) might allow the underlying mechanisms to be unravelled. 5. Conclusions From a conservation perspective, the investigation raised an important cautionary note; the three small mammal species appeared to respond to anthropogenic habitat loss and fragmentation at landscape scales substantially greater than the area that an individual would occupy in its lifetime. For the particular mammal species and geographical region examined here at least, restoration of forest reserves may not nec-

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

essarily increase the animals’ local abundance if native vegetation levels in the general area are below those needed to sustain ‘healthy’ populations. The method of analysis used here could be employed to address the question that still concerns conservation managers with respect to reserve area, namely ‘How much is enough?’ (Fahrig, 2001). It is also of concern that the three mammal species studied are considered locally common in some areas (Menkhorst and Knight, 2004) and are generally not the focus of any conservation effort in the study area, throughout most of which bush rats and dusky antechinus occurred at abundances that were below their apparent ‘norms’ recorded in the least-disturbed sites. Small and isolated populations are susceptible to stochastic threats (Fischer and Lindenmayer, 2007), but the gradual decline to extinction of populations after fragmentation and degradation of their habitat may take years or decades (Diamond et al., 1987). Without longitudinal data, it is difficult to know if the native species in the current study might be undergoing gradual, but deterministic, decline which is characteristic of extinction debt (Tilman et al., 1994), but our results seem to raise this possibility. It is easy to overlook such a cryptic decline until it is too late. Acknowledgments Trapping and sampling were conducted under Monash University Biological Sciences Animal Ethics Committee approvals BSCI/ 2008/03 and BSCI/2006/05 and Department of Sustainability and Environment permit 10003798. This research was supported by the Holsworth Wildlife Fund and access was kindly granted by private landowners throughout the South Gippsland region. Field accommodation was provided by Parks Victoria, J. & S. Bell, G. & J. Wallis, D. & M. Hook and D. Farrar. We also thank C. Rankin for access to South Gippsland Shire council reserves. The support, co-operation and enthusiasm of many individuals and groups helped to facilitate this project, notably the South Gippsland Conservation Society, Venus Bay Landcare and Anders Inlet Landcare. The following are a small fraction of the many people who deserve special thanks and recognition: Eric Cumming, John and Sue Bell, Rick and Marion Bowron (and Johnny), Mary Ellis, David Farrar, Ian Gunn, Daryl and Margaret Hook, Geoff Hutchinson, David Kelly, Martin Newman and Alex and Herb Wilde. We would also like to thank the two anonymous reviewers of this paper who made numerous thoughtful and helpful suggestions for improvement. References Australian Bureau of Meteorology, 2009. Climate Statistics for Australian Sites. . Andrén, H., 1994. Effects of habitat fragmentation on birds and mammals in landscapes with different proportions of suitable habitat: a review. Oikos 71, 355–366. Arndt, S., Turvey, C., Andreasen, N.C., 1999. Correlating and predicting psychiatric symptom ratings: spearmans r versus Kendalls tau correlation. J. Psychiatr. Res. 33, 97–104. Atchley, W.R., Gaskins, C.T., Anderson, D., 1976. Statistical properties of ratios. I. Empirical results. Syst. Biol. 25, 137–148. Baker, P.J., Bunyavejchewin, S., Oliver, C.D., Ashton, P.S., 2005. Disturbance history and historical stand dynamics of a seasonal tropical forest in western Thailand. Ecol. Monogr. 75, 317–343. Ball, S.J., Lindenmayer, D.B., Possingham, H.P., 2003. The predictive accuracy of population viability analysis: a test using data from two small mammal species in a fragmented landscape. Biodivers. Conserv. 12, 2393–2413. Banks, P.B., Dickman, C.R., 2000. Effects of winter food supplementation on reproduction, body mass, and numbers of small mammals in montane Australia. Can. J. Zool. 78, 1775–1783. Barnett, J.L., How, R.A., Humphreys, W.F., 1979. Blood parameters in natural populations of Trichosurus species (Marsupialia: Phalangeridae). II. Influence of habitat and population strategies of T. caninus and T. vulpecula. Aust. J. Zool. 27, 927–938. Beauvais, G.P., Buskirk, S.W., 1999. Modifying estimates of sampling effort to account for sprung traps. Wildl. Soc. Bull. 27, 39–43.

97

Bennett, A.F., 1990a. Habitat corridors and the conservation of small mammals in a fragmented forest environment. Landscape Ecol. 4, 109–122. Bennett, A.F., 1990b. Land use, forest fragmentation and the mammalian fauna at Naringal, south-western Victoria. Aust. Wildlife Res. 17, 325–347. Bennett, A.F., 1993. Microhabitat use by the long-nosed potoroo, Potorous tridactylus, and other small mammals in remnant forest vegetation of southwestern Victoria. Wildl. Res. 20, 267–285. Boonstra, R., Hik, D., Singleton, G.R., Tinnikov, A., 1998a. The impact of predatorinduced stress on the snowshoe hare cycle. Ecol. Monogr. 68, 371–394. Boonstra, R., Krebs, C.J., Stenseth, N.C., 1998b. Population cycles in small mammals: the problem of explaining the low phase. Ecology 79, 1479–1488. Braithwaite, R.W., Lee, A.K., 1979. A mammalian example of semelparity. Am. Nat. 113, 151–155. Bureau, A., Dupuis, J., Falls, K., Lunetta, K.L., Hayward, B., Keith, T.P., Van Eerdewegh, P., 2005. Identifying SNPs predictive of phenotype using random forests. Genet. Epidemiol. 28, 171–182. Cardillo, M., Bromham, L., 2001. Body size and risk of extinction in Australian mammals. Conserv. Biol. 15, 1435–1440. Caughley, G., 1977. Analysis of Vertebrate Populations. Wiley New York, New York, USA. Caughley, G., Grice, D., Barker, R., Brown, B., 1988. The edge of the range. J. Anim. Ecol. 57, 771–785. Cheal, P.D., Lee, A.K., Barnett, J.L., 1976. Changes in the haematology of Antechinus stuartii (Marsupialia), and their association with male mortality. Aust. J. Zool. 24, 299–311. Cockburn, A., Lazenby-Cohen, K.A., 1992. Use of nest trees by Antechinus stuartii, a semelparous lekking marsupial. J. Zool. (Lond.) 226, 657–680. Cockburn, A., Scott, M.P., Dickman, C.R., 1985a. Sex ratio and intrasexual kin competition in mammals. Oecologia 66, 427–429. Cockburn, A., Scott, M.P., Scotts, D.J., 1985b. Inbreeding avoidance and male-biased natal dispersal in Antechinus spp. (Marsupiala: Dasyuridae). Anim. Behav. 33, 908–915. Connor, E.F., McCoy, E.D., 1979. The statistics and biology of the species-area relationship. Am. Nat. 113, 791–833. Cutler, D.R., Edwards Jr, T.C., Beard, K.H., Cutler, A., Hess, K.T., Gibson, J., Lawler, J.J., 2007. Random forests for classification in ecology. Ecology 88, 2783–2792. Czúcz, B., Gálhidy, L., Mátyás, C., 2011. Present and forecasted xeric climatic limits of beech and sessile oak distribution at low altitudes in Central Europe. Ann. For. Sci. 68, 99–108. Daley, C., 1962. The Story of Gippsland. Whitcombe & Tombs for the Gippsland Municipalities Association. Davies, J.B., Oates, A.M., Trumbull-Ward, A.V., 2002. Ecological Vegetation Class Mapping at 1:25,000 in Gippsland: Final Report. Victorian Department of Natural Resources and Environment, Melbourne, Australia. Davis, A.K., Maney, D.L., Maerz, J.C., 2008. The use of leukocyte profiles to measure stress in vertebrates: a review for ecologists. Funct. Ecol. 22, 760–772. Dhabhar, F.S., Miller, A.H., Stein, M., McEwen, B.S., Spencer, R.L., 1994. Diurnal and acute stress-induced changes in distribution of peripheral blood leukocyte subpopulations. Brain, Behav. Immun. 8, 66–79. Diamond, J., 1986. Overview: laboratory experiments, field experiments, and natural experiments. In: Diamond, J., Case, T.J. (Eds.), Community Ecology. Harper & Row, New York, pp. 3–22. Diamond, J.M., 1983. Ecology: laboratory, field and natural experiments. Nature 304, 586. Diamond, J.M., Bishop, K.D., Van Balen, S., 1987. Bird survival in an isolated Javan woodland: island or mirror? Conserv. Biol. 1, 132–142. Dickman, C.R., 1986. An experimental manipulation of the intensity of interspecific competition: effects on a small marsupial. Oecologia 70, 536–543. Dickman, C.R., 1988. Body size, prey size, and community structure in insectivorous mammals. Ecology 69, 569–580. Dickman, C.R., 1989. Demographic responses of Antechinus stuartii (Marsupialia) to supplementary food. Austral Ecol. 14, 387–398. Dickman, C.R., Parnaby, H.E., Crowther, M.S., King, D.H., 1998. Antechinus agilis (Marsupialia: Dasyuridae), a new species from the A. stuartii complex in southeastern Australia. Aust. J. Zool. 46, 1–26. Dormann, F.C., McPherson, M.J., Araújo, B.M., Bivand, R., Bolliger, J., Carl, G., Davies, G.R., Hirzel, A., Jetz, W., Kissling, D.W., 2007. Methods to account for spatial autocorrelation in the analysis of species distributional data: a review. Ecography 30, 609–628. Dunstan, C.E., Fox, B.J., 1996. The effects of fragmentation and disturbance of rainforest on ground-dwelling small mammals on the Robertson Plateau, New South Wales, Australia. J. Biogeogr. 23, 187–201. Environment, D.o.t., 2014. EPBC Act List of Threatened Fauna, ed. D.o.t.E. , . Fahrig, L., 1997. Relative effects of habitat loss and fragmentation on population extinction. J. Wildl. Manage. 61, 603–610. Fahrig, L., 2001. How much habitat is enough? Biol. Conserv. 100, 65–74. Fazey, I., Fischer, J., Lindenmayer, D.B., 2005. What do conservation biologists publish? Biol. Conserv. 124, 63–73. Fischer, J., Lindenmayer, D.B., 2007. Landscape modification and habitat fragmentation: a synthesis. Glob. Ecol. Biogeogr. 16, 265–280. Fletcher, J.R.J., Ries, L., Battin, J., Chalfoun, A.D., 2007. The role of habitat area and edge in fragmented landscapes: definitively distinct or inevitably intertwined? Can. J. Zool. 85, 1017–1030.

98

C.P. Johnstone et al. / Biological Conservation 176 (2014) 80–98

Fletcher, Q.E., Boonstra, R., 2006. Impact of live trapping on the stress response of the meadow vole (Microtus pennsylvanicus). J. Zool. 270, 473–478. Foley, J.A., DeFries, R., Asner, G.P., Barford, C., Bonan, G., Carpenter, S.R., Chapin, F.S., Coe, M.T., Daily, G.C., Gibbs, H.K., 2005. Global consequences of land use. Science 309, 570–574. Gibbons, P., Lindenmayer, D.B., 2007. Offsets for land clearing: no net loss or the tail wagging the dog? Ecol. Manage. Restor. 8, 26. Green, A.J., 2001. Mass/length residuals: measures of body condition or generators of spurious results? Ecology 82, 1473–1483. Herring, G., Cook, M.I., Gawlik, D.E., Call, E.M., 2011. Food availability is expressed through physiological stress indicators in nestling white ibis: a food supplementation experiment. Funct. Ecol. 25, 682–690. Holland, G.J., Bennet, A.F., 2011. Recolonization of forest fragments by a native rodent following experimental ‘extinctions’. Austral Ecol. 36, 521–529. Holland, G.J., Bennett, A.F., 2007. Occurrence of small mammals in a fragmented landscape: the role of vegetation heterogeneity. Wildl. Res. 34, 387–397. Holland, G.J., Bennett, A.F., 2009. Differing responses to landscape change: implications for small mammal assemblages in forest fragments. Biodivers. Conserv. 18, 2997–3016. Homyack, J.A., 2010. Evaluating habitat quality of vertebrates using conservation physiology tools. Wildl. Res. 37, 332–342. Horner, G.J., Baker, P.J., Nally, R.M., Cunningham, S.C., Thomson, J.R., Hamilton, F., 2010. Forest structure, habitat and carbon benefits from thinning floodplain forests: Managing early stand density makes a difference. For. Ecol. Manage. 259, 286–293. Hothorn, T., Hornik, K., Zeileis, A., 2006a. Party: A Laboratory for Recursive Partytioning. R Package Version 0.9-0. . Hothorn, T., Hornik, K., Zeileis, A., 2006b. Unbiased recursive partitioning: a conditional inference framework. J. Comput. Graph. Stat. 15, 651–674. Janin, A., Lena, J.P., Joly, P., 2011. Beyond occurrence: body condition and stress hormone as integrative indicators of habitat availability and fragmentation in the common toad. Biol. Conserv. 144, 1008–1016. Johnstone, C.P., Lill, A., Reina, R.D., 2011. Response of the agile antechinus to habitat edge, configuration and condition in fragmented forest. PLoS ONE 6, e27158. Johnstone, C.P., Lill, A., Reina, R.D., 2012a. Does habitat fragmentation cause stress in the agile antechinus? A haematological approach. J. Comp. Physiol. [B] 182, 139–155. Johnstone, C.P., Reina, R.D., Lill, A., 2010. Impact of anthropogenic habitat fragmentation on population health in a small, carnivorous marsupial. J. Mammal. 91, 1332–1341. Johnstone, C.P., Reina, R.D., Lill, A., 2012b. Interpreting indices of physiological stress in free-living vertebrates: a review. J. Comp. Physiol. [B] 182. http:// dx.doi.org/10.1007/s00360-00012-00656-00369. Kelly, L.T., Bennett, A.F., 2008. Habitat requirements of the yellow-footed antechinus (Antechinus flavipes) in box-ironbark forest, Victoria, Australia. Wildl. Res. 35, 128–133. Kelt, D.A., Van Vuren, D.H., 2001. The ecology and macroecology of mammalian home range area. Am. Nat. 157, 637–645. Knight, E.H., Fox, B.J., 2000. Does habitat structure mediate the effects of forest fragmentation and human-induced disturbance on the abundance of Antechinus stuartii? Aust. J. Zool. 48, 577–595. Kraaijeveld-Smit, F.J.L., Ward, S., Temple-Smith, P., 2002a. Multiple paternity in a field population of a small carnivorous marsupial, the agile antechinus, Antechinus agilis. Behav. Ecol. Sociobiol. 52, 84–91. Kraaijeveld-Smit, F.J.L., Ward, S.J., Temple-Smith, P.D., 2003. Paternity success and the direction of sexual selection in a field population of a semelparous marsupial, Antechinus agilis. Mol. Ecol. 12, 475–484. Kraaijeveld-Smit, F.J.L., Ward, S.J., Temple-Smith, P.D., Paetkau, D., 2002b. Factors influencing paternity success in Antechinus agilis: last-male sperm precedence, timing of mating and genetic compatibility. J. Evol. Biol. 15, 100–107. Krebs, C.J., 2009. Population dynamics of large and small mammals: Graeme Caughley’s grand vision. Wildl. Res. 36, 1–7. Krebs, C.J., Boutin, S., Boonstra, R., Sinclair, A.R.E., Smith, J.N.M., Dale, M.R.T., Martin, K., Turkington, R., 1995. Impact of food and predation on the snowshoe hare cycle. Science 269, 1112–1115. Laurance, W.F., Cochrane, M.A., 2001. Synergistic effects in fragmented landscapes– introduction. Conserv. Biol. 15, 1488–1489. Lawler, J.J., White, D., Neilson, R.P., Blaustein, A.R., 2006. Predicting climate-induced range shifts: model differences and model reliability. Glob. Change Biol. 12, 1568–1584. Lazenby-Cohen, K.A., Cockburn, A., 1988. Lek promiscuity in a semelparous mammal, Antechinus stuartii (Marsupialia: Dasyuridae)? Behav. Ecol. Sociobiol. 22, 195–202. Legendre, P., 1993. Spatial autocorrelation: trouble or new paradigm? Ecology 74, 1659–1673. Lewis, S.M., Bain, B.J., Bates, I., Dacie, J.V., 2006. Dacie and Lewis Practical Haematology. Churchill Livingstone, London. Lindenmayer, D., Lacy, R., 2002. Small mammals, habitat patches and PVA models: a field test of model predictive ability. Biol. Conserv. 103, 247–265.

Lindenmayer, D., Possingham, H., Lacy, R., McCarthy, M., Pope, M., 2003. How accurate are population models? Lessons from landscape-scale tests in a fragmented system. Ecol. Lett. 6, 41–47. Mac Nally, R., Bennett, A.F., 1997. Species-specific predictions of the impact of habitat fragmentation: local extinction of birds in the box-ironbark forests of central Victoria, Australia. Biol. Conserv. 82, 147–155. Mac Nally, R., Horrocks, G., 2002. Habitat change and restoration: responses of a forest-floor mammal species to manipulations of fallen timber in floodplain forests. Anim. Biodivers. Conserv. 25, 41–52. Martínez-de La Puente, J., Merino, S., Tomás, G., Moreno, J., Morales, J., Lobato, E., Martínez, J., 2011. Nest ectoparasites increase physiological stress in breeding birds: an experiment. Naturwissenschaften 98, 99–106. Martínez-Mota, R., Valdespino, C., Sánchez-Ramos, M.A., Serio-Silva, J.C., 2007. Effects of forest fragmentation on the physiological stress response of black howler monkeys. Anim. Conserv. 10, 374–379. Mazerolle, D.F., Hobson, K.A., 2002. Physiological ramifications of habitat selection in territorial male ovenbirds: consequences of landscape fragmentation. Oecologia 130, 356–363. Menkhorst, P., Knight, F., 2004. A field guide to the mammals of Australia. Oxford University Press, Melbourne. Nagy, K., Reiczigel, J., Harnos, A., Schrott, A., Kabai, P., 2010. Tree-based methods as an alternative to logistic regression in revealing risk factors of crib-biting in horses. J. Equine Veter. Sci. 30, 21–26. Peig, J., Green, A.J., 2009. New perspectives for estimating body condition from mass/length data: the scaled mass index as an alternative method. Oikos 118, 1883–1891. Peters, R.H., Raelson, J.V., 1984. Relations between individual size and mammalian population density. Am. Nat. 124, 498–517. Prasad, A.M., Iverson, L.R., Liaw, A., 2006. Newer classification and regression tree techniques: bagging and random forests for ecological prediction. Ecosystems 9, 181–199. Quinn, G.P., Keough, M.J., 2002. Experimental Design and Data Analysis for Biologists. Cambridge University Press, Cambridge, UK. Robinson, J.G., Redford, K.H., 1986. Body size, diet, and population density of Neotropical forest mammals. Am. Nat. 128, 665–680. Schulte-Hostedde, A.I., Zinner, B., Millar, J.S., Hickling, G.J., 2005. Restitution of mass–size residuals: validating body condition indices. Ecology 86, 155–163. Segurado, P., Araújo, M.B., Kunin, W.E., 2006. Consequences of spatial autocorrelation for niche-based models. Ecology 43, 433–444. Selwood, K., Mac Nally, R., Thomson, J.R., 2009. Native bird breeding in a chronosequence of revegetated sites. Oecologia 159, 435–446. Short, J., Smith, A., 1994. Mammal decline and recovery in Australia. J. Mammal. 75, 288–297. Smith, B., Wilson, J.B., 1996. A consumer’s guide to evenness indices. Oikos 76, 70– 82. Stewart, A., 1979. Trapping success in relation to trap placement with three species of small mammals, Rattus fuscipes, Antechinus swainsonii and A. stuartii. Wildl. Res. 6, 165–172. Stokes, V.L., Pech, R.P., Banks, P.B., Arthur, A.D., 2004. Foraging behaviour and habitat use by Antechinus flavipes and Sminthopsis murina (Marsupialia: Dasyuridae) in response to predation risk in eucalypt woodland. Biol. Conserv. 117, 331–342. Strobl, C., Boulesteix, A.L., Zeileis, A., Hothorn, T., 2007. Bias in random forest variable importance measures: Illustrations, sources and a solution. BMC Bioinformatics 8, 25. Sumner, J., Dickman, C.R., 1998. Distribution and identity of species in the Antechinus stuartii–A. flavipes group (Marsupialia: Dasyuridae) in southeastern Australia. Aust. J. Zool. 46, 27–41. Suorsa, P., Helle, H., Koivunen, V., Huhta, E., Nikula, A., Hakkarainen, H., 2004. Effects of forest patch size on physiological stress and immunocompetence in an areasensitive passerine, the Eurasian treecreeper (Certhia familiaris): an experiment. Proc. R. Soc. Lond.: Ser. B: Biol. Sci. (Lond.) 271, 435–440. Tews, J., Brose, U., Grimm, V., Tielborger, K., Wichmann, M.C., Schwager, M., Jeltsch, F., 2004. Animal species diversity driven by habitat heterogeneity/diversity: the importance of keystone structures. J. Biogeogr. 31, 79–92. Tilman, D., May, R.M., Lehman, C.L., Nowak, M.A., 1994. Habitat destruction and the extinction debt. Nature 371, 65–66. Turner, I.M., 1996. Species loss in fragments of tropical rain forest: a review of the evidence. J. Appl. Ecol. 33, 200–209. Wingfield, J.C., Sapolsky, R.M., 2003. Reproduction and resistance to stress: when and how. J. Neuroendocrinol. 15, 711–724. Wright, D.H., 1983. Species-energy theory: an extension of species-area theory. Oikos 41, 496–506. Zanette, L., Smith, J.N.M., van Oort, H., Clinchy, M., 2003. Synergistic effects of food and predators on annual reproductive success in song sparrows. Proc. R. Soc. Lond.: Ser. B: Biol. Sci. (Lond.) 270, 799–803. Zuur, A.F., Ieno, E.N., Smith, G.M., 2007. Analysing Ecological Data. Springer Verlag, New York.