Climate change leads to differential shifts in the timing of annual cycle stages in a migratory bird

Shifts in reproductive phenology due to climate change have been well documented in many species but how, within the same species, other annual cycle stages (e.g. moult, migration) shift relative to the timing of breeding has rarely been studied. When stages shift at different rates, the interval between stages may change resulting in overlaps, and as each stage is energetically demanding, these overlaps may have negative fitness consequences. We used long‐term data of a population of European pied flycatchers (Ficedula hypoleuca) to investigate phenological shifts in three annual cycle stages: spring migration (arrival dates), breeding (egg‐laying and hatching dates) and the onset of postbreeding moult. We found different advancements in the timing of breeding compared with moult (moult advances faster) and no advancement in arrival dates. To understand these differential shifts, we explored which temperatures best explain the year‐to‐year variation in the timing of these stages, and show that they respond differently to temperature increases in the Netherlands, causing the intervals between arrival and breeding and between breeding and moult to decrease. Next, we tested the fitness consequences of these shortened intervals. We found no effect on clutch size, but the probability of a fledged chick to recruit increased with a shorter arrival‐breeding interval (earlier breeding). Finally, mark–recapture analyses did not detect an effect of shortened intervals on adult survival. Our results suggest that the advancement of breeding allows more time for fledgling development, increasing their probability to recruit. This may incur costs to other parts of the annual cycle, but, despite the shorter intervals, there was no effect on adult survival. Our results show that to fully understand the consequences of climate change, it is necessary to look carefully at different annual cycle stages, especially for organisms with complex cycles, such as migratory birds.


Introduction
During the past decades, the increases in temperature have affected biological systems in diverse ways (Walther et al. 2002). One of the most evident effects of climate change are the alterations in phenology or timing of annual events across different animal and plant taxa (Crick et al. 1997;Brown et al. 1999;Parmesan & Yohe 2003;Parmesan 2006). For example, timing of flowering, breeding and migration are some of the events known to have advanced in response to the increased temperatures (Crick et al. 1997;Parmesan 2007;Charmantier & Gienapp 2014;Thackeray et al. 2016). The important question is whether these shifts in timing are fast enough to track the changes in the optimal timing for stages to happen (Visser & Both 2005;Visser 2008).
The use of standardised long-term datasets, especially of the same population, has allowed researchers to track how organisms respond to climate change and to identify potential causes and consequences of these responses (Visser 2008(Visser , 2016. A relatively large number of studies have reported the effects of climate change on timing of breeding, particularly in birds (e.g., Crick et al. 1997;Both et al. 2004). A few studies also have explored if optimal breeding time shifted with climate change and if animals were able to track such changes and respond accordingly (Visser et al. 2006;Reed et al. 2013a, b;Plard et al. 2014;Stoks et al. 2014;Phillimore et al. 2016). These studies are crucial to our understanding of whether species can adapt to such rapid changes in the environment or not (Gienapp et al. 2013).
While some organisms present quite simple annual cycles with only a breeding/nonbreeding transition, others have much more complex cycles (Jacobs & Wingfield 2000;Wingfield 2008). For example, many species of birds and mammals also migrate, moult their plumage/pelage and/or hibernate. These additional stages of the annual cycle are likewise reported to shift due to climate change (Both & te Marvelde 2007;Ozgul et al. 2010;Charmantier & Gienapp 2014;Morrison et al. 2015;Zimova et al. 2016). However, all of these stages (including breeding) are not necessarily affected in the same way by changes in temperatures (Serreze & Francis 2006;Visser et al. 2006;Visser 2008;Both et al. 2009). Moreover, global temperatures do not change at the same rate in time or space (Easterling et al. 1997;Vose et al. 2005;Serreze & Francis 2006;Stocker et al. 2013), so it is possible that within the same population, different parts of the annual cycle change at different rates (Crozier et al. 2008). In this scenario, we can expect a mistiming among annual cycle stages, with increased or decreased intervals between them. Therefore, climate change will also alter the time constraints across the annual cycle. This is particularly problematic for organisms with complex annual cycles since they are supposedly more time constrained than organisms with simpler cycles (Jacobs & Wingfield 2000;Wingfield 2008).
Analyses showing how intervals between annual cycle stages are changing are rare in comparison to studies dealing with changes in single stages. To detect such patterns it is necessary to collect long-term data of not only one, but multiple annual cycle stages within a population. It has been reported that some species gained more time in key intervals. For example, marmots (Marmota flaviventris) advanced both the termination of hibernation and weaning, resulting in more time for growth and fattening for their offspring (Ozgul et al. 2010). In red deer (Cervus elaphus), on the other hand, reproductive stages of males and females unequally advanced suggesting that males are unable to track advancements in oestrus of females. Moreover, the termination dates of males' rut advanced more than the initiation dates, which resulted in a shorter breeding window and likely reduced breeding performance (Moyes et al. 2011). Barnacle geese (Branta leucopsis) that rapidly expanded their breeding range to temperate areas more than doubled the interval between breeding and moult, resulting in changed body mass dynamics (Van der Jeugd et al. 2009;Eichhorn et al. 2010). Finally, long-term data on date of arrival and breeding of migratory birds suggest that timing of breeding and migration may not respond the same way to climate change. There are reports of both shorter (Both & Visser 2001) and longer (Ahola et al. 2004;Valtonen et al. 2016) intervals between arrival and breeding. Thus, for animals with more complex annual cycles, the assessment of the impacts of climate change becomes considerably more challenging, because unequal changes in the timing of stages may impose fitness costs (e.g., reduce the interval between moult and breeding causing more overlap; Hemborg & Lundberg 1998).
Here, we analysed how climate change affects the timing of different annual cycle stages of a long-distance migratory bird, the European pied flycatcher. We used long-term data  collected from a Dutch population of flycatchers, looking at three important annual cycle stages: spring migration (arrival dates), breeding (egg-laying and hatching dates) and post-breeding moult onset. Furthermore, we included initial information on the variability in timing of autumn migration. Apart from describing the changes in patterns through time, we also investigated the potential environmental causes and fitness consequences of the variation in timing of these important stages.

Study species and study area
Pied flycatchers (Ficedula hypoleuca ([Pallas], 1764); Muscicapidae) are long-distance migratory birds that reproduce in Europe and winter in West Africa (Ivory Coast in the case of our study populations). These birds readily accept artificial nest boxes and have low nest desertion rates, allowing the precise monitoring of their reproduction. We used long-term data from the breeding population of the forests of the Hoge Veluwe National Park (The Netherlands; 5°51'E 52°02'N). Forested areas in the park are dominated by pedunculate oaks (Quercus robur), northern red oaks (Quercus rubra), Scots pines (Pinus sylvestris), Larches (Larix spp.) and birches (Betula spp.). For more than 60 years we provided nest boxes year-round in an area of 171 ha, nest boxes were occupied in spring by cavity-nesting passerines such as pied flycatchers, great tits (Parus major), blue tits (Cyanistes caeruleus), nuthatches (Sitta europaea) and coal tits (Periparus ater). Voucher material of this pied flycatcher population was deposited in the ornithology collection of the Naturalis Biodiversity Center (Leiden, The Netherlands) under the inventory numbers RMNH 592347, RMNH 592348 and RMNH 592349.

Data collection
Data on timing of breeding (egg-laying, chick hatching) was collected for this pied flycatcher population since 1959, timing of migration (arrival dates and, more recently, departure time) and timing of moult since 2005. We also obtained 35 years of arrival data of pied flycatchers breeding in a nearby location (from the Vogelwerkgroep Arnhem, a similar dataset as the one used in Both & Visser 2001). For our final dataset, we used 35 years of egg-laying, chick hatching dates and female arrival dates (to match the number of years available for arrival dates from the Arnhem dataset), 11 years of male arrival dates collected on our study population, 10 years of moult onset dates and 3 years of departure dates and arrival at the wintering grounds.

a) Timing of breeding
Data for timing of breeding was available for 1959 (when the first pied flycatchers started to breed in our study area) and then from 1962 to 2015, but as mentioned above, we used data only from 1980 until 2015 because male arrival data (Arnhem) was available from that year onwards. Every year nest boxes were checked weekly from early March until late July and information on the progression of nest building and on the date when first eggs were laid was collected. Since eggs are laid in one-day intervals and pied flycatcher clutches typically have six eggs, we had accurate measurements of individual laying dates. Pied flycatcher females typically incubate for 14 days, usually starting at the date when the last egg is laid, thus, after around 13 days of incubation, nests were inspected daily to identify the hatching dates of chicks. In some cases when we missed the actual hatching date, the date could be determined by ageing the chicks based on size and plumage development. When nestlings were seven days old, they were identified with a uniquely numbered aluminium ring and parents were also caught and identified.

b) Arrival dates
From 2005 until 2015, individual arrival date of males was assessed by daily scoring newly arriving males in our study area from early April onwards (Potti 1998;Visser et al. 2015;Both et al. 2016). Birds choose a territory upon arrival and advertise their cavity or nest box to the females by singing continuously at or close to the potential nest site. Two or three trained observers walked independently pre-established routes covering the whole study area and visiting all boxes. Routes and direction of the routes were alternated daily among observers in order to prevent any potential bias among them. Detected birds were described in terms of plumage and aluminium/colour ring combinations. Male pied flycatchers display relatively large individual variation in plumage characteristic which, combined with colour ring combinations, allows an initial recognition in the field without the need of capturing the birds. In our study site, plumage coloration varied from female-brown or light grey to almost entirely black with intermediates of increasing blackness (Drost 1936). The forehead white patch also varied in size, from absent to a large patch covering most of the forehead, and also in shape from two distinct dots to a rectangular-shaped patch. We associated singing males to the closest nest-box in the vicinity. During the chick-rearing phase, those males were caught and described again in terms of plumage characteristics and ring combination. In most years (except 2005 and 2012) we also collected data on "bachelors", i.e., males still singing one week after the first eggs were found. These latter males were captured, identified and blood sampled and a few became breeding birds later in the season.
Apart from this more detailed dataset on individual arrivals, we also used a second dataset from the Vogelwerkgroep Arnhem, a bird observation group that collected data on arrival dates of male pied flycatchers from 1980 to 2015, allowing the calculation of much longer term temporal shifts. This data collection occurred around 10 km from our study area and there was a high correlation (Pearson = 0.83, n = 11 years; Fig. S3.1) between their mean arrival data and the mean arrival dates calculated based on the individual arrival dates for the Hoge Veluwe. Therefore we are confident that this is a reliable and representative dataset for longer term trends in the studied pied flycatcher population.
Female individual arrival date was not obtained directly as in the males' case. Instead, we used data from the start of nest building as a proxy, following the same procedure of Visser et al. (2015). Female pied flycatcher nest building reflects well their arrival dates since they choose a male and start nest building shortly after arriving (Dale et al. 1992;Dale & Slagsvold 1995). Nest building of individual pied flycatcher females was collected in our study area from 1980 to 2015, allowing the observation of long-term shifts in timing of arrival of females.

c) Timing of moult
From 2005 to 2015 we had information on whether birds were or were not moulting when they were caught for identification (when they had 7 days-old chicks). From 2009 to 2015 we also had information on moult score, i.e., which feather was dropped and how much its replacement feather had grown. Finally, from 2013 to 2015, we had the actual date when the first feather was dropped for most of the breeding birds (Chapter 6). We could thus use the latter to define the accuracy of each type of measurement. We only used male data for this, since for most years we only had data on one or two moulting females. In 2012 no information on moult was collected and thus we could not use this year in our analysis.
We used the R "moult" package (Erni et al. 2013) to calculate the population average moult onset based on 2 analyses: (1) using presence/absence data, we calculated the average starting date with a "probit model", a generalised linear model with a binomial distribution and probit link function (Erni et al. 2013).
(2) Using the moult scores of different individuals (based on which feather was missing or re-growing and how much it had grown) converted to a value of new feather mass grown (Dawson & Newton 2004;Erni et al. 2013) and regressing a line through the values of all individual new feather mass grown to obtain the population moult onset per year. When we compared these values with the averages of the actual observed individual moult onsets, we noticed that the first presence/absence model provided a better estimate of moult onset than the feather mass regression (2013: observed = 76.93, presence/absence model = 73.09, feather mass regression = 59.22; 2014: observed = 71.70, presence/absence model = 66.65, feather mass regression = 61.81; 2015: observed = 74.42, presence/absence model = 71.14, feather mass regression = 62.81). This was probably due to the fact that moult scores were mostly collected earlier in the season, with very few individuals with larger moult scores, making the linear regression less reliable. It is also important to mention that we could use this "probit model" because we have experimental data showing that the male moult onset is independent of their termination of breeding, occurring more or less on the same calendar day in a given year for most individuals (Chapter 6). Thus, it is not so problematic that only mid to late breeding birds were caught later in the season for moult scorings (as captures happen in a standardised way when chicks are 7 days old), and we can extrapolate their onset to the rest of the population. We could then use 10 years of moult onset data based on presence/absence of moulting birds. Geolocators were recovered in the subsequent years (2014 to 2016) when birds returned for breeding and 26 tracks (out of 98, 26%) could be analysed (12 from 2013, 7 from 2014 and 7 from 2015). Data was processed similarly to Åkesson et al. (2012), but due to the imprecision of latitude data, we only used the information collected for longitude (inferred from local solar noon/midnight). Twilight transitions were determined using TransEdit (British Antarctic Survey, Cambridge) with a single threshold value of five, minimum daylight periods of one hour and minimum night period of four hours. Positions were obtained using the software BirdTracker, which gave us two positions per day (noon and midnight). Data was then visually inspected to detect large changes from the study area longitude, indicating a departure from breeding grounds and then arrival at the wintering grounds, since pied flycatchers move to the west, following the African coast during autumn migration.

d) Departure dates and arrival at African wintering grounds
If the logger was still working upon recapture and a full track could be downloaded, we used the data corrected for clock drift, otherwise, clock drift effects on longitude were tested as described in Ouwehand et al. (2016). There was no noticeable clock drift effect (always <1 min).

e) Temperatures
Daily temperature values were collected from the Dutch meteorological institute database (KNMI -https://www.knmi.nl/nederland-nu/klimatologie/, accessed February 2016) for the Dutch temperatures and from the US National Oceanic and Atmospheric Administration database (NOAA -ftp://ftp.ncdc.noaa.gov/pub/data/gsod/, accessed February 2016) for African temperatures. Data from the NOAA database was converted to match the KNMI database so all values in Fahrenheit were converted to Celsius for temperatures.
For the Dutch weather variables, we used data from the Deelen weather station which is directly adjacent to the study area. For African temperature data, we used information on the pied flycatchers' wintering location using the data from the geolocators that we deployed and also from the literature (Ouwehand et al. 2016) to identify the closest weather station from their wintering grounds. Pied flycatchers from the Netherlands winter in the Ivory Coast, where Daloa (6°27'W 6°53'N) is the closest weather station with a reasonable amount of data. Because the dataset still had large gaps we also used information from two other nearby locations (Gagnoa, 5°56'W 6°08'N, and Yammousoukro, 5°17'W 6°49'N). We still ended up with a few gaps in the data, that we interpolated using the average of data at the boundaries of the gap. This was seen as a minimal issue due to the way this data is used in the statistical analysis (see below). We obtained complete data from 1980 until 2015 for Dutch temperatures.
We also obtained data on photoperiodic variation of the Netherlands from the NOAA. We considered the civil twilight as the boundary of the effective light phase important for the birds (Gwinner 1989). Because photoperiodic variation is completely correlated throughout all locations of the Earth, it was not necessary to obtain and model day length data for Africa separately.

f) Adult survival
We used two datasets with data on individual capture histories; one with 11 years for which we had data on the average interval between breeding and moult for each year (2005 to 2015, 1252 individuals) and a second one with 35 years for which we had data on the average interval between arrival and breeding for each year (1980 to 2015, 3887 individuals). Individuals were not included in the analyses when nestlings, but only when breeding for the first time. Thus, we did not include the nestling survival in this analysis. We used these datasets to estimate adult survival and recapture probability and whether the change in the two intervals had an effect on adult survival, while taking into account effects of sex, capture occasion (two categories: first or later) and age at first capture (three categories: (i) second calendar year birds (hereafter SY) ringed as nestling in the previous year; (ii) after second calendar year birds (ASY), ringed as nestling two or more years prior to the capture; and (iii) after first calendar year but otherwise of unknown age (Unknown) for birds that were not ringed as nestlings and consequently their age could not be reliably determined).

Data analysis
All analyses were performed in R version 3.2.1 (R Core Team 2015). To define the minimal models, we always used backwards model selection, dropping non-significant terms in each step. Survival analyses were performed with program MARK (White & Burnham 1999).

a) Shifts in timing
We used simple multiple regressions fitting year as linear and quadratic terms to test for shifts in timing through time for the average value of each annual cycle component (including both longer-Arnhem and shorter-Hoge Veluwe datasets for arrival dates). Then to test whether there has been any change in the amount of time available for arrival and moult in relation to breeding we calculated the differences between arrival and breeding (arrival date of males and females and egg-laying dates) and between chick fledging and moult. Chick fledging was calculated from egg-laying dates by adding 6 days of egg-laying, 12 days of incubation and 15 days of chick care until fledging.
We also tested whether the different slopes that we obtained from the separate regressions were significantly different from each other. We used multiple regression analyses and fitted date against the interaction of annual cycle stage (arrival, moult, egg-laying and hatching dates) and year (linear and quadratic). We ran two separate analyses, one including all stages and a second one only including those stages that we had data since 1980 (arrival data Arnhem, egg-laying and hatching dates).

b) Causes of variation in timing
We tested whether the variation of each of the annual cycle's components (with the exception of hatching dates, which depend mostly on egg-laying dates) could be explained by variation in temperature cues alone or in interaction with changes in day length. Following the method described in Gienapp et al. (2005), we used proportional hazard models (Cox 1992) implemented in the R "survival" package (Therneau 2015) to model the relationship between the climatic event and the occurrence of the event. Proportional hazard models calculate the daily probability of an event to occur. They, therefore, allow including time-dependent variables, i.e. variables that change their value during the time an individual is "at risk". Modelling effects of weather variables on annual cycle stages, as arrival time, is biologically more realistic than using fixed time windows over which these variables are averaged (Gienapp et al. 2005). The value of this time-dependent weather variable at day t was calculated as the average over periods of various lengths (5 to 30 days) ending at day t. For African weather variables, we also used lagged shifting windows of the same length (of 20 days) but ending 20 to 80 days before day t. We defined intervals that we deemed as biologically significant, thus African temperatures were only tested for arrival dates because for egg-laying and moult it is unlikely that African temperatures prior to these events would have affected them as the birds were already in the Netherlands. Therefore, in total, we compared 24 possible combinations for arrival dates, 5 for egg-laying dates and 5 for moult (Table S3.3). After the best window was defined for each stage, we tested what temperatures significantly explained the variation in timing of different stages using different proportional hazard models for each annual cycle stage. We fitted African temperatures (with and without lag), Dutch temperatures and the interactions of day length and Dutch temperature and day length and lagged African temperature, depending on the stage. The temperatures selected in the best models for each stage were also fitted against year as a trend (linear and quadratic) to test whether these temperatures also changed through time.

c) Consequences of variation in timing
We used multiple regressions to test whether the time intervals between arrival and breeding and breeding and moult explained fitness components. We looked at two components related to breeding success (the average clutch size and the proportion of fledged chicks that recruited per year) and at adult survival.
We used generalised linear models with binomial (Bernoulli) response and logit link function to test whether the proportion of fledged chicks that recruited was explained by either the annual difference in time between arrival and breeding or breeding and moult. For the interval between arrival and breeding, we only used the longer (Arnhem) dataset, since it correlated with the shorter (Hoge Veluwe) dataset, showing a similar pattern, but included many more years. Clutch size was similarly analysed using multiple regressions testing for year and interval effects.
To analyse adult survival we performed a Cormack-Jolly-Seber (CJS) mark-recapture analysis in the software MARK (White & Burnham 1999). The CJS model estimates annual local survival probabilities (Φ) based on live recaptures only while controlling for capture probability (P). In this analysis, we used Akaike's Information Criterion (AICc) for the model selection and goodness of fit was tested using the bootstrap procedure in the MARK software. We first defined the best model including time (year as a factor), capture occasion (first or later), sex (male or female) and age at first capture (SY, ASY or unknown), for both survival and recapture probabilities. In the most complex models, birds that were in their SY or of unknown age at their first capture moved to the ASY age class in the following year and remained in this age class for the rest of the years. Birds first caught in their ASY never moved to another age class in subsequent captures. We could not use plumage characters as ageing criteria because they were not deemed as precise enough and also not always recorded for all captured individuals. On the other hand, in the simplified models that did not include the age at first capture, the capture occasion variable only explained differences between the first or later captures of the same individual independent on their age. We first fitted the model with the interaction between time, capture occasion and the interaction between sex and age at first capture and then used backwards comparison to define the best model. After the best model was defined we replaced the time variable by the interval between arrival and breeding (longer dataset) or breeding and moult (shorter dataset) and compared these new models with the best model to investigate specifically whether variation in survival among years could, in fact, be explained by variation in these intervals.
Because arrival date, egg-laying dates and moult dates shifted unequally, we calculated the interval between each stage and tested whether they changed over time. As expected from the previous analyses, the intervals between male arrival and breeding and between breeding and male moult changed over time. In contrast, the interval between female arrival and egg-laying dates did not significantly change since 1980 (Table S3.2, Fig.  3.1e). The interval between male arrival and egg-laying date changed non-linearly: it became shorter until around 2008 and then started to increase again (estimate for the quadratic term = 0.02 ±0.01, F 1,33 = 8.70, p-value <0.01; Table S3.2). A post hoc broken stick analysis suggested that two separate regressions, one before 2008 and one after 2008, had a better fit than the quadratic model (adjusted R² for quadratic term = 0.38, adjusted R² for two regressions before and after 2008 = 0.43). Dividing the dataset before and after this year and testing separately we found that the first regression was significant (slope = -0.45 ±0.09, F 1,26 = 22.62, p-value <0.01), but the second was not (F 1,6 = 3.93, p-value = 0.09) (Fig. 3.1e). In contrast, even within this short time, we observed that the interval between calculated fledging date (termination of breeding) and date of male moult onset significantly decreased over time. As a consequence, there was a larger overlap between moult and breeding in recent years (slope = -1.19 ±0.24, F 1,8 = 23.64, p-value <0.01; Table S3.2, Fig. 3.1e).

Causes of variation in timing
Variation in male arrival date was significantly explained by Dutch temperatures in both the Hoge Veluwe and Arnhem datasets, with higher temperatures in the Netherlands being related to earlier male arrival (Arnhem: coefficient = 0.31, χ² = 8.40, p-value <0.01;  the interaction = -0.01, coefficient of the temperature as main term = 5.10, χ² = 23.36, p-value <0.01; Table 3.1, Table S3.4) suggesting that this temperature effect gets weaker later in the season. Moreover, African temperatures immediately prior to arrival were also significantly related to arrival, but in the opposite direction (coefficient = -0.21, χ² = 9.72, p-value <0.01; Table 3.1, Table S3.4). None of these temperatures significantly changed since 1980 (Table S3.5, Fig. S3.2).  In both cases, higher temperatures were related to earlier arrival. Dutch temperatures correlating with female arrival changed non-linearly, with an increase until around 2000 and then a decrease in recent years (quadratic estimate = -0.01 ±0.001, F 1,33 = 9.85, p-value <0.01; Table S3.5, Fig. S3.2). A post hoc broken stick analysis suggested that this model with the quadratic term had a better fit than any two regressions.
Variation in egg-laying date was similarly explained by Dutch temperatures in interaction with day length, with earlier laying at higher temperatures (coefficient of the interaction = -0.005, main term for temperature = 4.73, χ² = 220.58, p-value <0.01; Tables 3.1 and S3.4), however this temperature did not change through time (Table S3.5, Fig. S3.2).
Finally, variation in male moult onset was not related to Dutch temperatures (Tables 3.1 and S3.4) thus the advancement observed for the moult onset cannot be related to any recent increase in Dutch temperatures.

Consequences of variation in timing
There was no effect of the interval between arrival and breeding and breeding and moult on clutch size. Clutch size, however changed through time since 1980 in a non-linear way, increasing until around 2003 and later on decreasing again (quadratic estimate = -0.001 ±0.0005, F 1,33 = 9.55, p-value <0.01; Table S3.6). A post hoc broken stick analysis suggests that two regressions before and after 2008 have a better fit than the model with the quadratic term (adjusted R² for quadratic term = 0.40, adjusted R² for two regressions before and after 2008 = 0.48). Dividing the dataset before and after this year and testing them separately we found that the first regression was significant (slope = 0.03 ±0.01, F 1,26 = 17.43, p-value <0.01), but the second was not (F 1,6 = 4.90, p-value = 0.07) (Fig. 3.2a).
The proportion of fledged chicks that recruited also changed non-linearly in time (quadratic estimate = -0.001 ±0.0002, χ²= 10.49, p-value <0.01; Fig.3.2b, Table S3.6), but it was also significantly related to the interval between arrival and breeding. A higher proportion of chicks fledged when this interval was shorter (slope = -0.02 ±0.01, χ² = 5.30, p-value = 0.02; Fig. 3.2c, Table S3.6). A shorter interval is also related to earlier egg-laying dates of the females, and earlier laying was significantly related to a higher proportion of fledged chicks that recruited (slope = -0.03 ±0.02, χ² = 4.04, p-value = 0.04; Fig. 3.2d, Table S3.6). Thus the improved chick recruitment may be related to the earlier breeding.
Neither the interval between arrival and breeding or breeding and moult explained the variation in adult survival. When analysing adult survival probability (Φ), none of the models including the interval between arrival and breeding and almost none containing the interval between breeding and moult was among the best models (delta AICc always larger than 2; Table 3.2). Even those models containing the breeding/moult interval, with a delta AICc smaller than 2, however, had negligible effects of this interval (beta estimate in the model only including that interval = 0.0001 ±0.0008).

Discussion
Our results show that climate change affected the annual cycle stages of the European pied flycatcher, making them shift at different rates from one another. This was probably due to the different changes in the various temperatures that correlate with the phenology of these annual cycle stages. An exception was moult, for which we could not detect a correlating temperature period. Our results also suggest that the advancement of breeding allows more time for events such as chick development, which could explain at least in part the increase in the proportion of fledged chicks that recruited when the breeding -moult interval got longer. The shortening of the interval between arrival and  breeding and the larger moult/breeding overlap would be expected to incur fitness costs, but we did not detect effects on adult survival. Therefore, at least for the aspects studied so far, climate change has actually lead to an improvement in breeding conditions for this species, potentially by allowing more time for chick development (Chapter 8).

Other consequences of unequal shifts
While costs of shortening of stages could not be detected in terms of adult survival, it is still possible that costs are present in other aspects we have not investigated. Females, for instance, may suffer costs that are much more subtle. For example, a decrease in the time between arrival and breeding may also mean that females need to be much faster in choosing a male (Dale et al. 1992;Dale & Slagsvold 1995), building the nest and laying eggs. Indeed, in recent years male and female arrival are happening almost at the same moment, so when the first females arrive, a good number of males is not settled yet and territories are still being claimed (Alatalo et al. 1984). This could, for example, result in a reduced time to assess male quality (Alatalo et al. 1984) and increase the probability of a female to pair with a male that already has another female. Polygyny is costly for females since the number of unfertilised eggs and chick mortality is higher when a male has more than one female (Lubjuhn et al. 2000). This could be problematic for late-arriving females, as they would supposedly be even more constrained by the increased competition with other females late in the season (Dale et al. 1992). This shorter time span could also decrease the females' body condition and in the long run be detrimental to their breeding success, since egg-laying is particularly costly (Visser & Lessells 2001). This may have been reflected in the previously increasing clutch size trend, which changed in recent years (although there was no significant decline detected by the broken stick analysis) (Fig. 3.2a).
The shifts observed in the timing of moult in males may also result in a shortening of the total time available for breeding, similarly to what was reported by Moyes et al. (2011). Males advanced their timing of moult at a higher rate than the breeding time, which suggests they are regressing their gonads earlier as well. Experimental studies in captive starlings (Sturnus vulgaris) show that onset of moult is related to the gonadal regression in males (Dawson 2006); moreover, both gonadal regression and moult advanced in great tits (Parus major) experimentally exposed to higher spring temperatures in captivity (Visser et al. 2011). In the present study, there was no effect of temperatures on moult onset, but regardless of the factor that is causing advancements in the timing of moult, an uncoupling between onset and termination of breeding is possible.

Underlying causes of unequal shifts
A curious outcome of our analysis is the faster advancement of moult in relation to breeding. Moult onset in males seems to be determined earlier in the season and not determined by the termination of breeding (Chapter 6). If moult is not connected to timing of breeding, it could advance independently. Another possibility is that moult is set when the individual is born. It is known, for example, that the photoperiod when the animal is born can modify the timing of events (Lee & Zucker 1988;Coppack et al. 2001;Coppack & Pulido 2009). Thus, it has been proposed that advancements in birth date (as laying dates advance) could modify the average timing of events at the population level in particular for timing of migration (Both 2010;Gill et al. 2014). If the timing of moult is set when the birds are born, this could explain the change in moult onset at the population level (but see Larsson 1996).
If the dates when birds are born affect both their arrival and moult, we should also observe an advancement in arrival dates, but it was not the case. Other migratory species, however, were reported to advance their arrival dates (Usui et al. 2016) and this was the case for different populations of pied flycatchers as well (Ahola et al. 2004;Both et al. 2016;Valtonen et al. 2016). Arrival dates, in comparison to the onset of moult, are much more susceptible to modulations by environmental conditions, such as weather, body condition and wind patterns (Ahola et al. 2004;Erni et al. 2005;Sinelschikova et al. 2007;Bauer et al. 2008;Eikenaar & Schmaljohann 2014;Both et al. 2016;Teitelbaum et al. 2016). Moreover, arrival dates can usually only be assigned to individuals that survived the migration and thus there might be a bias if early arrival increases mortality (Brown & Brown 2000). In comparison to arrival, moult is a relatively cleaner expression of the individuals (endogenous) timing (Gwinner 1996), thus potential changes related to birth effects are supposedly more detectable in timing of moult than timing of arrival.
In terms of causes of the observed differences, it is worth noting that none of the temperatures important for timing of stages actually changed over the past years, even those correlating with egg-laying dates. The exception was the Dutch temperatures that correlated with female arrival (Fig. S3.2). It is important to remember that our measure of arrival time of females actually corresponds to nest building dates, which are closely related to egg-laying dates. The important Dutch temperatures for female arrival and breeding are almost the same and partly overlap (Table 3.1) as the average arrival is 10 days earlier than the average laying date, but the window size of the temperature important for laying date is 10 days larger. One possibility is that the temperatures important for female arrival (nest building) and egg-laying are nearly the same and trends are not detected for the egg-laying temperatures because they are more variable across years (as is noticeable in Fig. S3.2). Alternatively, the advancements in timing of breeding are actually driven by the temperature effects on arrival date (nest building decisions), that highly correlate with laying dates and not by the temperature effects on laying dates per se.
In the present study, while we included multiple annual cycle stages, we were limited to the stages that occur on the breeding grounds -for which we have long-term data available. Even so, there was one stage at the breeding ground that we could not include: the timing of autumn migration. Date of departure in late summer seems to be correlated with the timing of egg-laying, but not so much with the timing of chick hatching (Chapter 6). Departure dates also seem to be variable across years. Thus, it is possible that, together with the advancement of breeding and moult, the birds are now also departing earlier. On the other hand, if there is no change in timing of migration, males have more time to moult. It has been reported that conditions late in the season may improve for some birds, such as for some short-distance migrants (Jenni & Kery 2003). For now, how climate change has affected the timing of autumn migration in our population must remain speculative.
Another aspect that we are unable to assess is the timing of stages at the wintering grounds in Africa. It is unknown if birds remain time constrained at their wintering grounds or if the wintering grounds serve as "time buffers". Early arrival at the wintering grounds may, for example, be important to secure resources, which could result in improved body reserves that carry-over to the next season (for example in terms of earlier arrival and/or improved individual quality). This is the case for American redstarts (Setophaga ruticilla;Marra et al. 1998Marra et al. , 2015aNorris et al. 2004). In pied flycatchers, an experimental delay of hatching dates imposed a larger moult-breeding overlap and made males from this group winter at a different location than Controls and Advanced males (Chapter 6). This suggests that the selection of wintering territory also depends on what birds experience at the breeding grounds, but whether wintering will later impact breeding remains unknown (but see Ouwehand & Both 2017).

Conclusions
Climate change unequally affects the annual cycle stages of bird (Van der Jeugd et al. 2009;Eichhorn et al. 2010;Valtonen et al. 2016) and mammal species (Ozgul et al. 2010;Moyes et al. 2011). Such shifts may lead to positive fitness consequences in some cases, for example, in marmots` offspring that gain more time. But it may also have negative consequences, in the case of the red deer in which not only a mismatch between male and female timing happened, but also a shortening in their breeding window. In the present study, there were also positive and negative fitness consequences of unequal shifts, but they depended on the sex or life stage of the animal. In pied flycatchers, climate change seems to benefit males and offspring which gain time due to the advancement of breeding, but will potentially be costly for the females. The impacts of climate change, thus, are not only different for distinct trophic levels (Parmesan 2006), but also for stages (e.g., breeding, moult) and individuals (e.g., males, females, offspring) within the same species. It is well possible that such patterns are widespread, especially among organisms with complex annual cycles, meriting a more careful look.
For a broader understanding of the ecological consequences of climate change, different stages of the annual cycle should be considered, in particular for organisms with complex cycles, such as migratory birds (Small-Lorenz et al. 2013;Marra et al. 2015b). Our knowledge on climate change impacts on organisms will thus greatly benefit from continued standardised data collection that includes more than one stage.        Table S3.6. Model results for the simple and multiple regression analyses testing effects of the linear and quadratic year and intervals (arrival and breeding, breeding and moult) on clutch size and proportion of fledged chicks that recruited. Statistics are given for each term at the point of exclusion of the term from the model. Estimates and standard errors are presented only for significant terms.