Among-individual and within-individual variation in seasonal migration covaries with subsequent reproductive success in a partially migratory bird

Abstract

Within-individual and among-individual variation in expression of key environmentally sensitive traits, and associated variation in fitness components occurring within and between years, determine the extents of phenotypic plasticity and selection and shape population responses to changing environments. Reversible seasonal migration is one key trait that directly mediates spatial escape from seasonally deteriorating environments, causing spatio-seasonal population dynamics. Yet, within-individual and among-individual variation in seasonal migration versus residence, and dynamic associations with subsequent reproductive success, have not been fully quantified. We used novel capture-mark-recapture mixture models to assign individual European shags (Phalacrocorax aristotelis) to ‘resident’, ‘early migrant’, or ‘late migrant’ strategies in two consecutive years, using year-round local resightings. We demonstrate substantial among-individual variation in strategy within years, and directional within-individual change between years. Furthermore, subsequent reproductive success varied substantially among strategies, and relationships differed between years; residents and late migrants had highest success in the 2 years, respectively, matching the years in which these strategies were most frequently expressed. These results imply that migratory strategies can experience fluctuating reproductive selection, and that flexible expression of migration can be partially aligned with reproductive outcomes. Plastic seasonal migration could then potentially contribute to adaptive population responses to currently changing forms of environmental seasonality.

1. Introduction

Population responses to changing environments depend on dynamic relationships between components of fitness and within-individual and among-individual variation in expression of environmentally sensitive phenotypic traits [1–6]. Specifically, population outcomes depend on the degrees to which individuals' phenotypes, and the fitness consequences of those phenotypes, vary within and among seasons and years, representing labile plasticity and temporal variation in selection [1,6]. They also depend on the degree to which temporal variation in individual phenotype is aligned with temporal variation in the fitness consequences, representing adaptive plasticity [6,7]. Quantifying such effects is prerequisite for predicting individual and population responses to environmental variation and change, and for identifying mechanisms that maintain or constrain the critical phenotypic variation [5–8].

One key trait that directly mediates spatial escape from seasonally deteriorating local environments, and thereby directly causes spatial population dynamics on short (within-year) timescales, is seasonal migration (i.e. reversible cross-season movements). Phenotypic expression of seasonal migration, versus year-round residence at one location, commonly varies substantially among individuals within populations. Many populations are therefore ‘partially migratory’, including many fish, birds, mammals, and amphibians [9–15]. Furthermore, experiments show that expression of migration can be state-, condition-, and density-dependent, representing rapid labile plasticity (i.e. within-individual variation [16–18], also termed phenotypic flexibility [19]). Hence, individuals can switch between residence and migration both between years, and on shorter timeframes within years. For example, individuals may be ‘early’ (pre-emptive or anticipatory) migrants that move in advance of predictable local environmental deterioration (e.g. onset of winter or other forms of seasonality), or ‘late’ (responsive) migrants that move upon experiencing further environmental deterioration [12,14,20–22]. Dynamic relationships between such variable expression of migration versus residence and major fitness components must then be quantified in order to understand the maintenance of within-individual and among-individual variation and to consider the short-term and longer-term population consequences; but this is rarely achieved [13,14,23,24].

Recent studies show that survival probability can be higher in migrants [25,26] or residents [10,13], or be similar for both groups [27], and that such relationships can vary with environmental conditions [28]. Carry-over effects can then cause subsequent reproductive success to differ between surviving sympatric-breeding residents and migrants [29], but such effects are not always evident [13,24]. However, no studies have yet quantified variation in subsequent reproductive success across multiple coexisting phenotypes, for example, across early migrants versus late migrants versus residents (hereafter termed ‘migratory strategies’). Consequently, no studies have quantified the degree to which associations between such migratory strategies and reproductive success can vary between years, or hence examined whether migratory strategies that yield the highest success in any particular year are more frequently expressed. Evidence of any such adaptive labile plasticity would imply that dynamic seasonal migration could contribute to maintaining populations experiencing changing forms of environmental seasonality, by rapidly reshaping spatio-seasonal population distributions [14].

Quantifying such effects requires quantifying variation in reproductive success in relation to individual location(s) during the preceding non-breeding season, which is still very challenging. Retrieval of location data from non-transmitting archival loggers or bio-markers (e.g. stable isotopes) is often restricted to individuals that return and breed and can consequently be (re-)captured. Such state-dependent sampling could bias estimated relationships with reproductive success, and bio-markers may not clearly distinguish migratory strategies that differ in timing. Meanwhile, technologies that directly transmit individuals’ locations are still typically deployed on relatively few (often non-random) individuals. Systems where marked individuals can be directly observed all year therefore provide valuable opportunities to relate reproductive success to preceding non-breeding season locations, but require advanced statistical analyses to account for inevitably incomplete observations and resulting uncertainty.

Accordingly, we fitted novel capture-mark-recapture mixture models to year-round local (i.e. breeding area) resightings of colour-ringed European shags (Phalacrocorax aristotelis) in a partially migratory population to probabilistically assign individuals as ‘early migrant’, ‘late migrant’, or ‘resident’ in each of two winters. We thereby quantified variation in phenotypic expression of seasonal migration arising within and among individuals, within and between the 2 years, with associated uncertainty. We then tested whether subsequent reproductive success differed between the three migratory strategies, whether observed relationships differed between the 2 years, and whether between-year directional changes in individual migratory strategy occurred and were aligned with reproductive outcomes. We thereby consider the scope for adaptive labile plasticity in a key trait, seasonal migration, that directly shapes spatio-seasonal population dynamics.

2. Methods

(a). Overall approach

One approach to quantifying variation in reproductive success in relation to preceding non-breeding season location (and hence migratory strategy) is to mark individuals with field-readable tags, then undertake intensive non-breeding season resighting surveys followed by comprehensive reproductive monitoring. However, even in readily observable systems, not all present individuals are typically detected in any one survey. Using ‘capture-mark-recapture’ (CMR) analyses can then minimize bias in parameter estimates, but basic analyses assume that all individuals within defined model strata have identical occasion-specific detection probabilities (i.e. no heterogeneity). In partially migratory populations, this assumption will be strongly violated regarding local (breeding area) detection probability (P_L) during non-breeding season surveys. Here, individuals that have migrated away (i.e. temporarily emigrated) cannot be locally detected, while individuals that remain resident are locally detectable. However, this heterogeneity provides information with which to infer individuals' migratory strategies solely from local resightings; locally detected individuals are by definition currently resident, while individuals that are not locally detected on any occasion have a non-zero probability of being a departed migrant. Repeated surveys can then allow individuals to be assigned to strategies with high probability.

Finite mixture models, that consider a specified number of latent (i.e. unobserved) classes of individuals with different detection probabilities (and/or other parameters, [30–34]), provide one means of using heterogeneity in P_L to infer individual migratory strategy. Such models can assign individuals to latent classes (representing different migratory strategies) given their observed encounter history, yielding posterior probabilities of class membership [30]. Advantages are that classes need not be defined a priori but can emerge from analyses, and uncertainty in class assignments can be formally quantified and propagated through subsequent analyses [30,31,33] (e.g. relating individuals’ probabilistic class membership to subsequent reproductive success). Potential challenges are that biological interpretations of model-estimated classes may not be clear or necessarily consistent across different datasets. These challenges can be mitigated by formulating model structures with clear objectives and biological knowledge of the focal system; by using additional information and/or heuristic criteria to validate class interpretations; and by explicitly evaluating uncertainty in class assignments and consistency across datasets [30,32,35].

(b). Field data

We collected the required data for two full annual cycles (2017–2018, 2018–2019) for shags breeding at Bullers of Buchan, Aberdeenshire, Scotland (hereafter ‘BoB’, electronic supplementary material, S1 [36]). During 2009–2018, shags hatched or breeding at BoB were marked with individually coded colour-rings (licenced by British Trust for Ornithology), creating a sample of identifiable adults (age ≥3 years) alive during 2017–2019 (electronic supplementary material, S1). Each breeding season (April–July), the breeding area was intensively surveyed every 5–10 days (19, 18, and 20 surveys in 2017–2019, respectively). All nests were mapped and systematically monitored through to offspring fledging or nest failure following established protocols [37], and all adults were systematically identified as unringed or colour-ringed, and ring codes recorded (electronic supplementary material, S1). Coastal rocks were also surveyed for roosting shags, yielding observations of additional colour-ringed individuals that were apparently not currently breeding (sub-adults, skipping adults, or failed breeders). This systematic effort ensured overall breeding season P_L ≈ 1 (electronic supplementary material, S1).

These observations identified 121 and 127 colour-ringed individuals alive in summers 2017 and 2018 that would be aged at least 3 years, and hence deemed capable of breeding, in summers 2018 and 2019, respectively (conditional on over-winter survival, electronic supplementary material, S1). The reproductive success of all surviving individuals was measured as the number of chicks fledged (range: 0–4) following standard protocols [37] (electronic supplementary material, S1). Surviving individuals that apparently did not breed (or could have failed early) were assigned values of zero, thereby preventing bias due to excluding failed or non-breeders [38] (electronic supplementary material, S4).

Since shags have partially wettable plumage they must return to land daily to dry, facilitating year-round resightings [29,39]. Accordingly, during winters (1 September–28 February) 2017–2018 and 2018–2019, all daytime and pre-dusk roost sites within daily foraging range of BoB (ca. 16 km) were regularly surveyed for colour-ringed shags (electronic supplementary material, S1). This provided substantial data on local non-breeding season presence in the BoB area, overall 77 positive survey days (i.e. at least 1 colour-ringed individual locally resighted) totalling 1250 resightings of 111 individuals in 2017–2018; and 48 positive survey days totalling 942 resightings of 117 individuals in 2018–2019. Further winter surveys were undertaken across multiple other night roost sites and their adjacent day roosts spanning greater than 500 km of UK east coast, providing direct resightings of some colour-ringed individuals that bred at BoB and had definitely migrated away [28,29,39] (electronic supplementary material, S1). However, since surveys did not cover the entire UK coastline, not all departed migrants would be directly identified.

(c). Mixture model analyses

Individuals’ local encounter histories were compiled by defining 17 consecutive 10-day ‘occasions' spanning 1 September–18 February through each winter. This is the maximum number (i.e. shortest duration) of equal-duration occasions for which there was at least 1 local positive survey day within each occasion in both winters. Later (after 18 February) sightings were excluded because data inspection showed that some known (i.e. directly observed elsewhere) migrants had already returned, meaning that local breeding area resightings were no longer highly informative of an individual's migratory strategy (electronic supplementary material, S2). Two occasions representing the breeding seasons before and after each focal winter were also included. Each individual's full encounter history for each year therefore comprised 19 occasions, on which it was locally observed or not (electronic supplementary material, S3). Winter resightings from other locations (i.e. direct observations of migrants elsewhere) were not included in encounter histories, but were used to validate mixture model class interpretations (i.e. assess whether known migrants were assigned as migrants based on local resightings), as is good practice with mixture models (details below and electronic supplementary material, S3).

Resightings during 2010–2017 strongly suggested three broad categories of individuals: those regularly locally resighted throughout winter (residents); those never resighted locally in autumn or winter, or only very early in autumn (early September, putative ‘early migrants’); and those resighted locally through autumn (September–November) but not later in winter (late November–January, putative ‘late migrants’). Accordingly, CMR mixture models with three latent classes in P_L were fitted to data from 2017–2018 and 2018–2019 (full technical details, electronic supplementary material, S2). Initial analyses showed that three-class models were substantially better supported than two-class alternatives, providing formal support for the three-class structure (ΔAICc ≥ 20, electronic supplementary material, S2). Occasion-specific parameter structures for P_L were then defined to broadly capture the three postulated migratory strategies (full details, electronic supplementary material, S2). These constraints ensured that all estimated P_L values exceeded zero, allowing estimation of individual class assignment probabilities. They also facilitate inference since local resightings in different winter occasions are not equally informative regarding migratory strategy (e.g. residence is more strongly implied by local mid-winter resightings than by early September resightings). However, note that all P_L values are estimated from the local resighting data, not set a priori. Annual survival probabilities (ɸ_A) were known to be high in both years. Of 121 and 127 focal individuals observed in summers 2017 and 2018, 111 and 122 were resighted in summers 2018 and 2019, respectively. This implies ɸ_A ≥ 0.92 and ≥0.96, respectively, and hence ɸ_I ≥ 0.995 and ≥0.998 for each inter-occasion interval (ɸ_I = ɸ_A^1/18 given 19 occasions and 18 intervals, assuming time-independent mortality). Consequently, we fitted constant ɸ_I across all 18 intervals within each annual cycle, and did not attempt to estimate class-specific survival probabilities with the current dataset and analysis (electronic supplementary material, S2).

CMR mixture models were fitted in programme E-Surge, and individual class probabilities extracted [30,31,40] (electronic supplementary material, S2). Separate models were fitted to encounter histories for 2017–2018 and 2018–2019, facilitating model validation and estimation of class probabilities (electronic supplementary material, S2). General goodness-of-fit (GoF) tests for CMR mixture models are not yet available. We, therefore, structured the encounter histories into three fixed post hoc groups according to their most probable latent class, and assessed GoF of models fitted to the group-structured approximation. There was no evidence of substantial lack of fit (bootstrap p: 0.16 and 0.11 for 2017–2018 and 2018–2019, respectively) or additional overdispersion (median ĉ: 1.07 and 1.09, respectively; full details, electronic supplementary material, S2).

The biological interpretation of resultant mixture model latent classes as distinct ‘resident’, ‘late migrant’, and ‘early migrant’ migratory strategies was validated in four ways. First, we summarized the distributions of individual class membership probabilities and thereby quantified the degree to which individuals were assigned to distinct classes with high probabilities (versus more continuously distributed probabilities, [30]). Second, we examined whether estimated P_L values concurred with patterns expected for each postulated strategy (e.g. low mid-winter P_L for migrants, higher for residents). Third, we examined whether each individual's most probable class matched heuristic interpretation of its local encounter history based on its last autumn resighting occasion and total positive winter occasions (i.e. early last observation and few positive occasions for early migrants; late last observation and many positive occasions for residents; intermediate values for late migrants). Fourth, for directly observed migrants (i.e. individuals resighted at night roosts elsewhere in winter), we evaluated whether the total summed mixture model probabilities that these individuals were migrants (early plus late) were close to one. Additionally, to assess across-year consistency of class assignment probabilities, we quantified the differences in probabilities for the set of identical or very similar encounter histories that occurred in both 2017–2018 and 2018–2019.

Mixture models also directly estimate initial class probabilities [30,33], here interpretable as the proportions of early migrants, late migrants, and residents at the start of each annual cycle. To further quantify these proportions among individuals that survived each winter and whose reproductive success was observed, we drew each individual's class from the multinomial distribution defined by its simplex of estimated class probabilities (which sum to one), and recorded the total number of individuals assigned to each class. We tested whether estimated class frequencies differed between the 2 years using a robust χ² test on the 3-strategy by 2-year contingency table. This process was repeated for 10 000 independent realizations of each individual's class, thereby propagating uncertainty resulting from class probabilities less than 1. Resulting distributions of class frequencies and χ² statistics are summarized as the mean and 95% confidence interval across 10 000 iterations.

Finally, to quantify the degree and form of between-year change in migratory strategy within individuals that survived through both years, the frequencies with which individuals were assigned to the same or different classes across the 2 years were extracted. Robust χ² tests on the 3 × 3-strategy contingency table were used to test whether these frequencies differed from a null expectation generated by random resampling from the overall class frequencies estimated in 2018–2019, conditional on classes assigned in 2017–2018 (full details and other null models, electronic supplementary material, S3). These analyses were also repeated for 10 000 independent realizations, and distributions of frequencies and χ² statistics summarized as above.

(d). Reproductive success

We fitted a generalized linear model (GLM) to test whether mean annual reproductive success differed between the three mixture classes (specified as a three-level fixed effect), and hence between the three inferred migratory strategies. We explicitly tested for a class-by-year interaction, and hence for between-year variation in the relationship between reproductive success and migratory strategy. Models with and without the interaction were fitted assuming a quasi-Poisson distribution of reproductive success, and the comparative ANOVA test statistic $({\chi }_{\mathrm{Obs}}^2)$ extracted. This analysis was again repeated for 10 000 independent realizations of each individual's probabilistic class assignment. Each model was then refitted with individual reproductive success randomly resampled from the focal year's observations, yielding a null ${\chi }_{\mathrm{Rand}}^2$. The statistic ${\chi }_{\mathrm{Diff}}^2={\chi }_{\mathrm{Obs}}^2-{\chi }_{\mathrm{Rand}}^2$ was computed, where positive values imply that the class-by-year interaction effect was greater given observed versus randomized reproductive success. The proportion of iterations for which this difference was negative $(P({\chi }_{\mathrm{Diff}}^2<0))$ was extracted.

To broadly capture variation in reproductive success with age, and hence minimize the degree to which current analyses could potentially be confounded by age-specific variation, we also fitted a two-level factor defining individuals as young (aged 3–4 years) or older (5+ years) adults (details in electronic supplementary material, S4). While most focal colour-ringed individuals had unringed mates, there were 17 breeding attempts where both adults were colour-ringed. Their observed reproductive success consequently occurred in the dataset twice. To account for resulting non-independence, reproductive success was resampled at the level of breeding attempts (rather than individuals). Across 103 individuals that appeared in the dataset in both years, the between-year correlation in individual reproductive success was small (Spearman correlation r_s = −0.08). Individual reproductive success was therefore resampled independently for each year. In practice, since most individuals were assigned to one mixture class with very high probability (see Results), conclusions were the same if models were directly fitted taking each individual's most probable class and hence migratory strategy in each year, with random individual effects. Possible effects of sex on relationships between migratory strategy and reproductive success were not considered because approximately 55% of adults were of unknown sex. Across known-sex individuals, frequencies of assigned migratory strategies did not differ between females and males (electronic supplementary material, S4). Similarly, there was no consistent age-specific variation in migratory strategy across the current dataset (electronic supplementary material, S4). Analyses were run in R v. 3.5.1 [41].

3. Results

(a). Migratory strategies

The mixture models successfully assigned encounter histories, and hence individuals, to distinct classes that were clearly and consistently interpretable as resident, late migrant, and early migrant (figure 1). In each year, most individuals were assigned to one class with very high probability (overall ≥0.95 for ≥83% of individuals, ≥0.75 for ≥97% of individuals; figure 1; electronic supplementary material, S1). Estimated class-specific P_L for the 17 winter occasions concurred with patterns expected for the three migratory strategies. Specifically, P_L was always low for early migrants and close to zero through mid-winter, and initially high and then lower for late migrants (figure 1). P_L for residents varied among occasions, reflecting variable survey success, but was often substantially higher than for the migrant classes through mid-winter (figure 1). Consequently, encounter histories assigned with high probabilities clearly matched heuristic interpretations; early and late migrants were typically last seen before winter occasion 4 (early October) and 11 (early December), respectively, while residents were seen subsequently, with more positive occasions in total (figure 1). The few less certain assignments were probabilistically split between early and late migrant, or between late migrant and resident (figure 1; electronic supplementary material, S3). In total, 17 and 29 focal individuals were resighted at night roosts elsewhere in winters 2017–2018 and 2018–2019, respectively (46 encounter histories of 36 different individuals), and hence were known migrants (electronic supplementary material, S1). The mixture models assigned all 46 histories as migrant with very high total probabilities (mean: 0.999, range: 0.964–1.000). By corollary, zero individuals assigned as resident by the mixture models were resighted elsewhere. Comparisons of 15 identical or similar encounter histories observed in both years showed that class assignment probabilities were highly consistent across years (mean absolute difference: 0.008, range: 0.000–0.090, electronic supplementary material, S2).

Figure 1.

Summaries of mixture model latent class assignments for European shags observed in (a–e) 2017–2018 and (f–j) 2018–2019. (a,f) Frequencies of individual maximum class membership probability. (b,g) Ternary plots visualizing each individual's simplex of class probabilities. Vertices represent classes interpreted as resident (R, left), early migrant (EM, right), and late migrant (LM, top). Points represent individuals, line segments indicate multiple identical points. Most points are at or near the vertices, indicating individuals assigned to one class with very high probability. Points on intervening dashed lines indicate individuals assigned to two classes with non-zero probabilities. Few points are within the triangular space, indicating little three-way uncertainty. (c,h) Estimated local (breeding area) resighting probabilities (P_L, with 95% confidence intervals) across 19 occasions for the classes representing residents (black, solid line), late migrants (mid-grey, dashed line), and early migrants (light grey, dotted line). (d,i) Last autumn observation occasion for individuals assigned as resident, late migrant, or early migrant with probability greater than or equal 0.95 (96 and 101 individuals in 2017–2018 and 2018–2019, respectively, occasion zero is the initial summer). (e,j) Total positive winter occasions for these individuals.

(b). Strategy frequencies and between-year change

Estimated initial probabilities (±1 s.e.) for the three mixture classes representing residents, late migrants, and early migrants were 0.42 ± 0.05, 0.26 ± 0.04, and 0.32 ± 0.06, respectively, in 2017–2018, and 0.22 ± 0.04, 0.39 ± 0.05, and 0.39 ± 0.06, respectively, in 2018–2019. The relative class frequencies estimated across individuals with observed reproductive success were quantitatively similar (electronic supplementary material, S1). These frequencies differed between years (mean χ²: 11.2, 95% CI: 8.5–14.0; mean p: 0.005, 95% CI: 0.001–0.016 across 10 000 realizations). Overall, residents were most frequent in 2017–2018 but least frequent in 2018–2019, while late migrants, and to some degree early migrants, were more frequent in 2018–2019 than 2017–2018. The population sample was therefore more migratory in 2018–2019 than in 2017–2018.

These changes in frequencies between years arose partly because, of 103 individuals that survived through both years, a mean proportion of 0.36 (95% CI: 0.32–0.40) were assigned to different classes in the 2 years and hence changed strategy (table 1a, hence mean proportion assigned to the same class in both years: 0.64, 95% CI: 0.60–0.68). The frequencies of individuals assigned to the same and different classes differed substantially from null expectation (mean χ²: 76.2, 95% CI: 38.4–140.9, all p < 0.001 across 10 000 realizations, table 1; electronic supplementary material, S3). More individuals than expected retained the same strategy across both years, while few or no individuals changed from late or early migrant to resident, from early migrant to late migrant or from resident to early migrant (table 1b). Individuals that changed strategy therefore predominantly became slightly more migratory (i.e. resident to late migrant, and late migrant to early migrant, table 1; electronic supplementary material, S3). There was, therefore, evidence of substantial between-year individual repeatability in migratory strategy, but also of substantial directional change.

Table 1.

Frequencies of inferred 2-year migratory strategies for 103 European shags included in analyses for both focal years, given (a) mixture model-assigned strategies for both years, and (b) model-assigned strategy for 2017–2018 and a randomly resampled strategy for 2018–2019 (full details, electronic supplementary material, S3). Rows and columns index strategies in 2017–2018 and 2018–2019, respectively, showing the mean frequency and 95% confidence interval across 10 000 realizations. In (a), dark grey shading (leading diagonal) identifies individuals assigned to the same strategy in both years. Light grey and white identify individuals that became more and less migratory in 2018–2019, respectively. In (b), dark and light grey indicate 2-year strategies that occurred more and less frequently than expected, respectively (compared to frequencies in a). White indicates strategies whose frequencies did not differ from overall expectation, but occurred more frequently than expected conditional on change (electronic supplementary material, S3). In addition, 17 new individuals entered the dataset in 2018–2019 (and hence do not contribute to estimates of between-year individual variation): 2, 6, and 9 residents, late migrants, and early migrants, respectively.

			2018–2019
			resident	late migrant	early migrant
(a)	2017–2018	resident	22.4 (21,24)	19.1 (17,21)	3.1 (2,4)
		late migrant	2.0 (1,3)	17.3 (14,20)	8.7 (6,11)
		early migrant	0.0 (0,0)	4.1 (2,6)	26.3 (24,29)
(b)	2017–2018	resident	10.6 (5,17)	17.5 (11,24)	16.5 (10,23)
		late migrant	6.6 (2,11)	11.0 (6,16)	10.4 (5,16)
		early migrant	7.2 (3,12)	12.0 (7,17)	11.2 (6,17)

(c). Reproductive success

Mean reproductive success (±1 s.d.) across all observed focal individuals was 1.1 ± 1.1 and 1.0 ± 1.1 fledged chicks in 2018 and 2019, respectively. In 2018, individuals assigned as residents during the preceding winter had substantially higher mean reproductive success than early migrants, while late migrants had intermediate success (figure 2). In 2019, late migrants had higher mean reproductive success than residents and early migrants (figure 2). Consequently, there was a strong class-by-year interaction, largely reflecting that residents had high mean success in 2018 but low mean success in 2019 (figure 2; electronic supplementary material, S4). Residents and late migrants, and to some extent early migrants, therefore had highest relative reproductive success in the year in which they were most frequent (figure 2). As anticipated, young adults had lower success than older adults (figure 2; electronic supplementary material, S4).

Figure 2.

(a) Predicted mean reproductive success in 2018 and 2019 for individuals assigned as resident (R), late migrant (LM), or early migrant (EM) during the preceding winter. Open and filled symbols denote young (age 3–4 years) and older (age 5+) adults, respectively. (b) Relative (mean-standardized) predicted reproductive success for each migratory strategy versus the proportion of individuals assigned to that strategy in each year (shown for older adults only for clarity). On both panels, squares, circles, and triangles represent residents, late migrants, and early migrants, respectively. Black and grey symbols denote 2017–2018 and 2018–2019, respectively. Points denote back-transformed mean predicted values and vertical solid and dashed lines denote back-transformed mean standard errors and 95% confidence intervals estimated across 10 000 realizations. Horizontal solid and dashed lines denote mean standard errors and 95% confidence intervals for initial class probabilities.

The substantial decrease in mean reproductive success of residents in 2019 compared to 2018, and the corresponding relative increase in success of late migrants (figure 2), could potentially arise if the most successful resident individuals in 2017–2018 became late migrants in 2018–2019 and remained successful. High reproductive success would then be associated primarily with individuals rather than migratory strategies. However, additional analyses provided no evidence of such individual effects. Specifically, reproductive success of late migrants in 2019 did not differ between surviving individuals that had been residents or late migrants in 2017–2018 (table 2a). Furthermore, reproductive success of residents in 2018 did not differ between individuals that subsequently remained as residents or became late migrants in 2018–2019 (table 2b). If anything, individuals that became late migrants tended to have lower reproductive success in 2018, opposing the between-year change in mean reproductive success estimated across all individuals (figure 2).

Table 2.

Comparisons of (a) reproductive success in 2019 of individuals assigned as late migrants (LM) in 2018–2019 that were residents (R) or late migrants in 2017–2018, and (b) reproductive success in 2018 of individuals assigned as residents in 2017–2018 that remained as residents or became late migrants in 2018–2019. Statistics are the number of individuals in each group (N), the group mean and standard deviation (s.d.) of reproductive success in the focal year, and the Wilcoxon test statistic (W) for the difference in mean and associated p-value (all presented as means and 95% confidence intervals across 10 000 realizations). Statistics for reproductive success in 2018 of individuals assigned as late migrants in both years are shown for comparison, but do not contribute to current hypothesis tests.

year	classes	N	mean	s.d.	W	p-value
(a) 2019	R → LM	19.0 (17,21)	1.32 (1.21,1.42)	1.01 (0.95,1.08)	154.4 (118.0,194.0)	0.71 (0.30,1.00)
	LM → LM	17.2 (14,20)	1.42 (1.21,1.61)	1.07 (0.98,1.16)
(b) 2018	R → R	19.3 (18,21)	1.74 (1.65,1.80)	0.88 (0.86,0.96)	144.7 (123.0,166.5)	0.36 (0.11,0.74)
	R → LM	18.2 (16,20)	1.35 (1.21,1.50)	1.21 (1.15,1.26)
	LM → LM	17.2 (14,20)	1.02 (0.84,1.20)	1.09 (1.05,1.13)

4. Discussion

Population responses to environmental variation and change depend on patterns of variation and covariation in individual phenotypes and fitness, representing plasticity and selection [1,2,6,8]. Seasonal migration (versus residence) is one key phenotype that allows rapid spatial escape from temporarily deteriorating environments, but patterns of dynamic (co)variation in migratory phenotypes and fitness components have not been fully quantified. We show that substantial phenotypic variation, with individuals classified with high probabilities as early migrants (inferred to depart from the breeding area in early autumn), late migrants (depart in late autumn), and year-round residents, is evident within a European shag population, and that these ‘migratory strategies' can have substantially differing subsequent mean year-specific reproductive success. Further, strategy frequencies and associated reproductive success varied markedly between two study years, and in each year the strategy associated with the highest estimated success in each year was more frequently expressed. These results highlight that seasonal migration can be an individually and temporally variable non-neutral trait, and imply that flexible seasonal migration may have capacity to mediate adaptive plastic responses to currently changing forms of environmental seasonality.

(a). Migratory strategy

Objectively classifying individual migratory strategy is generally challenging, requiring year-round location data and appropriate analyses [22,24]. We used capture-mark-recapture mixture models to assign individuals to latent classes representing distinct migratory strategies, with explicit uncertainty, solely using local (breeding area) observations (further commentary, electronic supplementary material, S2). CMR mixture models were originally devised to overcome ‘nuisance’ heterogeneity in detection probability and minimize resulting bias in estimated survival probabilities and population sizes [30,32,35]. In that context, support for a model with K classes does not necessarily mean that K biologically meaningful distinct groups exist, but simply that K latent classes adequately capture heterogeneity [30,34,35]. However, in our analyses, most encounter histories were assigned to one of three mixture classes with very high probability, and these classes are clearly biologically interpretable as ‘resident’, ‘late migrant’, and ‘early migrant’ (figure 1). While there may not in fact be three entirely distinct strategies, this structure clearly provides a very good approximation that allows strong individual-level inference. The few less typical cases with higher class uncertainty (figure 1; electronic supplementary material, S3), might represent slightly different strategies (e.g. mid-autumn departure), or could reflect further individual heterogeneity in P_L conditional on presence, or simply chance sequences of local resighting failure. Nevertheless, our results strongly imply the existence of different ‘migratory strategies', such that some sympatric-breeding individuals spend considerably more time per year than others in the breeding area environment.

Our analyses also demonstrate substantial within-individual variation in non-breeding season location within years, and in defined migratory strategy between years. Defining individual state at any instant as resident (i.e. in the breeding area) or migrant (i.e. away), individuals were inferred to transition between these states at different times of year, or never, generating emergent early migrant, late migrant, and resident strategies that can be viewed as differential within-year plasticity. Furthermore, while 64% of surviving individuals were assigned to the same class in both study years, representing high between-year repeatability of migratory strategy, 36% changed class, demonstrating between-year plasticity. Indeed, evidence of labile plasticity in migration, both within and between years, is accumulating in diverse taxa (but see [15]). For example, European blackbirds (Turdus merula) can show early or late migration alongside residence [21]; storms can induce altitudinal migration in white-ruffed manakins (Corapipo altera, [20]); and adult roach (Rutilus rutilus, [42]), elk (Cervus elaphus, [12]), and skylarks (Alauda arvensis, [24]) can switch between migration and residence between years.

In European shags, the form of between-year variation differed substantially from null expectations, and resulted in directional change that contributed to making the population more migratory in 2018–2019 than 2017–2018. Specifically, approximately 42% of surviving residents became late migrants and 31% of surviving late migrants became early migrants, but other transitions were infrequent to the degree that zero surviving early migrants became residents (table 1; electronic supplementary material, S3). Overall, these observed patterns of within-year and between-year migratory plasticity and repeatability are consistent with a ‘threshold model’, which proposes that discrete phenotypes are expressed when an underlying continuously distributed ‘liability’ exceeds some threshold [43] (electronic supplementary material, S5). This model was previously suggested to apply to migration versus residence in birds and fish [23,44], and to other discrete phenotypes including wingless versus winged (dispersive) morphs in hemimetabolous insects [45]. The patterns of phenotypic variation observed in shags could arise given among-individual variation and time-of-season dependence in liability, and increased population mean liability in 2018–2019 compared to 2017–2018 (electronic supplementary material, S5). Any effects of migratory strategy on subsequent reproductive success would then impose selection on a threshold trait, and hence on underlying migratory liability.

(b). Reproductive success

Despite recent interest in partial migration and its consequences [9,11–29], no studies have yet quantified within-year and between-year variation in sympatric reproductive success across diverse migratory strategies (more than simply resident versus migrant). For example, a leading study that used stable isotopes to assign individual skylarks as resident or migrant found no difference in breeding success across 4 years, but sample sizes were small and individuals that were not clearly assigned as simply resident or migrant were excluded [24]. Previous analyses of a different shag population breeding on the Isle of May, Scotland, showed that residents had higher breeding success than migrants across 3 years, but did not distinguish early and late migrants [29].

Our comprehensive reproductive monitoring of shags that survived through each winter revealed substantial variation in mean reproductive success with migratory strategy, and showed that the most successful strategy differed between years (figure 2). The estimated mean between-class differences exceeded 0.7 fledged chicks in both years, constituting notable effects compared to the grand mean success of approximately 1 fledged chick. Direct causal effects of migratory strategy on subsequent reproductive success cannot, of course, be proved by phenotypic associations. However, there was no evidence that observed differences in relative reproductive success between residents and late migrants across the two years arose because individuals that switched strategies had consistent reproductive success (table 2). Instead, the observed relationships could potentially reflect varying carry-over effects of winter location(s), perhaps including lasting effects on physiology or acquisition of breeding sites or local ecological information. Such mechanisms can be investigated in future. Our results then imply time-lagged indirect selection on migratory liability, and imply that such selection can act on both the occurrence and within-year timing of migration, and can also differ markedly between consecutive years.

Further, in each study year, more individuals expressed the migratory strategy that yielded the highest mean reproductive success; residents were most frequent and had high reproductive success in 2018, while late migrants were more frequent and had the highest reproductive success in 2019 (figure 2). Switching from residence to late migration is therefore expected to yield the highest mean reproductive success across the 2 years (electronic supplementary material, S4), and this was indeed the most frequently observed form of between-year switching. Strong evidence of adaptive plasticity of course cannot be definitively inferred across 2 years. However, the observed patterns could indicate adaptive labile plasticity expressed by some individuals, which is still surprisingly infrequently demonstrated for any trait or system [3,5,7,23]. This would in turn imply selection for migratory liability values near the threshold, thereby maintaining both within-year and between-year environmental responsiveness and generating substantial phenotypic variation.

However, this interpretation would raise the question of why plasticity was far from complete. Many individuals were early migrants in both years, even though repeated early migration is expected to yield low mean reproductive success (electronic supplementary material, S4), and some 2017–2018 residents remained resident in 2018–2019. Proximately, this implies that some individuals have high or low migratory liabilities, whether due to genetic or permanent environmental effects, and may consequently be quasi-obligate migrants and residents [e.g. 42]. High migratory liabilities could be maintained in a population if early migration yields the highest probability of over-winter survival, at least in some years, as recently shown for blackbirds (Turdus merula, [26]) and shags breeding on the Isle of May [28].

Overall, our current results provide evidence of among-individual variation, within-individual plasticity, and fluctuating reproductive selection on a key phenotype across 2 years, and indicate that plasticity and selection could be aligned. More years of data are clearly now required to ascertain which, if either, currently observed pattern of variation in reproductive success with migratory strategy (figure 2) is more typical; whether early migration is ever associated with high reproductive success; to identify underlying environmental drivers; and to quantify relative reproductive success of individuals that do and do not change migratory strategies between years. Partially migratory systems then provide exciting opportunities to quantify the magnitude and form of adaptive labile plasticity, the direct and constitutive costs, and hence the direct implications of such processes for spatio-seasonal population dynamics (e.g. [6,14]).

Ethics

Bird ringing was licensed by the British Trust for Ornithology (permit A4389).

Data accessibility

Data are available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.pvmcvdnhv [46].

Authors' contributions

J.M.R. conceived the ideas, undertook the analyses, and drafted the manuscript. J.M.R., M.S., and S.F. collected the data. All other authors contributed substantially to general conceptual and technical developments, and edited manuscript drafts.

Competing interests

We declare we have no competing interests.