Behavioral responses to spring snow conditions contribute to long-term shift in migration phenology in American robins

Migratory birds have the capacity to shift their migration phenology in response to climatic change. Yet the mechanistic underpinning of changes in migratory timing remain poorly understood. We employed newly developed global positioning system (GPS) tracking devices and long-term dataset of migration passage timing to investigate how behavioral responses to environmental conditions relate to phenological shifts in American robins (Turdus migratorius) during spring migration to Arctic-boreal breeding grounds. We found that over the past quarter-century (1994–2018), robins have migrated ca. 5 d/decade earlier. Based on GPS data collected for 55 robins over three springs (2016–2018), we found the arrival timing and likelihood of stopovers, and timing of arrival to breeding grounds, were strongly influenced by dynamics in snow conditions along migratory paths. These findings suggest plasticity in migratory behavior may be an important mechanism for how long-distance migrants adjust their breeding phenology to keep pace with advancement of spring on breeding grounds.

One of the most pronounced and well-documented effects of global climate change has been the advancement of spring at high northern latitudes (Gray 2007), where temperatures are rising nearly two to three times faster than the global average (Gray 2007, Stocker et al 2013. This has caused alarm because of the important links between seasonal environmental dynamics and the phenology, reproductive success, and overall population health of many species (Moller et al 2008, Meller et al 2018. With some exceptions (e.g. Post and Forchhammer 2008, Clausen and Clausen 2013, Gauthier et al 2013, phenological responses are proving especially strong in Arctic and boreal ecosystems. Plant growing seasons have advanced by nearly two days per decade at the pan-Arctic scale (Park et al 2019), and some animal species have shifted their migratory (e.g. Le Corre et al 2016) or reproductive phenologies (e.g. Réale et al 2003, Høye et al 2007 to keep pace. There is widespread evidence that many migratory bird species are shifting the timing of their arrival to breeding grounds to match changes in local climate and phenology (Cotton 2003, Marra et al 2005, Rubolini et al 2007, Barrett 2011. These shifts are likely driven by intense selection for philopatry (Winger et al 2019) coupled with the benefit of increased reproductive success associated with early arrival to the breeding grounds (Norris et al 2004). However, for many long-distance migratory passerines, overall shifts in migratory phenology are complicated by the uncoupling between rapidly changing conditions in their Arctic and boreal breeding grounds and the relative stability of photoperiod and environmental cues that trigger migration in their temperate and tropical wintering grounds (Farner and Follett 1966, Dawson et al 2001, Ramenofsky and Wingfield 2007. In addition, the high energetic costs of migration, pressure to breed, and subsequent consequences for reproductive success (Paxton and Moore 2015) force migrants to make numerous movement decisions as they travel to their breeding grounds (Alerstam 2011).
Behavioral responses to environmental conditions along migratory routes could contribute to observed phenological shifts on breeding grounds by constraining the overall window in which subsequent breeding activities may occur. Yet, we lack critical understanding of how these decisions relate to environmental conditions along migratory routes, limiting our ability to predict potential constraints on birds' ability to shift arrival timing to breeding grounds in response to ongoing and future change (Chmura et al 2019). This is particularly true for migratory songbirds because, while tremendous insight into songbird migration has been gleaned from light-level geolocators, until recently, global positioning system (GPS) tracking devices were too large to be worn by most species (Kays et al 2015), precluding direct study of individuals' fine-scale movements (McKinnon et al 2016). Recent advances in GPS tracking technology have generated units small enough to be placed on some songbird species, providing opportunities to test how individuals adjust their migration in response to environmental dynamics.
Here, we combined long-term observations  of the timing of spring migration and GPS tracking data of individuals (n=55, 2016-2018) to understand potential mechanisms underlying phenological shifts of one of the most widespread, common, and recognizable passerine species in North America, the American robin (Turdus migratorius) (figures 1(a), (b), 2) as they migrated to Arctic-boreal breeding grounds. First, we tested whether the timing of spring migration of American robins through the Canadian boreal forest has significantly advanced over the past quarter-century. Second, we tested whether inter-and intra-seasonal dynamics in environmental conditions along migratory routes influence the migration behavior and phenology of individual American robins, including the passage timing of migration over our tagging site (Slave Lake, Alberta), timing and likelihood of stopovers, migration rates, and timing of arrival to breeding sites.

Study area
American robins were captured in the Lesser Slave Lake Provincial Park (7700 ha; 55°26' N, 114°49'W) near the town of Slave Lake in northern Alberta, Canada. The park is bordered by Lesser Slave Lake to the west and Marten Mountain (elevation: 1020 m) to the east. These two natural features serve as a funnel that concentrates bird populations over the park as they migrate through the area.

Study species
American robins are the most abundant and broadly distributed thrush species in North America. Robin breeding ranges extend from northern Alaska, Yukon, Northwest Territories, Manitoba, and Quebec to southern California, west-central Texas, the panhandle of Florida and even into Mexico along interior slopes from western Sonora and eastern Chihuahua to eastern Oaxaca (Armstrong 2016). Robins overwinter in much of North America, but are seen only rarely north of the Canada-United States border and southeastern Alaska during winter months (Armstrong 2016). Large robin migrations are witnessed over much of North America, with flock sizes ranging from a dozen to several hundred individuals. Observations at the Lesser Slave Lake Bird Observatory (LSLBO) show that robins typically arrive in the region in late April.

Field methods and robin location data
Daily estimates of individuals from the LSLBO were based on daily constant effort mist netting, visual migration counts, census for each day during spring migration over the past 25 years . Daily constant effort mist netting was based on twelve 30 mm mesh panel mist nets (12 m × 2.6 m) which have been established in the same location over the entire study period. Nets were deployed daily 30 min prior to sunrise and checked every 30 min. Visual migration counts were based on hourly 5 min visual overhead migration count conducted by an LSLBO staff member through binoculars from the same location daily starting 30 min prior to sunrise and ending 7 h later for a total of 8 counts per day. Census counts are based on an established 600 m census route which has remained in the same location over the entire study period. During the first three hours after sunrise, an LSLBO staff member walked the route south for 30 min, recording all birds seen or heard. Daily estimates of the total number of robins migrating through Slave Lake was estimated as the sum of estimated individuals from the three protocols.
American robins were captured in mist nets from 2016 to 2018 at the LSLBO and Boreal Centre for Bird Conservation (BCBC). Mist netting occurred between April 15 and May 8 each year. The LSLBO is a part of the Canadian Migration Monitoring Network and deploys twelve mist nets (12 m × 3 m) which are opened daily for 7 h beginning 30 min before sunrise. In order to increase opportunities for capturing robins, we deployed on average six additional mist nets which were also opened daily beginning 30 min before sunrise. Weight, wing length, age, and sex of captured individuals were determined following relevant guidelines by Pyle (1997). Weight of captured individuals was determined to the nearest hundredth of a gram by placing individuals in a tared weighing tube on a digital scale. Wing chord length was measured using a flat rule with a perpendicular stop at zero. The stop was placed against the carpal joint of the wing and the length was read at the natural tip of the wing extending down the length of the rule with no downward pressure applied to feathers. Second year birds were determined by the presence of a moult limit within the greater coverts where inner replaced formative feathers contrast with the outer retained juvenile feathers. After second year birds were determined by a lack of said moult limit. Sex was determined after age by plumage.
Individuals were then banded and outfitted with a Lotek ARGOS PinPoint GPS (Lotek Wireless, Ontario, Canada) unit via a nylon harness. GPS units weighing either 3.5 or 4 g were fitted on individuals weighing greater than 70 or 80 g, respectively, representing less than 5% of the bird's mass. The mean body mass of individuals outfitted with 4 g tags was 84.1 g, with a mean percentage of body mass of 4.7% (n=8). The mean body mass of individuals outfitted with 3.5 g tags was 79.8 g, with a mean percentage of body mass of 4.3% (n=47) (figure S1 is available online at stacks. iop.org/ERL/15/045003/mmedia).
GPS units were scheduled to acquire and store their location every 48 h beginning the day after deployment and continuing for a total of 30 fixes (approximately late April to early July) and eventually upload location data via the Argos satellite system. Each unit is limited to a maximum of 30 fixes with a battery lifetime of approximately 3 months. Therefore, no individuals were tracked for more than a single spring migration.

Environmental data
We used the environmental-data automated track annotation (Env-DATA) system (Dodge et al 2013) to access environmental datasets for our analysis of robin migratory movements and habitat selection. Env-DATA is housed within Movebank (www.movebank. org), an open-access, online system for the management and analysis of animal movement data (Kranstauber et al 2011, Dodge et al 2013). Daily surface wind velocity data are provided by the European Centre for Medium-Range Weather Forecasts global reanalysis dataset. Daily total accumulated precipitation at the surface data are provided by the NOAA National Centers for Environmental Protection (NCEP) through the North American Regional Reanalysis (NARR) dataset. We characterized habitat used by robins based on the GlobCover Land Cover dataset, provided by the European Space Agency.
MicroMet (Liston and Elder 2006a) and SnowModel (Liston and Elder 2006b) were used to simulate daily air temperature and snow depth distributions on a 5 km grid, over the robin migration routes, for the period 1 September 1980 through 30 June 2018. The simulated daily snow-depth distributions were processed to extract the annual snow-free date following (Liston and Hiemstra 2011). The model simulations required 3 hourly inputs of air temperature, relative humidity, wind speed and direction at 10 m above the ground, and precipitation. These were provided by NASA's Modern Era Retrospective-Analysis for Research and Applications (MERRA-2 Gelaro et al 2017) atmospheric reanalysis datasets. In addition, SnowModel requires spatially distributed topography and land-cover data. These were provided by the United States Geological Survey Global Digital Elevation Model (GTOPO30; 30 arc seconds, or~1 km) dataset, and the GlobCover Land Cover (v2.2; 10 arc seconds, or~300 m) dataset, and re-gridded to the 5 km simulation grid.
Monthly values of the strength of the Pacific decadal oscillation (PDO) were obtained from the National Oceanic and Atmospheric Administration.
Robin migration passage timing over Slave Lake, Alberta (1994Alberta ( -2018 We tested for the presence of trends and the influence of environmental conditions on robin migration timing to Slave Lake, Alberta using long-term dataset collected by the LSLBO (1994LSLBO ( -2018. Timing of migration was described using three migration phases (date of first 5%, 50%, and 95% of total spring observations). We tested for the presence of trends in passage timing using a generalized least squares approach taking into account first order temporal autocorrelation, by incorporating first order autoregressive correlation structure. To test for the influence of environmental variables on passage timing, we performed model selection of linear models based on all possible combinations of environmental variables (monthly mean air temperature, snow depth, and precipitation for March and April, annual snow-free date, and mean strength of the PDO for the 12 months preceding arrival) and year. We restricted the inclusion of environmental variables which were highly correlated (Pearson's |R|>0.7). As a result, snow-free date was excluded from being included with March and April snow-depth and March air temperature, as well as March snow depth with March air temperature and April snow depth. We ranked candidate models based on Akaike's information criterion corrected for small sample sizes (AIC c ) (Burnham and Anderson 2002). We considered models with ΔAIC c <2 to have strong support for associations between variables of interest.

Stopover and breeding grounds arrival timing
We modeled robin stopover timing at our tagging site in Slave Lake, Alberta and their breeding grounds, and investigated factors influencing timing using the Anderson-Gill extension to the Cox proportional hazards (Cox PH) regression modeling (Therneau and Grambsch 2013). We tested the proportional hazards assumption of Cox PH using the formula test recommended by Therneau and Grambsch (2013) and only included predictor variables which did not have significant violations of the proportionality (P>0.05). We used a 365 d, recurrent, time scale to model the baseline hazard standardized to a year beginning on 1 January (Fieberg and DelGiudice 2009). In this context, 'hazard' does not represent the standard hazard of mortality, but rather the 'hazard' of arriving to a given location.
To understand factors influencing robin stopover arrival timing, we used the date on which we captured individuals at Slave Lake as a stopover event because all individuals were captured while foraging. We consider this to be stopover event as all individuals were captured while refueling, as opposed to migrating which is consistent with known flexibility in stopover behavior in passerines, as compared to other groups (Catry et al 2004). We generated daily records of environmental conditions for each individual starting on 1 January of the year they were observed and ending when they were captured at Slave Lake. We included sex (referenced to female), mass, wing length, and a binary age classification (second year or after second year, referenced to second year) as predictor variables, as well as snow-free date and daily mean precipitation, snow depth, and air temperature at Slave Lake were averaged at four spatial scales (5-, 25-, 105-, and 255-km grid cells centered around Slave Lake). Snow depth and air temperature at Slave Lake averaged within a 5 km grid cell were highly correlated with the same variables averaged across larger spatial scales (25, 105, and 255km grid cell; R 2 >0.95), indicating local conditions at Slave Lake are representative of regional conditions. All further analyses were based on environmental conditions averaged across the smallest spatial resolution. To test whether arrival timing differed between years we tested for an effect of categorical year (2016-2018). We considered years to be significantly different if the 95% confidence intervals of odds ratios (exponentiated β coefficients) of included year variables did not overlap.
To test for the influence of environmental, demographic, and morphological variables on stopover arrival timing based on capture data, we performed model selection of Cox proportional hazards models of all individuals captured as part of our tagging effort (n=77). We generated candidate models based on all possible combinations of variables. However, we restricted the inclusion of snow depth and air temperature as predictor variables in the same model based on a high correlation (Pearson's |R|>0.7). We additionally restricted models to contain seven or fewer covariates based on the overall sample size (n=77). We ranked candidate models based on Akaike's information criterion corrected for small sample sizes (AIC c ) (Burnham and Anderson 2002), where the number of individuals observed at Slave Lake was considered the sample size. We considered models with ΔAIC c <2 to have strong empirical support. We assessed significance of variables with robust z tests and 95% confidence intervals for odd ratios (Therneau and Grambsch 2013). We considered predictor variables significant if odds ratio confidence intervals did not overlap one. In this case, for categorical variables the odds ratio corresponds to the instantaneous odds of arrival at the location in question relative to the reference group. For continuous variables, the odds ratios correspond to a proportional change in the odds of arrival per unit change in the covariate. In all models, we estimated robust standard errors for parameter estimates based on data clustered by year. For variables that were retained in multiple top models, responses were similar across models, so we report results from the model with the lowest AIC C value.
We modeled robin arrival to breeding grounds based on GPS locations collected over our three study years (2016)(2017)(2018). Each GPS location was considered a record. We identified general breeding locations for individuals based on a sustained reduction in movement during known breeding months (June and July) (Young 1955). GPS fixes were identified as being in the general location an individual's breeding grounds based on the proximity to the final GPS location (> 80% of total distance traveled) and consistently low movement rates (<10 km d −1 ). Arrival to breeding grounds was designated by the first breeding ground location. In some cases, robins were not tracked completely to their breeding grounds due to GPS failure. We again considered demographic and morphological variables (sex, mass, wing length, and a binary age classification) as well as daily snow depth and air temperature in a 5 km grid cell around each GPS location. To account for the fact that individuals traveled to breeding grounds of variable distances from our tagging site, we included distance traveled from our tagging site to breeding grounds as a predictor variable. We also included an interaction term between snow-free date and distance traveled from tagging site because snow is likely to melt out later at higher latitudes. We used the same procedure for testing differences between years as we did for stopover timing to Slave Lake. To test for the influence of environmental, demographic, and morphological variables on arrival to breeding grounds, we performed the same model selection procedure as above. We tested for correlation among all environmental covariates to exclude any combinations of variables that were highly correlated (Pearson's |R|>0.7). However, no covariates were found to be highly correlated. We additionally restricted all models to containing five or fewer covariates based on the sample size (n=55). The overall number of individuals included in the model was 55, 42 of which had sufficient GPS data to suggest that the individual was tracked to the general breeding area, based on a sustained dramatic decrease in movement.

Migration movement rate
We investigated environmental, demographic, and morphological variables influencing movement rates during migration. We identified locations of each individual as occurring during migration, as opposed to on breeding grounds, based on the proximity to the final GPS location (<80% of total distance traveled) and the movement rate (>10 km d −1 ). To standardize our analysis we sub-sampled to only include sequential locations, hereafter referred to as steps, that were acquired 48 h apart. This resulted in 101 unique migratory steps from 17 individuals. Preliminary analysis revealed that movement rates were highly related to distance from Slave Lake and that individuals selected breeding areas throughout northwestern Canada and Alaska. The northern Rocky Mountains represent a significant geographic feature of the landscape, which several individuals traversed in the course of their migration. To test potential differences as individuals crossed the Rocky Mountains, we tested differences in migration movement rates for individuals that crossed 130°W longitude in the regions east and west of the barrier using two sample t-tests. We also tested differences in mean environmental conditions, such as snow depth, number of days before/after the snow free date, proportion of locations with snow present, and wind speeds, experienced at the locations selected by these two regions using two sample t-tests.
We investigated the influence of environmental conditions on migration movement rates for all individuals and the subset of individuals that crossed 130°W longitude over the entire study domain and in the regions east and west of the barrier using linear regression. We summarized environmental conditions along each step by averaging conditions at each start and end location using the following variables: snow depth, air temperature, precipitation, and zonal and meridional wind speeds and the interaction between air temperature and snow depth. We also included distance from Slave Lake in our model of migration movement rates over the entire study domain.
We tested for differences in movement rates and total migration rate between sexes and age classes using t-tests and tested for the influence of mass and wing lengths using linear regression. For all individuals that were tracked entirely to their breeding grounds we estimated total migration rates based on minimum distance traveled and number of days between capture at Slave Lake and arrival to breeding grounds.

Results and discussion
We found that over the previous 25 years, American robins have migrated northward over the boreal forest surrounding Slave Lake, Alberta approximately 5 days earlier per decade using a genearlized least squares approach taking into account first order temporal autocorrelation (figure 3, 5% passage date: coeffi-cient=−0.5 d yr −1 , se=0.25, t(23)=−2.19, p=0.039; 50% passage date: coefficient=−0.65 d yr −1 , se=0.34, t(23)=−1.89, p=0.070; 95% passage date: coefficient=−0.69 d yr −1 , se=0.19, t(23)=−3.52, p=0.002) Interestingly, our observed rate of advance in migration timing is comparable to the only other long-term study of American robin migration, that we are aware of, that found that robins advanced their arrival to a high altitude site in the Rocky Mountains by 14 d over a 19 year period (approx. 0.74 d yr −1 ) from 1981 to 1999 (Inouye et al 2000). Earlier passage through Slave Lake may reflect a shift in migration timing to maintain the same arrival time to the breeding grounds as ranges expand northward (Tingley et al 2009), or could reflect selection to reach breeding grounds earlier as habitat and prey phenology advance both along migratory corridors and within breeding grounds (Norris et al 2004).
Over the same 25 year period, we found interannual differences in the timing of robin migration over Slave Lake, Alberta were explained by interannual variability not only in environmental conditions at this specific locale, but also in the strength of a continental-scale climate oscillation, the PDO (Gedalof, Mantua and Peterson 2002). Specifically, the migration passage timing of robins was significantly earlier in years when the mean annual strength of the PDO was more negative (supplementary tables 1-4). Since negative phases of the PDO are associated with warm and dry winters in continental North America (Gedalof et al 2002), our results indicate that robins migrate earlier through the northwestern Canadian boreal forest following milder winters in the region.
Our results suggest that American robins may fine-tune their northward rates of migration based not only on local environmental conditions along the way, forging ahead when and where environmental conditions permit, but also adjust departure from overwintering grounds and migration rates according to broad-scale climatic conditions (as indexed by the PDO). Numerous previous studies have found that dynamics in migration phenology are often related in complementary ways to both local environmental conditions and large-scale climate oscillations (e.g. Marra et al 2005, Jonzen et al 2006, Macmynowski et al 2007. Rainio et al (2006) found that the majority of 75 species migrated earlier to sites in northern Europe after winters in which the North Atlantic Oscillation was higher, suggesting large-scale climate oscillations may shape inter-annual differences in seasonality along avian migratory routes more broadly than local environmental conditions (Hallett et al 2004). Further, several studies suggest that long-distance migrants may be able to shift arrival timing to breeding grounds by taking advantage of broad-scale climatic connections between southern overwintering areas and breeding grounds Ambrosini 2008, Pancerasa et al 2018).
The GPS tagging data collected from individual robins migrating through Slave Lake during three consecutive springs (2016-2018) not only revealed the migratory flyway and breeding habitats used by populations which migrate through this locale (figures 1(a), (d)), but also allowed us to relate dynamics in the migratory behavior and phenology of individual birds to environmental conditions along their migratory paths. Analysis of stopover arrival timing from the 77 robins captured in Slave Lake as part of our tagging effort, using Cox proportional hazards regression models, showed that arrival timing at this known stopover site was delayed in years with persistent local snow cover, such as occurred in 2018 ( figure 4(a), supplementary tables 5-7). However, once in the general region, individuals were more likely to stopover during periods of inclement weather (i.e. deeper snow and higher precipitation) (figure 4(a), supplementary tables 5-7). By considering multiple behavioral responses (i.e. timing versus likelihood of stopovers), our results demonstrate multi-directional responses of robins to dynamics in snow conditions.
Our results indicate that predicting future migration phenology of long-distance migratory passerines to Arctic-boreal regions will require understanding of both broad-scale patterns in the advancement in spring environmental conditions, as well as the potential for increasing storm frequency and intensities (Maloney et al 2014). Although long-distance migrants are under pressure to reach their breeding grounds, the elevated energetic demands imposed by adverse environmental conditions (Wikelski et al 2003) may force them to moderate their migration behavior. For example, Briedis et al (2017) found that semi-collared flycatchers (Ficedula semitorquata) departed from their overwintering grounds in Eastern-Central Africa on approximately the same dates in two consecutive years. However, a persistent cold snap caused them to spend twice as long in the Mediterranean Basin relative to a warmer year, before continuing on to their breeding grounds. In addition, Both (2010) found that pied flycatchers (Ficedula hypoleuca) spring departure dates advanced, but not arrival dates  . Colored lines each correspond to one of three phases of migration during which 5% (light blue), 50% (dark blue), and 95% (black) of the total number of robins was observed passing northward over Slave Lake during the spring migration monitoring period in a given year. The dashed lines show statistically significant the trends in 5% and 95% passage dates. to breeding grounds due to environmental constraints along the migratory route.
Our analysis of the influence of environmental conditions along the migration paths taken by individual robins, using Cox proportional hazards regression models, revealed that the timing of arrival at their breeding grounds (n=42, figure 2, see Materials and Methods for estimation of general breeding location) was significantly delayed in response to later snow-free dates, deeper snow depths, as well as the distance of a given breeding site from Slave Lake (figure 4(c), supplementary tables 8-10). Like most northern-breeding passerine populations, the robins studied herein were likely under pressure to reach their breeding grounds as their nesting season is approximately half that of populations breeding in much of the conterminous United States (James and Shugart 1974). The compressed breeding season results in a narrow window in which robins must initiate clutches to ensure sufficient time for young, including potential second broods (Young 1955), to develop, and before photorefractoriness and the onset of molt (Dawson et al 2001), and harsh environmental conditions cue fall migration (Verhulst and Nilsson 2008).
Our results demonstrate, that despite this pressure, persistent spring snow cover may force birds to delay movement northward, likely by limiting access to food resources en route and inhibiting passerine flight (Richardson 1978). Delayed arrival to breeding grounds can have a significant impact on the breeding success of passerines (Smith andMoore 2003, Newton 2004) and may constrain their ability to adjust to climatic change on their breeding grounds (Both and Visser 2001). American robins have vast differences in migration distance and timing across their range (Jahn et al 2019), suggesting shifts in the populations passing through Slave Lake could play a role in long term advancement. Alternatively, advanced migration date may result from inter-generational advances, coupled with static individual migratory timing, as in Icelandic black-tailed godwits (Gill et al 2014), selection in response to changing environments (Charmantier and Gienapp 2014), or evolution in migratory timing programs as recently shown in pied flycatchers (Helm et al 2019). However, our results suggest that responses to environmental conditions en route may, at least in part, contribute to shifts in timing.
The responses to environmental conditions along robins' migratory route that we found are consistent with other Arctic studies that found both passerine, shorebird, and waterfowl arrival timing and breeding phenology is delayed or advanced in response to spring snow dynamics (Liebezeit et al 2014, Boelman et al 2017, Lameris et al 2018. In this way, our study is novel in quantitatively relating the movement of individually GPS-tagged robins to the environmental conditions along each migratory path. In lieu of this type of data, the majority of passerine migration studies have had to rely on environmental data collected at the specific locale where, and leading up to the time when, arrival data was collected (e.g. Ahola et al 2004, Boelman et al 2017. Alternatively, these studies have relied on regional-scale environmental data products that provide estimates of the conditions for large areas, and windows of time, through which birds may have migrated (e.g. Huin andSparks 2000, Haest et al 2018). In contrast, the fine scales and spatiotemporally matched nature of the robin movement and environmental datasets analyzed in our study has strengthened inferences regarding factors responsible for trends and variations in the timing of when robins arrive at their breeding locations. Our results Figure 4. Environmental, demographic, and morphological variables influencing spring migration phenology of American Robins across our three study years. (a) Odds ratios of variables retained in top models of robin arrival timing at a known stopover site at Slave Lake, AB. (b) Coefficient strengths of variables found to significantly predict migration rates over the entire study domain (upper section) and only in the region east of −130°longitude (i.e. east of the Rocky Mountains) (lower section). No variables were found to be significant in the region west of −130°longitude. (c) Odds ratios of variables retained in top models of robin arrival timing at their respective breeding grounds throughout Alaska and northwestern Canada. Error bars represent 95% confidence intervals. The dashed red lines indicate no significant response. demonstrate when, where and how dynamics in specific environmental conditions along migratory routes were related to the migration behavior and phenology of American robins, revealing that dynamics in multiple variables related to snow conditions play an overwhelmingly important role.
In addition to responding to localized and temporally transient environmental conditions, robins responded strongly to permanent features of the landscape. Impressively, while migrating between the stopover site at Slave Lake and their respective breeding grounds, robin migration rates between consecutive GPS fixes, which are influenced by both flight speed and potential stop-overs, averaged 84 km d −1 and reached up to 410 km d -1 (figure 1(c)). Migration rates decreased with deeper snow depths (coeffi-cient=−0.19, se=0.09, t(90)=−2.11, p=0.038, 95% CI=[−0.38 to −0.01]). Unexpectedly, we also found that migration rates increased with distance from Slave Lake (coefficient=0.69, se=0.08, p<0.001, 95% CI=[0.52-0.86]). Further investigation revealed that this acceleration is explained by the fact that the Alaska-breeding robins in our tracking study nearly doubled their migration rates (from 85 to 137 km d −1 , t(27.5)=−1.8, p=0.079) after traversing from the east to the west side of the Rocky Mountains, potentially suggesting individuals responded to local conditions once arriving in the region (supplementary table 10). Although meridional wind speeds did not differ between locations selected by Alaska-breeding individuals, strong easterly winds prevailed west, but not east, of the Rockies (from 0.59 to −0.22 m s −1 , t(25.8)=1.72, p=0.09) (supplementary table 10). The latter could be involved in the significant acceleration of robin migration west of the Rockies. Tailwinds have been shown to increase a bird's ground speed and decrease the overall energy needed to travel the same distance, thereby conserving energy reserves (Liechti 2006). This may be especially important for spring migrants at high latitudes where food may be scarce upon arrival to breeding grounds (Richardson 1978).
Further, we found that while on the eastern side of the Rockies, the migration rates of Alaska-breeding individuals were significantly, positively related to snow depth (coefficient=0.77, se=0.26, p=0.005, 95% CI=[0.24-1.29]), air temperature (coefficient=0.47, se=0.17, p=0.007, 95% CI=[0.13-0.81]), and the interaction between snow depth and air temperature (coefficient=0.48, se=0.22, p=0.036, 95% CI= [0.032-0.92]), as well as negatively related to zonal wind velocity (coefficient=−0.28, se=0.13, p=0.043, 95% CI=[−0.55 to −0.01]) ( figure 4(b)). However, we found no relation between their migration rates and environmental conditions once these same individuals had passed over to the west side of the Rockies (supplementary table 10). When we considered the migration rates of Alaska-breeding robins in combination with snow disappearance dates, we found that while on the eastern side of the Rockies, individuals were migrating through landscapes that had been snow-free for approximately two weeks (supplementary table 10). In contrast, once these individuals had passed over to the western side, the more persistent snow cover coupled with their faster rates of migration resulted in their passage over landscapes that had become snow-free less than only two days prior (supplementary table 10). Our analysis demonstrates that as robins pass through the Rocky Mountains en route to Alaskan breeding grounds, they take a significant step backwards in 'seasonal time' by entering into a different climatic region where spring has only just begun, and favorable winds propel them towards their breeding grounds. Our findings suggest that together, these environmental conditions may play a large role in keeping the breeding phenology of long-distance migratory passerines wellmatched to the inherently short window of suitable breeding and young rearing conditions inherent to Arctic-boreal habitats each year. In fact, as climate warming continues, fortuitous regional climatic conditions such as this may provide long-distance migrants with natural buffer zones that prove critical to overcoming both increasingly unpredictable and disjointed spatiotemporal dynamics, as well as long-term shifts, in seasonality on overwintering grounds, along migratory corridors, and on breeding grounds.

Conclusions
We leveraged advanced GPS tracking technology to quantitatively link spatiotemporally-explicit environmental conditions to the migration behavior and phenology of American robins en route to their rapidly warming breeding grounds at high northern latitudes. Our results demonstrate that behavioral responses to environmental conditions along migratory routes may contribute to the advance in migration timing we observed through the Canadian boreal forest over the past quarter century. The improved understanding of migration behavior and phenology of passerines we have developed is critical to creating predictive models that integrate habitat preferences and movement traits into forecasts of phenological responses as the climate continues to warm.