Introduction

Human manipulation of natural landscapes via agriculture, grazing, deforestation, and urbanization has greatly reduced the availability of terrestrial wildlife habitats worldwide and has resulted in increasing isolation of remaining habitat patches. Not surprisingly, habitat loss, fragmentation, and related anthropogenic modifications (hereafter referred to as “modifications”, for brevity) are among the most important challenges for wildlife conservation worldwide (Foley et al. 2005; Fischer and Lindenmayer 2007; Fleishman et al. 2011). However, wildlife species differ in their sensitivity to human modification of habitat. Differences in species responses may be explained by corresponding differences in life history traits, such as trophic level, dispersal ability, reproductive potential, and niche breadth (Henle et al. 2004; Ewers and Didham 2006).

A species’ niche breadth can be quantified as some combination of the number of different habitat types used, number of different types of diet items, or other related metrics (Swihart et al. 2003; Devictor et al. 2008). Species with a greater niche breadth (“generalists”) have been shown to have larger range sizes (Slatyer et al. 2013) and should occur in more patches across a landscape than species with narrow niches (“specialists”) for several reasons. First, specialists can exploit a wider range of habitat types and diet items (Brown 1995; Swihart et al. 2006). As native habitat in a landscape is modified, declining availability of habitat types or diet items should, on average, threaten the viability of populations of specialists more quickly than generalists (Henle et al. 2004; Öckinger et al. 2010; Allouche et al. 2012). Second, human-created “matrix” habitat may be more hospitable for generalists relative to specialists (Andrén 1994; Swihart et al. 2003; Ewers and Didham 2006; Clavel et al. 2011; Betts et al. 2014). Finally, competition between generalists and specialists may be altered to favor generalist species in patchy, heavily modified landscapes (Tilman 1994; Cadotte 2007; Nagelkerke and Menken 2013). When landscapes are “coarse-grained” (i.e. contain relatively large patches of unmodified habitat), most organisms may encounter only one habitat type, thus favoring species that specialize on that habitat. In patchy landscapes with large amounts of habitat edge, in contrast, generalist vertebrates may have an advantage because they generally are more tolerant of edge and matrix and thus can disperse farther and/or colonize a larger proportion of patch types they encounter (Pfeifer et al. 2017). A “competition-colonization trade-off” has been observed in some natural systems, e.g. birds (Rodríguez et al. 2007) and insects (Yu et al. 2004), although empirical evidence remains relatively weak (Livingston et al. 2012).

In this paper, we focus on niche breadth as a predictor of vertebrate responses to landscape modification rather than other traits tied to mobility/dispersal (e.g., body size, morphology, ontogeny) for two reasons. First, direct comparisons in our study system using data for 32 species demonstrated that niche breadth was superior to traits tied to mobility/dispersal as a predictor of occupancy (Swihart et al. 2003). Second, empirical support for the prediction that specialist species should be more sensitive than generalists to habitat modification exists across multiple taxa and studies. For example, Öckinger et al. (2010) found specialist butterflies (defined by the number of plant species on which larvae fed) were more sensitive to habitat modification than generalists, and Bommarco et al. (2010) observed a similar relationship in bees. Penone et al. (2013) found that specialist Orthoptera declined to a greater degree with urbanization than generalists; this pattern was also observed by Deguines et al. (2016) for insect visitors to plants. In birds, Devictor et al. (2008) reported that specialist species had a more negative response to modification than generalists in France, and Wang et al. (2015) observed that vulnerability to fragmentation was greater for specialists in an island system in China. For non-volant small mammals, Swihart et al. (2003) found a significant association between niche breadth and occupancy of forest patches in an intensively agricultural landscape, and Pardini et al. (2010) noted landscape-wide loss of specialists in Brazilian forests subjected to high levels of habitat loss. However, variable results have been observed for some taxa. For instance in bats, Kerbiriou et al. (2018) found that intensive agriculture and other anthropogenic factors reduced community richness and promoted generalist species, while Safi and Kerth (2004) and Meyer et al. (2008) found little evidence for a relationship between diet or habitat specialization and extinction risk due to habitat modification.

Assessment of landscape-level responses to habitat modification requires replication across landscapes that vary in land use and cover characteristics (Bennett et al. 2006; Castillo et al. 2016). In addition, associations between life history traits and extinction risk in modified landscapes may vary across taxa (Atauri and de Lucio 2001). Thus, robust tests of species response to landscape modification should entail simultaneous consideration of multiple landscapes and taxa. Unfortunately, few studies incorporate replicated landscapes and multiple taxa; Atauri and de Lucio (2001), Swihart et al. (2006), and Herrera et al. (2016) are notable exceptions.

Even in instances where multiple taxa and landscapes are studied, estimating the effect of niche breadth on species responses to habitat modification is complicated by at least three methodological challenges. First, response to modification typically is quantified as a change in abundance or occupancy estimated in the field. These metrics can be biased by imperfect detection, particularly for cryptic species or in cases when detection probability varies among habitat types (Kellner and Swihart 2014). Second, multi-species analyses generally have treated species as independent data points, ignoring phylogenetic relationships among species that may correspond to correlations in response to landscape modification. The degree to which these factors have biased results from earlier studies is unknown; however, at least one study demonstrated that failure to account for imperfect detection may alter inference about response to landscape configuration (Wang et al. 2015). Finally, quantifying landscape modification itself is challenging, especially when separating components of habitat loss and configuration (Fahrig 2003; Prugh et al. 2008; Didham et al. 2012; Wang et al. 2014). Addressing potential sources of bias is an important first step in assessing the reliability of prior results linking niche breadth and species response to modification, and may facilitate future studies of how the relationship varies taxonomically and under different environmental and ecological conditions.

Here, we assess the effects of species specialization on occurrence and response to degree of landscape modification, defined using a combination of habitat abundance and fragmentation metrics, in an intensively agricultural region of the Midwestern US. Our assessment includes data for multiple vertebrate taxa (nonvolant small mammals [hereafter small mammals], bats, birds, and aquatic turtles) collected simultaneously across replicate landscapes within a major river basin. In addition, we analytically extend prior work (Swihart et al. 2006) by assessing the effect on predicted responses to landscape modification of (1) imperfect detection, and (2) phylogeny, thereby providing insight on the degree to which these factors influence accuracy, precision, and inference in this and similar studies.

We predicted that species with broader niches would exhibit greater average occurrence across all landscapes. Consistent with findings of Devictor et al. (2008), we also predicted that species niche breadth would be associated positively with response to landscape modification; that is, specialists would have more negative responses to modification, whereas generalists would generally have less negative or even positive responses. Finally, we expected the degree of uncertainty in these relationships to increase after accounting for imperfect detection and phylogenetic relationships among species, due to increased uncertainty in parameter estimation.

Methods

Study design

The study was conducted in the Wabash River basin in central Indiana, USA. The basin is approximately 20,000 km2 in area, representing 20% of the total area of the state (Swihart and Slade 2004). Land use in the basin is dominated by corn and soybean agriculture (88% of the area in 2003), with natural habitats (hardwood forest, wetland, grassland) highly fragmented and comprising less than 10% of the total area (Swihart et al. 2006). The amount of remaining natural habitats and the degree of fragmentation varied across the basin area. We selected 35 square cells, chosen to capture the range of levels of landscape modification in the region, across the basin area for inclusion in the study (Moore and Swihart 2005; Swihart et al. 2006; Fig. 1). Each cell (hereafter landscape) was 23 km2 in area.

Fig. 1
figure 1

Map of the Upper Wabash river basin in Indiana, USA, showing locations of the 35 landscape cells included in the study area as well as sample plot locations within two example landscapes

Full size image

Landscape modification

To quantify landscape modification, we sought to define a single score that incorporated components of both habitat availability and fragmentation across all natural habitat types. We began by deriving for each landscape a land-use raster at 3 m resolution from digitized USGS orthophotos taken in 1998 at 1 m resolution (Swihart et al. 2006). We assumed that natural habitats were converted to anthropogenic land-uses relatively randomly in these landscapes, and thus the degree of modification of one natural habitat type would be positively correlated with the degree of modification of the other natural habitat types in a given landscape. Thus, we reclassified all raster pixels as either natural habitats (forest, wetland, grassland) or human-dominated land-use types (e.g. agriculture, urban) according to the most common habitat type in the pixel. Using FRAGSTATS 4.2 (McGarigal et al. 2012), we calculated several metrics of natural habitat composition and configuration for each landscape raster including percent of the landscape area in natural habitat types (PA), total edge length of natural habitats (EL), mean patch area (MPA), average nearest neighbor distance among patches (ENN), degree of contagion of natural habitat (CLUMPY), and perimeter-area fractal dimension (PAFRAC), a measurement of patch shape complexity (Wang et al. 2014). Each of these metrics reflects a different aspect of landscape modification, but we expected that in many cases they would be highly correlated. To derive a final modification score for each landscape from the metrics, we used a two-step process. First, for each metric, we ranked the 35 landscapes by metric score and assessed the degree to which this ranking corresponded to an expert visual assessment of habitat loss and fragmentation. We then combined metrics that appeared to be closely related to degree of modification into a single factor score using factor analysis in R 3.3.1 (R Development Core Team 2017).

Field sampling

Within each landscape, we randomly established a series of sampling plots for aquatic turtles, small mammals, bats, and/or birds. Each plot was sampled once, in the months of May through August, over the 3-year period of the study (2001–2003). The number of sample plots and sampling protocols differed among landscapes and taxonomic groups as described below.

For turtles, we established 210 plots (range 2–18 per landscape) in wetland, river, and pond habitats. Each plot was sampled with 2–8 hoop traps, checked for 4 consecutive days.

For small mammals, a total of 832 plots (10–46 per landscape) were sampled in forest, grassland, agriculture, and corridor (primarily fencerow) habitats. In each plot, 9–25 trap stations were established in a grid. Trap stations were always separated by 15 m. When habitat patches were small (e.g., small woodlots) this resulted in a smaller number of traps per grid; thus, the number of traps used was proportional to patch area. Soda-can Fitch traps (2001), Sherman live traps (2003) or a combination of the two (2002) were placed at each station, and Tomahawk live traps were placed at every other trap station. Each grid was pre-baited for 3 days and then checked twice per day for 5 consecutive days (Moore and Swihart 2005; Swihart et al. 2006).

Bat sampling was conducted at 689 plots in forest, grassland, agriculture, and wetland habitats across 29 of the 35 landscapes (7–42 plots per landscape). At each plot, echolocation calls were recorded from dusk until dawn for four consecutive nights using 1–2 ANABAT II detectors (Titley Electronics, NSW, Australia). Recording for four nights falls into the range (2–5) recommended for detecting the common species present at a single sampling location (Skalak et al. 2012). Recorded call sequences were prepared for identification within program analook (v. 4.8; Corben 2001) by using filters to remove excess noise according to parameters outlined by Britzke and Murray (2000). Call sequences were identified to species via artificial neural networks that compared measured parameters of recorded calls to a reference library of bat calls from the eastern United States (Britzke et al. 2011). When tested against a set of reference calls, this method correctly identified 94% of call sequences to species (Britzke et al. 2011). More details are provided in Duchamp and Swihart (2008).

Birds were sampled at 1722 plots in forest, grassland, agriculture, wetland, and urban habitats (16–156 plots per landscape). Each plot was surveyed by one trained observer for either three (2001) or four (2002–2003) consecutive intervals of 2 min each. The species and time interval for the first detection (either auditory or visual) of each bird was recorded. Surveys were not conducted under conditions of rain or high wind.

Species niche breadth

We quantified the niche breadth of each species using taxon-specific approaches. For each small mammal, bat, and turtles species, we identified the number of common diet types consumed and the number of habitat types used. Then, within each taxon, we standardized each count to a z-score. The niche breadth for a species was the average of the diet and habitat z-scores (Swihart et al. 2003, 2006). For birds, niche breadth was the number of habitat types used by a given species based on Birds of North America descriptions (Rodewald 2015), also converted to a z-score.

Species phylogenetic relationships

For each taxonomic group, we obtained a pruned phylogenetic tree containing only the included species. For small mammals and bats, we obtained pruned trees from Bininda-Emonds et al. (2007) using Phylomatic (http://phylodiversity.net/phylomatic/). For turtles and birds, we obtained trees based on the NCBI taxonomy (Federhen 2012) using phyloT (http://phylot.biobyte.de/index.html). From each tree, we derived a phylogenetic correlation matrix D (assuming Brownian motion evolution) using package “ape” in R (Paradis et al. 2004). For turtles and birds, branch lengths were assumed to be 1 (Wang et al. 2015). The phylogenetic trees we used are available in Appendix S1 in Supporting Information (Figures S1.1–S1.4).

Analysis

We conducted the analysis separately for each taxonomic group using a two-stage approach. In the first stage we estimated occurrence for each species in each landscape. In the second stage, we used these estimated species x landscape occurrence values as the response variable in a new analysis. A detailed description of this two-stage approach is provided below.

In the first stage, we estimated occurrence of species with adequate data in each landscape under two different modeling scenarios: ignoring (N) or accounting for imperfect detection (I). For model scenario (N), a naïve value for occurrence Φij for species i in landscape j was calculated as the total number of sample plots in landscape j in which a species was observed divided by the total sample plots in landscape j located in suitable habitat types (e.g., “grassland” was not considered suitable habitat for birds associated with mature forest).

For scenario (I), we used the same approach except that we corrected occurrence Φij for imperfect detection using single-species occupancy models (turtles, bats, and small mammals) and Farnsworth time-removal abundance models (birds; Farnsworth et al. 2002) fit in a Bayesian framework using jags (Plummer 2003) and package ‘jagsUI’ (Kellner 2015). Estimated abundances from the Farnsworth models were truncated to occupancy (i.e., if a site had abundance > 0, it was occupied). We accounted for variability in sampling effort among plots for each taxa by including effort as a covariate on detection probability. We also included habitat type as a covariate on both detection probability and occupancy or abundance.

The models gave us posterior distributions for zikj, the unobserved “true” occupancy state of species i at plot k in landscape j. For each posterior sample of zikj, we calculated a corresponding realization of occurrence φij using the same approach as with the naïve occurrence in scenario (N), yielding a new posterior distribution for φij. Using this posterior distribution, we calculated a mean (Φij) and standard deviation (sij) for occurrence of species i in landscape j, with s representing uncertainty in Φ due to imperfect detection.

Values of occurrence Φij from scenarios (N) and (I) and sij from scenario (I) were used as inputs for the second stage of the analysis. As the response variable Φij was constrained [0,1], we first transformed it with an arcsine-square-root transformation. For scenario (N), we modeled the transformed occurrence of species i in landscape j according to the following equation:

$$\phi_{ij} = \alpha_{\phi } + \beta_{N} NICHE_{i} + \beta_{M,i} MOD_{j} + \varepsilon_{ij}$$
(1)

where αΦ is the mean occurrence for all species across all landscapes, βN is the effect of niche breadth for species i on occurrence, βM,i is the species-specific effect of modification on occurrence of species i in landscape j (see Eq. 4), and εij is the residual error. Under scenario (N) (ignoring imperfect detection), residuals came from species-specific normal distributions:

$$\varepsilon_{ij} N\left( {0,\sigma_{i}^{2} } \right)$$
(2)

Under scenario (I) (accounting for imperfect detection), residuals were weighted according to the uncertainty s in the posterior distribution of occurrence for species i in landscape j:

$$\varepsilon_{ij} N\left( {0,\sigma_{i}^{2} s_{ij} } \right)$$
(3)

In both model scenarios N and I, we modeled βM,i as a function of niche breadth:

$$\beta_{M,i} = \alpha_{M} + \beta_{NM} NICHE_{i} + \varepsilon_{M,i}$$
(4)

where αM is the mean response to modification for all species, βNM is the effect of niche breadth on response to modification, and εM,i is the residual error. We crossed scenarios N and I with two additional scenarios: no phylogenetic relationship in response to landscape modification among species (N) and possible phylogenetic relationships in response among species (P) yielding four total model scenarios (hereafter NN, IN, NP, and IP). Under the no-phylogeny scenario, residual errors in the response of species i to modification βM,i were modeled as coming from a common normal distribution:

$$\varepsilon_{M,i} N\left( {0,\sigma_{M}^{2} } \right)$$
(5)

Under the possible phylogenetic relationships scenario, the vector of residual errors for all species εM came from a multivariate normal distribution (Frishkoff et al. 2017):

$$\varepsilon_{M} MVN\left( {0,\varSigma_{M} } \right)$$
(6)

The variance–covariance matrix ΣM was calculated as

$$\varSigma_{M} = \sigma_{M}^{2} \times \lambda D \times \left( {1 - \lambda } \right)I$$
(7)

where D is the expected correlation matrix among species derived from phylogenetic distance, I is the identity matrix with dimensions i × i, and λ is Pagel’s λ (Frishkoff et al. 2017). Pagel’s λ is estimated by the model, controlling the strength of the phylogenetic signal (Revell 2010).

Separate models for all combinations of taxonomic groups (4) and model scenarios (4) were fit in JAGS using package ‘jagsUI’ (Appendix S1, Code S1.1–S1.4). Model runs consisted of three MCMC chains, each with 6000 total iterations, a burn-in of 4000 iterations, and a thinning rate of 20. Model convergence was assessed using the Brooks–Gelman–Rubin statistic (Brooks and Gelman 1998). For each model parameter, we calculated a value f, the proportion of the posterior distribution with the same sign as the mean. Values of f close to 1 represent a high confidence in the direction of the effect. We concluded that there was a strong effect of niche breadth on occurrence (hypothesis one) when f ≥ 0.95 for parameter βN. Similarly, we concluded there was a strong effect of niche breadth on response to landscape modification (hypothesis two) when f ≥ 0.95 for parameter βNM. This approach is analogous to setting α = 0.05 for one-way tests in a frequentist framework.

Results

Vertebrate sampling

Over the 3 years of the study, 31,980 total observations were collected of the 65 species of small mammals, bats, birds, and turtles that were included in the analysis. This included 10,823 captures of 11 species of small mammals, 2068 detection-nights of 7 species of bats, 18,469 observations of 39 species of birds, and 620 captures of 8 species of turtles. Total observed counts for individual species are given in Appendix 1 (Table S1.1). Occupancy and abundance models fit in stage one of the modeling process resulted in highly variable detection probabilities. For small mammal species, mean detection probability was 0.29 (range 0.08–0.60); for bats it was 0.37 (range 0.05–0.51); for birds 0.70 (range 0.32–0.98) and for turtles 0.17 (range 0.01–0.39). A total of 15 species had detection probabilities that varied among habitat types (Appendix S1, Figures S1.5–S1.7). Full output from the occupancy and time-removal models is provided in Appendix S2.

Landscape variables

Correlation among landscape variables was high, particularly correlations with percent natural habitat in the landscape (PA; Table 1). After considering these correlations and comparing landscape variable scores to an expert assessment of the degree of fragmentation in each landscape, we included PA, mean patch area (MPA), and nearest neighbor distance among patches (ENN) in a factor analysis to generate a single metric. Hereafter, we will refer to this as our “modification” metric, but acknowledge that it combines effects of both natural habitat amount and configuration (Devictor et al. 2008). The factor score modi for landscape i was calculated as:

$$mod_{i} = - \left( {1.00 \cdot PA + 0.76 \cdot MPA - 0.56 \cdot ENN} \right)$$
(8)

and in total explained 63% of the variation in the three landscape variables. Factor scores corresponded well with visual assessment of the degree of landscape modification (Fig. 2).

Table 1 Pearson correlations between metrics of natural habitat abundance and fragmentation in the 35 study landscapes
Full size table
Fig. 2
figure 2

Maps of natural habitat in landscapes with the minimum, median, and maximum factor scores for degree of landscape modification based on factor analysis

Full size image

Niche breadth effects on occurrence

Niche breadth had a positive relationship with occurrence for all four vertebrate groups (all f > 0.95, Fig. 3). The effect of niche breadth was strongest for small mammals and turtles; a one-standard-deviation increase in niche breadth resulted in an average increase in occurrence of 62% and 46%, respectively, in model scenarios accounting for imperfect detection. For bats and birds, the effect size was positive but small (Fig. 3). In most cases, the direction and size of the effect of niche breadth was similar between models that accounted for imperfect detection and those that did not (Fig. 3). For bats, accounting for imperfect detection reduced the effect size of niche breadth (defined as the value of the parameter βN) by 71%, while for turtles the effect size was increased by 60% (Fig. 3).

Fig. 3
figure 3

Effect of niche breadth on occurrence, after accounting for response to landscape modification, for (a, b), bats (c, d), birds (e, f), and turtles (g, h). The first column of panels (a), (c), (e), (f) represent the model scenario not incorporating imperfect detection (N), and the second column represents model scenarios incorporating imperfect detection (I). Individual points represent the mean residuals of a species across all landscapes ± SD. Parameter β is the effect size (i.e., slope) of niche breadth, and f represents the proportion of the posterior distribution with the same sign as the mean value of β. Values of f ≥ 0.95 represent a statistically important effect of niche breadth

Full size image

Niche breadth effects on response to landscape modification

For most taxonomic groups and model scenarios, there was a trend towards positive effects of niche breadth on response to landscape modification (Figs. 4, 5, 6, and 7). For small mammals, bats, and turtles, evidence was weak in all cases (f < 0.90). For birds, in contrast, there was stronger evidence of a positive relationship between niche breadth and response to modification, particularly after accounting for imperfect detection (f ≥0.95; Fig. 6). However, the effect size was relatively small. For example, under model scenario IP, a one-standard deviation increase in niche breadth increased the predicted value of the response to modification (i.e., the slope βM) by 0.020, or approximately 13% of the observed range of values for βM (− 0.072–0.087). Furthermore, even the largest absolute value of βM estimated in model scenario IP (0.087, for the American Robin) represents a relatively small effect of landscape modification on occurrence.

Fig. 4
figure 4

Effect of niche breadth (β) on response to landscape modification for small mammals, under four model scenarios: a NN, no imperfect detection, no phylogenetic relationships; b IN, imperfect detection, no phylogenetic relationships; c NP, no imperfect detection, phylogenetic relationships; and d IP, imperfect detection, phylogenetic relationships. Response values < 0 represent a negative relationship between degree of landscape fragmentation and species occurrence and vice versa. Parameter f represents the proportion of the posterior distribution with the same sign as the mean value of β. Values of f ≥ 0.95 represent a statistically important effect of niche breadth

Full size image
Fig. 5
figure 5

Effect of niche breadth (β) on response to landscape modification for bats, under four model scenarios: a NN, no imperfect detection, no phylogenetic relationships; b IN, imperfect detection, no phylogenetic relationships; c NP, no imperfect detection, phylogenetic relationships; and d IP, imperfect detection, phylogenetic relationships. Response values < 0 represent a negative relationship between degree of landscape fragmentation and species occurrence and vice versa. Parameter f represents the proportion of the posterior distribution with the same sign as the mean value of β. Values of f ≥ 0.95 represent a statistically important effect of niche breadth

Full size image
Fig. 6
figure 6

Effect of niche breadth (β) on response to landscape modification for birds, under four model scenarios: a NN, no imperfect detection, no phylogenetic relationships; b IN, imperfect detection, no phylogenetic relationships; c NP, no imperfect detection, phylogenetic relationships; and d IP, imperfect detection, phylogenetic relationships. Response values < 0 represent a negative relationship between degree of landscape fragmentation and species occurrence and vice versa. Parameter f represents the proportion of the posterior distribution with the same sign as the mean value of β. Values of f ≥ 0.95 represent a statistically important effect of niche breadth

Full size image
Fig. 7
figure 7

Effect of niche breadth (β) on response to landscape modification for turtles, under four model scenarios: a NN, no imperfect detection, no phylogenetic relationships; b IN, imperfect detection, no phylogenetic relationships; c NP, no imperfect detection, phylogenetic relationships; and d IP, imperfect detection, phylogenetic relationships. Response values < 0 represent a negative relationship between degree of landscape fragmentation and species occurrence and vice versa. Parameter f represents the proportion of the posterior distribution with the same sign as the mean value of β. Values of f ≥ 0.95 represent a statistically important effect of niche breadth

Full size image

Overall, accounting for imperfect detection resulted in few differences in model output, with two exceptions. First, there was slightly greater evidence for a relationship between niche breadth and response to landscape modification for birds when accounting for imperfect detection (f = 0.98, relative to f = 0.95 under naïve model scenarios; Fig. 6). Second, accounting for imperfect detection changed the direction of the relationship between niche breadth and response to modification for turtles, from negative to positive (Fig. 7). However, none of the model scenarios for turtles met our benchmark for statistically strong associations.

We found little evidence for a phylogenetically related structure in responses to landscape modification in any taxonomic group. Effect sizes and f values were similar between model scenarios that included the possibility of phylogenetic correlation and those that did not (Figs. 4, 5, 6, and 7). Furthermore, posterior distributions of Pagel’s λ (the strength of the phylogenetic signal) in model scenarios NP and IP either had modes at or near 0 (for birds) or were highly diffuse (mammals, bats, and turtles; Appendix S2), indicating that there was little support for strong, phylogenetically derived correlations among responses to landscape modification. Full output from the four model scenarios can be found in Appendix S2.

Discussion

Effects of niche breadth on occurrence

We found strong support for our first hypothesis: there was a positive relationship between occurrence and species niche breadth for all four taxonomic groups. Habitat generalists were prevalent among the species with the highest mean occurrence in each taxonomic group—the white-footed mouse (Peromyscus leucopus), American Robin (Turdus migratorius), big brown bat (Eptesicus fuscus), and snapping turtle (Chelydra serpentina). We observed the relationship between niche breadth and occurrence even under conservative assumptions about the habitat types in which a species was likely to be found—occurrence for a given species was calculated using only plots in habitat types the species would realistically use (e.g., plots in agricultural fields were excluded for flying squirrels). Furthermore, we observed this relationship across human-dominated landscapes in which natural habitat diversity was low. Even given these constraints, the ability of a species to make use of a wider range of potential habitat types and/or diet items corresponded with a wider distribution in the upper Wabash River basin.

The effect size of niche breadth on landscape-level occurrence was large for turtles and small mammals, and relatively small for bats and birds. The relative mobility of these latter taxa likely contributed to the difference in effect size. Specialists among birds and bats are capable of dispersing greater distances to reach suitable habitat resources relative to specialists among small mammals and turtles (Ewers and Didham 2006). Thus, in the context of environmental modification, the detrimental effects of a narrow niche breadth may be less profound for specialists in these groups, resulting in the reduced effect size we observed. Greater dispersal ability also may expose species to a greater risk of dispersal-related mortality, particularly in human-dominated landscapes (Fahrig 2007). The balance of dispersal capability versus dispersal-related mortality depends on the hostility of the matrix to a particular species (Fahrig 2007; Van Houtan et al. 2007). Among bats, the matrix in our study area was probably relatively hospitable to at least some species, such as big brown bat and red bat (Lasiurus borealis) (Duchamp et al. 2004; Duchamp and Swihart 2008). Among specialist forest birds, perceived predation risk in the agricultural matrix was likely greater than in remnant forest patches (Bélisle et al. 2001; Gobeil and Villard 2002). Our results emphasize that specialist species, particularly among taxa with relatively low dispersal ability, should be prioritized for monitoring and conservation in human-dominated landscapes, as their reduced occurrence and potentially increased isolation increase the risk of local extinction.

In addition to differences in dispersal ability among taxonomic groups, differences in dispersal ability within a taxon may influence response to landscape modification (Ewers and Didham 2006; Bommarco et al. 2010). Dispersal ability in the context modified landscapes includes both the distance a species is capable of dispersing, as well as its tolerance of matrix land-use types; both of these aspects of “functional mobility” should be related to a species’ response to modification. In both bats (Meyer et al. 2008) and birds (Lees and Peres 2008) there is empirical evidence of a positive relationship between species dispersal capability and occurrence in fragmented landscapes.

Effects of niche breadth on response to landscape modification

We found inconsistent support for our second hypothesis that response to landscape modification would increase with niche breadth. We observed a generally positive correlation between response to modification and niche breadth across taxonomic groups as predicted, but the relationship was strong only for birds. Even for birds, occurrence of most species showed a trend of a negative relationship with landscape modification; only 13 of 39 species (33%) actually had a response to modification > 0 (i.e., increasing occurrence with increasing levels of modification). Positive responses characterized some generalists and synanthropic species, e.g. the American Robin, House Wren (Troglodytes aedon), and Blue Jay (Cyanocitta cristata). The remaining species, including many with above-average niche breadth, still exhibited negative responses to landscape modification—they just weren’t affected as negatively as more specialized species. Our results are similar to those of Devictor et al. (2008), who found that responses to fragmentation (correlations between species abundance and habitat edge length), were more negative for French breeding bird species with greater degrees of specialization. Similarly, Carrara et al. (2015) found that forest specialist birds showed a stronger response to landscape structure than generalists in a Mexican rainforest. Torrenta et al. (2018) observed that, among 12 forest bird species in Canada, only habitat specialists experienced significantly reduced abundance and negative responses to reduced forest cover in highly modified landscapes. Interestingly, they also showed that forest loss and fragmentation reduced the structural habitat breadth of territories for all 12 species, suggesting density-dependent selection operating at a regional scale to reduce spillover into suboptimal habitat. In less mobile taxa, reduced functional connectivity may limit opportunities for selection of optimal habitat irrespective of density (van Langeveld 2015).

For the other taxa, the relationship was generally positive but weak; for turtles the direction of the relationship varied with model scenario. This inconsistent response is likely related to the range of the habitat configurations in the landscapes we sampled, and the history of human dominance in those landscapes. Ultimately, all 35 study landscapes in the upper Wabash River basin were heavily impacted by humans. Even the “least disturbed” of the landscapes was dominated by intensive row-crop agriculture and urban land use (51% by area). Agriculture has been a predominant feature of the basin for more than a century (Swihart et al. 2003), coincident with dramatic declines in amount and quality of grassland, wetland, and to a lesser degree, forest habitat (Hartman 1994; Smith et al. 1992; Martin et al. 2008; Urban and Swihart 2009). The duration and magnitude of landscape modification have served as powerful environmental filters. Among species historically present in our study area, those with the narrowest niches and greatest sensitivity to landscape modification have either been extirpated (e.g., Whooping Crane, Grus americana; fisher, Pekania pennanti) or were so rare that we could not include them in our analysis (e.g., Cerulean Warbler, Setophaga cerulea). The remaining pool of extant and relatively common species by definition have some resistance to landscape modification, thereby attenuating the relationship with niche breadth compared to an assessment that included the original complement of species present at the onset of modification (Swihart et al. 2003; Swihart and Verboom 2004; Ewers and Didham 2006; Allouche et al. 2012).

Low taxonomic diversity may have contributed to the failure to detect relationships for small mammals, bats, and turtles. Richness for these taxonomic groups was not as high in our study areas as for birds, resulting in a more limited set of niche breadths. If the true effect size of niche breadth on response to landscape modification for these groups is different from zero but similar in magnitude to what we observed for birds, our power to detect it is probably insufficient. Unfortunately, the species pool is fixed and limits our ability to improve statistical power. Testing for cross-species differences in response to landscape modification, as we did in this study, will likely become increasingly difficult as the pool of remaining species is depleted by anthropogenic change, with surviving species biased towards a more homogenous subset that exhibits broader niche dimensions and greater resistance to landscape modification (Ewers and Didham 2006; Clavel et al. 2011; Allouche et al. 2012).

In addition to low taxonomic diversity, our approach to quantifying modification in each of the replicated landscapes may have impacted our power to detect effects of modification. We pooled all land-use types into a single binary variable (classified as either “natural habitats” or “anthropogenic land-use”) before calculation of landscape metrics. We took this approach in order to obtain a single score for each landscape that represented the degree to which that landscape had been impacted by humans. We expected that the degree of modification of one natural habitat type would be positively correlated with the degree of modification of the other natural habitat types in a given landscape. However, reliance on a single modification metric may fail to accurately represent habitat modification as perceived by species with divergent habitat and dietary needs.

Effects of imperfect detection and phylogenetic relationships

Our third hypothesis was generally not supported. We did not see a consistent pattern of changes in inference or in degree of uncertainty when accounting for imperfect detection nor when accounting for phylogenetic relationships. That model inference was robust to imperfect detection was surprising, given that the detection probabilities for individual species were often substantially less than 1 (depending on taxonomic group) and in many cases detection varied with habitat type. However, the group with the lowest detection probabilities (aquatic turtles) did show evidence of a reversal in the direction of the relationship between niche breadth and response to landscape modification. If the other taxonomic groups (particularly birds) had equally low detectability we may have observed a stronger effect. Our results contrast with Wang et al. (2014), who found that incorporating imperfect detection had a modest effect on the ranking of a set of models of bird species vulnerability to fragmentation.

The propagation of uncertainty from the scale at which data were collected (occupancy of individual species at individual sample plots in a landscape) to the scale at which we fit the models (occurrence for a species in a landscape, the proportion of sample plots occupied) may also have contributed to the weak effects of imperfect detection that we observed. Although occupancy of individual sample plots had high uncertainty for many species due to low detection probabilities, estimated landscape-wide occurrence was not as variable; some of the variability was “averaged out”. Scale-dependent effects of detection on inference are not well understood, though a review of the literature indicated that studies covering large spatial scales and/or multiple species were less likely to use methods that accounted for imperfect detection (Kellner and Swihart 2014). Although we did not find major effects of imperfect detection on inference at the landscape scale in our study, the low detection probabilities we estimated imply that ignoring imperfect detection would have greater consequences for studies of wildlife response to habitat modification at smaller spatial scales (i.e., individual sample plots).

As with imperfect detection, we found little evidence that accounting for phylogenetic relationships among species affected inference. In models that allowed for correlation based on phylogeny, Pagel’s lambda values were either relatively small (< 0.35) or were poorly estimated (i.e., had diffuse posterior distributions). In the upper Wabash River basin, and for the vertebrates we considered, it appears that closely related species are not necessarily more likely to respond to landscape modification in similar fashion. It is unclear if this result is generalizable outside our study system. Few studies have attempted to account for phylogeny in comparative studies of species responses to disturbance and landscape pattern, despite the availability of approaches like phylogenetic least-squares (PGLS) regression and phylogenetic occupancy models (Orme et al. 2015; Frishkoff et al. 2017). A recent study of fragmentation effects on a bird community in China obtained results similar to ours from sets of models accounting for and ignoring phylogeny, but did not report values of Pagel’s lambda (Wang et al. 2015). Future work in different systems and with different taxonomic groups should quantify the degree to which phylogeny drives similarities in responses to disturbance.

Selection of metrics to quantify habitat modification

A wide variety of different metrics that quantify habitat loss and fragmentation can be derived from landscape composition data. Wang et al. (2014) provide a good summary of correlations among different metrics and highlight several approaches to measuring habitat fragmentation that are independent of habitat amount in a landscape. In many cases, disentangling the two forms of modification may be difficult or unproductive (Didham et al. 2012; Hanski 2015). Researchers should be careful to select measurements of modification that are appropriate for the species and landscape of interest rather than relying on metrics previously used in the literature. For example, we initially attempted to quantify modification in different landscapes as the total edge length of all habitat types (Devictor et al. 2008). However, anthropogenic land use was so dominant in our study landscapes (51–94% by area) that sites with the lowest amount of edge were also the sites with the lowest proportion of natural habitats. For a species that in truth was sensitive to habitat loss and/or fragmentation, use of total edge length as a proxy for degree of fragmentation in the landscape could result in model output indicating a relationship in the opposite direction (i.e., a negative relationship between response to modification and species specialization). Instead, we chose to combine several metrics using factor analysis to obtain a single landscape modification score, and then visually compared this score to maps of land use in our study landscapes to insure the scores matched our knowledge of the landscapes. Thus, metrics of landscape modification must be chosen carefully to align with existing knowledge of the area and organisms being studied; a species-centered approach to defining habitat suitability is a sensible and increasingly feasible alternative (Betts et al. 2014).