Early returning long-distance migrant males do pay a survival cost

2.1 Modeling the effects of timing on survival

The data set consists of data from eight passerine migrant species in spring (three short- and five long-distance migrants) and eight in autumn (three long- and five short-distance migrants) ringed in Denmark during migration from 1950 to 2002 (Bønløkke et al., 2006; Thorup et al., 2013) and recovered dead at any time at any location (see Table 1). The birds were ringed across the country by volunteers under the national Danish ringing scheme (Copenhagen Bird Ringing Centre) and the primary breeding area is in Scandinavia and Finland (Bønløkke et al., 2006). For the recoveries included in our analyses, the date of death was recorded as accurate to the day.

Survival of young is often lower than adults (Grüebler & Naef-Daenzer, 2010), but in our data, it was generally not possible to distinguish between young and adults. Instead, we excluded birds dead in the first year after ringing ensuring that only adult birds were compared. In this case, one year was subtracted from all life spans to avoid bias from the truncated distribution. We found overall same results when including birds found dead in the first year, but excluding this reduced noise in the data set (see Appendix 1 for the model results with individuals dead in the first year included).

Migration periods were defined individually for each species. Our focus was to determine overall patterns across species groups rather than species-specific patterns. To allow for standardization of migration timing within species, we included species with at least five birds recovered after the first year. For species with a large overlap between migration and breeding (mainly hole-nesting species such as pied flycatcher Ficedula hypoleuca), we chose a shorter period than the full migration period to make sure that actively breeding birds were excluded. For the species included, Denmark is near the breeding grounds (Bønløkke et al., 2006) and migration timing is highly correlated throughout northern Scandinavia (Bakken, Runde, & Tjorve, 2006). Thus, we regard capture dates as reasonable proxies for timing of return to and departure from the breeding grounds, respectively, as songbirds stopovers in Europe are generally few and relatively short, often 1–2 days (Goymann, Spina, Ferri, & Fusani, 2010; Wernham et al., 2002) but up to 15 days (Chernetsov, 2012).

We converted capture date to day of the year for all individuals (1 January = 1) and used it as a proxy for arrival (spring) or departure (autumn) day. Migration day was standardized within species subtracting the species-specific mean and dividing by the species-specific standard deviation. The species were divided in short- (European) and long-distance (African) migrants (referred to as migration distance).

We analyze our data in a capture–mark–recapture framework using a modification of the dead recovery model for ringing data from the British Trust for Ornithology (BTO dead recovery model; Cooch & White, 2016). The BTO dead recovery model uses the time between ringing and dead recovery to estimate survival probability. The probability of finding a dead individual and reporting its ring number to a ringing scheme, hereof “recovery probability,” is thereby assumed to be independent of the age at which the individual died. The model allows estimating survival when the total number of ringed birds is unknown (Burnham, 1990) as is the case for the Danish ringing data, where only annual totals are known—not numbers ringed in specific migratory periods. Further, because survival estimates are based on the exact and known life span of the individuals, survival estimates are independent of recovery probability, that is, among-individual variance in recovery probability does not affect survival estimates. The proportion of recoveries in the j^th year after ringing is S₁S₂S₃…S_j−1(1 − S_j)/(1 − S₁S₂S₃…S_k), with k years being the maximal duration (in years) between ringing and recovery. The number of recoveries in the j = 1, …, k years between ringing and recovery follows a multinomial distribution with cell probabilities dependent on j as defined above. We used the logit link function to include a linear predictor for annual survival probability S_i for each individual i. We included spring or autumn migration date (linear and quadratic), the migration distance (short- vs. long-distance migrants), and their interaction as predictors for survival probability. We further included year (linear and quadratic) in order to account for biases produced by inhomogeneous temporal distribution of our data and possible temporal changes in both survival and timing of migration. There are a number of well-known biases in ringing and recovery data, especially given the large spatial and temporal scale of our data. However, heterogeneity of recovery probability (such as lower recovery probabilities in the northernmost latitudes compared to central Europe; Saurola, Valkama, & Velmala, 2013) is not expected to affect the robustness of our results. To confirm the lack of influence from variation in recovery probability and investigate potential effects of variation in survival, we plotted latitude against year and timing of return migration, and return migration against year (Appendix 2a–c). These plots revealed no correlations beside a weak trend toward earlier spring migration in recent years. Therefore, we included year but not latitude as a covariate in our models.

The assumption that recovery probability is independent of the age at which an individual died may be violated in our case because it is known that recovery probabilities have almost constantly declined between 1960 and 1998 (Robinson, Grantham, & Clark, 2009). Declining recovery probabilities within the life span of an individual results in an underestimation of survival. However, because we are not interested in the absolute survival estimates but in the effect of a covariate (return date) on survival, we do not expect the model assumption to bias our conclusions. As in our data set, years from ringing to recovery and year of ringing do not seem to be correlated (r = 0.09, SE = 0.04; Appendix 2e); a change in recovery probability over time is indeed unlikely to be influencing our survival estimates. Year was included as a covariate (linear and quadratic effect) in our models to account for systematic changes in survival over time, see below. Despite spring and autumn migration dates having changed over time (Tøttrup, Thorup, & Rahbek, 2006a, 2006b), such changes were only slight in our data (Appendix 2c,d). Thus, we did not include more complex interactions of survival with return date over time.

To partially pool information between the different species and to account for pseudoreplication, we included species as a random factor in the model. We modeled the intercept, and the linear and quadratic effects of day by species and assumed a normal distribution of the species-specific deviations from the parameter means. The model was fitted to the data using Markov chain Monte Carlo simulations as implemented in JAGS that we used via the R2jags package (Su & Yajima, 2015) in R 3.2.0 (R Core Team, 2016). We used weakly informative prior distributions that should not have a recognizable effect on the results but that give a priori implausible values small probabilities as recommended by Gelman et al. (2014). We used normal distributions with means of zero and standard deviations of 5 for all the fixed effects coefficients. For the among-species variances of the intercept, the linear and quadratic date effects, we used a folded-t distribution with two degrees of freedom for the standard deviation (recommended by Gelman, 2006). Two Markov chains of length 10,000 were simulated with burn-in set to 1,000 and thinning to 5. Convergence was evaluated visually and by the Brooks–Gelman–Rubin statistics (Brooks & Gelman, 1998). Spring and autumn data were analyzed separately (See JAGS code for the model in Appendix 4). A prior sensitivity analysis showed that the normal (0, 5) prior for the fixed effects had a negligible effect on the results (Appendix 5).

2.2 Estimating timing in relation to maximum survival

To compare the survival statistics to timing of return migration, we use percentage of ringed birds per day in the migration season on Christiansø (55.32°N, 15.2°E) from year 1976 to 1997. Christiansø is a 0.22 km² rocky island located in the Baltic Sea 18 km from the nearest land source. Due to the small size of the island and limited breeding numbers, birds caught in the migration seasons are with high certainty migrants. The primary breeding grounds of these birds are North Scandinavia and Finland. We use data from Christiansø because of the long time span and a fully standardized trapping effort. The time span for the standardized trapping is shorter than our full data set and slightly biased toward late years. During the period of standardized trapping, trapping dates advanced on average by 0.26 and 0.18 days/year in spring and autumn, respectively (Tøttrup et al., 2006a, 2006b). These changes are comparatively small compared to the span of return and autumn migration timings considered here, and given that only minor changes in return time were revealed in the ringing data, we do not expect pronounced effects on our results.

The percentage is derived from the mean number of birds caught per day (calculated as birds per day summed over all years divided by total number of birds). Further, for species where data on sex are available, we derive percentages for each sex separately. To test for gender effects in our survival estimates which could confound the comparisons with timing, we ran models separately for males and females to generate gender-specific survival curves. The gender-specific models did not support differences among sexes in survival optimum (Appendix 3). Credible intervals (CrI) were smaller for the combined models, and these were used throughout. We compare the timing of return to the survival patterns by calculating the percentage of birds migrating before the maximum survival day for long-distance migrants only because short-distance migrants show no clear relationship between timing and maximum survival. We do this for each sex separately in species where males and females could be sexed.

3.1 Effects of timing on survival

We find that short-distance migrants had the highest survival if returning early and their survival decreases with migration day in spring. Long-distance migrants returning early have low survival, and individuals migrating slightly later than the species-specific average return time have highest survival. For late returning long-distance migrants, the mean survival seems to decrease again. However, the uncertainty of these survival estimates is large so that we do not know whether the survival patterns for late returning individuals differ between long- and short-distance migrants (Figure 2a).

For the North European migrants, we find that maximum survival is associated with spring migration dates around 19 May for pied flycatcher, 21 May for redstarts Phoenicurus phoenicurus, 26 May for common whitethroat Sylvia communis, and 20 May for lesser whitethroat Sylvia curruca which is before the maximum for willow warbler Phylloscopus trochilus on 28 May (Figure 2b). There are also indications of a survival optimum for robins Erithacus rubecula around 1 May, whereas survival in song thrush Turdus philomelus and dunnock Prunella modularis was the highest early in the season.

In autumn, short-distance migrants migrating early and late have the highest survival while long-distance migrants show the opposite pattern with a survival optimum on intermediate dates. However, 95% CrIs for long-distance migrants are large. Hence, it is difficult to determine the survival at the beginning and end of the migration periods (Figure 2c). Short-distance migrant species show varying patterns with robin and dunnock driving the pattern of high survival for early and late autumn migration days found in the initial model (Figure 2d).

3.2 Timing in relation to maximum survival

In long-distance migrants, peak spring migration occurs before the date of maximum survival. In species where males and females can be separated, only males arrive before the optimum and females come close to coincide with maximum survival (Figure 3). We find that 81% of the males and 56% of the females pass before maximum day of survival in pied flycatchers, for redstarts 79% of males and 51% of females pass before maximum survival and for common whitethroat 72% of males and 46% of females pass before the survival maximum (Figure 3). Timing in relation to survival in spring for species that could not be sexed is shown in Figure 4, and timing in relation to survival in autumn is shown in Figure 5.