7
8
9
10
11
12
13
14
15
16
17
18
5
6
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
3
4
1
2
By Craig DeMars, Robert Serrouya, and Tyler Muhly
Caribou Spatial Data
We used GPS location data from 152 adult female caribou distributed among eight independent herds: East Side of the Athabasca River (ESAR), West Side of the Athabasca River
(WSAR), Cold Lake, Slave Lake, Nipisi, Richardson, Red Earth and Chinchaga. These data were collected by Alberta Environment and Sustainable Resource Development (AESRD),
Government of Alberta, during their caribou monitoring program from 1998 to 2012. Prior to analyses, the data were screened to remove all locations with low positional accuracy (e.g. < three-dimensional fixes and a horizontal dilution of precision > 12, M. Russell, Government of
Alberta, pers. comm.; Lewis et al.
2007). We also used VHF location data collected during
1991-2012 from 606 individual adult female distributed among all northern Alberta boreal caribou herds. Positional accuracy of the VHF data was estimated to be < 250-m and this data set totalled 4,098 locations.
Sampling Framework
We evaluated for seasonal differences in resource selection by first partitioning the data into two seasons: summer, from May 15th to September 30th and winter, from October 1 to May
14. The summer data consisted of 280,970 GPS locations from 123 individuals with the number of locations per individual varying from 27 to 10,394. The winter data set had 327,070 GPS locations from 150 individuals with the number of locations per individual varying from 50 to
12,212. To assess how female caribou selected resources within a herd’s range (i.e. second-order selection, sensu
Johnson 1980), we compared caribou GPS locations (i.e. “used” locations) to randomly generated locations within herd ranges (i.e. “available” locations). For each GPS location, we generated one random location within the range of the individual's herd.
Resource Variables
We modelled resource selection using explanatory variables representing vegetation characteristics (cover type and normalized difference vegetation index [NDVI]), fire history and anthropogenic footprint. NDVI is a measure of green vegetation in the landscape and has been used as an index of forage quantity and/or quality for ungulates. For land cover, we used remotely sensed vegetation cover data from the Alberta Biodiversity Monitoring Institute (ABMI
Remote Sensing Group 2012). This data indicates dominant vegetation cover type (categories: conifer forest (lc_confor), broadleaf forest (lc_decfor), mixed forest(reference category),

63
64
65
66
67
68
69
70
71
57
58
59
60
61
62
42
43
44
45
46
47
37
38
39
40
41
48
49
50
51
52
53
54
55
56
72
73
74
75 shrubland (lc_shrub), grassland (lc_grass), water, and other (lc_other2), which included agriculture, exposed land, rock/rubble, developed and snow/ice) derived from Landsat data collected in year 2000 at 0.5 ha to 2.0 ha spatial resolution. To specifically model wetlands, an important land cover for caribou, we used Alberta peatland inventory data (Halsey et al. 2003).
This data contains information about wetland class, vegetation and landform mapped as polygons from aerial photos. For our analysis, we identified polygons with >50% bog as bogs
(bog) and >50% fen as fens (fen). NDVI was measured from MODIS satellites at 16 day intervals at a 250 m pixel resolution
(https://lpdaac.usgs.gov/products/modis_products_table/myd13q1). Average NDVI values were calculated at each pixel for each season (ndvi_act). We modelled fire history using mapped fire inventory polygons (Alberta Sustainable Resource Development 2012), which we further classified into fires ≤40 years old (recent burns) and fires >40 years old (fire40). To characterize anthropogenic footprint, we measured distance of each location to five different human footprint types: industrial (i.e., polygonal industrial sites; d_ind_act), hard linear (i.e., roads), soft linear
(i.e., seismic and pipelines; d_sft_act), agriculture (i.e., cultivated and pasture lands) and forestry cutblock (d_cut_act). Footprint was obtained from the ABMI, which mapped the location of all footprint across Alberta in 2007 and 2010 using data from a variety of sources. Caribou locations from 2008 and earlier were attributed to 2007 human footprint data while caribou locations from
2009 and later were attributed 2010 human footprint data.
Data Analysis
We evaluated caribou resource selection by developing resource selection functions
(RSFs; Manly et al.
2002). Prior to RSF development, we assessed for collinearity among resource variables by calculating Pearson correlation coefficients. For those variables that were highly correlated ( r > 0.6) we removed the covariate least correlated with used locations. We also standardized all continuous variables prior to model fitting to facilitate comparison of relative effect sizes among resource variables.
We estimated RSFs using generalized linear mixed effect models (GLMMs; Zuur et al.
2009), which account for the hierarchical structure inherent in GPS location data. In all GLMMs we assigned individual caribou as a random grouping effect, which creates a random intercept for each caribou. We further developed two model structures that accounted for functional responses in selection (i.e., differential selection as a function of resource availability; Mysterud
& Ims 1998). For the first, we specified random slopes for those variables identified as having a functional response (see below). This model structure has the form
2006) g(x) = β
0
+ βx ij
+ ... + β n x nij
+ γ nj x nj
+ γ j
(Gillies et al.
where γ j
is the random intercept for caribou j and γ nj
is the random slope (or coefficient) for caribou j with respect to covariate x n
. The fixed-effects, or marginal, coefficients (βx ij
) yield population-level inferences that can be interpreted within the classic use-availability design of
ω(x i
) = exp(β
1 x
1
+β
2 x
2
+ ...β n x n
) (Manly et al. 2002)

94
99
100
101
102
103
104
95
96
97
98
87
88
89
90
91
92
93
83
84
85
86
76
77
78
79
80
81
82
105
106
107
108
109
110
111
112
113
114
115
116 where ω(x i
) is the relative selection value of a sample unit (or pixel) in category i as a function of the explanatory covariates. For the second model structure, we developed generalized functional response models (GFRs; Matthiopoulos et al.
2011). GFRs account for functional responses in selection by including additional covariates describing resource availability for those resources identified as having a functional response. We assessed the performance of GFR variables in
GLMMs that included random-intercepts only as well as in models that specified random slopes.
Thus, in total we evaluated four RSF models for each season: i.
A random-intercept only GLMM without GFR variables ii.
A random-intercept only GLMM with GFR variables iii.
A random-intercept and random slope GLMM without GFR variables iv.
A random-intercept and random slope GLMM with GFR variables
To identify covariates to include as functional responses, we first estimated individual
RSFs by calculating logistic regression models with all covariates for each individual caribou included in the model. We then plotted the relationship between the availability of each covariate and the selection coefficient to determine if there was a functional response. A functional response was identified as a relationship that could be defined by a linear or quadratic equation with an R
2
value > 0.1 and crossed zero (i.e., went from selection to no selection or vice-versa).
RSF Model Performance and Validation
We evaluated model fit and parsimony of the four seasonal RSF models using Bayesian
Information Criterion (BIC) scores. We validated each RSF model with the independently collected caribou VHF locations following a method similar to k -fold cross-validation (Boyce et al.
2002). Specifically, we measured each RSF model score at each VHF location then binned them into ten bins with more-or-less equal sample sizes. We then divided the number of locations by the total area of that bin in the RSF map (i.e., the area-adjusted frequency). Finally, we calculated a Spearman correlation coefficient ( r
S
) between the area-adjusted frequency and bin rank. Models with higher predictive power will have larger correlation coefficients. We identified the best seasonal model for describing caribou resource selection by comparing by model fit and parsimony (BIC scores) with predictive performance ( r
S
score).
The RSF model indicated that caribou selected vegetation with intermediate NDVI values in summer, but selected higher NDVI values in winter. Caribou selected fen and bog wetland types over other wetland types, selected conifer forest (lc_confor) and shrub (lc_shrub) landcover types, and avoided deciduous forest (lc_decfor), water (lc_water) and other (lc_other2) landcover types relative to mixed forest landcover types in the in summer and winter. Caribou also selected grassland (lc_grass) cover types in the winter, but not in the summer. In both seasons, caribou avoided burns less than 40 years old (fire40). Caribou selected soft linear features (d_sft_act, e.g., seismic lines and pipelines) and avoided forestry cutblocks (d_cut_act) in the summer and winter. Caribou avoided industrial features (d_ind_act) in the summer but selected them in the winter.

117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
Results of the RSF model are consistent with other second-order (i.e., home range) boreal caribou habitat selection studies. Boreal caribou typically avoid habitats expected to have higher predation risk, i.e., higher vegetation food quality and quantity habitats that support more ungulates and thus predators. Therefore caribou typically select home ranges consisting of low productivity, older conifer forest and peatlands (Bradshaw et al. 1995; Rettie and Messier 2000;
McLoughlin et al. 2005; Courbin et al. 2009; Moreau et al. 2012), which is consistent with our findings. Consequently, caribou typically avoid forestry activity that creates early seral forests
(Smith et al. 2000; DeCesare et al. 2012), which is also consistent with our findings. The selection of soft linear features that we found is counter to the theory that caribou may avoid linear features to mitigate predation risk. However, other studies of woodland caribou have found that avoidance of linear features may occur at smaller scales (i.e., within home range) then our study (DeCesare et al. 2012).
For the summer season, the predictive performance was highest for the full model, random intercept, but with no GFR ( r
S
= 0.70). This model had a predictive performance that was 9 units higher than the next best model (i.e. the random slope model that included a GFR). In winter however, 3 models had similar predictive performance scores: the random intercept full model with no GFR ( r
S
= 0.82), the full model with the GFR ( r
S
= 0.82), and random slope full model with the GFR ( r
S
= 0.83). For parsimony and for consistency we chose to use the same model structure as the summer model: i.e. the full model with random intercept and no GFR. Model details including parameter estimates are provided below.

138
139
140
Variable
(Intercept) ndvi_act_s ndvi_act_s_sq fen bog lc_other2 lc_water lc_shrub lc_grass lc_confor lc_decfor fire40 d_sft_act d_ind_act d_cut_act d_sft_E ndvi_act_s:d_sft_E ndvi_act_s_sq:d_sft_E d_sft_act:d_sft_E d_ind_act:d_sft_E d_cut_act:d_sft_E
Validation [r] area-adj freq
Summer model: Yellow indicates the model that was used for the RSF mapping
Random Intercept Models
Full Model, No GFR
(BIC=626771; r af
=0.96; r ea
=94)
Estimate
-0.72
4.20
-4.36
1.44
1.63
-1.16
-3.74
0.08
0.02
0.16
-0.97
-0.81
-0.18
0.18
0.04
SE
0.08
0.05
0.04
0.01
0.01
0.04
0.07
0.01
0.02
0.01
0.02
0.01
0.01
0.01
0.01
Full Model, GFR (BIC =
624142, r af
=0.73, r ea
=94)
Estimate
-0.96
SE
0.06
8.86
-8.28
1.44
1.62
0.12
0.10
0.01
0.01
-1.13
-3.52
0.08
0.002
0.13
-0.93
-0.82
-0.36
0.28
-0.005
-0.13
-2.47
2.03
0.11
-0.09
0.09
0.01
0.07
0.07
0.07
0.01
0.01
0.01
0.04
0.07
0.01
0.02
0.01
0.02
0.01
0.01
0.01
Veg Covariates Only (BIC
= 627343)
Estimate
-0.71
4.16
-4.33
1.44
1.63
-1.18
-3.75
0.09
0.02
0.15
-0.96
-0.80
SE
0.08
0.05
0.04
0.01
0.01
0.04
0.07
0.01
0.02
0.01
0.02
0.01
Human Covariates Only
(BIC = 769889)
Estimate
-0.03
SE
0.04
-0.06
-0.01
0.26
-0.04
-0.09
0.02
-0.02
0.01
0.01
0.01
0.05
0.01
0.01
0.01
0.96
0.73
Full Model, no GFR (BIC =
599561, r af
=0.70, r ea
=0.95)
Estimate
-0.79
SE
0.16
4.19
-4.38
1.46
1.62
0.05
0.05
0.01
0.01
-1.19
-4.00
0.12
0.04
0.17
-0.96
-0.82
-0.23
0.19
0.07
0.04
0.08
0.01
0.02
0.01
0.02
0.01
0.54
0.01
0.01
Random Slope Models
Full Model, GFR (BIC =
599561, r af
=0.87,
0.02
0.01
0.56
0.02
0.01
0.18
0.07
0.07
0.65
0.01
0.01
SE
0.16
0.12
0.10
0.01
0.01
0.04
0.08
0.01
0.02
0.01
r ea
=0.95)
Estimate
-0.99
8.84
-8.29
1.45
1.61
-1.17
-3.78
0.12
0.02
0.15
-0.93
-0.83
-0.10
0.13
0.02
-0.55
-2.42
1.93
-0.58
0.02
0.17
0.7
0.87
Human Covariates
Only (BIC = 741050)
Estimate
-0.14
SE
0.14
-0.30
-0.13
0.30
-0.31
-0.42
0.13
0.05
0.54
0.01
0.01
0.16
0.64
0.01
0.01
GFR means generalized functional response

141
142
143
Variable
(Intercept) ndvi_act_w ndvi_act_w_sq fen bog lc_other2 lc_water lc_shrub lc_grass lc_confor lc_decfor fire40 d_sft_act d_ind_act d_cut_act d_sft_E ndvi_act_s:d_sft_E ndvi_act_s_sq:d_sft_E d_sft_act:d_sft_E d_ind_act:d_sft_E d_cut_act:d_sft_E
Validation [r] area-adj freq
Winter model: Yellow indicates the model that was used for the RSF mapping
Full Model, No GFR
(BIC=682888; r af
=0.82, r ea
=0.95)
Random Intercept Models
Full Model (BIC =
679758; r af
=82; r ea
=0.96)
Veg Covariates Only
(BIC = 690412)
Human Covariates
Only (BIC = 888321)
Estimate
-1.05
-2.73
3.51
1.73
1.75
-0.29
-2.83
0.65
0.50
0.45
-1.00
-1.27
-0.26
-0.15
0.24
SE
0.08
0.02
0.02
0.01
0.01
0.03
0.04
0.01
0.02
0.01
0.02
0.01
0.01
0.01
0.01
0.46
1.51
-1.82
-0.08
0.02
-0.05
0.44
-1.03
-1.24
-0.09
-0.18
0.30
Estimate
-1.18
-2.76
3.49
1.71
1.73
-0.31
-2.96
0.59
0.46
0.08
0.03
0.05
0.01
0.01
0.01
0.01
0.02
0.01
0.01
0.01
0.01
SE
0.08
0.02
0.02
0.01
0.01
0.03
0.04
0.01
0.02
Estimate
-1.06
-2.65
3.43
1.77
1.79
-0.23
-2.83
0.73
0.56
0.48
-0.96
-1.37
SE
0.08
0.02
0.02
0.01
0.01
0.03
0.04
0.01
0.02
0.01
0.02
0.01
Estimate SE
-0.03
0.04
-0.06
-0.01
0.26
-0.04
-0.09
0.02
-0.02
0.01
0.01
0.01
0.05
0.01
0.01
0.01
0.82
0.82
Full Model, no GFR (BIC =673074
, r af
=0.78, r ea
=0.98)
Random Slope Models
Full Model (BIC =
670543, r af
=0.83 r ea
=0.98)
Human Covariates
Only (BIC = 874892)
Estimate
-1.19
-2.75
3.54
1.75
1.74
-0.30
-2.81
0.64
0.48
0.45
-1.00
-1.23
-0.97
-0.12
0.24
SE
0.11
0.02
0.02
0.01
0.01
0.03
0.04
0.01
0.02
0.01
0.02
0.01
0.29
0.01
0.01
0.60
1.45
-1.76
0.13
0.04
-0.07
0.44
-1.03
-1.22
-0.95
-0.17
0.29
Estimate SE
-1.37
-2.76
3.49
0.10
0.02
0.02
1.73
1.73
-0.32
-2.89
0.59
0.44
0.01
0.01
0.03
0.04
0.01
0.02
0.11
0.03
0.05
0.34
0.01
0.01
0.01
0.02
0.01
0.31
0.01
0.01
Estimate SE
-0.25
0.07
-1.16
-0.32
0.33
0.30
0.22
0.11
-0.10
0.31
0.01
0.00
0.08
0.34
0.01
0.00
0.78
0.83

144
Name
Normalized Difference Vegetation Index (NDVI) squared NDVI value
Polygons with proportions of fen >50%
Polygons with proportions of bog>50%
Other
Water
Shrubland
Grassland
Conifer forest
Deciduous forest
Fires less than 40 years old
Euclidean distance to nearest soft linear feature
Euclidean distance to nearest industrial site
Euclidean distance to nearest cut block
Expected Euclidean distance to nearest soft linear feature
145
146
Description
NDVI value averaged over the season. Based on MODIS NDVI data calculated over a 16-day interval at a resolution of 250-m see above
Wetland classification GIS layer from Alberta Peatlands Inventory data
Wetland classification GIS layer from Alberta Peatlands Inventory data
Vegetation class from ABMI remotely sensed vegetation cover map (raster). Combines categories of agriculture, development, exposed land, rock/rubble, and snow/ice.
Vegetation class from ABMI remotely sensed vegetation cover map (raster)
Vegetation class from ABMI remotely sensed vegetation cover map (raster)
Vegetation class from ABMI remotely sensed vegetation cover map (raster)
Vegetation class from ABMI remotely sensed vegetation cover map (raster)
Vegetation class from ABMI remotely sensed vegetation cover map (raster)
Polygon GIS layer of mapped fire inventory from Alberta SRD
Soft linear features defined as seismic lines or pipelines. Derived from ABMI human footprint data.
Derived from ABMI human footprint data.
Derived from ABMI human footprint data.
Expected (or mean) distance to nearest soft linear feature calculated in a 51.5 km moving window analysis at a resolution of 30-m. 51.5 km is the mean radius of the caribou ranges considered.
ABMI = Alberta Biodiversity Monitoring Institute: abmi.ca

159
160
161
162
163
164
165
166
153
154
155
156
157
158
147
148
149
150
151
152
173
174
175
176
177
178
179
167
168
169
170
171
172
Boyce, M.S., Vernier, P.R., Nielsen, S.E. & Schmiegelow, F.K. (2002). Evaluating resource selection functions. Ecological Modelling, 157, 281–300.
Bradshaw, C. J., Boutin, S., Hebert, D. M., & Rippin, A. B. (1995). Winter peatland habitat selection by woodland caribou in northeastern Alberta. Canadian Journal of Zoology, 73(8),
1567-1574.
Courbin, N., Fortin, D., Dussault, C., & Courtois, R. (2009). Landscape management for woodland caribou: the protection of forest blocks influences wolf-caribou co-occurrence.
Landscape ecology, 24(10), 1375-1388.
DeCesare, N. J., Hebblewhite, M., Schmiegelow, F., Hervieux, D., McDermid, G. J., Neufeld, L., ...
& Musiani, M. (2012). Transcending scale dependence in identifying habitat with resource selection functions. Ecological Applications, 22(4), 1068-1083.
Gillies, C.S., Hebblewhite, M., Nielsen, S.E., Krawchuk, M.A., Aldridge, C.L., Frair, J.L., Saher, D.J.,
Stevens, C.E. & Jerde, C.L. (2006). Application of random effects to the study of resource selection by animals. Journal of Animal Ecology, 75, 887–898.
Halsey, L.A., D. Beilman, D.H. Vitt, S. Crow, S. Mehelcic and R. Wells. (2013). Alberta wetland inventory classification system 2.0. Alberta Sustainable Resource Development,
Edmonton, AB, Canada, Pub. No. T/031.
Johnson, D.H. (1980). The comparison of usage and availability measurements for evaluating resource preference. Ecology, 61, 65–71.
Lewis, J.S., Rachlow, J.L., Garton, E.O. & Vierling, L.A. (2007). Effects of habitat on GPS collar performance: using data screening to reduce location error: GPS collar performance. Journal
of Applied Ecology, 44, 663–671.
Manly, B.F.J., McDonald, L., Thomas, D.L., McDonald, T.L. & Erickson, W.P. (2002). Resource
selection by animals: statistical design and analysis for field studies, Second edition. Kluwere
Academic Publishers.
Matthiopoulos, J., Hebblewhite, M., Aarts, G. & Fieberg, J. (2011). Generalized functional responses for species distributions. Ecology, 92, 583–589.
McLOUGHLIN, P. D., DUNFORD, J. S. and BOUTIN, S. (2005), Relating predation mortality to broad-scale habitat selection. Journal of Animal Ecology, 74: 701–707
Moreau, G., Fortin, D., Couturier, S. and Duchesne, T. (2012), Multi-level functional responses for wildlife conservation: the case of threatened caribou in managed boreal forests. Journal of Applied Ecology, 49: 611–620.

180
181
182
183
184
185
Mysterud, A. & Ims, R.A. (1998). Functional responses in habitat use: availability influences relative use in trade-off situations. Ecology, 79, 1435–1441.
Rettie, W. J., & Messier, F. (2000). Hierarchical habitat selection by woodland caribou: its relationship to limiting factors. Ecography, 23(4), 466-478.
Zuur, A.F., Ieno, E., Walker, N., Saveliev, A.A. & Smith, G.M. (2009). Mixed effects models and
extensions in ecology with R. Springer, New York, NY.