A Cont = Continuous, Cat = Categorical

Supplement 1. Model variables used to create cost surfaces for individual American marten (Martes americana) in the Upper Peninsula of Michigan, USA.

Variable / Typea / Description / Rationale / Data Source / Weights
Roads / Cat / Roads / Behavioral avoidance and potential source of mortality (Robitaille and Aubry, 2000) / MI Geographic Framework All Roads dataset (http://www.mcgi.state.mi.us/) / 1:10, 1:100, 1:1000
Proportion Forested Area / Cont / Percentage of grid cell containing forest stands / Marten prefer forested areas of coniferous or deciduous and mixed coniferous/deciduous stands (Coffin et al. 1997; Zielinski and Duncan 2004; Steele 1998; Wilson and Ruff 1999; Potvin et al. 2000) where there may be a higher abundance of prey resources, resting sites, and escape from possible predators. / 2001 GAP Landcover dataset derived from Landsat satellite imagery (http://gapanalysis.usgs.gov/gaplandcover/data/), 30m x 30m resolution / 1-10, 1-100, 1-1000
Proportion Coniferous Area / Cont / Percentage of grid cell containing coniferous forest stands / Marten have been shown to select areas with conifer cover (e.g. Buskirk and Ruggiero 1994). / 2001 GAP Landcover dataset derived from Landsat satellite imagery (http://gapanalysis.usgs.gov/gaplandcover/data/), 30m x 30m resolution / 1-10, 1-100, 1-1000
Canopy Cover / Cont / Percent of a given area occupied by overhead cover / Protection from aerial predators (Drew 1995), associated with subnivean resting access (Corn and Raphael 1992), and prey (Thompson and Colgan 1994) / National Landcover Database 2001 Percent Tree Canopy dataset (http://www.mrlc.gov/nlcd2001.php), 30m x 30m resolution / 1-10, 1-100, 1-1000
Fisher Harvest Density / Cont / Density of harvested fishers / Fisher may predate on marten (Raine 1987) and represent a source of indirect competition for food resources, particularly small mammal prey (Krohn et al. 1997) and denning sites (Clem 1977). / Derived from Michigan Department of Natural Resources fisher harvest locations during 2000-2004 and ESRI ArcGIS 9.3 Kernel Density Tool / 1-10, 1-100, 1-1000

a Cont = Continuous, Cat = Categorical.

Buskirk SW, Ruggiero LF (1994) American marten. In: Ruggiero LF, Aubry KB, Buskirk SW, Lyon LJ, Zielinski WJ, (ed) The scientific basis for conserving forest carnivores: American marten, fisher, lynx, and wolverine in the western United States. Gen. Tech. Rep. RM-254. Fort Collins, CO: US. Department of Agriculture, Forest Service, Rocky Mountain Forest and Range Experiment Station. pp7-37

Clem MK (1977) Interspecific relationship of fisher and marten in Ontario during winter. In: Philips RL, Jonkel C (ed) Proceedings of the 1975 predator symposium, Missoula, MT, 16-19 June 1975. Montana Forest Conservation Experiment Station, University of Montana, Missoula. pp165-182

Coffin KW, Kujala QJ, Douglass RJ, Irby LR (1997) Interactions among marten prey availability, vulnerability, and habitat structure. Martes: taxonomy, ecology, techniques, and management. The Provincial Museum of Alberta, Edmonton, Alberta, Canada, pp199–210

Corn JG, Raphael MG (1992) Habitat characteristics at marten subnivean access sites. J Wildlife Manage 56:442–448

Drew GS (1995) Winter habitat selection by American marten (Martes americana) in Newfoundland: why old growth? PhD thesis, Utah State University

Krohn WB, Zielinski WJ, Boone RB (1997) Relations among fishers, snow, and martens in California: results from small-scale spatial comparisons. Martes: taxonomy, ecology, techniques, and management. Provincial Museum of Alberta, Edmonton, Alberta, pp211–232

Potvin F, Belanger L, Lowell K (2000) Marten habitat selection in a clearcut boreal landscape. Conserv Biol 14:844-857

Raine RM (1987) Winter food habits and foraging behaviour of fishers (Martes pennanti) and martens (Martes americana) in southeastern Manitoba. Can J Zool 65:745–747

Robitaille JF, Aubry K (2000) Occurrence and activity of American martens Martes americana in relation to roads and other routes. Acta Theriol 45:137–143

Steele MA (1998) Tamiasciurus hudsonicus. Mammalian Species, American Society of Mammalogists, pp1-9

Thompson ID, Colgan PW (1994) Marten activity in uncut and logged boreal forests in Ontario. J Wildl Manage 58:280–288

Wilson EBDE, Ruff S (1999) North American I Mammals. Smithsonian Institute, Washington, DC, USA

Zielinski WJ, Duncan NP. (2004) Diets of sympatric populations of American martens (Martes americana) and fishers (Martes pennanti) in California. J Mamm 85:470–477

Supplement 2. Hypothesized cost of American marten (Martes americana) movement in the upper peninsula of Michigan, USA in relation to marten harvest location and affiliation to one of three genetic clusters (indicated with different symbols). Cost surfaces represent a) presence of roads, b) canopy cover, c) percent forested area, d) percent coniferous forest, and e) fisher (Pekania pennanti) harvest density. Variables are described in Supplement 1.

Supplement 3. Results from sensitivity analysis to determine the influence of weighting scheme on the resulting cost distance from least cost path (LCP) analysis for each landscape feature.

Model / Weights / Mantel r / P value
Geographic Distance (Euc) / 1 / 0.209 / 0.002
Roads (2 levels) (Rd) / 1, 10 / 0.224 / 0.002
1,100 / 0.216 / 0.002
1, 1000 / 0.223 / 0.002
Canopy Cover (Can) / 1 to 10 / 0.190 / 0.002
1 to 100 / 0.211 / 0.002
1 to 1000 / 0.185 / 0.002
Proportion Forested Area (For) / 1 to 10 / 0.184 / 0.002
1 to 100 / 0.208 / 0.002
1 to 1000 / 0.052 / 0.035
Fisher Density (PP) / 1 to 10 / 0.198 / 0.002
1 to 100 / 0.192 / 0.002
1 to 1000 / 0.162 / 0.002
Proportion Coniferous Area (Con) / 1 to 10 / 0.243 / 0.002
1 to 100 / 0.221 / 0.002
1 to 1000 / 0.245 / 0.002

Supplement 4: Description of Causal Modeling methods. We used causal modeling (Cushman et al. 2006; Cushman and Landguth 2010b) based on partial Mantel tests to quantify support (Mantel r) for each of our models and to inform our boundary analysis (described in Supplement 5) by identifying the landscape variables (given the set of variables we quantified) that most highly correlated with genetic relatedness (Smouse et al. 1986; Cushman et al. 2006. We assessed statistical significance of each partial Mantel test at α = 0.0021 (Bonferonni correction of α = 0.05 based on multiple model comparisons). We conducted a series of partial Mantel tests between genetic distance and least cost distance corresponding to each of our landscape hypotheses, after partialing out Euclidean distance (e.g., Genetic Distance ~ Cost Distance from Model 1 | Euclidean Distance; Cushman et al. 2006). We also conducted a series of partial Mantel tests between genetic distance and Euclidean distance, partialing out the least cost distance corresponding to each of our landscape hypotheses (e.g., Genetic Distance ~ Euclidean Distance | Cost Distance from Model 1). If cost distance estimated based on one or more landscape features was significantly associated with genetic distance independent of Euclidean distance, we expected the former test to be statistically significant and the latter to be not significant (Cushman et al. 2006).

Cushman SA, McKelvey KS, Hayden J, Schwartz MK (2006) Gene flow in complex landscapes: testing multiple hypotheses with causal modeling. Am Nat 168:486–499

Cushman SA, Landguth EL (2010b) Spurious correlations and inference in landscape genetics. Mol Ecol 19:3592-3602

Smouse PE, Long JC, Sokal RR (1986) Regression and Correlation Extensions of the Mantel Test of Matrix Correspondence. Syst Zool 35:627-632

Supplement 5: Defining genetic clusters (additional details)

The clustering algorithm implemented in program Geneland incorporates the spatial coordinates of the multilocus genotype for each individual to determine posterior probabilities of membership to a genetic cluster. We defined the boundaries of genetic clusters as sharp gradients in posterior probabilities among different clusters. We assumed that the locational error associated with marten harvest location reporting was randomly distributed around the centroid of each township section (1 section = 2.6 km2), and included coordinate uncertainty of 1 km associated with marten harvest locations.

We estimated the number of genetic clusters (K) using a Markov Chain Monte Carlo (MCMC) algorithm. We first allowed K to vary and subsequently ran the algorithm with K fixed at the value most supported by the data (i.e., K from the initial run with the highest average posterior probability; Guillot et al. 2005b). We iterated this process four times with the following parameters: 250,000 MCMC iterations, maximum rate of Poisson process at 495 (i.e., equal to the number of individuals as suggested by Guillot et al. 2005a), minimum K = 1, maximum K = 5, and maximum number of nuclei of the Poisson-Voronoi tessellation at approximately 3 times the maximum rate of the Poisson process (Guillot et al. 2005b). We used an uncorrelated allele frequency model (Guillot et al. 2005a). We calculated the mean logarithm of the posterior probability by re-running the MCMC algorithm 10 times (thinning = 10) with K fixed at 3 (previously inferred number of genetic clusters when allowing K to vary).

Guillot G, Estoup A, Mortier F, Cosson JF (2005b) A spatial statistical model for landscape genetics. Genetics 170:1261–1280

Supplement 6: Mantel and partial Mantel tests, comparing the relative influence of landcover variables in predicting American marten (Martes americana) gene flow. Significant partial Mantel results are in bold.

Mantel / partial Mantel2
Model1 / r / P / r / P3
Con + Rd / 0.226 / <0.002 / 0.076 / <0.002
Con / 0.221 / <0.002 / 0.056 / <0.002
Can + Con / 0.220 / <0.002 / 0.055 / <0.002
Rd / 0.216 / <0.002 / 0.037 / <0.002
Can + Con + Rd / 0.220 / <0.002 / 0.056 / 0.004
Con + PP / 0.219 / <0.002 / 0.053 / 0.004
PP + Rd + Con / 0.220 / <0.002 / 0.052 / 0.004
Can + Con + PP + Rd / 0.218 / <0.002 / 0.046 / 0.004
Can + Con + PP / 0.218 / <0.002 / 0.045 / 0.006
PP / 0.192 / <0.002 / 0.028 / 0.076
Can / 0.211 / <0.002 / -0.021 / 0.093
Can + PP / 0.208 / <0.002 / -0.017 / 0.127
PP + Rd / 0.201 / <0.002 / 0.019 / 0.135
Can + Rd / 0.210 / <0.002 / -0.017 / 0.158
PP + Rd + For / 0.209 / <0.002 / -0.017 / 0.166
Can + For + Rd / 0.209 / <0.002 / -0.017 / 0.171
PP + Rd + Can / 0.209 / <0.002 / -0.015 / 0.176
Can + For / 0.208 / <0.002 / -0.017 / 0.181
For + PP / 0.197 / <0.002 / -0.017 / 0.211
Can + For + PP / 0.209 / <0.002 / -0.010 / 0.292
Can + For + PP + Rd / 0.210 / <0.002 / -0.008 / 0.298
For / 0.208 / <0.002 / 0.010 / 0.314
For + Rd / 0.212 / <0.002 / 0.007 / 0.323
Euc / 0.209 / <0.002 / NA / NA

1 Euc = Euclidean distance, Rd = Roads, Con = Proportion of coniferous forest, For= Proportion of area that is forested, PP = Fisher (Pekania pennanti) density, Can = Canopy cover.

2 Partial Mantel tests quantify the residual variation between each model and genetic distance, after accounting for variation associated with Euclidean distance (e.g., Can + Con | Euc).

3Bonferroni correction of α = 0.05/24 = 0.0021).

All univariate Mantel models correlating genetic relatedness to least cost paths were significant, with the highest Mantel r value corresponding to the model that included Con + Rd (r = 0.226, P 0.002; Supplement 6). After accounting for the variation associated with Euclidean distance (Euc) using partial Mantel tests, we found 4 significant models that described the boundaries of genetic relatedness (α < 0.0021): 1) Coniferous, 2) Roads 3) Canopy + Coniferous, and 4) Coniferous + Roads (Supplement 6). Additionally, the relationship between genetic relatedness and Euclidean distance was not significant after partialing out the variation associated with each of the above 4 models (Supplement 7). The Coniferous + Roads model had the most support compared to all alternative models based on the magnitude of the partial Mantel r values (Supplement 6).

Supplement 7. Partial Mantel tests comparing the relative influence of several landcover variables in predicting American marten (Martes americana) gene flow. Significant partial Mantel results are in bold.

1 Partial Mantel tests quantify the residual variation between Euclidean distance and genetic distance, after accounting for variation associated with each alternative landscape model.

2Bonferroni correction of α = 0.05/24 = 0.0021)