Understanding an organism's shifting resource needs throughout its life cycle is crucial for developing effective conservation strategies. Northern bobwhite (Colinus virginianus) populations have been declining for nearly a century due largely to habitat loss/degradation. Many management regimes are based on analyses of resource selection conducted at coarse temporal scales, often within regions with productive soils. Although prior research provides a foundation for the development of management guidelines, patterns derived from finer-scale temporal analyses and varying ecological contexts may reveal novel information. We compared inferences derived from analyses conducted at varying temporal scales (full year, two-season, and daily [moving-window]; n = 1,302 bobwhites) in southeastern North Carolina, a region with poor edaphic conditions (i.e., nutrient-poor soils). We focused on four habitat attributes that can be affected by management: distance to fallow fields, distance to supplemental feed lines, distance to deciduous cover, and burn status. Habitats were strongly avoided for 18 weeks following a burn, and were not selected for until the following spring. Deciduous cover, primarily hardwood drains and Carolina bays, was selected year-round with the strongest selection from October - April (strongest distance-to-deciduous, 23 January β: -0.49 [95 % CI: -0.44 – -0.54]). Bobwhites selected for fallow fields strongest from March-October (strongest distance-to-fallow field, 10 August β: -0.47, [95 % CI: -0.39 – - 0.55]). Finally, bobwhites always selected for areas nearer feed lines with an increase outside brood-rearing (late September - March; distance-from-feedline 4 February β: -0.85, [95 % CI: -0.79 – -0.91]. Management for bobwhite populations in regions with poor edaphic conditions should consider alternative burn regimes to mitigate effects on bobwhites. Managers may consider fallow field enhancement for the dormant season and accommodate the increased selection for feed lines in winter. Fine-temporal analysis captured dynamic resource selection patterns in response to key habitat attributes specific to regions with poor edaphic conditions.

Study Area and Study Site. Our study occurred in the Middle Atlantic Coastal Plain Level III Ecoregion (Griffith et al., 2002; U.S. EPA 2013), on a private property in Brunswick County, North Carolina. Much of the property consisted of mature longleaf pine (Pinus palustris) and loblolly pine (P. taeda) savanna (48 % land cover) with rolling sandhills, flatwoods, and bottomlands interspersed throughout. Within the uplands, soils were well-drained and were of low productivity: Kureb fine sands with 1–8 % slopes (forest site index [ft]: 53 for longleaf pine, 64 for loblolly pine), Leon fine sands 0-2 % slopes (ft: 70 for longleaf pine, 75 for loblolly pine), Baymeade fine sands on 1–6 % slopes (ft: 63 for longleaf pine, 66 for loblolly pine), and Blanton fine sand of 0–5 % slopes (ft: 70 for longleaf pine, 85 for loblolly pine; USDA 2023; NCDACS 2024). The overstory consisted primarily of longleaf pine, loblolly pine, and southern live oak (Quercus virginiana). The understory was a mix of species such as sand live oak (Q. geminata), turkey oak (Q. laevis), yaupon holly (Ilex vomitoria), gallberry (I. glabra), and wax myrtle (Morella cerifera). Groundcover species included native bunch grass such as bluestems (e.g., Schizachyrium scoparium, Andropogon spp.), wiregrass (Aristida stricta), and a variety of dicots such as clusterspike false indigo (Amorpha herbacea), partridge-pea (Chamaecrista fasciculata), common bracken fern (Pteridium aquilinum), and greenbrier (Smilax spp.). Another important landcover type in our study area was deciduous cover (29 % land cover), some of which was dominated by trees (e.g., American sweetgum [Liquidambar styraciflua], black gum [Nyssa sylvatica], etc.) but more often was dominated by short-stature woody shrubs like swamp titi (Cyrilla racemiflora), sweetbay magnolia (Magnolia virginiana), and groundsel bush (Baccharis halimifolia). Finally, our study area was also punctuated by Carolina bays. Carolina bays are elliptical wetlands with a northwest orientation and hydrology driven chiefly by precipitation and evapotranspiration (Sharitz 2003). Although most Carolina bays in our study area were dominated by deciduous woody cover, several were characterized by open water and the largest one, in the center of our study area, was herbaceous in nature (Figure 1).

Uplands were managed using a 2–3-year fire-return interval to control understory hardwoods and to maintain open savanna conditions beneficial to fire-affiliated flora and fauna (Walker and Peet, 1984; Walker and Silleti, 2006). The entire upland portion of the property (> 80 %) was burned every 2-3 years (most every-other-year) with moderate-sized burn blocks (~5-45 ha) and burns mostly occurring in March. A primary focus of the land managers of our study area was to manage habitat for bobwhites and other game species, including white-tailed deer (Odocoileus virginianus) and mourning doves (Zenaida macroura). As such, intentional management practices included seasonal mowing, mesocarnivore (e.g. Virginia opossum [Didelphis virginiana] and common raccoon [Procyon lotor]) removal, and bi-monthly broadcast distribution of supplemental feed throughout the property at a rate of 2-4 bushels/acre/year consisting of milo (Sorgum bicolor) and wheat (Triticum aestivum) during the breeding season and integration of cracked corn (Zea mays) during the winter months. Fallow fields sown with common ragweed (Ambrosia artemisiifolia) and partridge pea were integrated into the landscape (3.7% land cover) with the goal of enhancing habitat quality for bobwhite broods (Palmer et al., 2012). Our study area was divided into seven discrete management units (767 - 1,280 hectares of area within each unit; Fig. 1in associated manuscript) based on ecological similarity and management history.

Capture and Monitoring. We captured and radio-tagged bobwhites in fall (October-November) and spring (March-April) from October 2021 - April 2024. We baited (with cracked corn, wheat, and milo) candidate trapping locations across the study area (n ~ 240 per season, > 70 m apart) in patches of woody vegetation (Stoddard, 1931). After two weeks of baiting, walk-in funnel traps were deployed, baited, and checked nightly. On a subset of captured birds (maximum of 5 birds per covey), we deployed very high frequency (VHF) pendant-style radio transmitters (weight: 6 g, American Wildlife Enterprises, Monticello, Florida, USA) aiming for a monitored sex ratio of 60 % females in spring and 50 % females in fall. We monitored tagged individuals 2-5 times a week using the homing method with a 3-element Yagi antenna and a Lotek BioTracker receiver (Lotek Wireless Inc., Ontario, CA) to approximately 20-25 m. Capture and handling of all animals complied with the approved protocol by the Institutional Animal Care and Use Committee of the University of Kentucky (IACUC protocol #2022-4105).

Land Cover Data. We assessed bobwhite resource selection with respect to four variables related to important management practices/habitat elements: distance to nearest fallow field (m), prescribed fire (burn status; binary covariate with 1 representative of areas that had been burned within the most recent burn cycle), distance to nearest supplemental feed line (m), and distance to nearest deciduous cover (m). Using QGIS, we manually digitized all fallow fields and deciduous cover across the study area using aerial imagery (QGIS development team 2022). Supplemental feed line shapefiles were obtained from the landowner of the study area. Through the course of our study, spatial arrangement of deciduous cover, fallow fields, and feed lines were static, however, we used a new burn map for each year (see Supplemental Material).

Resource Selection Analyses. We used mixed-effects logistic regression models to assess resource selection patterns (Hebblewhite and Merrill, 2008). We filtered our dataset to remove birds with fewer than 5 locations since time of capture (Kenward et al. 2001, White and Garrot 1990) to avoid violating the assumption that collars did not affect birds’ movement patterns (Guthery and Lusk, 2004). ‘Used’ points were the known locations of radio-tagged bobwhite, which were assigned a value of “1”. Paired, random-generated locations were assigned a value of “0”. Random locations were generated at a ratio of 2:1 (2 random:1 used), with the randoms generated within the management unit within which the bird resided (Northrup et al. 2013). Individuals were assigned to a management unit by taking the centroid of all its locations. Prior to analyses, we confirmed that none of the four resource covariates (burn status, distance to nearest deciduous cover, distance to nearest fallow field, and distance to nearest supplemental feed line) were highly correlated (> 0.7; Sokal and Rohlf, 1969). All random points were generated using the st_sample function from the sf package in the program R (Pebesma, 2018; R Core Team, 2022).

With the binary used/random data, we constructed 368 models, each with the same structure (developed a priori) but using different subsets of the data to explicitly test hypotheses related to resource selection of northern bobwhites at different time scales. All models in all analyses were ‘global’ models including all four covariates, additively. We fit all models using the glmer function within the lme4 package in R (Bates et al., 2015; R Core Team, 2022). All models included a random intercept for individual bird identity (u_i) and a fixed effect for management unit identity (v_j). In sum, every model, fit with different subsets of data, had the following structure:

logit(P(Y_ijk = 1)) = β₀ + β₁ * Burn_ijk + β₂ * Deciduous_jk + β₃ * FeedLine_ijk + β₃ * BroodField_ijk + u_i + v_j

We created models with data from three different temporal scales: 1) Full year (“Full year” model = all data lumped together), 2) Two-season (“breeding” model = 1 April – 30 September, “non-breeding” model = October 1 - March 31), and 3) Daily (“Daily moving window”; one model representative for each day, total = 365 models). The models for the daily temporal analysis were fit using a daily moving window where each day’s data consisted of data from that day and also those from ± 14 days (totaling 29 days’ worth of data). We selected this window size because it struck a balance between a reasonable sample size of observations from which to fit models while still allowing high temporal resolution across the lifecycle. We examined β coefficients 95 % confidence intervals to assess whether parameters were informative and considered those overlapping with zero to be uninformative (Nakagawa and Cuthill, 2007; Stoltzfus 2011).