1. In structured populations, phenotypic change can result from changes throughout an individual’s lifetime (phenotypic plasticity, age-related changes), selection and changes in population composition (environment- or density-driven fluctuations in age-structure).
2. The contribution of population dynamics to phenotypic change has often been ignored. However, for understanding trait dynamics, it is important to identify both the individual- and population-level mechanisms responsible for trait change, because they potentially reinforce or counteract each other.
3. We use 22 years of field data to investigate the dynamics of a sexually selected phenological trait, the timing of nuptial moult in superb fairy-wrens Malurus cyaneus.
4. We show that trait expression is both climate- and age-dependent, but that phenotypic plasticity in response to climate variability also varies with age. Old males can acquire nuptial plumage very early after high rainfall, but 1- to 2-year-olds cannot. However, males of all ages that defer moult to later in the year acquire nuptial plumage earlier when conditions are warmer.
5. The underlying mechanism appears to be that old males may risk moulting in the most challenging period of the year: in autumn, when drought restricts food abundance and during the cold winter. By contrast, young males always moult during the spring transition to benign – warmer and generally wetter – conditions. Temperature changes dominate this transition that heralds the breeding season, thereby causing both young and late-moulting older birds to be temperature sensitive.
6. Climate and age also affect trait dynamics via a population dynamical pathway. The same high rainfall that triggers early moulting in old males concurrently increases offspring recruitment and thereby reduces the average age of males in the population. Consequently, effects of rainfall on trait dynamics through phenotypic plasticity of old males are dampened by synchronous rejuvenation of the age-structure.
7. A long-term trend towards drier environments prompted phenotypic change because of plasticity, but this was masked by climate-driven demographic change (causing apparent stasis). This suggests a novel explanation for why trait change may fail to reflect the observed pattern of directional selection or phenotypic plasticity.
A likely reason for our sometimes limited understanding of trait dynamics is that many different mechanisms can cause the mean trait value of a population to change over time, which makes it difficult to identify them all and quantify their contribution to phenotypic change (Coulson & Tuljapurkar 2008; Ozgul et al. 2009, 2010). This problem is particularly pervasive when studying trait dynamics in species with age- or stage-structured life cycles (i.e. most iteroparous species with overlapping generations), because in structured populations, phenotypic change in labile traits can occur because of three types of mechanisms, which each can take various forms (Charlesworth 1994; Coulson & Tuljapurkar 2008).
First, trait expression can change during the lifetime of an individual, resulting from the processes of age-related improvement, growth or senescence as well as from plastic responses of an individual’s phenotype to a variable environment (phenotypic plasticity, Pigliucci 2001; Nussey, Wilson & Brommer 2007). Secondly, the selective (dis)appearance of certain phenotypes because of selection (Vaupel, Manton & Stallard 1979) or dispersal (Postma & van Noordwijk 2005) can alter the population’s trait distribution. For example, viability selection can remove specific phenotypes, and fecundity selection can cause the offspring of certain phenotypes to be overrepresented in the next generation. Fecundity selection, however, primarily results in trait change when there is inheritance of trait values from parent to offspring, implying that the mode of inheritance and genetic architecture can further influence a population’s trait dynamics (Merilä, Sheldon & Kruuk 2001).
Thirdly, variability in demographic rates (survival, reproduction and dispersal) because of environmental stochasticity or density dependence can cause the age-structure of the population to fluctuate over time (an equilibrium age-distribution is never reached; e.g. Engen, Lande & Saether 2009). When trait expression depends on the age of an individual, such fluctuations in age-structure will also result in a change in the population’s trait distribution. Note that we do not refer to cases where the relative fitness of a phenotype depends on the population composition (King & Anderson 1971; Charlesworth & Giesel 1972), as we consider those to be cases of frequency- or density-dependent selection. This third mechanism is not restricted to age-dependent processes, because similar mechanisms can operate when trait expression depends on any state variable (e.g. life stage, social status, body size) and environmental, density- or frequency-dependent processes directly affect the distribution of such a state variable in the population.
Historically, both theory and empirical studies investigating trait change have focussed on the mechanisms of selection (Falconer 1960; Lande & Arnold 1983) and on age-related changes and phenotypic plasticity (Charlesworth 1994; de Jong & Bijma 2002), while the contribution of fluctuations in the population composition to trait change has been overwhelmingly ignored (but see Coulson & Tuljapurkar 2008; Engen, Lande & Saether 2009). Population-level processes have possibly been overlooked because the classical theory on evolution in age-structured populations generally assumes equilibrium population dynamics (Lande 1982; Charlesworth 1994). However, for understanding trait dynamics, it is important to identify both the individual- and population-level mechanisms responsible for trait change, because they potentially reinforce each other, resulting in an exaggerated effect, or counteract each other, resulting in dampening or apparent stasis. By also considering the population dynamical contributions to trait change, we thus may be better able to understand why phenotypic change may fail to reflect the observed pattern of directional selection or phenotypic plasticity, and we can start to identify the conditions under which this may occur.
Here, we use 22 years of field data collected in a rapidly changing Australian environment to explore the trait dynamics of a sexually selected phenological trait, the timing of nuptial moult in male superb fairy-wrens (Malurus cyaneus). The timing of expression of this labile trait was previously shown to be highly plastic within individuals, dependent on both their age and the prevailing climatic conditions (Mulder & Magrath 1994; Cockburn, Osmond & Double 2008; van de Pol & Cockburn 2011). We now investigate in detail the environmental factors determining this phenotypic plasticity and quantify how plasticity changes with age. Subsequently, we investigate what causes fluctuations in the age-structure and how this in turn affects the population’s trait dynamics. We show that climate-induced fluctuations in population composition not only produce additional stochasticity in trait dynamics, but that climatic conditions can cause a negative covariance between the fluctuations in the population’s age-structure and individual’s plastic responses, which dampens the impact of climate change on phenotypic change. We consider whether this novel mechanism may explain why trait change often fails to reflect the observed pattern of phenotypic plasticity or directional selection.
Materials and methods
Trait data collection and study system
Superb fairy-wrens are sexually dimorphic birds in which the males can moult from a female-like brown eclipse plumage to a blue nuptial plumage, as characterized by a full glossy light-blue crown, ear coverts and mantle, as well as starkly contrasting dark lores, back and throat (Mulder & Magrath 1994). In contrast to the post-breeding moult that both sexes go through, the nuptial moult is thought to be a sexually selected male signal; the timing of nuptial moult predicts male extra-pair mating success, which dominates reproductive success in this species (Dunn & Cockburn 1999; Cockburn, Osmond & Double 2008).
From 1986 on, a population of superb fairy-wrens was marked with unique combinations of colour bands as part of a long-term demographic study in Canberra, Australia (35°16′S, 149°6′E). Moult data were collected during year-round weekly censuses from 1988 to 2009 (n =22 years, 1465 males, 3530 male years, 88 979 weekly moult records). Superb fairy-wrens live in open habitat and are territorial year-round and easy to approach. At least once a week, each male was observed with binoculars (weekly resighting probability >0·95) to determine whether he was still in eclipse plumage (score 5), just started moulting (score 4), moulted half of his feathers (score 3), moulted most of his feathers (score 2) or fully completed the nuptial moult (score 1; Mulder & Magrath 1994). The start (completion) date of nuptial moult was calculated as the midpoint between the last day a bird was scored with moult score 5 (score 2), and the first day, it was observed with score 4 (score 1). The duration of moult – the period between the start and completion – typically spans 2–3 weeks.
Occasionally, we had to extrapolate the completion date of moult. In 0·5% of all cases, the male died during moulting, and we estimated their date of completion of moult based on their last stage of moulting and the time it usually takes to complete moult from that stage (moult duration is quite invariable, especially compared to the huge individual heterogeneity in timing of moulting; see Results). Furthermore, in 1·5% of the cases, males moulted directly from their nuptial plumage of the previous year into a new nuptial plumage during the post-breeding moult; their completion date was estimated to be 1 March (2 days earlier than the first male to acquire nuptial plumage after having started to moult into eclipse plumage).
Males are extremely philopatric; 98% of males die on their natal or neighbouring territory (Cockburn et al. 2008a), and most males were followed during their lifetime. Thus, the potential for selective dispersal is low. The vast majority of males (89%) were of known age, because colour banding started 2 years before we started collecting moult data, and the known-age fraction increased rapidly during the course of the study. Few males lived longer than 7 years; therefore, we grouped the age-classes of 7 years and older into a single age-class (7+) for sample size reasons. Superb fairy-wrens are cooperative breeders in which the dominant male may be assisted by male helpers. However, we do not explicitly investigate the effect of dominance status on timing of moult here, because dominance status is tightly correlated with age in this species (Cockburn et al. 2008a).
Data on daily weather at Canberra airport (8 km east of the study area) over the period 1940–2010 were obtained from the Australian Bureau of Meteorology; missing precipitation and temperature values (<0·1% of all values) were linearly interpolated. Because it rarely snows or hails in Canberra, we henceforth refer to precipitation as rainfall.
Analyses of phenological traits
The timing of the nuptial moult is highly variable within and between individuals and can occur from March until November (Mulder & Magrath 1994). We used time-to-event models to analyse the climate and age-dependency in the phenological trait moult date (e.g. Fox 1993). Even though most studies have analysed phenological traits similar to quantitative traits using standard linear regression approaches (van de Pol & Cockburn 2011), there are two key advantages in using time-to-event models to analyse phenological traits.
First, in standard regression models, the critical climatic time-window is assumed to be the same for all individuals, which is unrealistic for phenological events that can occur over long periods. For example, March rainfall may be likely to affect moulting rates in April, but unlikely to affect moulting rates in September. In time-to-event models, time-dependent covariates can be used that allow the critical climatic time-window to vary over time (Gienapp, Hemerik & Visser 2005; van de Pol & Cockburn 2011), such that moulting rates throughout the year may depend on the rainfall in for example the preceding month.
Second, time-to-event models specifically allow for the inclusion of individuals that did not express the trait of interest (26% of males died before moulting, and 2% failed to fully complete moult in a given year). Standard linear regression approaches typically exclude individuals with missing response data, which can lead to biased results (Fox 1993) because such data may not be missing completely at random (Nakagawa & Freckleton 2008), which is doubly important where selection is involved (Hadfield 2008).
We used two types of time-to-event models. The timing of moulting during the year was visualized using Kaplan–Meier curves based on product-limit estimators (Kleinbaum & Klein 2005), which account for censoring (e.g. death). We statistically analysed which factors affected the timing of moult using an extended Cox model (Kleinbaum & Klein 2005):
in which the weekly moulting hazard rate depends on a time-dependent baseline hazard (h0), time-dependent weather variables (Irainfall and Itemperature) and a time-independent age-predictor (see van de Pol & Cockburn (2011) for details). Following previous work (Mulder & Magrath 1994; Cockburn, Osmond & Double 2008), we focussed on the weather variables minimum temperature and rainfall, which are hypothesized to affect the condition of these insectivorous birds directly through energetics and indirectly through their effects on insect abundance (van de Pol & Cockburn 2011). We included age as a factor (levels 1, 2, 3, 4, 5, 6, 7+), because moulting time was previously suggested to be age-dependent in a nonlinear way (Mulder & Magrath 1994). Moreover, a gamma distributed shared frailty term was included to account for individual heterogeneity in moulting hazard (individual males moulted in multiple years).
Although time-to-event models allow for time-dependent weather covariates, such that moulting rates throughout the year may depend on the temperature in for example the preceding month, we performed additional analyses to identify what the precise critical climatic time-window is. In other words, should Irainfall in eqn 1 be quantified as the rainfall in the preceding week, preceding month or the rain that fell 2 months ago? We identified the critical climatic time-window that triggers the nuptial moult using a recently developed method that extends existing sliding window approaches (van de Pol & Cockburn 2011). For each week that individuals were at risk of moulting, we calculated a weighted average of a weather variable over the preceding period p, for example:
In contrast to classical sliding window approaches, each of the preceding weeks can contribute differentially to our weighted average of the weather variable according to a weight function wj, which can take many forms and also allows for time-lags. Maximum-likelihood optimization methods were subsequently used to determine which of the many possible weight functions produced the rainfall or minimum temperature variable that best predicts variation in moulting hazard (van de Pol & Cockburn 2011).
In the time-to-event analyses, we excluded the 11% of males whose birth date was unknown, generally because they were adults when the study was started, or more rarely because they moved on to territories on the perimeter of our study area. In addition, by adding a bird’s life span as a covariate to eqn 1, we accounted for potential selective disappearance within cohorts, which assures our age-dependent moulting hazards reflect within-individual changes (van de Pol & Verhulst 2006). Finally, we tested whether age-classes differed in their climate-dependency by including an interaction term between the weather variables and age, in which age was coded as a continuous covariate. Note that a hazard ratio (HR) of one implies no effect (e.g. ) parameter estimates are given with 95% confidence intervals between squared brackets (e.g. HR = 1·01 [0·97,1·05]).
We constructed an asexual population model to project fluctuations in the male age-structure (n1, n2, n3, n4, n5, n6, n7+) as a function of environmental variability. Males born in the 1988 breeding season (running from October 1988 to February 1989) are considered to turn 1 year old at the end of the breeding season (1 March), et cetera. We used a post-breeding-census age-structured matrix model (Caswell 2001) in which male adult survival (S) is constant with age (Cockburn et al. 2008b), and annual productivity (F) was defined as the number of male fledglings produced per year that survives to the age of 1 year old:
where F is modelled as a function of a climatic variable (see Results), such that fecundity varied between years (Fy). Note that Fy equals the annual number of offspring produced per territory per year divided by 1·5, reflecting that there are on average 1·5 males per territory (Cockburn et al. 2008a). The observed sizes of age-classes averaged over the second half of the study period were used as starting values in all simulations, while the historical climatic data (1940–2010) were used to generate annual fecundities (y =0 refers to the year 1940).
Male superb fairy-wrens complete their moult earlier as they age, with 1- to 2-year-old males moulting in spring just before the onset of the breeding season, with 3- to 4-year-olds moulting in both winter and spring, while males over 5 years old mostly moult in autumn and winter (Fig. 1). Ignoring a few pathological exceptions, only 1-year-olds sometimes failed to complete their moult, while only birds older than 3 years moulted directly from their nuptial plumage of the previous year into a new nuptial plumage at the end of the breeding season (week 1; Fig. 1). Consequently, not only the mean time of moulting changes with age, so does the shape (variance and skew) of the trait distribution (Fig. 1). There was little evidence that moult duration was age-dependent (slope −0·5 [−1·3,0·3] days per year, F1,3529 = 1·2, P =0·27), consequently results are virtually identical for start dates of moulting (not shown).
The critical climatic period for the timing of moulting in superb fairy-wrens was strikingly different for the two climate variables considered. The hazard of moult completion in any given week was best predicted by the rainfall in the preceding 25 weeks, without much variation in contribution among time intervals across this 25-week period (Fig. 2a). When rainfall was high in the preceding 25 weeks, the moulting hazard increased considerably ( [1·25,1·46], = 63·2, P <0·001), implying that males moult earlier when rainfall was high. By contrast, the hazard of moult completion in any given week was best predicted by the temperature during the preceding 15 weeks, with large variation in contribution among time intervals across this 15-week period (Fig. 2b). Specifically, the temperature in the two most recent preceding weeks (week 0 and 1) was found not to predict any of the variation in moulting hazards, implying a time-lag, while the preceding 3–7 weeks contributed most strongly to our weighted temperature index, suggesting that the effect of temperature on moulting gradually builds up in the 8 weeks before (weeks 8–15; Fig. 2b). When temperatures were high in the preceding 15 weeks, the moulting hazard increased considerably ( [2·06,3·30], = 63·8, P <0·001), suggesting that males moult earlier when minimum temperatures are high.
Similar results were obtained when analysing the climate-dependency of the hazard of starting nuptial moult, with the key difference that the weight function describing the critical period shifted about 2–3 weeks (Fig. 2), reflecting the typical duration of moult. Henceforth, we consider only the time of moult completion, because the time of attaining full nuptial plumage is thought to be the cue that females use in mate choice (Cockburn, Osmond & Double 2008).
So far, we have assumed that the dependency of weekly moulting hazard on weather variables is equally strong throughout the year. However, because male fairy-wrens can moult any time between March and November (Fig. 1), this means some individuals moult in the driest, while others moult in the wettest period of the year (Fig. 3a). Likewise, birds moulting in early autumn and late spring do so when nights are mild, while birds moulting in winter and early spring typically experience minimum temperatures around freezing (Fig. 3b). Further analyses showed that both of the climate variables only affected the moulting hazard in a given week when the climatic conditions were most severe. The effect of preceding rainfall on weekly moulting hazard was only detectable in autumn and early winter, generally the driest period of the year; in the remaining wetter periods of the year, moulting hazard was not dependent on rainfall (Fig. 3c). In other words, the weekly moulting hazard only depended on the rainfall of the preceding 25 weeks during periods when rainfall was typically a limiting factor. Similarly, the effect of minimum temperature on moulting hazard was only detectable in winter and early spring, the coldest period of the year; in the remaining warmer periods of the year, the weekly moulting hazard was insensitive to preceding temperatures (Fig. 3d). Thus, the hazard of early moulting is primarily dependent on preceding rainfall, while the hazard of moulting later in the year is mainly temperature-dependent.
Age-dependent phenotypic plasticity
The dependency of moulting on preceding rainfall also varied between age-classes. The rainfall effect on moulting hazard was negligible in 1–2 year olds, but increased in older age-classes, especially for birds ≥7 year old (Fig. 4a; [1·03,1·09], = 11·2, P <0·001). However, the effect of minimum temperature did not vary systematically with age (Fig. 4b; [0·97,1·05], = 0·4, P =0·52). The predictions of this final model were highly correlated with both the observed annual mean and standard deviation of the time of completion of moult (respectively r =0·74 and r =0·65, n =22 years), suggesting that our model is able to explain 55% and 42% of the interannual variability in the population’s mean and standard deviation of the timing of moult, respectively.
Because effect sizes expressed as hazard ratios are not always intuitive to interpret, we also visualized the above age × rainfall interaction in terms of how the observed timings of moult completion depend on the annual rainfall (i.e. ignoring individuals that died before or failed to moult). Again, we can see that the phenotypic plasticity to rainfall is virtually absent in young males, but gradually increases with age (Fig. 5a). An alternative way to view the age × climate interaction is that the apparent strength of age-related improvement in timing of moult is environmentally dependent, because the age-trajectory was much steeper in the five wettest years than in the five driest years of study (Fig. 5b).
Fluctuations in population composition
The mean age of males of known age fluctuated strongly between years in our population (Fig. 6a). However, the initial increase in male age during the first 10 years of study arises in part because at the start of the study, not all birds were of known age (Fig. 6a). During this period, known-age birds are logically expected to be relatively young. However, in the last 12 years of the study, 98% of birds were of known age, and fluctuations in age-structure were thus likely to be driven by population dynamics rather than sampling bias. From 1998 to 2009, the mean male age fluctuated between 2 and 3 years old and could increase as much as 0·8 years (+39%) or decrease as much as 0·6 years (−22%) from one year to the next (Fig. 6a).
The annual changes in the population’s mean male age were dependent on the rainfall in the preceding year (R2=0·26, F1,11 = 4·9, P =0·05), with above average rainfall typically being associated with reductions in the mean male age in the next year (vice versa for below average rainfall; Fig. 6b). We found no evidence for additional effects of annual mean minimum temperature on the dynamics of mean male age (R2 = 0·05, F1,11 = 0·5, P =0·49).
Rejuvenation of the male age-structure can be caused by increased production of young and by low survival of adult males (dispersal is negligible; Cockburn et al. 2008a). The annual production of yearlings is positively associated with annual rainfall in this population (inset of Fig. 6b; R2 = 0·28, F1,11 = 8·8, P =0·01), while adult survival is not (R2 = 0·02, F1,11 = 0·01, P =0·94, see also Cockburn et al. 2008b), indicating that juvenile recruitment is the main agent of population rejuvenation in response to environmental variability.
Given that there is a clear trend towards less rainfall over the 22-year study period (−12 mm per year; inset of Fig. 5b), and that most age-classes moult later when conditions are dry (Figs 4 and 5), one may have expected the mean timing of moult completion to have delayed during the study. Strikingly, there was no evidence for a positive linear trend over the years in the hazard of moult completion, if anything, the trend was negative (HRyear = 0·99 [0·96,1·02], = 0·3, P =0·58). The fact that – notwithstanding substantial annual fluctuations in moulting time – the population’s mean timing of moult did not increase from 1988 to 2009 (Fig. 5c) suggests that another mechanism dampened the impact of phenotypic plasticity on trait change. Because the same rainfall that causes old males to moult early also caused synchronous rejuvenation of the male age-structure, we hypothesized that the negative temporal covariance between the contribution of phenotypic plasticity and the contribution of fluctuations in population composition to trait change could be such a dampening mechanism.
To investigate this hypothesis, we determined the impact of rainfall-dependent plasticity and rainfall-dependent fluctuations in age-structure on trait dynamics using a population model and the previously described moulting hazard model with the daily weather data from 1940 to 2010 as input. We considered four scenarios to disentangle the influence of phenotypic plasticity and population dynamics on trait dynamics. In the first, only the seasonal patterns of minimum temperature varied between years, while the seasonal rainfall patterns (determining plasticity to rainfall) and the demographic rates (determining the population’s age-structure) were the same in all years. In the second, we also allowed the fecundity and thereby the age-structure to fluctuate between years as a function of annual rainfall, while in the third, we instead allowed the intra-annual rainfall pattern and therefore phenotypic plasticity to vary between years. Finally, in the fourth scenario, both the demographic rates and the intra-annual rainfall patterns were allowed to fluctuate between years, such that both phenotypic plasticity and population dynamics affected trait dynamics.
As expected, the interannual variance in the population’s mean moulting time was lowest when fluctuations in trait values were solely because of plastic responses to annual variation in minimum temperature patterns (Fig. 7a,e). Additional fluctuations in the population’s age-structure because of rainfall variability increased the interannual variance in moulting time with a factor of two (Fig. 7b,e). Alternatively, additional phenotypic plasticity in response to rainfall variability resulted in an increase of the interannual variance in moulting time with a factor of four (Fig. 7c,e). Strikingly, when both individual and population-level processes were allowed to affect trait dynamics, the interannual variance did not increase further, but was even reduced (Fig. 7d,e). Thus, the combined effects of rainfall on phenotypic plasticity and population composition result in less variable trait dynamics than because of the effects of phenotypic plasticity alone (Fig. 7d vs. 7c).
Moreover, because of the strong decrease in rainfall from 1988 to 2009 (Fig. 5b), the model containing phenotypic plasticity in response to rainfall variability predicted a delay in moulting time over the study period (0·05 weeks per year, grey area in Fig. 7c), while the model containing fluctuations in the population’s age-structure because of rainfall variability predicted an advancement in moulting time over the study period (−0·05 weeks per year, Fig. 7b). Consequently, a model including both these mechanisms predicted the virtual absence of a trend in moulting time (−0·01 weeks per year, Fig. 7d), which is much more consistent with the observed stasis in moulting time (Fig. 7d).
Female superb fairy-wrens choose males that acquire nuptial plumage early as the sires of the extra-group offspring that dominate parentage in this species (Dunn & Cockburn 1999; Cockburn, Osmond & Double 2008). We used a novel method of detecting climate sensitivity (van de Pol & Cockburn 2011) to show that acquisition of nuptial plumage occurs earlier among older males, particularly when rainfall is high, while for late moulters – typically younger males, moult is best predicted by warm temperatures. The population’s age-structure is also affected directly by climate. High rainfall enhances production and recruitment of young males, reducing the average age of the population, while drought increases the average age of the population. Interestingly, these two environmental effects on trait dynamics counteract each other. For example, high rainfall facilitates earlier moulting among older males, but concurrently increases the proportion of young males, which always moult late. In the remainder of the paper, we discuss how understanding the interplay between the different mechanisms affecting trait change is fundamental to predicting the response of moulting time to climate change, and the opportunity for sexual selection.
Changes during a lifetime
We showed that the expression of a sexually selected trait, the timing of acquisition of nuptial plumage, exhibits strong phenotypic plasticity in response to two climatic variables, and that this sensitivity was strongest when the climate is most harsh (the driest and coldest part of the year; Fig. 3). There are at least two factors restraining early acquisition of the nuptial plumage that is favoured by females. First, moulting is energetically costly in small birds because feathers are replaced, and the growing pin feathers are likely to lose heat rather than provide insulation (Payne 1972). Moreover, acquisition of nuptial plumage by superb fairy-wrens requires elevated levels of the immunosuppressant hormone testosterone (Peters et al. 2000). Second, maintenance of nuptial plumage before and during the breeding season is also likely to be costly, because maintenance too depends on high testosterone levels, and once males acquire nuptial plumage, they advertise their status by vigorously displaying to females living up to seven territories away from their own (Dunn & Cockburn 1999; Double & Cockburn 2003). As a likely consequence, males are expected to moult at times of the year when costs are ameliorated, and thus timing to be highly sensitive to climate.
Many studies have reported intra-population heterogeneity in phenotypic plasticity (e.g. Coulson et al. 2001; Schiegg et al. 2002; Réale et al. 2003; Oro et al. 2010; van de Pol et al. 2010a). Our year-round monitoring of trait phenology allowed a unique insight into the mechanism that can cause some individuals to respond differently to climate change than others. Only few males risked moulting early in what is generally the most challenging period of the year: in autumn when drought can restrict plant growth and in winter when it is always cold. Rainfall over the preceding summer is expected to prolong plant growth and reproduction by the arthropods on which fairy-wrens feed (Lowman 1982), and as a result, the probability of early moulting was rainfall dependent (Figs 2 and 3). By contrast, most males moult during the generally more benign (warmer and often wetter) conditions in spring, and this major moulting pulse appears to be triggered by the seasonal warming that heralds the onset of the reproductive season, which is accelerated in years when conditions are warmer (Fig. 3). Because temperature changes dominate this spring transition, late-moulting males are insensitive to rainfall, but are temperature sensitive instead. Interestingly, only older males undertake the risk associated with moulting early, and consequently, they are susceptible to variation in rainfall, while younger males are not (Figs 4a and 5a). Our study thus suggests that for moulting superb fairy-wrens, heterogeneity in climate sensitivity results from differences among age-classes in both the ability and rewards associated with risk-taking in the mating game.
Our understanding of when exactly climatic conditions are critical for trait expression likely also helps frame predictions of responses to climate change. For example, regional climate models suggest that in future, our study population will experience reduced rainfall primarily in winter and spring (Shi et al. 2008), precisely the period when moulting hazard is relatively insensitive to rainfall (Fig. 3). Furthermore, climate change is expected to result in both reduced rainfall and higher temperatures (Shi et al. 2008), and expected plastic responses to each of these changes would affect trait change in opposite directions. Although their effects on the population’s mean moulting time might (partially) cancel each other out (i.e. another mechanism causing stasis), drier and warmer years are likely to result in reduced intra-annual variability in moulting time, because low rainfall delays moulting time in early males and high temperatures advance moulting time in late males. Less variation among males in moulting time in a given year is expected to have direct consequences for the opportunity for female choice and therefore the strength of sexual selection (Cockburn, Osmond & Double 2008), especially because there is little evidence that females use alternative or backup cues for mate choice (Green et al. 2002; A. Cockburn, A.H. Dalziell, T.Y. Ward, unpublished data).
An interesting corollary of the observed age-climate interaction is that spatiotemporal heterogeneity in climate is expected to cause population variation in the apparent strength of age-related improvement. An increasingly drier Australian climate seems to have led to a shallower age-trajectory in recent years (Fig. 5b), and populations in dry areas may also be expected to exhibit shallower age-trajectories than populations in wet areas. Such populations only differ in their ‘apparent’ strength of age-related improvement, as differentiation is the sole result of environmental heterogeneity and not of underlying genetic differences.
Population composition, trait dynamics and stasis
In our study, the mean age of males fluctuated between years, driven largely by the greater recruitment of young males in years with high rainfall (Fig. 6). When – as in our study – there is additional age-dependency in trait expression, this implies that similar weather in different years can have variable impacts on a population, because of a different population composition (Coulson et al. 2001). We have demonstrated that climate-induced fluctuations in population composition not only produce additional stochasticity in trait dynamics, but that climatic conditions can cause a negative covariance between the fluctuations in age-structure and the plastic response of individuals. Consequently, the effects of rainfall on trait dynamics through phenotypic plasticity of old males are counteracted by synchronous rejuvenation of the age-structure, which dampened the impact of climate change on phenotypic change (Figs 5 and 7).
We have thus identified a novel mechanism based on population dynamics that may help explain (partial) trait stasis, the frequent observation that phenotypic change does not reflect the observed pattern of directional selection or phenotypic plasticity. Previous attempts to explain trait stasis have focussed on individual-level mechanisms (see Merilä, Sheldon & Kruuk (2001); Coulson & Tuljapurkar (2008) for an overview). These include the following: (i) opposing selection forces, such as fluctuating selection in time and space and opposing selection during a lifetime; (ii) misidentification of the targets of selection, such as non-heritable components of the phenotype; (iii) plasticity counteracting selection, as environmental change may cause plastic responses opposite to the direction of selection and (iv) genetic constraints arising from genetic correlations with other traits under selection, or sexually antagonistic selection.
It is interesting to consider whether the dampening mechanism observed in our study might be general. For this, it is important to be aware that although we focussed on age-dependent processes, similar mechanisms can operate when trait expression depends on any state variable (e.g. life stage, social status, body size) and environmental, density- or frequency-dependent processes directly affect the distribution of such a state variable in the population. Furthermore, the three conditions for the mechanism to occur may not be uncommon. The first condition – that the study population is structured according to some state variable – is very general, as reflected by the frequent use of (st)age-structured demographic and population models in the literature (e.g. Caswell 2001). The second condition – that the strength of phenotypic plasticity (or environment-dependent selection) depends on the state or age of an individual – has only recently received the attention it deserves (e.g. Coulson et al. 2001; Schiegg et al. 2002; Réale et al. 2003; Oro et al. 2010; van de Pol et al. 2010a). The third condition – that the same environmental variability that causes phenotypic plasticity or selection to occur synchronously affects the demographic rates that determine the population composition in terms of the state variable of interest (e.g. age-structure) – has not been well explored, but may also not be uncommon. Single environmental variables may often affect various aspects of life history as well as demography. For example, in Eurasian oystercatchers (Haematopus ostralegus), cold winters affect the timing of egg laying (Heg 1999) as well as the strength of selection on diet specialization (van de Pol et al. 2010b) and also concurrently result in population rejuvenation because of both low adult survival and high fecundity (van de Pol et al. 2011). Ectotherms are other likely candidates in which temperature variability drives aspects of both trait expression and demography. Finally, an environmental covariance between phenotypic plasticity and demography is implicit in the dynamics of demographic traits.
While in our study individual-level contributions to trait dynamics were counteracted by population-level contributions, it may also be possible that the synchronous changes in plasticity (or selection) and in population composition reinforce each other’s contribution to trait dynamics. For example, in Eurasian oystercatchers, reductions in food stocks over time because of commercial shellfisheries (Ens 2006) and global warming (van de Pol et al. 2010a) appear to have caused birds breeding in high-quality habitat to delay egg laying with a week over two decades, but there was no delay by breeders in low-quality habitat (van de Pol 2006). Concurrently, reduced food stocks also strongly affected the population composition by halving the number of birds inhabiting low-quality habitat (van de Pol et al. 2010a). Consequently, the environmentally driven changes in the proportion of birds occupying high-quality habitat (from 39% to 66% over a 20-year period) may have reinforced the impact of phenotypic plasticity on the trait dynamics of laying date in the entire population.
More generally, we would like to emphasize that, independent of whether population-level processes counteract or reinforce individual-level processes contributing to trait change and independent of whether fluctuations in population composition will generally be large enough to cause complete trait stasis, it is worthwhile to always check whether fluctuations in population composition obscure at least part of the patterns of trait dynamics caused by individual-level processes.
Finally, in our study, we have ignored the contribution of selection to trait dynamics, because our goal was to show that population-level processes may explain why phenotypic change does not reflect the observed pattern of individual-level processes (either phenotypic plasticity or directional selection). The timing of moult is known to be under directional selection, with positive selection on early moulting being strongest in years with high rainfall (Cockburn, Osmond & Double 2008), which suggests that there is also a negative temporal covariance between the contribution of selection and fluctuations in population composition to trait dynamics. However, future studies will have to show whether moulting time has a heritable basis and how any additive genetic variance might vary with age. Furthermore, a retrospective decomposition of all contributing factors to trait change will be required to quantify the relative contribution of selection, plasticity and population demography to trait dynamics (Coulson & Tuljapurkar 2008; Ozgul et al. 2009, 2010).
We thank the many people that contributed to data collection over the years. Data were collected under a series of Discovery Grants from the Australian Research Council. Permits were provided by the ANU Animal Experimentation Ethics Committee, and the Australian National Botanic Gardens. M.v.d.P. was supported by an Australian Postdoctoral Fellowship of the Australian Research Council (DP1092565).