Revealing species-specific habitat drivers of turtle occupancy in Northwest Oregon using Bayesian low-detectability models.

Revealing species-specific habitat drivers of turtle occupancy in Northwest Oregon using Bayesian low-detectability models. | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 21 January 2026 V1 Latest version Share on Revealing species-specific habitat drivers of turtle occupancy in Northwest Oregon using Bayesian low-detectability models. Authors : Laura Guderyahn 0000-0001-8828-1203 [email protected] , Jacob Schultz , Catherine de Rivera , Marc Hayes , and Daniel Taylor-Rodriguez Authors Info & Affiliations https://doi.org/10.22541/au.176900601.11448987/v1 123 views 88 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Estimating occupancy is difficult when detection is low and predictors are numerous and collinear—conditions typical of rare, cryptic turtles in human-modified landscapes. We surveyed 253 sites three times each across the northern Willamette Valley (Oregon, USA) and modeled occupancy for Actinemys marmorata (northwestern pond turtle), Chrysemys picta bellii (western painted turtle), and the non-native Trachemys scripta elegans (red-eared slider). Using Bayesian models with novel inherited shrinkage horseshoe priors, we identified tractable functional relationships between presence or detection probabilities and their corresponding predictors. Model estimates of the proportion of sites occupied (PSO) closely matched confirmed presence rates for A. marmorata (5.9% for each) and T. scripta elegans (22.5% for naïve estimates, 22.9% PSO). However, PSO was ~1.4× higher than naïve estimates for C. picta bellii. Detection varied with field conditions (e.g., overcast skies improved detection). Connectivity and edge configuration structured occupancy: C. picta bellii increased near other turtle-occupied waters, whereas A. marmorata was more likely at comparatively isolated sites. Submerged/floating aquatic vegetation favored C. picta bellii but seemed to disfavor T. scripta elegans; rather, gravel/mud shorelines predominated with T. scripta elegans. At the landscape scale, T. scripta elegans tracked road density, consistent with peri-urban introductions. In contrast, landscape associations for C. picta bellii weakened after accounting for within-site structure, suggesting that local site characteristics better explained occupancy. Our findings suggest clear actions: increase submerged/floating vegetation and soften bare edges for C. picta bellii, the former which may suppress T. scripta elegans; create or protect semi-isolated refugia for A. marmorata; and prioritize prevention in road-dense networks. These management suggestions merit testing within an adaptive management framework. The analytical approach shows promise for application to other taxa facing low detectability and high-dimensional habitat data. INTRODUCTION Determining where and how commonly species occur is especially challenging when detection is low and the candidate set of predictor variables is large—a frequent condition for rare or cryptic species in complex landscapes. Naïve presence/absence biases inference. Conventional maximum-likelihood occupancy models can struggle to converge or can return imprecise estimates when detection is low, predictors are collinear, or when using compositional data—predictors that are parts of a whole and sum to 1 (e.g., land-cover fractions)—especially with interaction terms (Aitchison 1986; MacKenzie et al. 2002; Dormann et al. 2013). In preliminary estimation rounds, low detectability and many predictors led to non-convergence and unstable/ill-conditioned standard errors (e.g., unmarked ; Fiske & Chandler 2011). This lack of reliability begs an approach that (1) explicitly accounts for imperfect detection, (2) shrinks weak or redundant effects to stabilize estimation, and (3) treats compositional habitat predictors without inducing multicollinearity. These conditions typify the Pacific Northwest’s two native freshwater turtles— Actinemys marmorata (northwestern pond turtle) and Chrysemys picta bellii (western painted turtle)—which are rare, cryptic, and occupy human-modified aquatic habitats. In the Willamette Valley of Oregon (USA), both turtle species face sustained pressures from habitat loss, altered hydrology, invasive species, and climate change (e.g., Gervais et al., 2009; ODFW, 2020). They serve as indicators of aquatic ecosystem condition and remain central to agency-led conservation (Cameron & St. Clair, 2002; ODFW, 2015). Actinemys marmorata is listed as Endangered in Washington and was petitioned for U.S. ESA listing as Threatened in April 2024 (Velasco, 2023; WDFW, 2024); both are designated “Sensitive–Critical” in Oregon (2020). Their ranges overlap narrowly in the upper Willamette Valley—Oregon’s most urbanized region—creating a critical setting to evaluate shared and species-specific habitat associations (Barela & Olsen, 2014). The valley has lost >57% of wetlands and 2011); channelization reduced floodplain connectivity (Flitcroft et al., 2023), and remaining still/slow-water habitats are fragmented and vulnerable to degradation (ODFW, 2020; Zamora-Marín et al., 2021). Projected warmer, wetter conditions will further alter habitat distribution, structure, and quality (Morgan et al., 2021; Shirk et al., 2021; Ren et al., 2023). Despite substantial investment in study and management, the features at site and landscape scales that most influence occupancy remain uncertain. A targeted review (see Chapter 1 of this dissertation) identified 172 studies linking turtle habitat to occupancy/abundance, but only 54 of these explicitly modeled how habitat variables influence detection and just 13 occurred in the PNW (12 on the two native species). Across detection-aware studies, positive signals were strongest for less surrounding development, greater depth/hydroperiod, more underwater refugia, warmer nesting microclimate/less dense near-edge vegetation, shorter distances to other ponds, more open and shallow water area, slower lotic flow, greater distance to forest, higher productivity, and more potential nesting area. Limited regional replication and mixed edge/cover results are exacerbated by the fact that many studies modeled detection rather than the underlying occupancy state, even though detection is imperfect and variable. These gaps, together with the statistical challenges above, motivate our approach. Using a region-wide occupancy dataset from the Willamette Valley, we mapped hypothesized drivers to operational predictors: pond size and surface composition (proxies for depth/hydroperiod and in-pond structure); shoreline composition, floating dead wood, and gravel/mud (edge microclimate and haul-out); distances to occupied sites, streams, and forests (connectivity and thermal/light context); and land-cover clusters, roads, and disturbance ranks (urbanization/disturbance). These predictors are expected to matter because hydroperiod and depth govern thermal stability, overwintering, and drought risk; in-pond and edge structure provide basking/refuge and influence body temperature and predator exposure (and thus detectability); haul-outs like wood and gravel bars increase basking opportunities and observer encounters; connectivity to occupied sites and streams facilitates dispersal and recolonization while forest context alters shading and water temperature; and urbanization/roads shape human disturbance, introductions, and mortality risk. Because many predictors were proportions, we grouped correlated land-cover components into interpretable habitat bundles to limit collinearity. We modeled detection separately from occupancy to account for imperfect detection, then applied shrinkage to isolate robust habitat signals relevant to management. Specifically, we aimed to: 1. Test local habitat structure and edges; 2. Test landscape connectivity and matrix; 3. Evaluate interactions that qualify connectivity (e.g., with forest/stream context); 4. Model detection to correct bias and inform survey design; and 5. Compare species responses in sympatry to derive actionable guidance suggestions for conservation and restoration across the PNW. METHODS Study Area. —We worked across twelve HUC-8 watersheds in northwestern Oregon (Fig. 1), centered on the Willamette River and its primary tributaries (Clackamas, Molalla–Pudding, Tualatin, Yamhill). The region spans ~1.44 million ha from near sea level in metropolitan Portland to >1,100 m in the Cascade foothills and has a Mediterranean climate (cool, wet winters; hot, dry summers; 2023). Aquatic habitats—oxbows, sloughs, emergent marshes, reservoirs, ponds, gravel pits—cover ~4,900 ha (≈3.3% of the landscape; USGS, 2023). Historic fluvial complexity persists in abundant off-channel features, though many sites are altered or hydrologically isolated by land use and flow modification. Occupancy Data Collection. —We estimated occupancy from regional visual-encounter surveys (VES) during the 2021–2023 basking seasons (see Chapter 2, this dissertation for details). The design combined high-priority sites with historical records and randomly selected sites drawn from a GIS inventory of suitable aquatic habitat to represent both known turtle locations and the broader habitat spectrum. We defined suitable aquatic habitat as any pond, wetland, slough, or slow-moving stream with a minimum depth of 0.5 meters for at least nine months of the year. Surveys followed established protocols (Bury et al., 2012; Samara Group, 2020): each site was surveyed 3 times under basking-conducive conditions from geo-referenced observation points using optical equipment. We constructed site-level detection histories and fit single-season occupancy models that correct for imperfect detection (MacKenzie et al., 2002). Habitat Variable Pre-Selection. —Guided by the targeted synthesis (see Chapter 1, this dissertation) and a priori hypotheses, we mapped ecological constructs to operational predictors as summarized in Table 1. Where direct measures (e.g., depth) were unavailable, we used ecologically justified proxies (pond size, surface composition). Variables are organized as field-collected (local) versus GIS-derived (landscape; full dictionary in Table S1). GIS-Derived Landscape Variables. —We computed metrics in ArcGIS Pro (version 3.5; ESRI, Redlands, CA, USA) from Oregon GEOHub sources (Table S2). The ordinary high-water (OHW; see Table S3 for list of acronyms) line—our reference geometry—was delineated in the field as the consistent line of recurrent inundation using convergent indicators (breaks in slope/vegetation, soil/color change, scour, wrack/staining, persistent debris). When indicators conflicted, we applied a weight-of-evidence rule—prioritizing indicators most directly linked to hydrologic action (e.g., staining/scour > wrack > vegetation break > slope) and the most recent ground-truth—and set the OHW conservatively at the higher-elevation, landward position when uncertainty remained. For example, where staining and scour were landward of the vegetation break, we adopted the landward (higher) line as OHW. From the OHW we calculated (Table S3): • Land cover within a priori 250-m buffers (chosen to avoid nested collinearity across concentric buffers), summarizing NLCD 2021 classes (open water; developed/impervious; barren; forest; grassland/pasture/open space; shrubs/crops; vegetated wetland). Because the total of land cover classes within each 250 buffer sum to 1, we applied small zero-replacement, centered log-ratio (CLR) transforms, then Ward’s hierarchical clustering (Ward, 1963) to obtain a few interpretable land-cover classes used as unordered factors (reference = most common class). • Road density (total road length/km²) within 250-m. • Distances (Euclidean from OHW) to the nearest stream (Strahler order 1–4), nearest forest patch (≥900 m²), nearest wetland (any size, >0.2 ha, >0.4 ha), and the nearest surveyed wetland occupied by any turtle species. Skewed continuous variables were transformed as log(x+1) and all continuous predictors were z-standardized before modeling. Only the 250 m scale entered final models; larger buffers are listed in Table S1. Field-Collected Local Variables. —During basking-season surveys, trained observers recorded site attributes from geo-referenced vantage points using standardized protocols (see Chapter 2, this dissertation): • Aquatic footprint (planform area at OHW; used to compute Pond Area). • Surface vegetation composition (% open water, emergent, submerged/floating). • Shoreline vegetation within 5 m landward of OHW (% trees, shrubs, graminoids/forbs, barren). • Basking structure availability and type, classified as near-shore (≤10 m), far-from-shore (>10 m), and island-associated (counts/coverage; live/dead natural materials vs. human-built structures). • Human disturbance (Disturbance Rank 0–3: 0 = public/livestock access to water; 1 = access behind barrier/paved trail ≤25 ft; 2 = ≥50 ft separation and/or dense buffer; 3 = no public/livestock access within 500 m). Surface and shoreline compositions were treated like landscape compositions (zero-replacement → CLR → Ward clustering) to derive surface and shore classes used as factors (reference = balanced/open-grass). Local continuous covariates used in models included DeadBaskLogs and GravelMudShore. Detection covariates (effort, temperature, cloud, etc.) were recorded concurrently. Data Preparation and Screening. —We reviewed all local, landscape, survey condition variables (Table S1, Fig. S1). Sites with no viable habitat (defined as ponds/wetlands/sloughs/slow streams with fitting but were retained as zeros for calculating proportion of sites occupied (PSO) as zeros. Air temperatures recorded as ranges were replaced by midpoints (e.g., 21–24 °C → 22 °C). Detection covariates (temperature, wind, cloud cover, survey effort in minutes, time of day, day-of-year, observer) were standardized to common units and checked for outliers. Heavily skewed continuous predictors (e.g., distances, road density) used log(x+1); all continuous predictors were z-standardized (mean 0, SD 1). We removed near-zero-variance terms and screened the remainder using pairwise |r|. Variables collinear with multiple others (|r| ≥ 0.70 with ≥2 predictors) were removed first; for collinear pairs, we retained the predictor with higher variance and/or stronger a priori relevance. For land-cover proportions (unit-sum), closure issues were handled via the composition→cluster workflow above (Table S1). Predictor Summaries and Compositional Predictors. —Continuous predictors were summarized with mean, SD, range; categorical with counts/percent. Three predictors were compositional: water-surface cover, shoreline vegetation (≤5 m), and surrounding land cover (250 m). For each, we (1) converted parts to proportions (0–1), (2) applied zero-replacement (0.0001 unit mass) where needed, (3) computed CLR values, (4) performed Ward’s minimum-variance clustering, and (5) selected k using average silhouette width (k = 2–8) as a starting point, adjusting by ±1 to improve ecological interpretability (Fig. S2). We labeled clusters by dominant components (e.g., open-water dominated, emergent-dominated, tree-lined, ag/pasture mosaic, urban/impervious). Cluster membership entered models as unordered factors (effects coding; reference = most common class; exemplar plots in Fig. S3). Data Analysis (Bayesian occupancy modeling) — An occupancy model was employed to account for the non-zero chance of “false negative” observations, also called imperfect detection (MacKenzie, et al 2002). The combination of sampling methodology and target species was such that false detections were not reasonably expected. We also assume no changes in site occupancy during the two-season study period, which is supported by the slow-changing population dynamics of the target species. These assumptions give rise to binomial distributions in both presences z i for sites i = 1, …, N and the observations y ij for visits j = 1, …, n i . We used a logistic link for the binary regression, giving the model structure shown in Equation 1. , (1) where \(\beta\) and \(\alpha\) are vectors of regression coefficients corresponding to occupancy and detection predictors x i and w ij , respectively, and the ( β 0 , α 0 ) are the corresponding intercept terms. Initial occupancy models were fit within the frequentist framework using the unmarked package in R (R Core Team 2025, Fiske and Chandler 2011). However, the combination of a large model space (i.e., many predictors) and low confirmed presence across species led to numerical instability in unmarked , including errors in standard error estimation. A Bayesian approach was better suited in this case because priors help ensure the covariance matrix remains invertible, especially when using shrinkage priors, which “shrink” the effect of unimportant predictors toward 0, helping to further stabilize covariance estimation. The Bayesian framework provides several further advantages: (1) nuanced parameter estimation through posterior sampling, (2) straightforward quantification of uncertainty via credible intervals, and (3) coherent propagation of uncertainty to derived quantities such as the proportion of sites occupied (PSO). To identify meaningful predictors among the many candidate variables, while accounting for higher order terms and interactions, we made use of a novel version of horseshoe priors, which enable accounting for the hierarchical structure of the predictors (forthcoming in Schultz et al., 2026). The traditional horseshoe prior (Carvalho, et al., 2010) provides adaptive shrinkage by strongly shrinking coefficients of weak predictors toward zero while allowing coefficients of relevant predictors to remain large, thus facilitating effective variable selection. However, regression models including interactions and higher-order polynomial terms must also include all of the associated lower-order terms (Nelder, 1998; Peixoto, 1990; Chipman, 1996), in other words, they must respect the polynomial hierarchy. For example, a model that includes the terms A*B and C 3 ought to include A, B, C, and C 2 as well. Briefly, this rule is required in the regression context to avoid imposing unwanted constraints on the regression model, and more importantly for the purpose of this analysis, this requirement makes variable selection invariant to how predictors are coded (e.g., selection should not depend on whether temperature is in degrees Fahrenheit or Celsius). In this principle, known as strong heredity (Chipman, 1996), lower-order terms denote the parents of higher-order terms (i.e., the children terms). In the example above, C 2 would be a child of C and a parent of C 3 . The interaction term A*B has both A and B as parents. Under strong heredity, a child can only be included in the model if all of its parents are also included. In terms of the horseshoe prior, a term can be thought of as “excluded” when the variance for the prior on the regression coefficient is shrunk towards 0 – such that distribution of the regression coefficient is tightly concentrated around 0. Therefore, we enforce the strong heredity principle here by bounding a child’s prior variance to be smaller or equal to the smallest variance among all of its parents (see Equation 2). We call this principle inherited local shrinkage (Schultz et al., 2025). This ensures that if the regression coefficient of a higher-order term is not “excluded” (i.e., is not shrunk to 0), neither can be the coefficients for any of its parents. Furthermore, to correctly penalize categorical predictors in our model (accounting for their multiple levels), all categories of a given predictor share a single local shrinkage parameter (Chipman, 1996). Formally, the horseshoe priors with inherited local shrinkage used in our occupancy models are given by: , (2) where N(μ, σ 2 ) and \(C^{+}(0,\ 1)\) are (mean \(\mu\)and variance \(\sigma^{2}\)) normal and standard half-Cauchy distributions, respectively. The indicator functions describe truncations on the local shrinkage priors, with the upper bound m r being the smallest local shrinkage parameter in the r -th predictor’s set of parents. If the r -th predictor has no parents, m r \(=\infty\). Note that the regression coefficients’ prior variances are products of global and local shrinkage terms. The global shrinkage parameter \(\tau^{2}\) controls how much shrinkage applies to the entire set of predictors, with small values indicating that most predictors yield a weak signal, and vice versa. The local shrinkage parameters \(\lambda^{2}\) take large values to allow individual regression coefficients to escape the influence of the global shrinkage when the data supports this notion, and to strongly shrink towards zero coefficients if predictors behave as noise in the data. The horseshoe prior was applied to all predictor coefficients except the intercepts. In logistic regression, shrinkage of the intercept toward zero can bias the baseline probability toward 0.5. To avoid this, intercept terms were treated as part of the baseline model and assigned flat priors. We fitted three occupancy models with logistic link using a data augmentation approach with Polya-Gamma latent variables (Polson et al. 2013; Clark and Altwegg 2019). corresponding to the T. scripta elegans, C. picta bellii, and A. marmorata data using the nimble package in R (de Valpine et al. 2017). We also fitted a model for all turtles collectively but those results provided limited information and are omitted here. The package uses a Markov-chain Monte Carlo method (MCMC) to draw samples from the posterior distributions of all of the parameters of interest, including regression coefficients and shrinkage parameters. MCMC algorithms require a starting point from the user, and must be run for some number of iterations (burn-in) to allow the chain to approach the high-probability region of the posterior distribution. Chains were initiated at the maximum likelihood estimations (MLEs) obtained from the unmarked models . We obtained 10 5 post-burn-in samples for each of the four models, and achieved convergence, as shown in the trace plots (Fig S4). Effect Importance Measures .—We used two measures to identify important predictors: probability of direction (PD, denoting the maximum probability of effect) and pseudo posterior inclusion probability\(P(Inc)\). PD is the maximum probability that the parameter value is either positive or negative, computed as max ( Makowski et al. 2019). If the bulk of the posterior distribution is far from 0, PD ≈ 1, but if the posterior is centered at zero PD = 0.5. In the context of horseshoe priors, a pseudo posterior exclusion probability or “shrinkage factor” ĸ j measures the amount of shrinkage the model applies to the j th predictor, where ĸ j =1 corresponds to strong shrinkage towards 0. As such, P(Inc) = 1 – ĸ j (see supplemental information for details). Because the horseshoe prior imposes continuous rather than discrete shrinkage, it does not yield a formal posterior inclusion probability as in spike-and-slab models. Instead, 1- ĸ j provides a pseudo inclusion probability, quantifying the extent to which each coefficient escapes shrinkage, and thus serves as a continuous measure of predictor importance. This interpretation follows established practice in the literature on global–local shrinkage priors (Carvalho, et al., 2010; Piironen & Vehtari, 2017). PD and \(P(Inc)\) differ in two key aspects. First, PD is computed directly from the posterior distributions of regression coefficients, yielding unique values for each coefficient. In contrast,\(P\left(\text{Inc}\right)\ \)is derived from the shrinkage parameters, resulting in a single \(P\left(\text{Inc}\right)\ \)value shared across all levels of a categorical predictor (Table S4). Consequently, \(P(Inc)\) is well-suited for identifying important categorical predictors, whereas PD enables the detection of influential levels within these predictors. Second,\(P\left(\text{Inc}\right)\ \)offers an intuitive and defensible interpretation with a natural decision threshold at 0.5. If the\(P\left(\text{Inc}\right)\ \)exceeds 0.5, the predictor is more likely to be included in the model than excluded, indicating its relevance to the outcome variable (Bhadra et al. 2019; Carvalho et al., 2010). The PD lacks well-defined decision thresholds, as its values reflect the proportion of the posterior distribution lying on one side of zero, making interpretation context-dependent and less standardized than the \(P(Inc)\). This absence of a universal cutoff complicates direct comparisons of PD across predictors within a model or across species models (Makowski et al., 2019). PSO Calculation .— We estimated overall occupancy by adding the number of confirmed-presence sites to our estimate of the occupied non-detection sites, and dividing by the total number of sites (MacKenzie et al. 2003). In order to correctly account for habitat loss, the total number of sites includes 25 “No Habitat” sites that meet all the requirements to be studied, but could not be sampled due to a lack of suitable aquatic habitat. The PSO estimator is given in Equation 3. A more in-depth derivation can be found in the Supplemental Information. where z i =1 corresponds to true presence at site i , are the regression coefficient estimates at the b ’th MCMC iteration, and n abs , n pres , and n unc are the number of sites in the following three categories: no habitat (confirmed absence, z i = 0), sites with detection (confirmed presence, z i = 1), and sampled sites with no detections (uncertain). A single PSO estimate was computed for each species at every MCMC iteration, producing the PSO posterior densities in Fig. S5. RESULTS Survey effort and detections. —We surveyed 253 sites across the Willamette Valley. A summary of site counts, visits, and total surveys appears in Table 2. Twenty-five sites lacked viable habitat (see Methods) and were excluded from model fitting but were retained as zeros when estimating the PSO. Actinemys marmorata was observed at 5.9% of sites, whereas C. picta bellii (18.6%) and T. scripta elegans (22.5%) were ~3.2–3.8× more frequently detected (Table 3; Fig. S5). Posterior PSO medians were 5.9% ( A. marmorata ), 26.4% ( C. picta bellii ), and 22.9% ( T. scripta elegans ), with overlapping credible intervals. Detection model. —For C. picta bellii , nine covariates had \(P(\text{Inc})\geq 0.5\), led by road density within 250 m (negative; P[Inc] = 0.913), time spent surveying (positive; 0.895), pond area (positive; 0.842), and cloudy conditions (positive; 0.824) (Fig. 2, 3; Table S4). For A. marmorata , no detection covariate exceeded \(P(\text{Inc})\)=0.186, suggesting no variable or combination of variables strongly predicted detection of this rare species. For T. scripta elegans , medium and low disturbance ranks and the full-survey indicator each reached \((\text{Inc})\approx\ 0.616\) . Occupancy model. —We evaluated 19 predictors for occupancy (Fig. 2, 3; Table S5). For C. picta bellii , four predictors met\(P(\text{Inc})\) ≥ 0.50; distance to nearest occupied turtle site showed the strongest support (\(P(\text{Inc})\) = 0.926; negative coefficient), and surface-composition clusters 2 (open water and emergent veg dominant) and 3(open water and submerged veg dominant) also had support (each P[Inc] = 0.788; full cluster definitions in Fig. S3). For A. marmorata , distance to the nearest occupied turtle site had the highest inclusion (P[Inc] = 0.593) with the opposite sign (positive coefficient). For T. scripta elegans , fifteen predictors reached \(P(\text{Inc})\) ≥ 0.50; the strongest were % shoreline covered in mud/gravel (\(P(\text{Inc})\) = 0.877; positive), road density (0.873; positive), and disturbance ranks (each ≈ 0.756), with additional contrasts indicating lower occupancy where surface cover is submerged/floating-dominated and where shoreline cover is a mixed forest/shrub/grass mosaic. DISCUSSION Across 253 sites in northwest Oregon, our survey and occupancy modeling identified clear habitat-use signals for C. picta bellii and A. marmorata and distinguished true absences from non-detections due to low detectability. Jointly modeling presence and detection is essential for turtles, where low encounter rates and heterogeneous observation conditions can bias naïve presence/non-detection (MacKenzie et al., 2002). Consistent with this framework, occupancy varied most with within-site structure and broader landscape context, whereas detection responded strongly to field conditions and site setting. Interpreting the PSO in light of study design is critical. Because “no-habitat” sites contribute zeros by definition and confirmed presences contribute ones, PSO posteriors are bounded below by naïve occupancy and above by the fraction of sites with viable habitat (see Methods). The resulting right-skewed posteriors—with medians close to the proportion of confirmed presence for A. marmorata and Trachemys scripta elegans —are consistent with either genuinely low occupancy at non-detection sites or relatively high detection in the conditions we sampled. The bimodality in C. picta bellii PSO reflects uncertainty about whether open-water- or submerged aquatic vegetation–dominated surfaces truly influence occupancy: posterior draws that include a positive effect boost predictions at non-detection sites with those features; draws that exclude it do not. This is expected behavior when propagating coefficient uncertainty to a derived quantity and not evidence of misfit (MacKenzie et al., 2002). Presence does not guarantee population function. Even though 98 sites (38.7%) showed turtles (and PSO medians exceeded naïve detections; Table 3, Fig. S5), long-lived turtles can persist for decades as adult-heavy remnants when recruitment is chronically low—producing populations that are effectively “functionally extinct” despite continued occupancy (Congdon, et al, 1993; Congdon et al., 1994; Heppell, 2000). In such systems, population growth is most sensitive to adult female survival, but sustained viability still requires periodic recruitment; hence occupancy alone is an incomplete metric of conservation status. A practical extension is to pair site-use PSO with a “PSO-recruiting” state by treating evidence of recruitment (e.g., nests/hatchlings, small size classes) as a second biological state in a multi-state occupancy model, or by modeling a recruitment index within the detection framework (Royle & Nichols, 2003; MacKenzie et al., 2009). Detection patterns were ecologically and operationally coherent. Longer survey times increased detection for C. picta bellii , as expected for visual encounter methods (MacKenzie et al., 2002), and overcast conditions likely aided observers by reducing surface glare. In contrast, road-adjacent sites showed depressed detection for C. picta bellii , plausibly due to noise/traffic effects on basking behavior and/or restricted lines of sight along armored, steep banks. Roads are also well known to constrain turtle movement and survival (e.g., Gibbs & Shriver, 2002; Aresco, 2005), underscoring the need to distinguish detection penalties from true absence in developed settings. Three occupancy signals carried the strongest management relevance. First, connectivity to other occupied sites increased occupancy for C. picta bellii —consistent with short-distance exchange among nearby ponds and sloughs in still/slow-water networks (Bodie, 2001). In contrast, A. marmorata was more likely at comparatively isolated sites, aligning with its regional sensitivity to human disturbance and to propagule pressure from nearby waters that support T. scripta elegans (Gervais et al., 2009; ODFW, 2020; WDFW, 2024). A. marmorata status assessments in the Pacific Northwest have repeatedly emphasized the species’ vulnerability where human access and non-native turtles are common (Gervais et al., 2009; ODFW, 2020). Second, within-site edge configuration mattered. Submerged/floating aquatic vegetation (SAV) favored C. picta bellii , which frequently uses vegetated, still waters for foraging and cover (Bodie, 2001), whereas T. scripta elegans was associated with bare gravel/mud margins and woody-dominated banks that provide sun-exposed, stable haul-outs. Experimental and field studies have long shown that T. scripta elegans excel at securing basking structure—often to the detriment of native turtles—especially in simple, open-edge habitats (Cadi & Joly, 2003; Cadi & Joly, 2004). Regionally, T. scripta elegans in Oregon continues to nest and thrive in human-proximate, open microhabitats (Delaney et al., 2017; Lloyd & Warner, 2019), matching our edge-configuration signal. Taken together, these patterns support a working hypothesis: increasing SAV should directly elevate occupancy for C. picta bellii , be approximately neutral for A. marmorata (given present uncertainty), and simultaneously suppress T. scripta elegans by reducing the sun-exposed, unobstructed basking and haul-out edges T. scripta elegans exploit. This prediction is readily testable via adaptive trials (SAV additions and edge-softening) with before–after monitoring of basking use, detection, and occupancy. Third, landscape context aligned with known invasion pathways. T. scripta elegans occupancy increased in roaded, peri-urban settings—where introductions (pet releases) and engineered shorelines are common (Kraus, 2009; Gervais et al., 2009)—while native responses to roads were weaker or negative once within-site structure was considered. This is consistent with broader evidence that human access and simplified edges promote T. scripta elegans establishment and that road networks both enable introductions and fragment wetland complexes (Gibbs & Shriver, 2002; ODFW, 2020). Taken together, our results—and their agreement with regional context in the Willamette Valley, where wetland loss, hydrologic alteration, and urbanization have been pronounced (Christy & Alverson, 2011; Flitcroft et al., 2023; ODFW, 2020)—suggest a practical synthesis: structurally “soft” edges with substantial submerged aquatic vegetation tend to support C. picta bellii and are less favorable to T. scripta elegans , whereas exposed, open, human-accessed shorelines tilt the balance toward T. scripta elegans . For A. marmorata , semi-isolated refugia with lower human pressure remain a priority within a landscape of ongoing stressors (Gervais et al., 2009; WDFW, 2024). 1. Favor vegetated surface structure for natives. In native-focused restorations, increase SAV and avoid creating long, bare gravel/mud margins; such designs are consistent with C. picta bellii ecology and appear less attractive to T. scripta elegans (Bodie, 2001; Cadi & Joly, 2004; Guderyahn, 2025a). 2. Reduce T. scripta elegans-friendly edges. In T. scripta elegans -dominated ponds, break up continuous bare shoreline, remove artificial basking platforms, and add continuous herbaceous/emergent benches at the waterline; these steps directly target the features T. scripta elegans exploit (Cadi & Joly, 2003; Delaney et al., 2017; Lloyd & Warner, 2019). 3. Create and protect semi-isolated refugia for A. marmorata . Prioritize wetlands buffered from high-access nodes and from known T. scripta elegans source waters; pair habitat work with early-detection/rapid-response near parks, campuses, and stormwater ponds (Gervais et al., 2009; ODFW, 2020; WDFW, 2024). 4. Track “PSO-recruiting,” not just PSO : add standardized nest/hatchling checks or size-class tallies so monitoring can separate remnant adult sites from self-sustaining populations (Congdon et al., 1993, 1994; MacKenzie et al., 2009). 5. Triage effort in peri-urban networks. Focus prevention and control where roads and access are dense, not only in urban cores; design monitoring to offset detection penalties at road-adjacent and vegetated shores (standardized minimum dwell times, multiple vantages) (Gibbs & Shriver, 2002; MacKenzie et al., 2002). Finally, we emphasize uncertainty and the need for adaptive management. Several effects are modest with overlapping credible intervals, model transferability is limited by our spatial/temporal scope, and detection varied with conditions. Accordingly, we view the management implications above as testable hypotheses rather than prescriptions. Implement actions as small-scale experiments (e.g., submerged and aquatic vegetation additions, edge-softening, peri-urban prevention) with BACI or rotating-control designs, pre-specified decision thresholds (e.g., ΔΨ or Δ p ), and power-informed sample sizes. Update decisions with monitoring (posterior updating of PSO and p ; checks for recruitment), report learning, and revise or halt interventions when targets are not met. This adaptive, hypothesis-driven approach is best suited to refining tactics as pertinent new information appears. LITERATURE CITED Aitchison, J. 1986. The statistical analysis of compositional data. Chapman & Hall, London, UK. Aresco, M.J. 2005. Mitigation measures to reduce highway mortality of turtles and other herpetofauna. Journal of Wildlife Management 69:549–560. Barela, K., and D. Olsen. 2014. Native turtle monitoring in the Pacific Northwest: 2013 results. Oregon Department of Fish and Wildlife Report, Portland, Oregon, USA. 26 pp. Bhadra, A., J. Datta, N.G. Polson, and B. Willard. 2019. Lasso meets horseshoe: a survey. Statistical Science 34:405–427. Bodie, J.R. 2001. Stream and riparian management for freshwater turtles. Journal of Environmental Management 62:443–455. Bury, R.B., H.H. Georges, and D.A. Ross. 2012. Protocols for visual encounter surveys of turtles. U.S. Geological Survey, Amphibian Research and Monitoring Initiative (ARMI), Open-File Report 2012–1234:1–24. Cadi, A., and P. Joly. 2003. Competition for basking sites between the endangered European pond turtle and the introduced red-eared slider. Canadian Journal of Zoology 81:1392–1398. Cadi, A., and P. Joly. 2004. Impact of the introduction of the red-eared slider on survival of the European pond turtle. Biodiversity and Conservation 13:2511–2518. Cameron, M.A., and R. St. Clair. 2002. COSEWIC status report on the Pacific pond turtle Clemmys marmorat a in Canada. Committee on the Status of Endangered Wildlife in Canada, Ottawa, Ontario, Canada. 1–17 pp. Carvalho, C.M., N.G. Polson, and J.G. Scott. 2010. The horseshoe estimator for sparse signals. Biometrika 97:465–480. Chipman, H. 1996. Bayesian variable selection with related predictors. Canadian Journal of Statistics 24:17–36. Christy, J.A., and E.R. Alverson. 2011. Historical wetlands of the Willamette Valley, Oregon. Northwest Science 85:93–107. Clark, A.E., and R. Altwegg. 2019. Efficient Bayesian analysis of occupancy models with logit link functions. Ecology and Evolution 9(2):756–768. Congdon, J.D., A.E. Dunham, and R.C. van Loben Sels. 1993. Delayed sexual maturity and demographics of Blanding’s turtles ( Emydoidea blandingii ): implications for conservation and management. Ecology 74:1744–1754. Congdon, J.D., A.E. Dunham, and R.C. van Loben Sels. 1994. Demographics of common snapping turtles ( Chelydra serpentina ): implications for conservation and management of long-lived organisms. American Zoologist 34:397–408. Crouse, D.T., L.B. Crowder, and H. Caswell. 1987. A stage-based population model for loggerhead sea turtles and implications for conservation. Ecology 68:1412–1423. Delaney, D.M., F.J. Janzen, and D.A. Warner. 2017. Nesting stage and distance to refuge influence terrestrial nesting behavior of painted turtles ( Chrysemys picta ). Canadian Journal of Zoology 95:837–841. de Valpine, P., D.T. Turek, C.J. Paciorek, C. Anderson-Bergman, D. Temple Lang, and R. Bodik. 2017. Programming with models: writing statistical algorithms for general model structures with NIMBLE. Journal of Computational and Graphical Statistics 26:403–413. Dorazio, R.M., and D. Taylor-Rodríguez. 2012. A Gibbs sampler for Bayesian analysis of site-occupancy data. Methods in Ecology and Evolution 3:1093–1098. Dormann, C.F., J. Elith, S. Bacher, C. Buchmann, G. Carl, G. Carré, T. Diekötter, J.G. García Márquez, B. Gruber, B. Lafourcade, P.J. Leitão, T. Münkemüller, C. McClean, P. Osborne, B. Reineking, B. Schröder, A.K. Skidmore, D. Zurell, and S. Lautenbach. 2013. Collinearity: a review of methods to deal with it and a simulation study evaluating their performance. Ecography 36:27–46. Esri. 2024. ArcGIS Pro, version 3.5. Environmental Systems Research Institute, Redlands, California, USA. Fiske, I.J., and R.B. Chandler. 2011. Unmarked: an R package for fitting hierarchical models of wildlife occurrence and abundance. Journal of Statistical Software 43(10):1–23. Flitcroft, R., L. Whitman, J. White, R. Wallick, L. Stratton Garvin, C. Smith, R. Plotnikoff, M. Mulvey, T. Kock, K. Jones, P. Gruendike, C. Gombert, G. Giannico, A. Dutterer, D. Brown, H. Barrett, B. Bangs, and R.M. Hughes. 2023. Science to support conservation action in a large river system: the Willamette River, Oregon, USA. Water Biology and Security 2:100203. Gervais, J.A., D.K. Rosenberg, S. Barnes, C. Puchy, and E. Stewart. 2009. Conservation assessment for the western painted turtle in Oregon ( Chrysemys picta bellii ), version 1.1. Oregon Wildlife Institute, Corvallis, Oregon, USA. 61 pp. Gibbs, J.P., and W.G. Shriver. 2002. Estimating the effects of road mortality on turtle populations. Conservation Biology 16:1647–1652. Kraus, F. 2009. Alien reptiles and amphibians: a scientific compendium and analysis. Springer, New York, New York, USA. Lambert, M.R., J.M. McKenzie, R.M. Screen, A.G. Clause, B.B. Johnson, G.G. Mount, H.B. Shaffer, and G.B. Pauly. 2019. Experimental removal of introduced slider turtles offers new insight into competition with a native, threatened turtle. PeerJ 7:e7444. MacKenzie, D.I., J.D. Nichols, G.B. Lachman, S. Droege, J.A. Royle, and C.A. Langtimm. 2002. Estimating site occupancy rates when detection probabilities are less than one. Ecology 83:2248–2255. MacKenzie, D.I., J.D. Nichols, J.E. Hines, M. Knutson, and B. Franklin. 2003. Estimating site occupancy, colonization, and local extinction when a species is detected imperfectly. Ecology 84:2200–2207. MacKenzie, D.I., J.D. Nichols, N. Sutton, K. Kawanishi, and L.L. Bailey. 2009. Modeling multi-state site occupancy. Ecology 90:823–835. Makowski, D., M.S. Ben-Shachar, and D. Lüdecke. 2019. Indices of effect existence and significance in the Bayesian framework. Frontiers in Psychology 10:2767. Nelder, J.A. 1998. The selection of terms in response-surface models—how strong is the weak-heredity principle? The American Statistician 52:315–318. ODFW (Oregon Department of Fish and Wildlife). 2020. Oregon Conservation Strategy: painted turtle species account. Oregon Department of Fish and Wildlife, Salem, Oregon, USA. ODFW (Oregon Department of Fish and Wildlife). 2021. Sensitive species list. Oregon Department of Fish and Wildlife, Salem, Oregon, USA. 11 pp. ODOT (Oregon Department of Transportation). 2025. OR-Trans road centerlines (statewide). Salem, Oregon, USA. Available at: https://geohub.oregon.gov/datasets/4b4f87b1fdef49319884cafb1bad827d_0 (accessed 20 Sep 2025). ORBIC (Oregon Biodiversity Information Center). 2019a. Oregon wetlands database—LWI Wetlands layer, version 2019_10. Institute for Natural Resources, Portland State University, Portland, Oregon, USA. Available at: https://geohub.oregon.gov/datasets/oregon-geo::oregon-wetland-database (accessed 20 Sep 2025). ORBIC (Oregon Biodiversity Information Center). 2019b. Oregon wetlands database—More Oregon Wetlands (MOW) layer, version 2019_10. Institute for Natural Resources, Portland State University, Portland, Oregon, USA. Available at: https://geohub.oregon.gov/datasets/oregon-geo::oregon-wetland-database (accessed 20 Sep 2025). Oregon GEOHub. 2023. Oregon Spatial Data Library. Available at: https://geohub.oregon.gov (accessed 20 Sep 2025). Peixoto, J.L. 1990. A property of well-formulated polynomial regression models. The American Statistician 44:26–30. Piironen, J., and A. Vehtari. 2017. Sparsity information and regularization in the horseshoe and other shrinkage priors. Bayesian Analysis 12:969–1008. Polson, N. G., J. G. Scott, and J. Windle. 2013. Bayesian inference for logistic models using Pólya–Gamma latent variables. Journal of the American Statistical Association 108:1339–1349. PRISM Climate Group. 2023. PRISM climate normals/precipitation. Oregon State University, Corvallis, Oregon, USA. Available at: https://prism.oregonstate.edu (accessed 20 Sep 2025). R Core Team. 2025. R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. Ren, Q., C. He, Q. Huang, D. Zhang, P. Shi, and W. Lu. 2023. Impacts of global urban expansion on natural habitats undermine the 2050 vision for biodiversity. Resources, Conservation & Recycling 190:106834. Rogers, M., and G.S. Mauger. 2021. Pacific Northwest Climate Projection Tool. University of Washington, Climate Impacts Group, Seattle, Washington, USA. Royle, J.A., and J.D. Nichols. 2003. Estimating abundance from repeated presence–absence data or point counts. Ecology 84:777–790. Samara Group. 2020. Appendix N: western pond turtle visual encounter survey protocol. In: Visual encounter survey protocol for western pond turtles. Oregon Department of Fish and Wildlife Report, Salem, Oregon, USA. Schultz, J., D. Taylor-Rodríguez, A.J. Womack, de Rivera, 2026. The Heredity Horseshoe: Hierarchical Shrinkage for Functional-Form Discovery. Manuscript in Preparation. Shirk, A., H. Morgan, M. Krosby, C. Raymond, G.S. Mauger, and L. Helbrecht. 2021. Preparing Washington Department of Fish and Wildlife for a changing climate: assessing risks and opportunities for action. Washington Department of Fish and Wildlife and University of Washington Climate Impacts Group, Olympia and Seattle, Washington, USA. USFWS (U.S. Fish and Wildlife Service). 2023. Endangered and threatened wildlife and plants; threatened species status with section 4(d) rule for the northwestern pond turtle and the southwestern pond turtle. Federal Register 88:68006–68032. USFWS (U.S. Fish and Wildlife Service). 2025. National Wetlands Inventory geospatial data for Oregon. U.S. Department of the Interior, Washington, D.C., USA. Available at: https://www.fws.gov/program/national-wetlands-inventory (accessed 20 Sep 2025). USGS (U.S. Geological Survey). 2025a. National Land Cover Database (NLCD) 2021 products. Multi-Resolution Land Characteristics Consortium, Sioux Falls, South Dakota, USA. Available at: https://www.usgs.gov/data/national-land-cover-database-nlcd-2021-products (accessed 20 Sep 2025). USGS (U.S. Geological Survey). 2025b. National Hydrography Dataset (High Resolution)—Oregon. Reston, Virginia, USA. Available at: https://geohub.oregon.gov/maps/oregon-geo::national-hydrography-dataset (accessed 20 Sep 2025). Ward, J.H., Jr. 1963. Hierarchical grouping to optimize an objective function. Journal of the American Statistical Association 58:236–244. WDFW (Washington Department of Fish and Wildlife). 2024. State listed species. Washington Department of Fish and Wildlife, Olympia, Washington, USA. Zamora-Marín, J.M., C. Ilg, E. Demierre, N. Bonnet, A. Wezel, J. Robin, D. Vallod, J.F. Calvo, F.J. Oliva-Paterna, and B. Oertli. 2021. Contribution of artificial waterbodies to biodiversity: a glass half empty or half full? Science of the Total Environment 753:141987. Table 1 . Summary of predictors retained for analysis, showing domain, variable family, derived predictors used in the models, role (p = detection; Ψ = occupancy), units/scale, preprocessing (e.g., log1p, standardization, CLR + Ward for compositional bundles), tested interactions, and brief notes. Compositional variables were aggregated into interpretable “habitat bundles” via centered log-ratio transform and Ward clustering to limit collinearity; among correlated road-density buffers, the 250-m buffer was retained for parsimony. Abbreviations: CLR = centered log-ratio; OHW = ordinary high-water. Domain Variable family Derived predictor(s) Role (p/Ψ) Units/scale Preprocessing / transform Interactions tested Reason Local (survey) Air temperature Air Temp p °C Averaged start/end; standardized AirTemp × Cloudy Local (survey) Cloud cover Cloudy (Y/N) p 0/1 (binarized) Binarized from proportion; standardized AirTemp × Cloudy Local (site) Pond size Pond Area Ψ m² log1p; standardized In-pond Surface vegetation Surface cluster (open/emergent/sub+float) Ψ proportions CLR + Ward; standardized Avoid collinearity Edge Shoreline vegetation Shore cover cluster (forest/shrub/grass/open) Ψ proportions CLR + Ward; standardized Avoid collinearity Edge Basking near shore (dead wood) DeadBaskLogs Ψ, p proportion log1p; standardized Edge Gravel/Mud at OHW Gravel/Mud Shore Ψ proportion log1p; standardized Disturbance Disturbance rank DisturbRank (1–3) Ψ ordinal (levels 1–3) Treated as factor Recoded: removed rare ’0’ Landscape (connectivity) Distance to occupied wetland DistTurtle Ψ m log1p; standardized × log(DistToStream); × log(DistToForest) Landscape (hydro/thermal) Distance to stream (1–4 order) DistStream Ψ m log1p; standardized with DistToTurtle Landscape (cover context) Distance to forest (≥900 m²) DistForest Ψ m log1p; standardized with DistToTurtle Landscape (urbanization) Road density Road Density Ψ km/km² Standardized High corr across buffers → chose 250 m Landscape (cover) Land cover (NLCD, 250 m buffer) Land-cover cluster score Ψ proportions CLR + Ward (compositional bundle); standardized Compositional bundle of NLCD classes: Table 2. Sampling frame and survey effort Item Count / Value All sampled sites (N total) 253 Sites with no viable habitat (excluded from modeling) 25 Modeled sites (used in occupancy fits) 228 Total surveys (all visits across sites) 709 Visits per site, median [IQR] 3, 3 [3] Table 3. Summary per species of naïve occupancy (proportion of sites with observed turtles) and Proportion of sites occupied (PSO) for 277 sampled sites (confirmed presence n, naïve proportion, the median of the posterior distribution of PSO for each species, and 95% CrI, bounded by the constrained by the observed minimum). Species Confirmed Presence (n) Naïve Occupancy PSO Median PSO 95CrI Lower PSO 95CrI Upper C. picta bellii 47 0.186 0.264 0.186 0.716 A. marmorata 15 0.059 0.059 0.059 0.497 T. scripta elegans 57 0.225 0.229 0.225 0.817 Any Turtle Spp. 98 0.387 0.417 0.387 0.782 Figure 1. Study area in northwest Oregon comprising 12 HUC-8 watersheds (Nehalem, Lower Columbia–Clatskanie, Lower Columbia–Sandy, Lower Willamette, Tualatin, Yamhill, Molalla–Pudding, Clackamas, North Santiam, Middle Willamette, Upper Willamette, and Siletz–Yaquina). Gray polygons show HUC-8 boundaries; the red outline delineates the study boundary; gray points mark the 253 survey sites. Inset map locates the study area within the Pacific Northwest. Panel A Panel B Figure 2. Panel A: Detection regression coefficient posteriors colored by Probability of Direction (PD). Points represent posterior medians, bars indicate 95% credible intervals, and the dashed vertical line marks zero. Panel B: Occupancy regression coefficient posteriors colored by PD. Xs represent posterior medians, bars indicate 95% credible intervals, and the dashed vertical line marks zero. Species shown: Chrysemys picta bellii , Actinemys marmorata , and Trachemys scripta elegans . See Methods and Table S1 for model specifications; full coefficient summaries appear in Tables S2 and S3. Panel A Panel B Figure 3. Detection and occupancy model covariate effects for three freshwater turtle species. Panel A (Detection): Bars represent posterior median coefficients (logit scale), indicating the direction and strength of each variable’s influence on detection probability (positive = increased detection; negative = decreased detection). Text labels show posterior inclusion probabilities (P(Inc)); only predictors with P(Inc) ≥ 0.50 are shown. Actinemys marmorata had no covariates exceeding the P(Inc) ≥ 0.50 threshold and is therefore not shown. See Table S2 and Fig. 3 for full coefficient summaries. Panel B (Occupancy): Bars represent posterior median coefficients (logit scale), showing variables that increase (right) or decrease (left) occupancy probability. Text labels indicate P(Inc); only predictors with P(Inc) ≥ 0.50 are displayed. Full summaries are provided in Table S3 and Fig. 4. Information & Authors Information Version history V1 Version 1 21 January 2026 Copyright This work is licensed under a Non Exclusive No Reuse License. Keywords actinemys marmorata bayesian hierarchical models chrysemys picta bellii connectivity habitat associations urbanizing landscapes Authors Affiliations Laura Guderyahn 0000-0001-8828-1203 [email protected] Portland State University View all articles by this author Jacob Schultz Portland State University View all articles by this author Catherine de Rivera Portland State University View all articles by this author Marc Hayes Aquatic and Herpetological Cooperative, Eagle Point, OR View all articles by this author Daniel Taylor-Rodriguez Portland State University View all articles by this author Metrics & Citations Metrics Article Usage 123 views 88 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Laura Guderyahn, Jacob Schultz, Catherine de Rivera, et al. Revealing species-specific habitat drivers of turtle occupancy in Northwest Oregon using Bayesian low-detectability models.. Authorea . 21 January 2026. DOI: https://doi.org/10.22541/au.176900601.11448987/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.176900601.11448987/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9fe1eb9bfb73ad07',t:'MTc3OTE4MTMyMQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();