Pacific Barrow's Goldeneye refine migratory phenology in response to overwintering temperatures and annual snowmelt

ABSTRACT Timing of seasonal bird migrations is broadly determined by internal biological clocks, which are synchronized by photoperiod, but individuals often refine their migratory timing decisions in response to external factors. Using 11 years of satellite telemetry data, we show that Pacific Barrow's Goldeneye (Bucephala islandica) at higher latitudes initiated spring and molt migrations later and fall migration earlier than individuals at lower latitudes. We further show that individuals refined migratory timing in response to interannual variation in environmental conditions. Individual Barrow's Goldeneye initiated spring migration earlier in years with warmer springs at their overwintering locations and concluded spring migration earlier in years with earlier annual snowmelt on their breeding grounds. Because individuals respond to conditions both where they initiate and where they conclude spring migration, our results suggest that Barrow's Goldeneye update their migratory decisions en route. For all 3 migrations in their annual cycle, birds delayed initiating migration if they had been captured and tagged prior to that migration. Birds that initiated migration late for their latitude were less likely to include a stopover and completed that migration faster, partially compensating for delayed departures. Our results are consistent with the hypothesis that Barrow's Goldeneye uses a combination of endogenous cues and environmental cues in migratory decision making. Sensitivity to environmental cues suggests that Barrow's Goldeneye may have behavioral plasticity that is adaptive when faced with ongoing climate change. LAY SUMMARY Migratory birds time their annual migrations to take advantage of temporal and spatial variation in resources, disease, and predation. Timing of these migrations is broadly determined by an internal biological clock set by day length, but individuals may refine when and how they migrate in response to environmental conditions. We used 11 years of satellite tracking data to investigate migratory cues in Barrow's Goldeneye, a species of sea duck. We found that birds used weather conditions as migratory cues, leaving their wintering grounds earlier in warmer springs and arriving on their breeding grounds earlier when the snow melted earlier in the year. Because individuals respond to conditions both where they start and where they end spring migration, our results suggest that Barrow's Goldeneye update their migratory decisions during migration. Barrow's Goldeneye's ability to respond to environmental cues suggests that they may be able to adapt to ongoing climate change. RÉSUMÉ Le calendrier des migrations saisonnières des oiseaux est largement déterminé par les horloges biologiques internes, qui sont synchronisées par la photopériode, mais les individus affinent souvent leurs décisions en matière de calendrier migratoire en réponse à des facteurs externes. En utilisant 11 années de données de télémétrie par satellite, nous montrons que les individus de Bucephala islandica de la population de l'Ouest ont commencé les migrations printanières et de mue plus tard et la migration automnale plus tôt aux latitudes plus élevées que les individus se trouvant aux latitudes plus basses. Nous montrons également que les individus ont affiné leur calendrier migratoire en réponse aux variations interannuelles des conditions environnementales. Les individus ont commencé leur migration printanière plus tôt les années où les printemps étaient plus chauds dans leurs lieux d'hivernage et ont terminé leur migration printanière plus tôt les années où la fonte annuelle des neiges était plus précoce sur leurs sites de reproduction. Puisque les individus réagissent aux conditions à la fois là où ils commencent et là où ils terminent leur migration printanière, nos résultats suggèrent que l'espèce adapte ses décisions migratoires en cours de route. Pour les trois migrations de leur cycle annuel, les oiseaux ont retardé le début de la migration lorsqu'ils ont été capturés et marqués avant cette migration. Les oiseaux ayant commencé leur migration tardivement pour leur latitude étaient moins susceptibles d'inclure une escale et terminaient cette migration plus rapidement, compensant partiellement les départs retardés. Nos résultats sont cohérents avec l'hypothèse selon laquelle B. islandica utilise une combinaison d'indices endogènes et environnementaux dans la prise de décision migratoire. La sensibilité aux indices environnementaux suggère que cette espèce peut avoir une plasticité comportementale qui s'adapte aux changements climatiques en cours.


INTRODUCTION
Timing of annual migrations, during which birds may move thousands of kilometers, result from a complex interaction between endogenous biological clocks and exogenous environmental conditions. Endogenous biological clocks generate circannual rhythms, which are synchronized primarily by photoperiod (Gwinner 1996(Gwinner , Åkesson et al. 2017, and are responsible for the broad seasonal timing of migrations (Berthold 1984, Gwinner 1996. Intraspecific variation in migratory timing often varies with latitude due to its relationship with photoperiod (Gow et al. 2018, Smith et al. 2020, and these latitudinal effects can carry over to affect the timing of subsequent events in the annual cycle (Gow et al. 2018, Forstner et al. 2022. Individuals also have flexibility to refine their timing decisions in response to external environmental factors, such as seasonal variation in weather conditions (Haest et al. 2020) or food availability (van der Graaf et al. 2006). For example, temperature can affect timing of spring migration (Haest et al. 2020) as well as departure from intermediate staging sites (Shariati-Najafabadi et al. 2016). Wind conditions can also affect migratory departures as well as flight duration and arrival timing (Gill et al. 2009, Drake et al. 2014, Haest et al. 2020. Additionally, seasonal variation in weather-mediated food conditions along the migration route can drive the timing of migration. For example, on northward spring migrations, some herbivorous geese are known to follow a green wave of progressively available forage along their spring migration route (van der Graaf et al. 2006, van Wijk et al. 2012. Similarly, on northward migrations, some Arcticbreeding raptors follow the latitudinal gradient of progressive snowmelt to maximize their ability to locate prey and as a cue to optimize their arrival timing on the breeding grounds (Curk et al. 2020). Further, northward progressions of both Pacific herring spawn events and migratory phenology of sea ducks that consume herring eggs are spatiotemporally correlated (Lok et al. 2012).
Migration and the corresponding conditions faced during migration can have effects on survival and reproduction. Migratory periods have been associated with significantly higher mortality than non-migratory periods, with more than half of annual mortality occurring during compara-tively short migratory periods for some species (Sillet andHolmes 2002, Klaassen et al. 2014). Adverse weather conditions may be responsible for some migratory mortality: for example, Black-tailed Godwits (Limosa limosa) experience higher mortality when migrating against prevailing winds than when migrating with favorable winds (Loonstra et al. 2019) and American Woodcocks (Scolopax minor) suffer increased mortality when faced with low clouds and strong winds on migration, leading to increased building collisions (Loss et al. 2020). Unfavorable conditions experienced during migration may further have non-lethal carry-over effects, such as reducing body condition or fat reserves available for the subsequent annual stage (Harrison et al. 2011). Given that timely arrival on the breeding grounds increases breeding success through mechanisms such as the acquisition of higher quality territories (Currie et al. 2000) and allowing time for replacement clutches (Morrison et al. 2019), weather-induced delays in arrival and nesting can have negative reproductive consequences. For example, unfavorable wind conditions experienced during spring migration delay breeding and reduce annual productivity of female Yellow Warblers (Setophaga petechia) (Drake et al. 2014). Similarly, rainfall during spring migration is correlated with reduced body mass in Barn Swallows (Hirundo rustica), which, in turn, could impact the subsequent breeding success (Robson and Barriocanal 2008). For some species, weather conditions during spring migration can have a larger effect on breeding phenology than weather conditions on the breeding grounds themselves (Finch et al. 2014).
Climate change is further complicating migratory strategies due to the growing disconnect between annually consistent photoperiods and shifting environmental cues. Additionally, there is geospatial variations in the rate of environmental changes occurring across species' migratory ranges. In response to warming temperatures, there exists a general trend in advancement of spring migration timing (Gienapp et al. 2007, Usui et al. 2017), but there is substantial variation within and between species (Usui et al. 2017). Advances in spring migration across decades vary across North American flyways but are more pronounced at higher latitudes (Horton et al. 2020). However, some migrants have not compensated for warming temperatures by advancing the timing of their migrations, which can lead to phenological mismatch between arrival on the breeding grounds or reproduction and optimal food availability, causing subsequent reductions in fecundity and population declines (Saino et al. 2010). Some populations may be especially vulnerable: long-distance migrants breeding in areas with high seasonality have shown pronounced population declines associated with phenological mismatch (Both et al. 2010). Even when species are able to advance their migratory timing, physiological constraints may impose further limits. For example, Barnacle Geese (Branta leucopsis) that advance their arrival on the breeding grounds are still unable to sufficiently advance egg laying in response to earlier snowmelt, resulting in a mismatch with the peak of food availability and therefore reduced reproductive success (Lameris et al. 2018). For Arctic-breeding shorebirds, earlier snowmelt can alter the timing of peak prey abundance relative to hatch, which can lead to inadequate food availability to support chick growth, although prey availability is highly variable (Kwon et al. 2019, Saalfeld et al. 2019. Where earlier hatching successfully responds to the general warming trend, mass mortality events can still occur if nestlings are exposed to cold snaps that directly reduce insect prey availability (Shipley et al. 2020). However, population declines are not consistently observed with advances in spring migration and laying date (Dunn and Møller 2014), and groups such as Arctic-breeding geese appear to benefit from earlier and warmer springs (Morrissette et al. 2010, Jensen et al. 2014. Sea ducks perform 3 annual migrations: spring migration from wintering to breeding grounds, molt migration from breeding to molting grounds, and fall migration from molting to wintering grounds (Salomonsen 1968). During migration, sea ducks can cover long distances and encounter environmental conditions ranging from temperate wintering grounds to sub-Arctic breeding grounds to Arctic molting grounds. Barrow's Goldeneye (Bucephala islandica) is a medium-sized sea duck whose global population is heavily concentrated west of the Rocky Mountains in North America (Eadie et al. 2020). Previous research has found that Pacific populations wintering in coastal Alaska and British Columbia have high migratory connectivity and that where adults overwinter is related to the timing of their spring, molt, and fall migrations (Forstner et al. 2022), consistent with latitudinally specific endogenous responses to circannual changes in photoperiod. Here we analyze 11 years of telemetry data for Barrow's Goldeneye to assess how interannual variation in environmental cues explains variation in timing of migratory initiation, stopover, duration, and conclusion, after accounting for latitudinal effects. We predicted that Barrow's Goldeneye, a short-distance migrant, would alter the timing of spring migration in response to temperature on their coastal wintering grounds, initiating migration earlier in warmer years. Because Barrow's Goldeneye breed and molt on high latitude inland lakes that are subject to annual freeze, we also expected that adults would adjust their migration behavior in response to conditions encountered en route, slowing their migration and delaying its conclusion in years when spring melt is late. Consistent with the frost wave hypothesis for fall migration (Xu and Si 2019), we further predicted that timing of freeze on molting grounds would influence timing of departure on fall migration.

Study Species
Pacific Barrow's Goldeneye winter along the west coast of North America from Washington to southern Alaska and breed inland from the Washington Cascades to central Alaska ( Figure 1). Females nest in tree cavities near freshwater lakes (Eadie et al. 2020). Barrow's Goldeneye, like many waterfowl, undertake an annual migration to molting grounds, where they undergo a simultaneous remigial molt and are temporarily flightless. Males depart for molting grounds while females are incubating. Unsuccessful females subsequently migrate to molting grounds, whereas successful females may molt on their breeding grounds or a discrete molting site nearby. For Pacific Barrow's Goldeneye, molting occurs at inland lakes from northern Alaska to the Northwest Territories and northern Alberta (Hogan et al. 2011, Forstner et al. 2022; Figure 1). Remigial molt leaves individuals flightless for ~30 days (Hogan et al. 2013a), but survival is high (Hogan et al. 2013b). Pairs are suspected to then reunite on their wintering grounds (Savard 1985). Both sexes display fidelity to wintering, breeding, and molting locations as adults (Savard 1985, Willie et al. 2020, although natal philopatry is higher in females than in males ). When overwintering, their diet consists largely of mollusks while during both breeding and molting stages, Barrow's Goldeneye consume primarily insects (Eadie et al. 2020 0°N, 122.5°W)), and one molting location (Cardinal Lake, Alberta (56.22°N, 117.78°W)) ( Figure 1). A total of 271 adult birds were surgically implanted with Argos platform terminal transmitters (hereafter "tags"), which transmitted locations for an average of 346 days (range: 18-1,171 days). Data quality and management are described in Forstner et al. (2022). Here, we use data for the adult Goldeneye for whom we were able to estimate the location and timing of initiation and conclusion of at least one migration, limiting the time interval between transmissions to < 30 days (spring migration: 67 females, 75 males, 208 tracks; molt migration: 55 females, 96 males, 205 tracks; fall migration: 53 females, 104 males, 208 tracks).
Stages and transitions within the annual cycle are described in Forstner et al. (2022), with the exception of initiation of fall migration (see below). Briefly, wintering, breeding, and molting locations for each year were determined as mean-center centroids, geographic averages of individuals' locations during each wintering, breeding, or molting stage (Willie et al 2020, Forstner et al. 2022. Initiation of spring migration was defined as the day an individual left the coast and moved > 100 km inland. Conclusion of spring migration was defined as the day an individual arrived and remained at an interior wetland in late spring or early summer. Initiation of molt migration was defined as the day an individual left an interior wetland and traveled > 20 km without returning. Conclusion of molt migration was defined as the day an individual arrived at a location from which they made no overland movements > 1 km for at least 30 days. Initiation of fall migration was defined as the first day an individual made a movement > 20 km. Finally, conclusion of fall migration was defined as the day an individual switched from making directional daily movements of > 100 km to non-directional daily movements of < 100 km. Different distance thresholds were used for the migrations because birds face unique movement constraints at each stage in their annual cycle. Following De La Cruz et al. (2009), if there was more than a day between transmissions before and after an initiation or conclusion, the initiation or conclusion date was estimated as the midpoint date between locations. The interval between transmissions ranged from 0.5 to 29.2 days (median: 4.08 days, IQR: 1.52 days). Distances of all migrations were calculated as great circle distances between the centroid from which the migration was initiated and the centroid at which the migration was concluded.

Data Acquisition
Minimum and maximum daily temperatures for wintering locations were obtained from Daymet version 4, a 1-km × 1-km resolution grid of weather conditions for North America (Thornton et al. 2020). The daymetr R package was used to obtain daily minimum and maximum temperatures for each bird's winter centroid. Daymet provides only terrestrial temperatures, so where a centroid occurred offshore, the coordinates for the nearest point of land were used in place of the centroid's coordinates. Temperatures were averaged across January and February to determine winter average daily minimum and maximum temperatures and across March and April to determine spring average daily minimum and maximum temperatures at each individual's wintering location.
We obtained snow and ice cover data for breeding and molting locations from the Interactive Multisensor Snow and Ice Mapping System (IMS); 4-km × 4-km grid cells are considered snow-or ice-covered when snow or ice covers >40% of the cell (U.S. National Ice Center 2008). We averaged daily snow and ice cover for all cells within a 5 km radius of breeding and molting centroid locations to determine daily snow cover for each centroid; from this, we determined the day of year on which each breeding and molting location melted and froze for a minimum of 7 days. We averaged snow and ice cover across cells to avoid potential bias due to topography. Melt and freeze were not assessed at wintering locations as these coastal locations do not freeze annually.

Statistical Analyses
We developed candidate linear mixed effects models using the lme4 R package to examine variation in day of migration initiation (i.e., departure day), inclusion of a stopover (0/1), migration duration, and day of migration conclusion (i.e., arrival day) for the 3 annual migrations undertaken by Barrow's Goldeneye. Dates were converted to ordinal days (January 1 = 1). A stopover was defined as a period of 3 or more days spent within a 20-km area that was not otherwise defined as a wintering, breeding, or molting location. Fixed variables were standardized for analyses by subtracting the mean and dividing by the standard deviation. We conducted our analyses in 2 stages using a build-up approach (Morin et al. 2020) to reduce the number of variables in any given model and used corrected Akaike Information Criterion (AIC c ) to identify top-performing models at each stage. First, most parsimonious base models without weather variables were identified. For base models for all 3 migrations, the initial candidate set assessed the roles of sex, latitude, migration distance, and the effect of having been tagged prior to that migration on 4 response variables: day of migration initiation, inclusion of a stopover, migration duration, and day of migration conclusion. For the stopover, duration, and conclusion analyses, we included relative departure day as a predictor, which is the residual of departure day (day of migration initiation) with respect to the linear regression of departure day on departure latitude. Relative departure day accounts for an individual departing early or late relative to others, independent of the effect of latitude on timing. Relative departure day from breeding grounds was sex-specific. Bird identity was included as a random effect in all candidate models to account for multiple tracks from some individuals. For weather-based predictions, the final candidate set assessed if the addition of a weather variable to the top performing base model improved the model. For spring migration, we assessed the importance of minimum and maximum winter and spring temperatures, as well as the day of melt on the breeding grounds, on day of migration initiation, inclusion of a stopover, migration duration, and day of migration conclusion. For molt migration, we assessed the importance of day of melt on the molting grounds. For fall migration, we assessed the importance of day of freeze on the molting grounds. When the top base model indicated that day of migration initiation, inclusion of a stopover, migration duration, or day of migration conclusion varied with latitude, the final candidate set examined whether weather conditions explained latitudinal effects, by including a model that replaced the latitude term with the appropriate weather variable, or had an additive effect, by including a model with both the latitude term and the weather variable. Weather variables on the wintering, breeding, and molting grounds were correlated with latitude (r p : -0.81 to 0.30, all P < 0.0001; Supplementary Material Table 1). We assessed whether multicollinearity was an issue using the variance inflation factor (VIF) for each explanatory variable. VIFs associated with all terms were consistently < 3 (Supplementary  Material Table 2), indicating that weather variables and latitudes can be included in the same model (Dormann et al. 2013). Model performance was assessed by AIC ranking and conditional and marginal R 2 values (Nakagawa and Schielzeth 2013). Analyses were performed with R version 4.0.2 (R Core Team 2016). All figures display models' partials residuals and fitted relationships.

Spring Migration
Across all latitudes, the median date of departure from the wintering grounds was April 28 (90% range: April 1 to May 21). Barrow's Goldeneye initiated spring migration later when wintering at higher latitudes and when they had to travel farther to reach their breeding grounds (Supplementary Material Figure 1A-B). Initiation was delayed by an average of 9 days for birds that had been tagged on their wintering grounds prior to initiation of spring migration (Supplementary Material Figure 1C). Environmental factors influenced the timing of departure and were independent of the effect of winter latitude: Barrow's Goldeneye departed earlier when spring temperatures on the wintering grounds were higher ( Figure 2). The top supported model in the final candidate set contained winter latitude (0.30 ± 0.08; 95% CI: 0.13-0.46), migration distance (0.24 ± 0.05; 95% CI: 0.13-0.33), tag (0.62 ± 0.10, 95% CI: 0.44-0.82), and spring average daily maximum temperature (-0.26 ± 0.08, 95% CI: -0.41 to -0.11) ( Table 1).
Birds were more likely to include a stopover on spring migration when they had left their wintering grounds early relative to the expected departure date for their latitude and when they had to travel farther to reach their breeding grounds (Supplementary Material Figure 2A-B). The top model in the final candidate set contained relative departure day (-0.61 ± 0.23; 95% CI: -1.16 to -0.19), distance (1.07 ± 0.29; 95% CI: 0.58-1.82), and snowmelt on the breeding grounds (0.36 ± 0.21; 95% CI: -0.07 to 0.79) ( Table 1). Although included in the top model, the timing of snowmelt on the breeding grounds had little effect on whether birds used a stopover site during spring migration (Supplementary Material Figure 2C).
Spring migration took 1-67 days. Spring migration was shorter when birds left their wintering grounds late relative to the expected departure date for their latitude ( Figure 3A). Spring migration was longer when birds used a stopover site and when they traveled farther to reach their breeding grounds (Supplementary Material Figure 3B-C). Environmental conditions on the breeding grounds did not affect duration of spring migration as the base model was not improved by including the snowmelt term (Table 1). The top model in the initial and final candidate sets contained stopover (1.17 ± 0.11; 95% CI: 0.95-1.40), distance (0.29 ± 0.06; 95% CI: 0.18-0.40), and relative departure day (-0.30 ± 0.06; 95% CI: -0.40 to -0.19).
Birds were more likely to use a stopover site during molt migration if they had left their breeding grounds early relative to the expected departure day for their sex and latitude and when they had to travel farther to reach their molting grounds (Supplementary Material Figure 6A-B). The timing of snowmelt on the molting grounds did not influence whether birds used a stopover site during their molt migration. The top model in the initial and final candidate sets contained relative departure day (-0.43 ± 0.21; 95% CI: -0.91 to -0.03) and distance (1.04 ± 0.29; 95% CI: 0.56-1.80) ( Table 2).
Molt migration took 1-93 days. Molt migration was shorter if birds had left their breeding grounds late relative to the expected departure day for their sex and latitude ( Figure  3B). Molt migration was longer if birds used a stopover site and when birds had to travel farther to reach their molting grounds (Supplementary Material Figure 7B-C). Duration of molt migration was not influenced by timing of snowmelt on the molting grounds. The top model contained stopover (1.09 ± 0.10; 95% CI: 0.89-1.30), distance (0.13 ± 0.05; TABLE 1. Model selection table for candidate base and final models explaining variation in spring migration parameters. Included are number of estimated parameters (K), corrected Akaike's information criterion (AIC c ), ΔAIC c , model probabilities (Akaike weight, w i ), log-likelihood (LL), and Nakagawa's conditional (Con R 2 ) and marginal R 2 (Mar R 2 ). All models include a random effect of bird identity. Candidate models with ΔAIC c < 2 and null models are shown. For final models, base models are shown regardless of ΔAIC c value.

Fall Migration
The median departure date from the molting grounds was October 8 (90% range: September 6-November 4) while the median day of freeze on the molting grounds was November 3. Barrow's Goldeneye initiated fall migration earlier when molting at higher latitudes and when they had farther to travel to reach their wintering grounds (Supplementary Material Figure 9A-B). Departure was delayed by an average of 5 days for birds that had been tagged on their molting grounds prior to initiation of fall migration (Supplementary Material Figure 9B). The top model in the final candidate set contained molting latitude (-0.54 ± 0.10; 95% CI: -0.74 to -0.34), distance (0.13 ± 0.08; 95% CI: -0.02 to 0.28), tag (0.31 ± 0.15; 95% CI: 0.02-0.59) and day of freeze on the molting grounds (0.12 ± 0.08; 95% CI: -0.03 to 0.28) ( Table 3). Although included in the top model, the timing of freeze had little effect on when Barrow's Goldeneye initiated  fall migration. Individuals initiated fall migration well before the day of freeze when molting at locations that froze relatively late in the year compared to other molting locations (Supplementary Material Figure 9D).
Birds were more likely to use a stopover site on fall migration if they had left their molting grounds early relative to the expected departure day for their latitude and when they had to travel farther to reach their wintering grounds (Supplementary Material Figure S10A-B). Birds were also more likely to stop if they had been tagged on their molting grounds prior to initiating fall migration (Supplementary Material Figure 10C). Timing of freeze on the molting grounds did not influence whether birds used a stopover site during fall migration. The top model in the initial and final candidate sets contained relative departure day (-2.25 ± 0.49; 95% CI: -3.90 to -1.48), distance (1.01 ± 0.36; 95% CI: 0.45-2.31), and tag (1.81 ± 0.84; 95% CI: 0.41-4.77) ( Table 3).
Fall migration took 1-90 days. Fall migration was shorter if birds had left their molting grounds late relative to the expected departure day for their latitude ( Figure 3C). Fall migration was longer if birds used a stopover site and when birds had to travel farther to reach their wintering grounds (Supplementary Material Figure 11B-C). The top model in the final candidate set contained stopover (0.78 ± 0.10; 95% CI: 0.58-0.99), distance (0.31 ± 0.05; 95% CI: 0.21-0.41), relative departure day (-0.53 ± 0.05; 95% CI: -0.63 to -0.43), and day of freeze on the molting grounds (0.09 ± 0.05; 95% CI: -0.01 to 0.18) ( Table 3). Although included in the top model, the timing of freeze at molting locations had little effect on the duration of fall migration (Supplementary Material Figure 11D).
The median arrival date on the wintering grounds was October 31 (90% range: October 2 to November 17). Barrow's Goldeneye concluded fall migration later if they wintered at lower latitudes, traveled shorter distances, and departed their molting grounds late relative to the expected departure day for their latitude (Supplementary Material Figure 12A-C). The timing of freeze on the molting grounds did not influence the timing of arrival on the wintering grounds. The top model in the initial and final candidate sets contained winter latitude (-0.75 ± 0.05; 95% CI: -0.84 to -0.65), distance (-0.18 ± 0.05; 95% CI: -0.28 to -0.09), and relative departure day (0.19 ± 0.05; 95% CI: 0.10-0.28) ( Table 3).

DISCUSSION
Our findings support the hypothesis that avian migration timing is governed by both endogenous biological clocks and environmental conditions (Åkesson et al. 2017). Our study is comprehensive in that it examines 3 annual migrations over the full latitudinal range of a species' distribution. Individual Barrow's Goldeneye migrate at a time that is seasonally appropriate for their latitude, but they further refine their migratory decisions in response to environmental conditions. Barrow's Goldeneye update their migratory decision-making en route, by responding to variation in spring temperatures when they initiate spring migration and to snowmelt when they conclude spring migration. Further, we found that Barrow's Goldeneye traveled faster if they initiated spring, molt, and fall migrations later than expected for a given latitude, partially compensating for delayed departures.
Migratory birds are expected to respond to warming temperatures by initiating spring migration earlier when the advancement in conditions on the breeding grounds is predictable from conditions on the wintering grounds and when there is a fitness benefit to arriving early to breed (Usui et al. 2017, TABLE 2. Model selection table for candidate base and final models explaining variation in molt migration parameters. Included are number of estimated parameters (K), corrected Akaike's information criterion (AIC c ), ΔAIC c , model probabilities (Akaike weight, w i ), log-likelihood (LL), and Nakagawa's conditional (Con R 2 ) and marginal R 2 (Mar R 2 ). All models include a random effect of bird identity. Candidate models with ΔAIC c < 2 and null models are shown. For final models, base models are shown regardless of ΔAIC c value.

Parameter
Model K AIC c ΔAIC c w i LL Con R 2 Mar R 2  Bauer et al. 2020). Across avian taxa, migratory birds are indeed advancing spring migration in response to warming temperatures associated with persistent climate change (Usui et al. 2017). Inter-and intraspecific comparisons suggest that shortdistance migrants are more responsive to temperature cues than long-distance migrants, who rely more heavily on endogenous rhythms (Gienapp et al. 2007, Usui et al. 2017. Within species, for example, American Kestrels (Falco sparverius) migrating short distances arrive on their breeding grounds earlier in warmer springs, whereas long-distance migrants do not adjust their arrival timing with temperatures on the breeding grounds (Powers et al. 2021). However, some long-distance migrants do respond to temperature: for example, temperatures at disparate wintering and stopover sites partly explain timing of arrival on the breeding grounds for long-distance Common Redstarts (Phoenicurus phoenicurus) and European Pied Flycatchers (Ficedula hypoleuca) (Haest et al. 2020). Similarly, we found that Pacific Barrow's Goldeneye departed on spring migration earlier in years with warmer spring temperatures where they overwinter. Their spring migrations averaged 370 km (Forstner et al. 2022) and their responsiveness to temperature is consistent with many short distance migrants. Novel conditions encountered en route can prolong migrations and delay arrivals through additional or lengthened stopovers (Li et al. 2020, Oliver et al. 2020. Further, when environmental conditions at the subsequent location in the annual cycle are largely unpredictable from the current location, modeling suggests that it benefits migrants to use stopover sites not only to refuel but also to acquire information about forthcoming environmental conditions (Bauer et al. 2020 Oliver et al. 2020). Tundra Bean Geese (Anser serrirostris) spend 25 more days at stopover sites when migrating along a flyway with more sustained snow cover compared to a flyway with earlier annual melt (Li et al. 2020). American Robins (Turdus migratorius) migrating long distances are more likely to include a migratory stopover when they encounter deep snow en route (Oliver et al. 2020). In our study, presence of snow on the breeding grounds delayed birds' arrival on the breeding grounds, in part due to a weak effect of snow increasing the likelihood of a stopover en route. Warmer spring average daily maximum temperatures on the wintering grounds, a cue used for migration initiation, were significantly correlated with earlier annual melt on the breeding grounds (r = -0.23, P < 0.001), but wintering temperatures were a poor predictor of the conditions on the breeding grounds. Presumably, individual Barrow's Goldeneye slow their migration or use a stopover site when necessary due to adverse conditions, thereby refining their arrival timing after encountering updated information about local environmental conditions. Arrival timing relative to melt can have important fitness consequences: for example, for Common Eiders (Somateria mollissima), another species of sea duck, reproductive success is highest not for pairs who lay earliest in the season, but when hatch and ice melt closely coincide temporally (Love et al. 2010). Here, individual Barrow's Goldeneye arrive on their breeding grounds before their breeding grounds have melted. This early arrival likely provides first access to high-quality nest cavities, which may provide fitness benefits as earlier clutches have higher hatching success (Savard 1988). Additionally, early arrival on the breeding grounds may allow birds to provide offspring with peak availability of seasonal foods.

Molt migration initiation
Barrow's Goldeneye, like many waterfowl, migrate to distinct molting grounds after breeding (Salomonsen 1968). 3. Model selection table for candidate base and final models explaining variation in fall migration parameters. Included are number of estimated parameters (K), corrected Akaike's information criterion (AIC c ), ΔAIC c , model probabilities (Akaike weight, w i ), log-likelihood (LL), and Nakagawa's conditional (Con R 2 ) and marginal R 2 (Mar R 2 ). All models include a random effect of bird identity. Candidate models with ΔAIC c < 2 and null models are shown. For final models, base models are shown regardless of ΔAIC c value.

Parameter
Model Molting sites provide a safe location for a sustained period of flightlessness, offering a reduction in disturbance by humans and predators (Fox et al. 2014) and access to nutrient-rich food resources after breeding locations may have been depleted (Salomonsen 1968). Male Pacific Barrow's Goldeneye travel hundreds of kilometers north or northeast of their breeding grounds to molt (Forstner et al. 2022). Because of the high seasonality of these northerly molting grounds, we hypothesized that the timing of melt and freeze might be important factors influencing migratory timing. However, because molting grounds melted, on average, 46 days before males arrived, timing of melt on the molting grounds did not affect birds' arrival timing. Because molt is a period of high survivorship (Hogan et al. 2013b), birds might be expected to remain on their molting grounds for as long as possible and depart only when necessary due to freeze. A dependency of departure on freeze would further align with the frost wave hypothesis for waterfowl on fall migration, which has been shown in Greater White-fronted Geese (Anser albifrons) and Swan Geese (Anser cygnoides) (Xu and Si 2019). Similarly, Mallards (Anas platyrhynchos) depart on fall migration and from intermediate stopover sites with the onset of snow cover regardless of latitude (Weller et al. 2022). Here, we found that the timing of freeze on the molting grounds had little effect on the timing of initiation of fall migration, influencing migration initiation only when Barrow's Goldeneye molted at locations that froze relatively early in the year compared to other molting locations. At molting locations that froze relatively late in the year, birds had typically already departed by the time of the annual freeze. Given the key importance of latitude in explaining departure timing, these birds may be cued to depart by endogenous rhythms, regardless of local weather conditions. Additionally, because molting grounds are heavily used by molting waterfowl, with as many as 4,000 Barrow's Goldeneye at a single molting location (Hogan et al. 2011), these sites may experience local resource depletion. It is possible that once they regained the ability to fly, some birds who had departed long before their specific molting grounds froze relocated to other northerly staging areas with more abundant food resources prior to beginning directed movements to southerly wintering locations. Previous studies show that the timing of one annual event can carry over to affect the timing of subsequent events (Breidis et al. 2018, Gow et al. 2018. For Tree Swallows (Tachycineta bicolor), the timing of breeding initiates a domino effect on events throughout the entire annual cycle (Gow et al. 2018). For Eurasian Hoopoes (Upupa epops) and Collared Flycatchers (Ficedula albicollis), the timing of a given annual event depends on the timing of the preceding event; for example, the timing of arrival on the breeding grounds strongly affected the timing of onset of breeding in both species (van Wijk et al. 2017, Breidis et al. 2018. Across many species, the nonbreeding period acts to reset timing (van Wijk et al. 2017, Breidis et al. 2018, Gow et al. 2018. We found that for each migration, Goldeneye that had initiated the migration later also concluded it later, indicating a degree of phenological carryover between stationary stages. However, individual Barrow's Goldeneye were able to partially compensate for delayed departures by consistently reducing the likelihood of stopover and the duration of that migratory leg. Delayed individuals thus updated their migratory decisions en route in an attempt to make up for time lost due to delayed departures. Across avian taxa, carrying a tag can have lethal (Bodey et al. 2018, Brlík et al. 2019) and sublethal impacts, ranging from bodily injury to a reduction in reproductive success (Bodey et al. 2018, Weiser et al. 2016. Effects are often most pronounced on smaller-bodied species (Brlík et al. 2019, Weiser et al. 2016 and for species with longer migratory distances (Bodey et al. 2018). In sea ducks, implantation with an intracoelomic transmitter can cause abnormal behavior for several days and subsequently delay the initiation of spring migration (Lamb et al. 2020). Implantation with a transmitter can also cause lethal effects, but fatalities are not always easily differentiated from transmitter failure (Forstner et al. 2022). Here, we show that capture and tagging have sublethal behavioral effects, delaying tagged birds' departure on the migration immediately following tagging by 9, 11, and 5 days, on spring, molt, and fall migrations, respectively. After controlling for this delayed departure, there were no additional effects of tagging on birds' migratory movements, suggesting that they were able to compensate for tag-induced delayed departure without resultant delays in timing of subsequent annual cycle events.
Our study shows that individual Barrow's Goldeneye refine their migratory decisions in response to environmental conditions encountered throughout their annual cycle. This flexibility in response to variable environmental conditions suggests that they have some capacity to adapt behaviorally to the changing climate. However, our study also shows that there is considerable unexplained variation in the migratory behavior of individuals. Our models explain 43-82% of the variation in migratory timing and stopover decisions. Additionally, migratory and breeding phenology do not necessarily respond to the same environmental cues (Åkesson et al. 2017). For Barrow's Goldeneye, it is unknown how changes in migratory phenology influence reproductive timing and success. Further work evaluating how individual variation in migratory behavior influences reproductive timing and success as well as survival across the annual cycle is warranted. Further, while plasticity, such as is seen here, allows migrants to quickly adapt to changes in their environment (Charmantier and Gienapp 2014), long-term persistence of populations in the face of ongoing climate change requires evolutionary adaptation in migratory phenology (Gienapp et al. 2007). To date, no studies have been able to demonstrate a definite link between advancing avian migratory phenology and evolution (Charmantier and Gienapp 2014). Barrow's Goldeneye display plasticity to adjust migratory timing in response to interannual variation in environmental conditions. However, it remains to be seen how well the species is able to adjust migratory timing and subsequent annual events, such as breeding, to persistent climatic changes. R. Scott, D. Shervill, K. Smith, R. Stayne, K. Tangen, B. Uher-Koch, C. Van Stratt, M. Wilson, K. Wright, C. Wohl, and R. Worcester for their assistance in sample collection; and J. Barrett, S. L. Lee, M. Willie, and C. E. Fuss for their assistance with Argos data management and GIS support. In addition to co-author M. McAdie, we also thank the following veterinarians who performed surgical implants: K. Doty, D. Mulcahy, and P. Tuomi. Any use of trade, firm, or product names is for descriptive purposes only and does not represent endorsement by the U.S. Government.

Funding statement
Funding and/or support for this research was provided by Environment and Climate Change Canada, U.S. Geological Survey, U.S. Fish and Wildlife Service, Alaska Department of Fish and Game, Alberta North American Waterfowl Management Plan Partnership, Alberta Conservation Association, Ducks Unlimited Canada, the Natural Sciences and Engineering Research Council, and the North American Sea Duck Joint Venture.

Ethics statement
This work was conducted in accordance with Environment and Climate Change Canada Scientific Permit to Capture and Band Migratory Birds 10673P and U.S. Fish and Wildlife Service Federal Fish and Wildlife Permit #MB789758. Capture and surgical procedures followed prescribed guidance by the Simon Fraser University Animal Care and Use Committee (Animal Care and Use Protocol 1121B-06).

Data availability
Argos PTT data used in this study are available from Movebank in the file "Migration Patterns of Pacific Sea Ducks" (https:// www.movebank.org/cms/webapp?gwt_fragment=page=studie s,path=study1441422788). Snow and ice cover data are from the Interactive Multisensor Snow and Ice Mapping System (IMS); (US National Snow and Ice Data Center 2008).