Whilst efficient movement through space is thought to increase the fitness of long-distance migrants, evidence that selection acts upon such traits remains elusive. Here, using 228 migratory tracks collected from 102 adult breeding common terns (Sterna hirundo) aged 3 – 22 years, we find evidence that older terns navigate more efficiently than younger terns, and that efficient navigation leads to a reduced migration duration and earlier arrival at the breeding and wintering grounds.
We additionally find that the age-specificity of navigational efficiency in adult breeding birds cannot be explained by within-individual change with age (i.e. learning), suggesting the selective disappearance of less navigationally efficient individuals.
This suggests that, at least in common terns, learning of navigational skills may be largely absent in adulthood, and limited to the pre-breeding phase of life where tracking is more difficult.
We propose that selection might explain parts of the age-specificity of navigational performance observed in migratory taxa more generally; discuss the causes and evolutionary implications of variation in navigational traits and the selective agents acting upon them; and highlight the necessity of longitudinal studies when considering changes in behaviour with age.
Study population
The common tern is a long-lived, migratory seabird (Becker and Ludwigs 2004). The data we present here come from a long-term study population of common terns located at the Banter See in Wilhelmshaven on the German North Sea coast (53°30’40” N, 08°06’20” E). Here, chicks have been ringed since 1980, while the presence and reproductive performance of individually marked adults has been monitored following a standard protocol since 1992. Hereto, 101 adult birds were caught and marked with individually numbered subcutaneously injected transponders, and, since 1992, all locally hatched birds have been ringed as well as marked with such a transponder shortly prior to fledging.
As part of the standard protocol, arrival date is monitored using antennae affixed to the walls of the colony (Moiron et al. 2024) and the colony site is checked three times a week to find nests and record laying dates, clutch sizes and egg sizes. During incubation, which is shared between partners, antennae that can read the individual transponder codes are placed around each nest to identify breeding individuals. Combined with further nest-checks to establish hatching, mark offspring and assess their growth and survival, these methods enable the documentation of individual life-history trajectories (e.g. Moiron et al. 2022; Zhang et al. 2015).
Tracking
Between mid-May and early July 2016-2021, 228 tracks were recorded from 102 birds (2016 n = 24 (22 retrieved); 2017 n = 36 (28 retrieved); 2018 n = 50 (40 retrieved); 2019 n = 54 (48 retrieved); 2020 n = 65 (54 retrieved); 2021 n = 54 (36 retrieved); total retrieval rate of geolocators with data = 80.5%) were equipped with an archival light-level geolocator (Intigeo-C65, Migrate Technology, UK) attached to the leg using a 10 mm aluminium ring. The total mass of the ring, glue and geolocator was 1.6 g, which equalled 1.2% ± 0.1 SD of the body mass of the birds at capture, and did not have a detectable effect on behaviour, reproductive performance or survival (Kürten et al. 2019). The geolocators were set to sample ambient light intensity every minute, with the maximum light intensity being stored every five minutes (‘mode 10’ on Migrate Technology devices).
Track processing
Upon return from migration, i.e. between mid-May and early July 2017-2022, birds were retrapped to retrieve their geolocators carrying the archived light-level data, from which their migratory trajectories were estimated using the R package ‘FLightR’ (Rakhimberdiev et al. 2017). Device failure mid-trip notwithstanding, we have both a spring and autumn trajectory from each bird (in total: 228 tracks of 102 individuals, with 228 autumn and 184 spring trajectories).
To analyse these trajectories, we extracted two daily positions using the ‘run.particle.filter’ function and used the ‘stationary.migration.summary’ function to detect individual-specific migration periods based on the dates of arrival to, and departure from, the breeding colony and wintering area. In essence, this involved using the posterior probability of movement for each day’s calculated position to ascertain whether the bird was likely to be moving (movement probability > 0.4). Based on these probabilities, geolocator fixes were then assembled into periods of movement and non-movement, from which the breeding and wintering periods could be calculated. Spring and autumn migratory periods were defined as the intervening periods (for more details, see Kürten et al. 2022).
Once the beginning and end of each migration period were estimated for each track, we sought to remove any stopovers, during which birds would not necessarily exert a navigational preference, by removing GLS fixes where the distance moved was estimated at <100 km between fixes (removing n = 3451 fixes). This was meant to ensure that all fixes included in our dataset described the behaviour of birds moving in directed flight, and robustness to the selected cut-off value was tested via reanalysis with the cut-off value set to 50 km (removing n = 4575 fixes; see Table S4).
Autumn migration occurred between the 24th of July and 15th of October, whilst spring migration occurred between the 10th of February and 9th of May, with birds moving from their colony in Northern Germany to sites distributed between West and Southern Africa (see Kürten et al., 2022). Common terns took remarkably direct migratory routes – characterised by relatively few stopovers – with autumn and spring migration lasting on average 17.95 (±14.60 days standard deviation) and 27.56 days (±18.94 days standard deviation), respectively.
Analysis of age-specific navigational performance
Given that GLS positional estimates are inferred using light levels, they are inherently less accurate than those obtained using other positioning technologies (e.g. GPS). However, since heavier GPS devices have been associated with changes in at-sea behaviour in similar taxa (e.g. Gillies et al. 2020), we assumed that GLS devices were best-suited to our purpose. The error inherent to light-level geolocation comes from device shading and/or electrical errors in the device itself, and this error can manifest in one of two ways: (i) noise – indiscriminately affecting all devices, increasing the number of false negatives (e.g. noise obscuring a correlation between GLS positional information and a predictor), or (ii) bias – impacting devices non-randomly, causing false positives (e.g. weather-induced device shading causing variation in positional estimate, leading to a spurious correlation between cloud cover and distance travelled; Lisovski et al. 2018).
To minimise the effect of noise, we used the FLightR algorithm to process our GLS data, as it is robust to equinox error and leads to substantially less overall error than more traditional thresholding methods (c.250 km error per fix; Halpin et al. 2021; Rakhimberdiev et al. 2017). In addition, our large sample size (228 migratory tracks from 102 individual terns) secured statistical power. For positional error to manifest as bias and in turn affect our study, the causes of positional error (i.e. device shading/electrical error) would have to correlate with age. To assess whether such bias was present in our dataset (i.e. whether birds of a given age were biased in their positional estimates), we tested whether the positional estimates of known-location birds varied with age. This we did by assessing whether longitude and latitude varied with age over the months of June and July – when all birds tracked were known to be at the breeding colony – using a linear mixed effects model (see below). Given that this wasn’t the case (see supplementary material), we assume our conclusions to be unaffected by bias.
Characterising navigational performance without knowing a bird’s preferred migratory route a priori is challenging, since whilst many birds take a direct migratory route (e.g. Prochazka et al. 2018; Schmaljohann et al. 2012), some do not (e.g. Alerstam 2001; González-Solís et al. 2009; Guilford et al. 2009; Lisovski et al. 2021; Mellone et al. 2013). We chose to characterise navigational performance for each GLS fix of each tern as the instantaneous deflection between the instantaneous migratory trajectory (i.e. the direction the bird is currently going in) and the goal (Padget et al. 2018; Wynn et al. 2020a). As such, instantaneous deflection was calculated twice-daily for each bird, and did not assume that the bird in question was following the Great Circle route (“beeline”). Whilst it is true that ‘improvements’ in navigational performance with age might reflect differing motivations to navigate efficiently towards the goal, we reasoned this was the less likely explanation of any change in performance since all birds were established breeders and, at least in spring, under selection to arrive and breed early (Dobson et al. 2017; Moiron et al. 2024).
For autumn migration the destination was considered to be the highly repeatable individual-specific wintering site (Kürten et al., 2022), whilst in spring the destination was considered to be the breeding site. Deflection angles were expressed as an absolute deflection between 0 and 180o, with 0o representing no difference between the bird’s trajectory and the beeline between the observed position and the destination, and 180o meaning complete reversal (Figure 2). Whilst the response variable was angular, it was not circular, since the beginning and the end of the scale were not the same value. As such, we used linear rather than circular statistics to analyse our data.
To assess the effect of age on navigational performance, we used the R-package ‘lme4’ (Bates et al. 2015) to run a linear mixed effects model with the GLS-fix-specific absolute instantaneous deflection as a response variable assuming a Gaussian error distribution, and with track identity, nested within individual identity, as random effects to account for pseudoreplication (Padget et al. 2018). Because multiple birds of different ages (expressed in years since hatching) were tracked per season – and no birds were kept in the study for its entire duration – year-on-year changes in migration caused by the environment were unlikely to confound with age in any analysis.
Age was partitioned into an ‘average age’ and ‘delta age’ component (van de Pol and Wright 2009), whereby an individual's average age was defined as the average of all ages at which it was tracked, while delta age was defined as the difference between an individual's age for a given track and its average age (i.e. delta age = age − average age). When both were added as a covariate to the model, average age reflected whether birds tracked at different ages differed in their navigational performance (i.e. an among-individual pattern), whilst delta age represented any within-individual change with age (e.g. learning; van de Pol and Wright 2009). A significant effect of delta age would suggest that learning (or senescence, depending on the direction of the effect) explained changes in navigational performance with age, whilst a significant difference between the effects of average age and delta age would indicate selective disappearance. A graphical illustration of this principle can be found in Figure S1. We also included the interaction between average age and delta age to test for non-linear effects of age, because learning could decelerate as birds grow older. Finally, we added season as a 2-level class variable (autumn and spring), both as a main effect and in interaction with all age components, in order to test whether (age-specific) navigational performance differed between autumn and spring migration. Since both average age and delta age were on the same scale, and there were no other predictors, we did not mean-centre or normalise the variable.
Significance was assessed using likelihood ratio tests, comparing the hypothesis model to a null model that was identical to the hypothesis model save for the exclusion of the interaction/term under investigation. Effect sizes and bootstrapped 95% confidence intervals, which were estimated using the R-package ‘arm’, are reported in Tables 1 and S1. The full model was simplified, allowing for the creation of a minimally adequate model with more immediately interpretable effect sizes. Since the three-way interaction between season, average age and delta age; the two-way interaction between average age and delta age; the two-way interaction between season and average age; and the two-way interaction between season and delta age were non-significant, these terms were excluded during a step-wise backwards elimination process (Table 1). The resulting minimally adequate model included the main effects of season, average age, and delta age. Delta age was included irrespective of significance so as the difference between the effects of the delta and average age components could be assessed. This we did by estimating the difference between both effects and assessing whether the associated bootstrapped 95% Confidence Intervals overlapped zero (van de Pol & Wright 2009).
We proceeded to re-run all analyses using only the 177 tracks (i.e. 78% of all tracks). We did so because birds migrating further south seemingly followed different, less efficient routes (see Figure 1), such that any difference in navigational performance might represent differences in destination rather than performance (see Tables S1 and S2).
Analysis of the effect of navigational efficiency on migratory timing and duration
To test whether the duration of migration, or the resulting arrival date (known to be under selection, at least in spring; see Moiron et al. 2024), varied with instantaneous deflection, for each route we regressed the overall time spent migrating (i.e. the time of the end of migration minus the time of migration onset, in days; assuming a Gaussian error distribution) or the Julian date of arrival against the average instantaneous deflection observed over the route, including individual identity as a random effect and season as a fixed effect.
As above, significance was assessed using likelihood ratio tests comparing the hypothesis model to a null model that was identical to the hypothesis model save for the exclusion of the interaction/term under investigation. All effect sizes and bootstrapped 95% confidence intervals are reported in Tables 2 and S2.