ORIGINAL PAPER

Capturing cross-scalar variation of habitat selection with gridsampling: an example with hazel grouse (Tetrastes bonasia L.)

Tommaso Sitzia & Matteo Dainese & Thomas Clementi &Silvano Mattedi

Received: 2 June 2013 /Revised: 28 August 2013 /Accepted: 2 September 2013 /Published online: 6 October 2013# Springer-Verlag Berlin Heidelberg 2013

Abstract The objective of this study was to investigate a grid-based sampling design to determine the cross-scalar selectionof habitat by a territorial animal species: the hazel grouse(Tetrastes bonasia L.). In each of three sites with increasinghazel grouse nest site density, three lattice grids were used tomeasure both the habitat variables and the species occurrencein 100 30×30 m cells. We calculated the average values forhabitat variables, as well as use versus non-use by the species,at three spatial scales: small (1×1 cell), intermediate (2×2cells) and large (3×3 cells). Generalised linear mixed modelswere integrated into a method of variation and hierarchicalpartitioning and used to assess the relationship between thehabitat variables and the species preferences at each scale. Inall scales, species selection was associated with ground layercomposition. Selection was also associated with the composi-tion of the woody layer and negatively associated with

dominance of tor grass (Brachypodium rupestre (Host)Roem. & Schult.) at the two larger scales. Both litter coverand thinning contributed positively to the habitat selection atthe two smaller scales. The other variables were significantonly at one scale or explained a relatively low proportion of thevariation at multiple scales. Neither the management nor thestand structure variables played a significant independent roleacross scales when compared with ground layer variables. Thetotal variation explained was highest (ca. 90 %) at the largescale. This finding indicates the possibility of obtaining cross-scalar hazel grouse preferences from grid-based sampling,provided that spatial autocorrelation in the data is handledappropriately.

Keywords Generalised linear mixedmodel . Habitatselection . Landscape ecology . Spatial scale . Variationpartitioning

Introduction

It is widely acknowledged that the distribution and abundanceof species are influenced by habitat quality, as well as by thespatial arrangement of suitable habitats (Turner 1989), at agrain size proportional to the resolution at which speciesperceive their environment (Kolasa 1989). However, the re-sults depend on both the scale of the data collection and theway the areal units are aggregated into zones or ‘patches’ ofhabitat (Jelinski and Wu 1996). This sensitivity to scale isknown as ‘the Modifiable Areal Unit Problem’ (Openshaw1977). One possible solution to this problem is a hierarchicalapproach that identifies the characteristic scale at which aspecies perceives habitat features and then focuses the studyon these scales (Jelinski andWu 1996). In fact, identifying howobservations change among domains of scale makes it possibleto extrapolate data among scales and improve cross-scalar

Communicated by C. Gortázar

Electronic supplementary material The online version of this article(doi:10.1007/s10344-013-0762-3) contains supplementary material,which is available to authorized users.

T. Sitzia (*) :M. DaineseDepartment of Land, Environment, Agriculture and Forestry,Università degli Studi di Padova, Viale dell’Università 16,35020 Legnaro Padua, Italye-mail: tommaso.sitzia@unipd.it

M. Dainesee-mail: matteo.dainese@unipd.it

T. ClementiDepartment for Nature, Landscape and Spatial Development,Bozen/Bolzano Province, Rittner Straße/Via Renon 4, 39100Bozen/Bolzano, Italye-mail: thomasclementi70@gmail.com

S. MattediStudio Ambiente, Via Marconi 12, 38121 Trento, Italye-mail: simattedi@tin.it

Eur J Wildl Res (2014) 60:177–186DOI 10.1007/s10344-013-0762-3

predictability. In addition, the most significant scale at which toobserve and manage a species-specific or community-specifichabitat features can be identified (Wiens 1989). A true multi-scalar approach should be able to capture the variation associatedwith habitat structure and thus observe the same phenomenon inthe same way at multiple spatial scales (Wheatley and Johnson2009). Other concepts of multiple scales that do not apply aspatial approach, such as the organisation levels of habitatselection, are equally valid but cannot identify the dominant orcharacteristic spatial scales of selection (Mayor et al. 2009).

The landscape-level study of the habitat selection by hazelgrouse (Tetrastes bonasia L.) has received much attention inrecent years because the bird's specialisation for forest environ-ments, poor dispersal ability and site tenaciousness (Jannsonet al. 2004) make it particularly sensitive to changes in land-scape composition (Saari et al. 1998), which are common in theEuropean Alps (e.g. Sitzia and Trentanovi 2011). A singlespatial scale was used in all the available landscape-levelstudies of grouse habitat selection within the alpine environ-ment (Mathys et al. 2006; Muller et al. 2009). To our knowl-edge, only Swenson (1993) and Åberg et al. (2000), workingon hazel grouse within the boreal forest, have addressed morethan one spatial scale. Åberg et al. (2000) conducted a com-prehensive habitat analysis, using multiple replicates of thesame plot size to assess the pattern of habitat selection.Although they expressed average habitat variation in severalways, the plot size uniformity masked changes among scales(Wheatley and Johnson 2009).

In this study, with the objective of extrapolating resultsacross scales, we present the evaluation of a grid-based multi-scalar sampling and analysis. This method allows us to deter-mine the n-order multiple of a micro-site scale at which thespecies responds. We tested the method at the individual levelto answer the following research questions within an arearepresentative of the species' home range: (i) What is therelative importance of management, stand structure and groundlayer for summer habitat selection by individuals in a fine-grained forested landscape of the Alps that is not intensivelymanaged? (ii) Does spatial scale affect the relative importanceof management, stand structure and ground layer factors?

Material and methods

Study area

Our studywas performed in the Trudner HornNature Park. Thepark offers a vast diversity of plant communities, ranging fromsub-alpine spruce woods to sub-Mediterranean coppice woods.The area alternates woodlands with meadows and steppe-likegrasslands. The park (6,866 ha) is located in the province ofBozen, North Italy (46°16′N, 11°18′E). Elevations range from215 to 1,817 m asl. The area belongs to the Mediterranean

mountains environmental zone sensu Metzger et al. (2005).The mean annual temperature is 6–7 °C, and the annual rainfallis 900–1,100mm (Rehder 1965). According to Peer (1995), thepredominant vegetation types are silver fir (Abies alba Mill.)woods, montane spruce (Picea abies L.) woods, and carbonaticand acid Scots pine (Pinus sylvestris L.) woods.

Field surveys

From August 1 to September 30, 2000, a route census of1,000 ha of hazel grouse habitat in the park was conductedby a surveyor and a dog to find all nest site locations and tomonitor their fate. Then, to address issues of accessibility andavailability (Jones 2001), we randomly selected three sites withvarious nest site densities: low (1/km2), medium (3/km2) andhigh (5/km2). During September and October, we sampledthese three sites in each of the 100 30×30 m cells of a 9 hagrid (Fig. 1). The size of the grid was chosen to be represen-tative of the home range (Montadert and Leonard 2004), andthe grain size was chosen according to the size of the feedingsites (Johnson 1980; Sachot et al. 2003).

Within each cell, we drew two orthogonal transects alongwhich we recorded herbaceous species, canopy cover and thepresence/absence of hazel grouse droppings every 1 m. Wealso recorded the stage of stand development every 5 m. Werecorded woody species cover and top height for the entirecell, as well as evidence of cuttings, the presence of red woodant (Formica rufa L.) nests, and mowing and silviculturalpractices (Table 1). It was assumed that the abundance of 1-m segments with droppings would proportionately reflect thelength of time spent in each cell (Bibby et al. 1992). Each cellrequired 20 min to be surveyed, and the grid required 8 daysfor survey by two surveyors.

Data analysis

We used a principal components analysis (PCA) to characterisethe variation of stand development stages (PCA_SDS), tree–shrub composition (PCA_TS) and ground layers (PCA_GL).PCAs were performed separately for each group of variablesand for each spatial scale. We considered principal compo-nents statistically significant only if they had eigenvalues thatexceeded those predicted by the broken stick model (Jackson1993; Legendre and Legendre 1998).

The effects of management, stand structure and groundlayer on the presence and abundance of hazel grouse drop-pings were analysed using generalised linear mixed models(GLMM). In this study, the data are hierarchically structuredin the sense that grid cells are nested within different nestingsites, and we used mixed effects models to account for thespatial dependencies within nesting sites. A suitable mixedeffects model for the purpose can be constructed by introduc-ing a random effect for site into the standard regression

178 Eur J Wildl Res (2014) 60:177–186

(Fieberg et al. 2010; Zuur et al. 2009). The GLMM structurealso incorporates non-normally distributed data through linkfunctions that account for the error in the presence/absence orabundance data; such errors rarely follow a normal distribu-tion. The likelihoods of the models were computed with theLaplace approximation, as suggested by Bolker et al. (2009),in the ‘lme4’ package (Bates and Maechler 2013) for the Rstatistical program (R Development Core Team 2009), assum-ing a binomial error distribution for the presence/absence dataand a Poisson error distribution for the abundance data.Separate GLMM analyses were performed within each of thethree sets of explanatory variables (management, stand struc-ture and ground layer) and for each spatial scale (1×1, 2×2 and3×3 grid cells). To check that the GLMM could adequatelyaccount for the spatial autocorrelation in the residuals, spatialcorrelograms for Moran's I were calculated for models contain-ing only study effort variables and also for the mixed effectsmodels to account for the spatial dependencies within sites (seeSupplementary material). The results suggested that the mixedmodel successfully accommodates spatial autocorrelation withinsites. Given that multicollinearity among explanatory variablescan hamper the identification of the most causal variables(MacNally 2000), the Pearson correlation matrix was performed.In the case of highly correlated variables (r>0.60), only onevariable was used to avoid multicollinearity (see Supplementarymaterial). An assumption of both binomial and Poisson distribu-tions is that the mean and variance are equal, resulting in adispersion parameter (∅ = deviance/residual degrees of freedom)approximating 1. We checked for overdispersion (ϕ > 1) in bothmodels, and we found they were not overdispersed (ϕ <).Modelselection was based on the Akaike's information criterion (AIC).

The AIC method selects models based on the explanatory vari-ables with penalties for over specification. Second-order bias-corrected Akaike's information criterion (AICc) was used forsmall sample sizes. We also calculated the relative probabilityof each model being the best model by calculating their Akaikeweights,wi. Thewi for eachmodel is themodel likelihood valuenormalised to sum to 1 across all R models being considered.This value can be interpreted as the probability that each modelprovides the best fit among all models to explain the observeddata (see Burnham and Anderson 2002). The wi was calculatedusing the ‘AICcmodavg’ R package (Mazerolle 2010).

The significant variables selected from each component inthe GLMM were then further analysed by variation parti-tioning (VP) to determine the unique and joint fractions ofvariation explained by the three sets of explanatory variables(management, stand structure and ground layer) for eachspatial scale. VP is implemented as function ‘varpart’ in the‘vegan’ R package (Oksanen et al. 2010). We report thevariation explained in each model as the adjusted R2 (R2

adj),which takes the number of predictor variables and sample sizeinto account to prevent the inflation of R2 values. When anegative R2

adj was obtained, we interpreted it as a zero value,which means that the fractions from one VP may not alwaysadd up to a perfect 100 % (Peres-Neto et al. 2006).

In addition to variation partitioning, hierarchical partitioning(HP) was also used to determine the relative importance of thevariables most likely to affect variation in the distribution ofhazel grouse (Chevan and Sutherland 1991). The analyses wereperformed by testing all of the significant variables selectedin the GLMM at each spatial scale in a unique model.HP addresses the presence of collinearity by determining the

Fig. 1 Picture summarising a the grid-based sampling design and b the case study

Eur J Wildl Res (2014) 60:177–186 179

independent contribution of each explanatory variable to theresponse variable and separates it from the joint contributionthat results from correlation with other variables. HP wasconducted using the ‘Hier.Part’ R package (Walsh and MacNally 2008). The independent effects were tested using arandomization routine (n =200), which gives Z scores for thegenerated distribution of randomised independent contribu-tions and a level of statistical significance (P) based on thisscore. We used a binomial error distribution for presence/absence data, a Poisson error distribution for abundance dataand log-likelihood as a measure of goodness-of-fit.

Results

In our study, the relative importance of three hazel grouse'shabitat component: management, stand structure and groundlayer was obtained using GLMM. Then, to assess whether

spatial scale affected the relative importance of each significantfactor selected by GLMM, we used VP. Finally, to distinguishwhich of these factors were likely to be most influential incontrolling variation in the species distribution, we used HP.

Importance of the habitat components

The results obtained using GLMM are shown in Table 2. Ascan be seen, the wi of the management and stand structuremodels, both in presence and abundance, was higher at thetwo larger spatial scales (2×2 and 3×3 grid cells), while thatof the ground layer model was similar at the small and medi-um scales (1×1 and 2×2 grid cells) and higher at the largescale (3×3 grid cells).

Within the models with the highest weights, habitat use byhazel grouse was positively associated with respacing (RES),dominance of two-storied stands (TWS) and blanks (BLA),presence of rowan (Sorbus aucuparia L.), aspen (Populus

Table 1 Predictors and their attributes

Variables Code Mathtype

Definition

Management

Recentness of tree cutting TC o 1: no signs of recent cuttings (stools); 2: >3 years ago; 3: 1–3 yearsago; 4: <1 year ago

Mowing MOW b

Type of treatment applied to the stands

Respacing RES b

Thinning ITC b

Final cutting FTC b

Selective cutting (uneven-aged system) STC b

Stand structure

Stand structure homogeneity SDS_H d Maximum number of contiguous 5-m transect segments belonging tothe same development stage

Dominant stand development stage see definition b The following stand development stage types were classified at every5-m transect segment. TWS, two storied; MST, multi-storied; BLA,blank; ADU, adult; POL, pole; SAP, sapling

PCA ordination of stand development stages PCA_SDS c PCs scores on the SDS composition

PCA ordination of tree and shrub layer composition PCA_TS c PCs scores on the composition of the vascular species growing>1 mabove the ground

Tree and shrub layer species richness TS_N c Richness of the abovementioned species

Percentage of deciduous species in TS_N % DEC c 0–100 %

Dominant height HEIGHT c Mean height of the dominant trees (m)

Ground layer

PCA ordination of ground layer species composition PCA_GL c PCs scores on the composition of the plant species growing not morethan 1 m above the ground

Ground layer plant species richness GL_N c Richness of the abovementioned species

Percentage of herbaceous species in GL_N % HERB c 0–100 %

Ground layer homogeneity GL_H d Maximum number of contiguous 1-m transect segments of the sameground layer species or type

Dominant species or type in the grid cell b LIT, leaf litter; WOOD, woody debris; BRA, Brachypodiumrupestre ; ERI, Erica carnea; MOL, Molinia coerulea ; VAC,Vaccinium myrtillus

The mathematical types are classified as o (ordinal), b (binary), d (quantitative discontinuous) and c (quantitative continuous)

Z

180 Eur J Wildl Res (2014) 60:177–186

tremula L.), birch (Betula pendula Roth), spruce and hazel(Corylus avellana L.) in the woody layers (PCA_TS1); bilber-ry (Vaccinium myrtillus L.), cowberry (Vaccinium vitis-idaeaL.), wavy hair-grass (Avenella flexuosa L.) and matgrass(Nardus stricta L.) in the ground layer (PCA_GL1), and withrichness of ground layer species (GL_N). Non-use was asso-ciated with mowing (MOW), a homogeneous distribution ofstand development stage (SDS_H), whitebeam (Sorbus ariaCrantz), snowy mespilus (Amelanchier ovalis Medik.), flyhoneysuckle (Lonicera xylosteum L.), barberry (Berberisvulgaris L.), alpine honeysuckle (Lonicera alpigena L.) andbeech (Fagus sylvatica L.) in the woody layer (PCA_TS1);white sedge (Carex alba Scop.), raspberry (Rubus idaeus L.)and bushgrass (Calamagrostis epigejos (L.) Roth) in theground layer (PCA_GL1), and with dominance of purple moorgrass (Molinia caerulea (L.) Moench) (MOL), tor grass

(Brachypodium rupestre (Host) Roem. & Schult.) (BRA)and woody debris (WOOD).

It can be observed that only some variables were relevant atall the spatial scales: SDS_H and the species scoring on thefirst PCA_GL were always associated with non-use whileTWS and BLA with use. MOL showed contrasting effectsmoving from the small to the two larger scales.

Spatial scale effects on the factors' relative importance

The VP indicated that at the small spatial scale (1×1), onlyground layer variables explained a part of the variation, whichis higher in the presence/absence model. Management had aneffect close to zero, and the stand structure variables were notsignificant (Fig. 2a, d). At the intermediate spatial scale (2×2),the presence/absence model showed that a large proportion of

Table 2 Summary of best-fitGLMM for the presence (P) andabundance (A) of hazel grousedroppings

Management, stand structure andground layer variables were test-ed. Tests of the variables wereperformed separately within eachgroup. The direction of the rela-tionship (− or +) is presented. Themodels are within each separatespatial scale (1×1, 2×2 and 3×3grid cells). Only the subsets ofpredictors included in the modelwith the lowest Akaike's infor-mation criterion (AIC) or second-order bias-corrected Akaike's in-formation criterion (AICc) valueare shown. See Table 1 for ab-breviations of the explanatoryvariablesa AICb Second-order bias-correctedAICc

Model AIC or AICc minus that ofthe best-fit model = 0 in eachresult

wi Model Akaike Weight

1×1 grid cells (n=300) 2×2 grid cells (n=75) 3×3 grid cells (n=27)

P A P A P A

Management

MOW − − −RES − + +

ITC + + +

AIC or AICc 277.49a 389.02a 57.52b 159.27b 13.44b 99.37b

wi 0.38 0.38 0.52 0.57 0.85 0.59

Stand structure

SDS_H − − −TWS + + + +

MS +

BLA + + +

ADU +

POL + +

PCA_SDS1 − −PCA_TS1 − − −PCA_TS2 + +

TS_N + + −% DEC +

AIC or AICc 269.9b 364.04b 47.99b 129.86b 21.6b 82.27b

wi 0.23 0.28 0.46 0.51 0.53 0.70

Ground layer

PCA_GL1 − − − − − −GL_N +

% HERB − − −GL_H −LIT + + +

WOOD + − −MOL + − −BRA − − –

AIC or AICc 257.06b 347.62b 48.44b 136.13b 9.82b 57.00b

wi 0.39 0.43 0.41 0.33 0.78 0.87

Eur J Wildl Res (2014) 60:177–186 181

variation was shared between stand structure and ground layer(28 %), while their pure effect was 7 %. The joint variationshared among the three groups was 15 % (Fig. 2b). Theabundance model (Fig. 2e) indicated different results: the pureeffects weremainly explained by ground layer, with additionalimpacts of joint variation (summing to 10 %). At the largespatial scale (3×3 grid cells), the largest part of the explainedvariation in the presence/absence model was related to thejoint variation between ground layer and stand structure(30 %) and among the three groups (29 %). The second largestfraction of the variation was accounted for by joint variationbetween management and ground layer (14 %); these vari-ables also showed significant pure effects (4 and 11 %, re-spectively) (Fig. 2c). In the abundance model, managementvariables had a pure effect (11 %), while the largest proportionof variation was shared between management and groundlayer (28 %) (Fig. 2f). In general, the presence/absencemodels (19–88 %) had higher explanatory power thanthe abundance models (10–43 %), which might be dueto an incomplete counting of droppings, which aresometimes difficult to distinguish in the undergrowth(Montadert, personal communication). Moreover, at thelarge scale, both of the models explained more of the

variation (88 and 43 %) than they did at the small scale(19 and 10 %).

Factors controlling variation in the species distribution

The results of the HP analyses generally reflected thoseascertained by the GLMM and VP analyses but revealedslightly different results concerning the relative importanceand significance of some of the variables (Fig. 3). At the smallspatial scale, PCA_GL1 was confirmed as an important vari-able explaining large fractions of variation (50–80 %) in bothof the models, while other explanatory variables producedsmall significant independent effects (Fig. 3a, d). At the inter-mediate scale, PCA_TS1 (35 %) and PCA_GL1 (32 %) madelarge independent contributions in the presence/absence mod-el (Fig. 3b), while PCA_GL1 (33 %) and LIT (22 %) were themore important variables in the abundance model (Fig. 3e).Additional important primary variables with relatively largeindependent effects were ITC, BRA and the percentage ofdeciduous woody species (% DEC) (Fig. 3b, e). At the largescale, PCA_GL1, PCA_TS1, RES and BRA had the strongestindependent contributions (15–30 %) in the presence/absencemodel (Fig. 3c). In the abundance model, BRA, RES and

Fig. 2 Venn diagrams describing variation partitioning of the GLMMmodels explaining the presence (a–c) and abundance (d–f) of hazelgrouse droppings within the separate spatial scales of the a , d 1×1 gridcell, b , e 2×2 grid cells and c , f 3×3 grid cells. Variation is explained bypure management [a ], stand structure [b ] and ground layer [c ]

components and their joint effects [d–g]. Values are adjusted R2 inpercentage. Null and negative fractions (always <1 %) are not shown.Adjusted fractions of total variation explained (TVE , in %) were estimat-ed following the procedure of Peres-Neto et al. (2006)

182 Eur J Wildl Res (2014) 60:177–186

PCA_GL1 produced the greatest independent effects (20–25 %) (Fig. 3f). The 2×2 and 3×3 models confirmed that arelatively large part of the explained variation was related tothe joint effects of the explanatory variables.

Discussion

Prior work has documented a general difficulty in detectinghabitat selection by hazel grouse at different scales. Åberg et al.(2000), for example, report that the only effect that was true atall scales, seasons and densities was that grouse avoided standsdominated by Scots pine. They began at 150-m interval censuspoint locations and included more 10-m radius sampling plotsof fine-scale habitat variables throughout four larger areas.However, to scale up and quantify habitat over larger scales,this study used multiple replicates of the same plot size, whichprecluded the authors from capturing the variation amongscales (Wheatley and Johnson 2009).

In this study, we tested the extent to which a multi-scalargrid-based approach improved the knowledge of habitat useby hazel grouse across scales. We used a grain size compara-ble to the typical size of feeding areas, and then, we multipliedthe extent twice, over a continuous lattice grid, to approxi-mately 1/20–1/40 of the size of a normal minimum convexpolygon hazel grouse home range in the study region. First,

we studied the variation in single habitat variables at the threescales separately with GLMM; second, we partitioned theirvariance into groups considering the three scales together; andthird, we hierarchically split the variation explained by eachhabitat variable into both a joint and an independent effect ateach scale.

We found that in all scales, habitat selection was associatedwith ground layer composition, while woody layer compositionand the absence of dominant cover by tor grass were importantat both of the two larger scales. Litter cover and thinningpositively contributed to habitat use at both of the two smallerscales. The other variables were important only at a single scaleor explained a relatively low proportion of the variation atmultiple scales, such as the composition and homogeneity ofstand structure or showed opposite signs depending on scale,like the dominance of purple moor grass. In addition, the vari-ation partitioning revealed that only the ground layer group ofvariables had a substantial effect across all scales independently,while the other two groups, management and stand structure,showed relevant effects only jointly or with both of the othergroups and only across the two larger scales. The total explainedvariation reached a remarkably high value of ca. 90 % at thelarge scale, while this value was quite low at the feeding itemsscale.

Many of the abovementioned habitat variables were signifi-cantly related to hazel grouse presence or absence, similar to the

TWS

TWS

Fig. 3 The independent and joint contributions (given as percentages ofthe total explained variation) of each explanatory variable for the presence(a–c) and abundance (d–f) of hazel grouse droppings within the separatespatial scales of the a , d 1×1 grid cell, b , e 2×2 grid cells and c , f 3×3grid cells, as estimated from hierarchical partitioning. Black bars

indicated independent effects; white bars, joint effects. Significant (P<0.05) independent effects are indicated by asterisks , resulting from the z-randomization procedure (n =200). The models are shown and are rankedaccording to decreasing independent effect. See Table 1 for abbreviationsof the explanatory variables

Eur J Wildl Res (2014) 60:177–186 183

results found in earlier studies. For instance, field layer, well-structured forests, dead wood, deciduous tree species and spruceproportion have each been found to play a role in previousstudies (Åberg et al. 2003; Eiberle and Koch 1975; Jannsonet al. 2004; Mathys et al. 2006; Muller et al. 2009). However,our general cross-scale findings contrast with those of Åberget al. (2000) in several aspects. The approach they used to studythe habitat variability within home range differed from ours inthat they adopted a multiple design study rather than a multi-scalar approach, where five concentric zones around speciesoccurrences were used to assess the grouse's preference of 76habitat variables surveyed around a fixed 10-m radius. Wesuspect that they did not find any clear patterns of habitatpreferences within the species' home range because they collect-ed habitat data at a too-small scale: they used only a single circa300 m2 (10-m radius) scale, which corresponds to one third ofour smallest scale. However, despite this absence of clear pat-terns, their results, even if at a different scale, are partiallycoincident with ours because according to these authors, hazelgrouse presence was positively related to shrubiness and fieldlayer and negatively related to the amount of Scots pine. Thesame lack of significant habitat variables characterises the workofMontadert (2005), who associated the animal localisation to asingle fixed 10-m radius vegetation relevé.

In addition, some variables and groups of variables weresignificantly preferred or avoided atmultiple scales, while otherswere only preferred or avoided at one or two scales. Our study,therefore, indicates that the use of a grid-based sampling designwhere grain and size are chosen according to the species ecologymay address species requirements across n-order multiples ofscale with a reasonable survey effort.

Management implications

To our knowledge, this is the first study to investigate the mostappropriate scale at which the hazel grouse habitat can beobserved and managed. Our results indicate that neither man-agement nor stand structure plays a significant independent roleacross scales when compared with ground layer variables.However, considered jointly, the two sets of explanatory vari-ables contribute substantially to the explanation of habitat pref-erence at the individual level. Ground layer seems also to beconnected to stand structure and some connections of standstructure to management was observed at the intermediate scale.In general, this suggests that ground layer variation is spatiallystructured by management actions, especially at the large scale.

Some variables and their interconnections became clearonly at specific scales. For example, mowing is not relatedto tor grass at the small scale, but it is related at the large scale.Additionally, management becomes substantially importantonly at the large scale. Operationally, this result means thatmowing becomes detrimental only when applied on surfaces

larger than approximately 1 ha (3×3 scale). Grouse alwaysavoid areas dominated by multi-storied structure, while openareas (blanks) are only avoided at the largest scale, when theyextend more than approximately half a hectare. Similarly,thinning was a positive practice when applied on a surfacesmaller than a half hectare. These results are in general agree-ment with previous evidence of hazel grouse avoiding areas ofopen land more than 200 m wide (Åberg et al. 1995). Thesefindings are not, in fact, contradictory to those of Åberg et al.(2003), who identified un-thinned stands as favourable habi-tats; we simply defined the scale at which hazel grouse per-ceive the habitat features related to thinning.

Conclusions and further research

A grid-based sampling design in which the basic unit corre-sponds to the typical size of the feeding area and the largest unitto the size of the home range could be a promising approach tostudy cross-scalar habitat selection at the individual level incontiguous landscapes because hazel grouse is a territorialspecialist that has limited movement capacity (With and Crist1995). Both the habitat variables and the species occurrencemust be surveyed at each of the cells in a continuous lattice gridto capture the variation of habitat preference in a continuum.One condition that must be met is that such studies shouldemploy appropriate analytical methods to account for non-normal distributions and spatial autocorrelation (Zuckerberget al. 2012). Our results are encouraging and should be validat-ed in a larger series of years, seasons and with different curren-cies of use and levels of organisation (Buskirk and Millspaugh2006; Mayor et al. 2009).

Normally, habitat selection studies must both design arandom sampling of the resource distribution according tothe heterogeneity of the sampled area and provide an adequatecoverage of the area (Buskirk and Millspaugh 2006). Thisstep, which generates bias (Gu and Swihart 2004), is avoidedin our method because we surveyed use versus non-use ineach cell of a sampling grid superimposed on areas availablefor use, and we adopted the single grid as a random effect toaccount for any spatial dependence within sites. In this ap-proach, the location of the grids must be restricted to withinregions of concentrated use, thereby assuring habitat accessi-bility (Jones 2001). Moreover, the application of such a gridallows for the comparison of results with many other availablespatial statistical methods, such as spatial eigenvector map-ping (Dormann et al. 2007).

Further research should address the replication of the singlelevels of use intensity in different landscape types and/orregions, with the aim of considering interactions among land-scape types, habitat variables and species density (Thomasand Taylor 2006).

184 Eur J Wildl Res (2014) 60:177–186

Acknowledgments We are grateful to Marc Montadert and ArianeBernard Laurent for useful comments on an earlier version of the manu-script. The research was financed by the Department for Nature, Land-scape and Spatial Development of the Bozen Province, Italy. We thank, inparticular, officials Renato Sascor and Rainer Ploner for promoting andfollowing the research and director Artur Kammerer for sustaining it.

References

Åberg J, Jannson G, Swenson JE, Mikusinski G (2000) Difficulties indetecting habitat selection by animals in generally suitable habitats.Wildl Biol 6:89–99

Åberg J, Jansson G, Swenson JE, Angelstam P (1995) The effect ofmatrix on the occurrence of hazel grouse (Bonasa bonasia) inisolated habitat fragments. Oecologia 103:265–269

Åberg J, Swenson JE, Angelstam P (2003) The habitat requirements ofhazel grouse (Bonasa bonasia) in managed boreal forest and appli-cability of forest stand descriptions as a tool to identify suitablepatches. For Ecol Manage 175:437–444

Bates D, Maechler M (2013) lme4: linear mixed-effects models using S4classes. R package version 0.999999-2. http://cran.r-project.org/web/packages/lme4/index.html. Accessed 19 Sept 2013

Bibby CJ, Burgess ND, Hill DA (1992) Bird census techniques. Aca-demic Press Limited, London

Bolker BM, Brooks ME, Clark CJ, Geange SW, Poulsen JR, StevensMH, White JS (2009) Generalized linear mixed models: a practicalguide for ecology and evolution. Trends Ecol Evol 24:127–135

Burnham KP, Anderson DR (2002) Model selection and multi-modelinference: a practical information-theoretic approach, 2nd edn.Springer-Verlag, New York

Buskirk SW, Millspaugh JJ (2006) Metrics for studies of resource selec-tion. J Wildl Manage 70:358–366

Chevan A, SutherlandM (1991) Hierarchical partitioning. Am Stat 45:90–96Dormann CF, McPherson JM, Araujo MB, Bivand R, Bolliger J,

Carl G, Davies RG, Hirzel A, Jetz W, Kissling WD, Kuhn I,Ohlemuller R, Peres-Neto PR, Reineking B, Schroder B, SchurrFM, Wilson R (2007) Methods to account for spatial autocor-relation in the analysis of species distributional data: a review.Ecography 30:609–628

Eiberle K, Koch N (1975) Die Bedeutung der Waldstruktur für dieErhaltung des Haselhuhns. Schweiz Z Forst 126:876–888

Fieberg J, Matthiopoulos J, Hebblewhite M, Boyce MS, Frair JL (2010)Correlation and studies of habitat selection: problem, red herring oropportunity? Philos T Roy Soc B 365:2233–2244

Gu WD, Swihart RK (2004) Absent or undetected? Effects of non-detection of species occurrence on wildlife-habitat models. BiolConserv 116:195–203

Jackson DA (1993) Stopping rules in principal component analysis: acomparison of heuristical and statistical approaches. Ecology 74:2204–2214

Jannson G, Angelstam P, Åberg J, Swenson JE (2004) Managementtargets for the conservation of hazel grouse in boreal landscapes.Ecol Bull 51:259–264

Jelinski D, Wu J (1996) The modifiable areal unit problem and implica-tions for landscape ecology. Landscape Ecol 11:129–140

Johnson DH (1980) The comparison of usage and availability measure-ments for evaluating resource preference. Ecology 61:65–71

Jones J (2001) Habitat selection studies in avian ecology: a critical review.Auk 118:557–562

Kolasa J (1989) Ecological-systems in hierarchical perspective—breaksin community structure and other consequences. Ecology 70:36–47

Legendre P, Legendre L (1998) Numerical ecology, 2nd edn. Elsevier,Amsterdam

MacNally R (2000) Regression and model-building in conservationbiology, biogeography, and ecology: the distinction between—andreconciliation of—‘predictive’ and ‘explanatory’ models. BiodiversConserv 9:655–671

Mathys L, Zimmermann NE, Zbinden N, Suter W (2006) Identifyinghabitat suitability for hazel grouse Bonasa bonasia at the landscapescale. Wildl Biol 12:357–366

Mayor SJ, Scheider DC, Schaefer JA, Mahoney SP (2009) Habitatselection at multiple scale. Ecoscience 16:238–247

Mazerolle MJ (2010) AICcmodavg: model selection and multimodelinference based on (Q)AIC(c). R package version 1.07, 1.07 ed.http://cran.r-project.org/web/packages/AICcmodavg/index.html.Accessed 30 July 2010

Metzger MJ, Bunce RGH, Jongman RHG, Mucher CA, Watkins JW(2005) A climatic stratification of the environment of Europe. Glob-al Ecol Biogeogr 14:549–563

Montadert M (2005) Fonctionnement démographique et sélection del'habitat d'une population en phase d'expansion géographique. Casde la Gélinotte des bois dans les Alpes du Sud, France. Universite deFranche-Comte, Besançon

Montadert M, Leonard P (2004) First results of a hazel grousepopulation study in the south-eastern French Alps. Grouse News28:15–20

Muller D, Schroder B, Muller J (2009) Modelling habitat selection of thecryptic hazel grouse Bonasa bonasia in a montane forest. J Ornithol150:717–732

Oksanen J, Blanchet FG, Kindt R, Legendre P, O'Hara RB, Simpson GL,Solymos P, Stevens MHH, Wagner H (2010) vegan: CommunityEcology Package. R package version 1.17-3, 1.17-3 ed.

Openshaw S (1977) Optimal zoning systems for spatial interactionmodels. Environ Plann A 9:169–184

Peer T (1995) La vegetazione naturale dell'Alto Adige. Note illustrative dellacarta della vegetazione naturale 1: 200.000. Ufficio pianificazionepaesaggistica, Bozen (Italy).

Peres-Neto PR, Legendre P, Dray S, Borcard D (2006) Variationpartitioning of species data matrices: estimation and comparison offractions. Ecology 87:2614–2625

R Development Core Team (2009) R: a language for statisticalcomputing. R Foundation for Statistical Computing, Wien(Austria)

Rehder H (1965) Die Klimatypen der Alpenkarte im Klimadiagramm-Weltatlas (Walter-Lieth) und ihre Beziehungen zur Vegetation. FloraB 156:78–93

Saari L, Aberg J, Swenson JE (1998) Factors influencing the dynamics ofoccurrence of the hazel grouse in a fine-grained managed landscape.Conserv Biol 12:586–592

Sachot S, Perrin N, Neet C (2003) Winter habitat selection by two sym-patric forest grouse in western Switzerland: implications for conser-vation. Biol Conserv 112:373–382

Sitzia T, Trentanovi G (2011) Maggengo meadow patches enclosed byforests in the Italian Alps: evidence of landscape legacy on plantdiversity. Biodivers Conserv 20:945–961

Swenson JE (1993) The importance of alder to hazel grouse inFennoscandian boreal forest—evidence from 4 levels of scale.Ecography 16:37–46

Thomas DL, Taylor EJ (2006) Study designs and tests for com-paring resource use and availability II. J Wildl Manage 70:324–336

Turner MG (1989) Landscape ecology: the effect of pattern on process.Annu Rev Ecol Syst 20:171–197

Walsh C, Mac Nally R (2008) hier.part: hierarchical partitioning. Rpackage version 1.0-3., 1.0-3 ed. http://crantastic.org/packages/hier-part/versions/1922. Accessed 19 Sept 2013

Wheatley M, Johnson C (2009) Factors limiting our understanding ofecological scale. Ecol Complex 6:150–159

Wiens JA (1989) Spatial scaling in ecology. Funct Ecol 3:385–397

Eur J Wildl Res (2014) 60:177–186 185

With KA, Crist TO (1995) Critical thresholds in species’ responses tolandscape structure. Ecology 76:2446–2459

Zuckerberg B, Desrochers A, Hochachka WM, Fink D, Koenig WD,Dickinson JL (2012) Overlapping landscapes: a persistent, but

misdirected concern when collecting and analyzing ecological data.J Wildl Manage 76:1072–1080

Zuur AF, Ieno EN,Walker NJ, SavelievAA, SmithGM (2009)Mixed effectmodels and extensions in ecology with R. Springer-Verlag, Berlin

186 Eur J Wildl Res (2014) 60:177–186