Remaining populations of an upland stream fish persist in refugia defined by habitat features at multiple scales

Conserving stream biota could require strategies that preserve habitats conveying resistance to ecological impacts of changing land use and climate. Retrospective analyses of species’ responses to anthropogenic disturbances can inform such strategies. We developed a hierarchical framework to contrast environmental conditions underlying persistence versus extirpation of an imperilled stream fish, Candy Darter (Etheostoma osburni), over decades of changing land use. The decline of E. osburni may broadly represent the challenge of conserving sensitive freshwater species in intensively used upland environments.


| INTRODUC TI ON
Understanding the factors maintaining suitable stream habitats is crucial for conserving sensitive biota, especially in intensively used landscapes (Weijters, Janse, Alkemade, & Verhoeven, 2009). Existing frameworks recognize stream habitats are hierarchically organized from catchment to microhabitat spatial scales (Frissell, Liss, Warren, & Hurley, 1986). Within this hierarchy, physical heterogeneity is created by interactions among broad-and fine-scale phenomena, giving rise to sets of filtering mechanisms collectively determining local biotic composition (Poff, 1997). Insight into dynamics at any single scale may be gained by examining the context imposed, and uniqueness provided, at broader and finer spatial scales, respectively (Poole, 2002).
The hierarchal organization of stream habitat often causes the effects of regional land use disturbance to propagate through spatial scales before impacting local habitat and biota (Burdon, McIntosh, & Harding, 2013;Isaak & Hubert, 2001;Maloney & Weller, 2011).
Alternatively, sensitive biota may persist within marginal landscapes because of the insulating effects of intermediate-scale features including unique channel geomorphology, groundwater inputs, deep pools and riparian cover (Sponseller, Benfield, & Valett, 2001;Torgersen, Price, Li, & McIntosh, 1999;Walters, Leigh, Freeman, Freeman, & Pringle, 2003). Knowing which intermediate-scale habitat features propagate or offset land use signals could help managers assess the potential for intensively used landscapes to support sensitive populations.
Many imperilled stream fishes have recently fragmented, localized distributions resulting from decades of insufficient recruitment and lost population connectivity (Schlosser & Angermeier, 1995). Land use change can facilitate this process and frequently culminates in generalists or cosmopolitan species replacing resource specialists. Moreover, many specialized fishes are restricted to upland regions (Hoagstrom, Ung, & Taylor, 2014), further heightening imperilment risks (Pritt & Frimpong, 2010).
For example, 38% of imperilled North American freshwater fishes are both narrowly distributed and threatened by habitat degradation (Jelks et al., 2008). Fine sediment and stream warming are two suspected primary local stressors as both generally increase with intensified land use (Poole & Berman, 2001;Scott, Helfman, McTammany, Benfield, & Bolstad, 2002), yet knowledge of the individual and combined effects of these stressors is overwhelmingly limited to sportfishes (Kemp, Sear, Collins, Naden, & Jones, 2011;Lynch et al., 2016). Insight into specific impacts of land use on upland fishes is especially germane to conservation given these impacts will be compounded by future climate (Martinuzzi et al., 2014;Staudt et al., 2013).
One approach to resisting the impacts of future land use and climate is proactively restoring ecosystems that support sensitive species (Stein et al., 2013). It may be untenable, however, to implement restorative actions at broad spatial scales encompassing entire ecosystems (e.g., reforesting entire catchments). Therefore, resistance-building strategies may instead target more manageable features that mitigate specific stressors including warming water and sediment loading (Folke et al., 2010). For example, stream temperatures can rapidly respond to shading provided by local riparian cover (Moore, Spittlehouse, & Story, 2005). Strategically deciding which local features to restore, augment, or preserve may be informed by retrospectively examining distributional changes of sensitive biota following historical disturbances, given these changes reflect underlying population dynamics. For example, a multi-decadal investigation by Harig and Fausch (2002) discovered deep pools were a key local habitat for translocated Cutthroat Trout (Oncorhynchus clarkii), a finding that currently guides conservation within the context of multiple regional stressors (Haak & Williams, 2012).
We developed a conceptual framework to examine distributional changes of an imperilled upland stream fish, Candy Darter (Etheostoma osburni, Hubbs & Trautman), over multiple decades.
The region's complex topography, conducive to refugia and heterogeneous land use history, make this context ideal for investigating biotic outcomes following intensive land use disturbance. We specifically asked, (a) what are the environmental correlates of refugia? (b) are the pathways by which land use impacts instream habitat constrained by catchment-and/or segment-scale features? and (c) are E. osburni distributional dynamics spatially structured and explained by fine sediment and warm stream temperatures?

| Hierarchical resistance framework
Our framework organized our investigation and may serve as a useful guide for other investigations into mechanisms underpinning distributional changes of stream biota. We integrated concepts from other hierarchical frameworks depicting the creation and maintenance of stream habitat (Hierarchical Patch Dynamics; Poole, 2002), the matriculating impacts of land use disturbance on streams (Land-Cover Cascades; Burcher, Valett, & Benfield, 2007), the scales at which management actions can interrupt regional disturbances (Multi-scale Resilience; Folke et al., 2010), and how multi-scale factors create environmental incompatibilities with species traits (Landscape Filters; Poff, 1997).
We hypothesized the effects of catchment-wide land use could be either propagated or offset at catchment (broadest) and segment (intermediate) spatial scales before affecting instream habitats (finest) and populations. We depict alternative outcomes of  Figure 1a). For example, intensive segment riparian land uses are unlikely to reduce the rates at which heat and fine sediment accumulate downstream in catchments with intensive land uses (Jones et al., 2001). This impact pathway would likely diminish the resistance of a local population at multiple scales.
Similarly, catchments and segments may be less amenable to, and/ or have habitats that are less impacted by, intensive land uses, causing features at both scales to jointly convey resistance to populations ( Figure 1b). We default to the term resistance when referring to persistent populations in changing landscapes but recognize persistence could also reflect resilient (recovered) habitat and populations following disturbance (Lake, 2000).
In contrast to propagating pathways, segment-scale features may offset signals originating at catchment scales (offsetting segment pathways). These segment features may mitigate disturbance signals originating from intensive catchment land uses, thereby contributing to local refugia within mostly degraded catchments ( Figure 1c). For example, particularly steep channel segments can limit siltation and preserve critical habitat for benthic biota (Walters, Leigh, & Bearden, 2003). Conversely, intensive near-stream land uses could impact instream habitat causing localized absences in catchments that are otherwise suitable (pathway not shown in Figure 1).
Biotic factors could also affect population resistance at multiple spatial scales. Within a catchment, frequent dispersal from neighbouring populations could enable a population to persist in suboptimal habitat (Figure 1d), or degraded corridor habitat could inhibit recolonization (Figure 1e; Albanese, Angermeier, & Peterson, 2009).
These pathways may be particularly important given dispersal ability, corridor suitability and network topology may influence stream fish responses to changing bioclimatic conditions (Comte & Grenouillet, 2015). Finally, species with certain traits may be especially sensitive to habitat changes linked to land use disturbance, including species F I G U R E 1 A conceptual framework depicting how physical and biotic factors (rectangles) at catchment and segment scales could interact with top-down land use disturbance signals (black arrows) to affect the persistence of a stream fish population. Minus signs (left positioning) and plus signs (right positioning) within factors indicate whether factors diminish or convey population resistance, respectively. (a) Resistance-diminishing catchment and segment environmental features propagate effects of a moderate land use disturbance leading to extirpation. (b) A population persists in a broad refugium shaped by resistance-conveying catchment and segment features. (c) Disturbance is first propagated by catchment features and then offset by segment features, allowing a population to persist in a local refugium. (d) High metapopulation connectivity maintains a population despite land use disturbance. (e) Absence of metapopulation connectivity prevents individuals from recolonizing a site with benthic, visual or coolwater dependencies, invertivory or limited mobility (Kirsch & Peterson, 2014;Scott, 2006).

| Study area
Our study occurred in the New River drainage (NRD), a northwardly draining system within the Appalachian Mountains of the eastern United States (Figure 2). Surveys were restricted to the historical range of E. osburni in the Appalachian Plateau (AP) and Valley and Ridge (VR) physiographic provinces. The AP is a more northern high-elevation province, typically underlain by clastic and shale geology. Located south and east of the AP, the VR comprises a mosaic of sedimentary rocks, creating steep continuous ridges separated by wide unconfined valleys (Messinger & Hughes, 2000). Mixed-deciduous forest would predominate the area, but European settlers (c. post-1750) cleared many valleys. Most remaining forests were subsequently logged (c.1880-c.1920), resulting in extensive wildfires, erosion and impacts to fish populations (Goldsborough & Clark, 1908). Current land uses include regrown forests (73%), agriculture (15%; namely, pasture) and urban development (7%).

| Focal species
Etheostoma osburni is a non-game fish endemic to the NRD that was recently proposed as "Endangered" under the Endangered Species Act (U.S. Federal Register 83: 58747-58754). The species is a member of a rich upland fish fauna concentrated in the temperate and subtropical regions of the south-eastern United States, a hotspot for freshwater biodiversity and imperilment (Burkhead, 2012). Sparse early surveys  documented E. osburni in several stream types, indicating its distribution before European settlement was wider than its current distribution suggests (Jenkins & Burkhead, 1994). However, recent absences of E. osburni from numerous streams suggest it has declined, especially in its southern range. Etheostoma osburni possess traits conferring low resistance to anthropogenic disturbance, including a narrow geographical range and specialized resource needs. Moreover, as a small benthic fish, it may be a poor disperser, sensitizing it to degraded corridor habitat and other constraints on colonization (Schlosser & Angermeier, 1995). Similar traits are shared by many other upland fish species (Hoagstrom et al., 2014).

| Site selection
Study sites were within 2nd-5th Strahler-order streams; most had records of E. osburni presence ( Figure 2; Table S1 has collection dates and site abbreviations). We avoided areas inundated by Bluestone Lake reservoir and downstream sections of some tributaries to the New, Gauley and Greenbrier rivers where extirpations may have resulted from Variegate Darter (E. variatum), a recently introduced competitor capable of hybridizing with E. osburni. We also surveyed sites in seven streams without historical records in E. osburni's southern range, where unknown localized populations would provide the greatest insight into refugium-forming habitats. These seven streams were within (five) or adjacent to (two) catchments with historical (pre-1970) records.
We determined E. osburni presence-absence within a stream segment (site), the length of stream between two consecutive tributary confluences (mean ± SD segment length = 4.1 ± 3.7 km). We surveyed one segment in streams with recent (post-1970) records (13 streams, 13 segments). We surveyed two segments in streams with only historical records (14 streams, 27 segments) in case declining populations were more patchily distributed. The exception was STN1, an adventitious stream with a historical record, where we surveyed only one segment due to its short length. If available, we selected segments overlapping previous collection localities. We randomly selected two segments of the same order in streams without historical records. Overall, we conducted a range-wide survey based on available records of E. osburni.

| Fish and habitat surveys
We surveyed 42 segments during May-July, 2012. We limited surveys to riffles, where E. osburni spends most of its life (Dunn & F I G U R E 2 Stream segments (study sites; n = 42) sampled for Etheostoma osburni in 2012 within the New River drainage, Virginia and West Virginia, USA. Map projection: Universal Transverse Mercator, zone 17 N Angermeier, 2016). Within each segment, we randomly selected three geographical coordinates within 300 m of a road or trail and sampled the closest riffle to each coordinate twice to limit imperfect detection (i.e., three riffles per segment). At the base of each riffle, two crewmembers held a 1.5-m × 3-m seine with 5-mm bar mesh and a double-weighted leadline. Beginning 3 m upstream of the seine, a third crewmember electrofished (Smith-Root LR-24 backpack, pulsed direct current) downstream while disturbing the substrate (one kick-seine). We repeated kick-seines along transects perpendicular to flow until the entire riffle was sampled. The estimated probability of failing to detect E. osburni within an occupied segment was <1% (C. Dunn, unpublished data). Individuals were widely distributed within occupied segments (84% of riffles), indicating segments were an appropriate spatial grain to examine relationships between habitat and population persistence.
After fish sampling, we estimated the means of five habitat variables in each sampled riffle. First, five equally spaced transects perpendicular to flow were placed in each riffle. Next, we recorded wetted-channel width and identified seven equidistant points along each transect to record stream depth, water-column velocity at 60% depth (Marsh-McBirney model 2000 flow meter), intermediate-axis width of a randomly selected substrate particle, and two visually estimated descriptors of fine sediment within the 0.5-m 2 area surrounding each point. Embeddedness was the average percentage of coarse substrates vertically covered by fine sediment; silt-cover was the percentage of substrate-surface area covered by silt. Percentages for both descriptors were categorized as 0 = ≤5%, 1 = 6%-25%, 2 = 26%-50%, 3 = 51%-75% and 4 = 76%-100%.
Finally, all 105 habitat observations were averaged within each segment to represent segment-wide riffle characteristics.
We also recorded hourly water temperatures in each stream in spring (1 March-31 May 2012) and summer (1 June-31 August 2012) with submersible temperature loggers (Hobo Pendant or Tidbit v2, Onset Computer Corporation, Pocasset, MA). Loggers were placed in well-mixed flowing water in each segment. Appendix S1 describes details of logger deployment and procedures for predicting missing temperatures.

| Organization of data analyses
Our analyses follow the hierarchical framework presented above.
We first used principal component analysis (PCA) and partial redundancy analysis (pRDA) to broadly investigate the correlates of refugia and pathways through which catchment-and segment-scale features influence instream habitat (i.e., temperatures and riffle habitat variables). Next, we used logistic regression to test whether two potential local stressors ultimately affected persistence and whether distributional patterns were spatially organized. These analyses collectively allowed us to gauge which impact pathways shaped distributional patterns ( Figure 1).
TA B L E 1 Groupings and data sources for environmental variables used in multivariate analyses. Appendix S2 has references for data sources

| Multi-scale correlates of refugia and impact pathways
We first visualized the similarity of instream habitats among segments via PCA. We standardized the means of seven instream variables within each segment: spring mean daily temperature (SPMDT), summer mean daily maximum temperature (SMDMX), depth, embeddedness, silt-cover, substrate size and velocity (

| Effects of temperature, fine sediment and space on persistence
We hypothesized warm temperatures and fine sediment would negatively affect E. osburni persistence. We selected two thermal metrics representing alternative hypothesized impacts on fitness: SPMDT represented temperatures during its spawning season, while SMDMX represented potentially heightened metabolic costs near its thermal tolerance. High collinearity between silt-cover and embeddedness (r = 0.75) precluded us from using both fine-sediment metrics. Therefore, we retained embeddedness because it better reflects substrate complexity needed for cover, spawning and foraging.
Variation in population connectivity among sites could also contribute to distributional change. However, population connectivity cannot be measured directly from distributional data without knowing whether intervening, unsampled segments are occupied. As a potential surrogate for population connectivity, we used distance-based eigenvector mapping (DBEM) to produce spatial eigenvectors, which represent the proximities of sites to one another along dispersal corridors within the fluvial network (Griffith & Peres-Neto, 2006 A randomization procedure with 999 replicates indicated presences and absences were spatially autocorrelated (Moran's I = 0.21, p < 0.01). Therefore, we used DBEM to convert the connectivity matrix into spatial eigenvectors and then retained the first two eigenvectors as spatial covariates in models.
We used logistic regression and an information-theoretic approach to quantify the relative support of models explaining E. osburni's distribution. Competing models were ranked by Akaike's information criterion corrected for small sample size (AICc), a metric for relative model fit penalized by model complexity (Hobbs & Hilborn, 2006). The 15 candidate models were limited to four predictors to prevent over-parameterization. Finally, to assess model performance, we also report the mean area under the receiver operating curve (AUC) of each model via sevenfold cross-validation (Manel, Williams, & Ormerod, 2001); values of 0.5 and 1.0 correspond to random and perfect predictive ability, respectively.

| Current distribution
We detected E. osburni in all 11 northern-range segments but in only 5 of 31 segments in the southern range ( Figure 2). We discov-

| Environmental correlates of refugia
The PCA revealed the main gradients in instream habitats among segments ( Figure 3a; Table 2). Principal component 1 (PC1, horizontal axis) explained 43.3% of variation and likely represented increasing habitat degradation towards the southern range (right side of Figure 3a is most impacted). Positively loaded segments on PC1 were warmer, more embedded and contained finer substrates and faster water velocities, likely due to greater summer baseflows in many southern VR streams. Principal component 2 (vertical axis) explained 24.8% of variation and reflected greater water depths and less silt-cover.
Axes were highly correlated with features at broader spatial scales (Table 2). Positively loaded segments on PC1 (more degraded) had more agricultural (Figure 3b,d) and developed land (Figure 3d), had carbonate rather than clastic geology (Figures 3b,c), and oc-

| Transmission of catchment disturbance to instream habitat
Partial RDA revealed the main pathways through which catchment features influenced instream habitat ( Figure 4;

| Local and spatial predictors of persistence
Population persistence was inversely related to temperature and embeddedness, and sensitive to a site's location within the fluvial network ( Variables explaining significant variation in scores from PC1 and PC2 via a permutation test (p ≤ 0.10) using package "vegan" (Oksanen et al., 2015). b Variables retained for partial redundancy analysis following stepwise selection using package "vegan" (Oksanen et al., 2015).  (Figure 5a). Similarly, the maximum rate of change in predicted presence for embeddedness occurred at 0.97 or approximately 25% riffle embeddedness (Figure 5b).
Spatial covariates indicated population persistence was spatially structured by latent processes or features. Interpreting spatial covariates is not straightforward because multiple non-mutually exclusive factors can contribute to spatial patterns. The first spatial covariate (hereafter "SP1"; ̂ = −3.80, 5.02 CI90) is geographically interpretable, with sites in northern and eastern portions of the Gauley and Greenbrier river subbasins having the highest probability of presence and sites in the southern range having the lowest.
High occupancy of E. osburni within the Gauley (100% of sites) and Greenbrier ( (Table 3) and imposed a spatial context on predictions. For example, an optimal spatial location counterbalanced negative effects of warm temperatures and embeddedness on persistence (Figure 5c,d). Comparisons between predictions from spatially explicit models and those using identical median scores for spatial covariates (i.e., controlling for space) indicated LWA2, SEC1 and SEC2 were unoccupied spatially suboptimal segments containing suitable spring temperatures and low embeddedness (Figure 5a,b) TA B L E 3 Akaike's information criterion (AICc) of models with effects of stream temperature, substrate embeddedness (EMB) and/or two spatial covariates (SP) on Etheostoma osburni presence in 42 stream segments in the New River drainage, USA. ΔAICc is the difference between top-and lower-ranked models (i). Model weight (W i ) is the probability of a model being the best-supported model. Evidence ratio (W 1 /W i ) is the number of times the top-ranked model is better supported over lower-ranked models. Area under the curve (AUC) is a thresholdindependent measure of cross-validation predictive ability (0.5 AUC = random, 1.0 AUC = perfect) spatially optimal locations in the Gauley River subbasin, GAL1 and LWV1, with high embeddedness (Figure 5b, Table S5).

| D ISCUSS I ON
We documented probable extirpations of E. osburni from at least seven streams in the species' southern range. This fragmented, localized distributional pattern is consistent with many other sensitive stream species' distributions, and we hypothesize these patterns result from at least three interacting phenomena: (a) spatial variation in landscape features, (b) propagation of regional land use impacts to instream habitat via indirect pathways and (c) insufficient corridor habitat to facilitate recolonization following disturbance-mediated extirpation.

| Multi-scale correlates of refugia and impact pathways
The distribution of catchment-wide abiotic features, including high elevations, rugged topography and clastic geology, seemed to govern the locations and variable sizes of refugia. These same features coincided with northern-range E. osburni populations and with most of the isolated populations in its southern range. High covariation with catchment land use indicates abiotic catchment features may initially influence population persistence by mediating the extent, intensity and duration of intensive land uses through time. For example, E. osburni was extirpated from catchments with more intensive agriculture and, to a lesser extent, urban development, which were enabled by amenable physical features, such as lower elevations and subdued topography associated with carbonate geology (Hack, 1957). Agriculture in these catchments (28.8%) exceeded levels that have impacted other upland fishes in the region [10%-20% non-forest (Sutherland, Meyer, & Gardiner, 2002); 12% agriculture ( Impacts of catchment land use disturbance appeared to transfer to instream habitat primarily via indirect propagating pathways ( Figure 1a). Multi-scale habitat relationships indicated several segment features potentially contributed to impact pathways, including easily weathered carbonate geology, which enables streamside pasture in low-relief valleys and channel widening in deforested unconfined channels (Hack, 1957). Rather than being subjected to the effects of a chronic press disturbance from agriculture (sensu Lake, 2000), many currently forested catchments with F I G U R E 5 Predicted probabilities, with 90% confidence intervals, of Etheostoma osburni presence along (a) temperature and (b) embeddedness gradients in the New River drainage, USA. The upper horizontal axis (b) presents the range in percentage embeddedness encompassed by categories of the embeddedness index (0-2 shown). Presences and absences in segments (n = 42) were plotted at 1 and 0, respectively. Filled and open circles represent segments where E. osburni was either documented or not documented by historical surveys, respectively. Panels c-d demonstrate the influence of spatial location on probabilities across temperature (c) and embeddedness (d) gradients at the 75th ("optimal" spatial context), 50th ("neutral" spatial context) and 25th ("suboptimal" spatial context) percentiles of two spatial covariates. We added codes for sites referenced in the main text.
resistance-conveying features were mainly exposed to an intense, but short-term, pulse disturbance from historical logging. Resistanceconveying features such as rugged topography and weathering-resistant clastic geology typically promote steep stream channels (Hack, 1957), which can limit fine-sediment deposition (Montgomery & Buffington, 1997) and afford access to high-elevation thermal refugia (Isaak & Rieman, 2013). Together, such features contribute to extensive habitat suitability for E. osburni, which may have enabled populations to resist shorter-term pulse disturbances (Figure 1b).
Unfortunately, no historical habitat data are available, beyond cursory descriptions, to more definitively demonstrate these impact and resistance pathways. However, unlike entirely static interpretations of land use disturbance impacts, which can be confounded by covariation with abiotic features (Allan, 2004), our retrospective investigation revealed that even intensive land uses operating within the context of abiotic features can result in distributional change. Generalizing our results, along with those from similar snapshot investigations (e.g., Jones, Helfman, Harper, & Bolstad, 1999;Sutherland et al., 2002;Walser & Bart, 1999), would benefit from additional retrospective investigations into distributional changes of upland fishes in areas with variable land use disturbance histories.

| Instream predictors of persistence
Ours is one of the first studies to link temporal distributional changes for a non-game fish with field-measured stream temperatures. Segments not supporting E. osburni were on average 1.6°C warmer in spring and had 0.6°C higher summer mean daily maxima (Table 2). Warmer temperatures in many of these segments partly reflect natural conditions that once supported populations at lower elevations and latitudes. However, agriculture could have also warmed stream temperatures by increasing surface run-off and solar exposure (Poole & Berman, 2001;Trimble & Mendel, 1995). Our models could not confidently determine which thermal metric was more influential. Because virtually nothing is known about of E. osburni's thermal ecology, we cannot discount any of several temperature-related hypotheses potentially explaining E. osburni's decline, including summer exceedances of thermal tolerances, mis-timed spring spawning (Krabbenhoft, Platania, & Turner, 2014), diminished performance at specific life stages (Turschwell, Balcombe, Steel, Sheldon, & Peterson, 2017) and interspecific interactions (Lawrence et al., 2014).  (Rice & Jastram, 2015).
Stream biota can be exposed locally to multiple stressors symptomatic of regional disturbances (Townsend, Uhlmann, & Matthaei, 2008). Our top models were consistent with this pattern and also included a negative effect of embeddedness, which can impact biota in several ways (Kemp et al., 2011;Wood & Armitage, 1997). Like many other upland fishes, E. osburni possesses several biotic traits potentially sensitizing it to fine sediment. For example, invertivory and lithophilous spawning could cause sensitivity to diminished substrate complexity across multiple life stages, or embedded substrates could reflect high levels of suspended fine sediment, which can diminish the foraging efficiencies of upland fishes (Zamor & Grossman, 2007). Overall, our observations at finer scales provided rare insight into distributional fragmentation and suggest the localized distributions of other upland stream species could reflect multiple local stressors.

| Potential factors spatially structuring populations
Inclusion of spatial covariates in our models partly explained rare mismatches between instream-habitat suitability and population and/or metapopulation connectivity (Figure 1d; Dunham & Rieman, 1999). Although we cannot discount the two former factors, genetic analysis indicates recent dispersal among respective E. osburni subpopulations within the Gauley and Greenbrier river subbasins (Gibson, 2017). This corroborates estimated effects of spatial covariates and suggests varying dispersal rates across the species' range contributed to population persistence.
Spatial covariates also explained multiple absences in southern-range segments with suitable habitat but spatially suboptimal locations. This pattern may have also resulted from latent environmental gradients, an inability to recolonize sites (Albanese et al., 2009; Figure 1e) or both. For example, southern-range segments in large corridor streams (catchments ≥100 km 2 ) were warmer (x SPMDT = 14.3°C ± 0.2 SE, n = 18) and more embedded (x emb.index = 0.9 ± 0.1 SE) than those in the northern range (x SPMDT = 11.8°C ± 0.4 SE; x emb.index = 0.4 ± 0.1 SE, n = 10).
Diminished corridor habitat could further isolate remaining southern populations, thereby potentially limiting recolonization. If distributional patterns arose from spatially variable population connectivity, translocation may be a necessary resistance-building strategy to mitigate diminished dispersal (Olden, Kennard, Lawler, & Poff, 2011).

| Introduced species as potential stressors
The replacement of specialized endemics by more generalized, and often introduced, species frequently contributes to biotic homogenization (Scott, 2006;Walters, Leigh, & Bearden, 2003 (Kovach et al., 2017;Labbe & Fausch, 2000). Moreover, the regional prevalence of degraded areas could mediate propagule pressure of more generalized competitors and predators on upland refugia (Lapointe, Thorson, & Angermeier, 2012), thereby leading to deleterious interactions at local scales (Merriam & Petty, 2016). Intensive catchment-wide land uses often improve habitat suitability for thermally and finesediment-tolerant species (Scott, 2006). Indeed, retrospective analyses by Hitt and Roberts (2012) and Buckwalter, Frimpong, Angermeier, and Barney (2018) in the NRD documented replacements of specialized upland natives with more generalized fishes; both studies implicated land use in these dynamics. Future research could extend our approach to examine whether multi-scale physical features and community dynamics jointly exacerbate or reduce effects of catchment-wide, top-down disturbances.

| Unique segment-scale contributions to local refugia
Although finer-scale features were primarily constrained by catch- 50.5% in CRP2), which may have helped maintain suitable habitat by limiting insolation and filtering fine sediment from run-off (Jones et al., 1999;Sponseller et al., 2001).
Discovering the population in CRP1 also demonstrated the con-

| CON CLUS ION
Land use is an important factor influencing the distributions of stream biota (Allan, 2004). However, investigations into distributional dynamics often overlook the indirect pathways through which land use disturbance either impacts or, just as importantly, fails to impact populations. Here, certain catchment and nested segment features were subjected to prolonged intensive land uses that likely compromised two local habitat variables important for E. osburni persistence: cool stream temperatures and unembedded substrates.
Cool stream temperatures will become increasingly rare as air temperatures rise (McDonnell et al., 2015), further threatening E. osburni and upland stream biota worldwide. However, our results indicate limiting effects of more manageable stressors (e.g., sediment loading) and potentially maintaining population connectivity may be viable strategies for resisting impacts of warming streams.
Conservation practitioners may benefit from our organizational framework in other cases where species distributions have become fragmented and localized. Retrospective analyses of distributional dynamics can reveal the primary scales through which features convey biotic resistance, and the variability of these relationships across a species' range. In turn, this knowledge could inform conservation strategies. For example, strategies may emphasize preserving population connectivity throughout broad refugia spanning multiple catchments, or opt to further insulate a single catchment refugium from introduced species. Alternatively, attempting to counterbalance impacts of regional disturbances with local resistance-building features may be the most appropriate strategy in catchments amenable to prolonged intensive land uses. Ultimately, the localization of species' distributions is a complex long-term process. By extension, conserving stream biota is equally complex but may be aided by clearer insights into the pathways through which anthropogenic disturbances impact stream habitats. government.

DATA ACCE SS I B I LIT Y
All data are provided in Tables S2-S4