Breeding phenology and winter activity predict subsequent breeding success in a trans-global migratory seabird

Inter-seasonal events are believed to connect and affect reproductive performance (RP) in animals. However, much remains unknown about such carry-over effects (COEs), in particular how behaviour patterns during highly mobile life-history stages, such as migration, affect RP. To address this question, we measured at-sea behaviour in a long-lived migratory seabird, the Manx shearwater (Puffinus puffinus) and obtained data for individual migration cycles over 5 years, by tracking with geolocator/immersion loggers, along with 6 years of RP data. We found that individual breeding and non-breeding phenology correlated with subsequent RP, with birds hyperactive during winter more likely to fail to reproduce. Furthermore, parental investment during one year influenced breeding success during the next, a COE reflecting the trade-off between current and future RP. Our results suggest that different life-history stages interact to influence RP in the next breeding season, so that behaviour patterns during winter may be important determinants of variation in subsequent fitness among individuals.

were broken. However, all these events are associated with parents failing to incubate the egg properly without gaps between incubation shifts -leading to the egg not hatching, or to the egg being left alone in the burrow and therefore being damaged by other birds.
Therefore this has to do with incubation quality and could be related to prior activity (e.g. parents could be in a poor body condition, requiring them to cut short their incubation stints to feed); it should not be adding any random noise to our results.
To minimise disturbance, we made sure that several steps were taken during our daily visits: 1) we used 'knock-down tags' to detect changes in occupancy [3,4]; 2) we used 'short-access tunnels', allowing us to have rapid access to chicks or adults, reducing over all disturbance [5,6]; and 3) many of our study birds already carry a radio frequency identification tag [7], allowing us to check bird identity without handling. Knock-down sticks allow us to reduce disturbance because we only need to inspect nests when the tags were disturbed at the entrance showing that an adult had entered or exited the burrow (these can be disturbed by prospectors, but this does not happen often). Short-access tunnels are a widely used method (e.g. [8]) to reduce disturbance as described by [5].
This system facilitates easy access to burrow chambers as we make a hole above the chamber and put a lid on which is covered by vegetation, hiding the holes completely [8].
The holes were made during daytime when adults are absent and had no negative effects on burrow occupancy [8]. The goal of our daily inspections was to identify 'egg neglect' and the breeding progress (egg lay date, start of incubation, hatch day, fledge day). Thus, we did not have to remove birds from their burrow during daily inspections, but only to confirm whether or not an adult was in the burrow.

Data analyses
To assess predictive power of phenology on individual RP, we employed a supervised machine-learning algorithm based on adaptive boosting [9]. The classifier was trained on a set of features to predict individual RP as one of these three categories: "CHICK", "EGG", or "SKIP". Nine features were included: we considered both prior-and postbreeding phenology (laying, hatching and fledging dates), as well as migratory phenology as extracted above (dates of colony departure, arrival at wintering grounds, departure from wintering grounds and colony arrival). The SAMME algorithm [10] was used as prediction is on > 2 (3 here) categories. The classifier's accuracy was determined by 10fold cross-validation, where the algorithm is trained on nine tenths of the data and the last decile is used to compute a confusion matrix, the procedure being repeated for all ten subsets. This process was repeated 1,000 times to assess classification accuracy.
To understand how wintering behaviour affects RP, we analysed behavioural patterns based on saltwater-immersion data. Because these data show high-frequency variability, a de-noising procedure was first used to extract nonlinear trends without any reference to breakpoints identified above. To this effect, a time-series additive decomposition was performed to extract nonlinear trends for each track. Cumulative distributions of denoised data were then extracted and averaged for "CHICK", "EGG" or "SKIP". The Kolmogorov-Smirnoff test was used to assess significance.
Year-to-year reproductive data can also be summarised as a transition matrix giving the frequencies of RP state changes from one year to the next. To this effect, sample size was expanded to include an additional 88 individuals (47 males, 41 females) whose breeding progress and breeding performance were monitored, but which had not been tracked with Geolocators. Multi-event capture-mark-recapture (MECMR) models [11] were used to estimate transitions rates among "SKIP", "CHICK", "EGG" and "DEAD" (= not recaptured). Survival rates were either constant or varying as a function of state and/or time; the best transition rate model was identified using Akaike Information Criterion with small sample correction.

A note on the breeding habits of Manx shearwaters
Unlike some albatrosses, Manx shearwaters do normally breed every year. In fact, their subsequent breeding success is dependent on the amount of investment that breeders made in the previous year. If Manx shearwaters use the "by-product breeding pattern", we should expect to see the simpler transition between SKIP/CHICK and CHICK/SKIP once they reach maturity, which we did not observe here. Thus, we suggest that Manx shearwaters do try to breed every year, and when their previous breeding event was costly, they try to regain body condition during winter as much as possible, but do not manage to regain their condition by the next spring and therefore decide to skip (or could not manage) breeding instead.

A note on the device effects
We have used a measure of 'breeding success' of individuals that carried a geolocator in our experimental colony (North Haven) to assess the overall impact of our tracking work by comparing it with a neighbouring unmanipulated plot (Isthmus) which has been monitored for many years by the National Nature Reserve staff (Bu che et al. 2013in JNCC Contract Report: available from 1995 to provide estimates of shearwater productivity on the island. Doing this allows a large sample of breeding attempts to be compared, but it restricts us to using the same measure of breeding success as in those studies. This measure was the number of chicks raised per egg laid, and is therefore not quite the same as that used for the carry-over analyses at the core of our study (which includes birds that skip). Thus, for the impact assessment comparison only we used the same measure as the island staff -chicks raised per egg laid -which across the duration of our study was 0.86, compared to 0.63 in the unmanipulated plot for the same period (2009)(2010)(2011)(2012)(2013).
We did not detect any measurable effect in breeding performance (number of chicks per laid egg) in this study (Table S1) as agreed with another geolocator study in Manx shearwaters [12]. However, negative effects of equipping year-round geolocator in hormonal responses have been reported [13]. Thus some caution in interpreting the results may be suggested.

Comparing breeding success at unmanipulated and manipulated plots
Breeding success of Manx shearwaters from 2009 to 2014 in unmanipulated plot (Isthmus: data are available until 2013: [14]) and in experimental plot (North Haven: individuals used in this study).

Plot
Year . Sampling distribution of the success rate of the SAMME classifier assessed by ten-fold cross-validation. This distribution was obtained by running the cross-validation 1,000 times.  200 300

Multi--event model framework for estimating state transition rates
Multi-event capture-mark-recapture (MECMR) is a modelling framework widely used to estimate state-dependent demographic rates of interest (e.g. survival) together with transition rates between different "states" individuals occupy (e.g. being infected or not), while explicitly accounting for imperfect (less than 1) and heterogeneous (biased) detectability of marked individuals, and uncertainty in the assignment of the state to an individual. Given their breeding status (i.e. skipped breeder, breeder with an egg, breeder with a chick), birds can be assigned to different states, and multi-event models provide an ideal framework for estimating transition rates (from state to state) and survival probabilities simultaneously in the same model. Furthermore, as different constraints can be imposed on state-dependent survival and transition parameters, multi-state models provide a rigorous method of evaluating the fitness consequences of transition rates.
In MECMR models, at each capture occasion an individual can occupy one amongst a finite set of mutually exclusive states. Between subsequent capture occasions, individuals move independently between these states [11]. However, a state is not always possible to assign when an individual is captured. Thus, at each capture occasion, we observe an event rather than a state. Events are related to the true, but not necessarily known, state of the individual through a series of conditional probabilities [15,16]. The MECMR model we develop here uses four exclusive states that an individual can occupy at each capture occasion (which is the breeding season): (1) being breeder with a chick (state "CHICK"), (2) being a failed breeder, producing only an egg (state "EGG"), (3) skipped breeder (state "SKIP"), (4) dead (state "DEAD").
An individual can occupy only one state in a given breeding season. Transitions among these four states (i.e. "CHICK", "EGG", "SKIP, "DEAD") happen between two subsequent breeding seasons, with the state 'DEAD' being an absorbing state (a dead individual cannot move to another state). Transitions are modelled as a two-step process composed of the probability of survival over the annual time interval, followed by the probability of transitioning among live states. The recapture of the marked individuals is described in the event matrix, where an individual that is alive (i.e. occupying alive state) can be either captured, or not captured. There are four possible events, related to one or more real underlying states: 1 = individual is captured at the nest, and it produced an egg in that breeding season ("EGG") 2 = individual is captured at the nest, and it produced a chick in that breeding season ("CHICK") 3 = individual is captured at the nest, but without an egg or chick ("SKIPPED") We coded the capture histories of males and females using the four event codes shown above (0, 1, 2, 3). We treated females and males in two separate analyses to avoid any problems related to non-independence between males and females (i.e. they may belong to the same breeding pair).

Specification of parameters and the model structure
Following notation in Pradel (reference [3]) our model is defined with three types of parameters: (1) initial state probabilities, represented in a vector of probabilities, (2) transition probabilities involving: survival probabilities (ϕ), and between-state transition probabilities (ψ); and (3) recapture probabilities (p).

Model covariates and model selection process
There is no specific goodness of fit (GOF) test for MECMR models. Thus, we assessed the fit of the general mark-recapture assumptions to our data by assessing the GOF of the single state Cormack-Jolly-Seber (CJS) model [17]. The CJS model assumes all animals present at the same sample occasion have equal future survival and recapture probabilities regardless of past history and capture in the current sampling occasion.
These assumptions were tested using program U-SURGE [18]. None of the components of the test returned significant results.
We considered four possibilities for the variation of each parameter (recapture, 32 Table S8. Initial state rates, transition rates and recapture rates and ± 95% Confident Interval: CI for female Manx shearwaters. Estimates were obtained from best-supported model in Table S6.