Skip to main content

Golden-winged Warbler post-fledging movement and stand-scale habitat selection

Cite this dataset

Fiss, Cameron et al. (2021). Golden-winged Warbler post-fledging movement and stand-scale habitat selection [Dataset]. Dryad.


Our understanding of songbird habitat needs during the breeding season stems largely from studies of nest success. However, growing evidence shows that nesting habitat and post-fledging habitat often differ. Management guidelines for declining species need to be revaluated and updated to account for habitat shifts that may occur across the full breeding cycle. The Golden-winged Warbler (Vermivora chrysoptera) is a declining songbird species for which best management practices (BMPs) are based overwhelmingly on nesting habitat. We studied stand-scale habitat selection by fledgling Golden-winged Warblers during May-July 2014-2017 in two landscapes (2 years of data for each landscape), 200 km apart in Pennsylvania. Across four years, we radio-tagged and tracked 156 fledglings. We used discrete-choice models to evaluate habitat selection during two post-fledging time periods (days 1-5, days 6-28). Fledglings used a variety of cover types, but most telemetry relocations (i.e. 85%) occurred in forest in the stand initiation stage, stem exclusion stage, or mature forest upland. Fledglings primarily selected stand initiation forest during the first five days, but preferred habitats differed between regions during days 6-28 post-fledging.

Fledglings in one landscape favored stands in the stem exclusion stage while fledglings in the other landscape continued to select stands in the initiation stage. Fledglings moved greater distances as they aged and dispersed approximately 750 m by day 28 post-fledging. These findings suggest the need to update Golden-winged Warbler BMPs to account for the broader habitat needs of fledglings during the breeding season. In addition, these results indicate that regional studies of habitat requirements can help guide management of dynamic forest landscapes for birds.


We searched for Golden-winged Warbler nests from May-June within early-successional forests and along edges of adjacent mature forest across both study areas. We used active searching techniques (e.g., parental behavior cues) to locate nests. For each nest discovered, we conducted checks on a three-day interval to monitor progress and to ensure accurate estimates of nestling age (Martin and Geupel 2016). As nestlings approached fledging (eight days old; Confer et al. 2011), we monitored nests daily.

Immature Golden-winged Warblers were usually marked as nestlings eight days after hatching. However, individuals that fledged prior to nest checks on day eight were caught by hand, typically within 10m of the nest. We randomly selected two members of each brood to be fitted with a VHF radio-transmitter (Blackburn Transmitters Inc., Nacogdoches, TX) with 95 mm antenna. Two fledglings were chosen because parents split broods shortly after fledgling (Peterson et al. 2016), and we wanted to increase the chance of monitoring separate sub-broods. Both birds received an aluminum USGS leg band and a radio transmitter affixed using the figure-eight harness method (Rappole and Tipton 1991). We constructed harnesses from <1mm black elastic thread to allow for growth (Streby et al. 2015). VHF radio transmitters used in this study weighed either 0.35 g or 0.40 g; and when combined with a harness and leg band, constituted <5% of each bird’s mass. There was no obvious indication that transmitters affected mobility or survival of fledglings, and radio-tagged individuals were often seen behaving in a similar fashion to brood-mates without radio transmitters. Handling time for each brood was ≤10 min and, upon completion of radio-tagging and banding, all birds were returned to their nest (nestlings) or perch (recently fledged young). In addition to fledglings from monitored nests, we opportunistically captured dependent fledglings that we encountered during nest searching and telemetry. We aged these birds to the nearest day by comparing their plumage characteristics to known-age fledglings.

Each radio-tagged fledgling was tracked daily between 06:00 and 16:00 hours using a Lotek STR 1000 (Lotek Wireless Inc., Newmarket, ON) receiver and Yagi three-element antenna. We tracked each fledgling once per day using the homing technique until we visually confirmed its location. Upon arriving at a fledgling’s location, we recorded the presence and behaviors of siblings and parents to determine fledgling independence. We recorded coordinates at the first location the fledgling was observed using a Garmin eTrex 20 GPS unit (Garmin Intl. Inc., Olathe, KS). We followed this tracking protocol until fledgling mortality or radio-transmitter battery failure (~30 days). When radio-signal was lost for an individual, we conducted systematic searches to determine if the fledgling had moved outside the normal detection range of our equipment. Searches were centered on the fledgling’s last known location and extended along 1-km transects in each cardinal direction. If a fledgling remained undetected, we conducted daily searches from automobile throughout the study area for ≥1 week before ceasing searches.

Movement and Space Use

We assessed fledgling movements and space use separately for each study area. Because Golden-winged Warblers are a brood-splitting species and multiple radio-tagged fledglings occasionally went with the same parent, we treated sub-broods as a random effect. To assess movement rate, we averaged daily straight-line movements across all sub-broods during two periods (low survival [~70% of mortalities]: day 1-5 post-fledging and high survival [~30% of mortalities]: day 6-28 post-fledging, McNeil 2019). We averaged Euclidean distance from each sub-brood location to its nest of origin to determine dispersal distance. During the high-survival period, we compared fledgling dispersal range between study areas using a Student’s T-test.

Cover Type Classification

We classified cover types in both study areas with ArcGIS 10.3 (Environmental Systems Research Institute 2015) using a combination of Pennsylvania State Forest and State Game Lands forest inventory data, ArcGIS online aerial imagery (Esri 2015), National Wetlands Inventory data, and records of recent (<10 years) timber harvests on public lands in PA. In addition, technicians visited >3800 randomly selected locations in our study areas and classified forest developmental stage. We used these ground-based samples to assist in classification of cover types. We classified most cover types based on tree size, stocking level (i.e. tree density relative to the stand’s capacity), and age class of the timber stand as described in the PA Department of Conservation and Natural Resources (DCNR) Bureau of Forestry Inventory Manual (PA DCNR 2016). We classified Stand Initiation (SI) cover as stands that had recently (approximately <10 years) undergone overstory removal harvest and were >50% stocked by trees <15 cm DBH. Stand Initiation cover closely represented Golden-winged Warbler nesting habitat and contained substantial shrub and herbaceous ground cover in addition to a diverse mixture of regenerating seedlings/saplings. We defined Stem Exclusion (SE) cover as older (approximately 10-25 years post-harvest) even-aged stands >50% stocked by trees <15 cm DBH. These stands were distinct from SI cover, due to the dominance of a dense sapling layer such that herbaceous vegetation and most shrubs were shaded-out by the overstory. Mature forest (i.e. stands in the understory reinitiation stage) was characterized by the dominance of trees >15 cm DBH. We divided mature forests into three sub-categories (Shelterwood/Understocked, Mature Forest Wetland, and Mature Forest Upland.  We classified Shelterwood/Understocked (SH) cover as mature forest <50% stocked. These stands were treated (e.g., shelterwood harvest), or had experienced non stand-replacing natural disturbance. We classified Mature Forest Upland (MU) as mature even- or uneven-aged stands that were >50% stocked. These stands were approximately 60-90 years old. We classified Mature Forest Wetland (MW; NE only) as mature palustrine stands >50% stocked. Mature Forest Wetlands were seasonally or perpetually inundated with water. We classified Shrub Wetland (SW; NE only) as stands dominated by shrubs and trees <15 cm DBH and, in many cases, perpetually inundated with water. We classified Upland Shrubland (US; NC only) as stands dominated by shrubs and <50% stocked with trees <15 cm DBH being dominant. Shrub cover in these stands was predominantly Vaccinium spp. or Gaylussacia spp. Upland Shrubland cover was largely derived from a forest fire which occurred in 1990.

Statistical Analyses

We used mixed-effects conditional logistic regression (i.e. discrete-choice) to model stand-scale habitat selection by fledgling Golden-winged Warblers and their parents (Thomas et al. 2006). As such, we created daily choice sets for fledglings beginning on the first day an individual was radio-tracked. Choice sets contained the fledgling’s observed location (used) and 19 available points. Similar ratios of used to available points have been used in local-scale habitat selection studies (Bonnot et al. 2011, Cheeseman et al. 2018). Available points were generated in ArcGIS using the “Create Random Points” tool. We restricted available points to a circle centered on a fledgling’s last used location, the radius of which was equal to the 75th percentile of all fledgling movements for a particular age, similar to Streby et al. (2016). As such, the range of available points expanded as fledglings developed and became more mobile. We measured Euclidean distance from all used and available points to each cover type to explain habitat selection (Conner et al. 2003). Specifically, use of a given alternative in the choice set acted as a binary response that varied as a function of the distance (continuous) to each cover type variable. Additionally, we included distance to edge to measure the influence of ecotones on habitat selection. Edge was calculated as the distance to the closest intersection between an early-successional stand (SI, SE, US, SW) and a mature stand (MU, MW, SH).

We fit habitat selection models within a Bayesian framework using JAGS (Plummer 2003) run from program R 3.5.1 (R Core Team 2018) with the jagsUI (Kellner 2015) package. Because individuals can respond differently to habitat, and because sub-broods occasionally had >1 radio-tagged fledgling, sub-broods were treated as random effects. We modeled each study area separately, and we modeled the post-fledging period in two parts for each study area (day 1-5 and day 6-28). Prior to model fitting we assessed collinearity using Pearson’s correlation coefficient with a cutoff of 0.6. One variable (distance to edge) was removed from the NC day 1-5 model due to collinearity. Because we were interested in evaluating habitat preferences for each cover type and edge, we constructed models for each study area and age class that included all variables, resulting in four models (O’Hara and Sillanpää 2009, Cheeseman et al. 2018; Appendix 1). We ran three concurrent Markov-chains for each model for 100 000 iterations of which 20 000 were allocated to a burn-in period. We assessed model convergence based on R values <1.1 (Gelman and Rubin 1996). We inferred selection for or against cover types based on regression coefficients with 95% credible intervals not overlapping zero (Kéry 2010).

Model Fit

Traditional goodness-of-fit (GOF) methods are not appropriate for discrete-choice models (Womack et al. 2013), so we adopted the k-fold cross validation approach to test the fit of our models (Boyce et al. 2002, Bonnot et al. 2009). Briefly, for each model, we randomly subset the data into a training set (80%) and a testing set (20%). We fit each model using the training set and then evaluated the rate at which the fit model accurately predicted used locations in the testing set versus 3 randomly selected available locations. We repeated this process 5 times for each model and report the average predictive-success as a measure of GOF. Given that we evaluated 4 choices, we would expect 25% predictive-success to be due to chance alone and predictive-success >25% suggesting adequate model fit (Bonnot et al. 2009).