Ecology, 98(10), 2017, pp. 2684–2697 © 2017 by the Ecological Society of America
Breeding habitat selection across spatial scales: is grass always greener on the other side? PAUL ACKER,1,2,4 AURELIEN BESNARD,2 JEAN-YVES MONNAT,3 AND EMMANUELLE CAM1 1 Laboratoire Evolution et Diversite Biologique (EDB), UMR 5174, Universite Paul Sabatier – Centre National de la Recherche Scientifique (CNRS) – Ecole Nationale de Formation Agronomique (ENFA), 118 Route de Narbonne, Toulouse F-31062 France 2 EPHE, PSL Research University, CNRS, UM, SupAgro, IRD, INRA, UMR 5175 CEFE, Montpellier F-34293 France 3 6 Pennarun d’An Traon, Goulien F-29770 France
Abstract. Habitat selection theory predicts that natural selection should favor mechanisms allowing individuals to choose habitats associated with the highest fitness prospects. However, identifying sources of information on habitat quality that individuals use to choose their breeding habitat has proved to be difficult. It has also proven difficult to identify dispersal costs that prevent individuals from joining the highest-quality sites. A synthesis that integrates dispersal costs and habitat selection mechanisms across space has remained elusive. Because costs of dispersal are generally distance-dependent, we suggest that a habitat selection strategy of sequential proximity search (SPS) can be favored by natural selection. This strategy requires that animals make decisions at multiple scales: whether to stay or leave the previous breeding site, depending on reproductive success; then, if dispersal is chosen, use information on neighborhood habitat quality to decide whether to stay in the neighborhood or leave, expanding the search area until the nearest suitable site is chosen. SPS minimizes distance-dependent dispersal costs while maximizing benefits of gaining a better habitat. We found evidence of breeding dispersal behavior consistent with this strategy in a kittiwake population stratified into a spatial hierarchy from colonies to nest sites. We used a mixed sequential regression model to study dispersal decisions, indexed by breeding dispersal movement, of 2,558 individuals over 32 yr. Scale-dependent dispersal propensities of kittiwakes varied according to breeding status, breeding experience, sex and individual identity. We suggest that distancedependent dispersal costs result from strong competition among kittiwakes for nest sites. Individual decisions regarding dispersal (whether to leave or not, and where to go) depend on nesting habitat quality as well as the competitive ability required to keep territory ownership in a previous site, or to acquire a new site; this ability varies according to distance between sites and individual characteristics. Additional studies are needed to establish the generality of SPS in habitat selection. Key words: colonial species; habitat quality; habitat selection; informed dispersal; life history; ordinal response; public information; seabird; shrinkage prior; spatial scales.
INTRODUCTION For most species, biotic and abiotic properties of breeding habitat are spatiotemporally variable at several scales, and strongly influence individual fitness (Stokes and Boersma 1998, Wilson 1998, van de Pol et al. 2006a, Nussey et al. 2007, Creighton et al. 2009). From an evolutionary viewpoint, breeding habitat quality is defined as fitness prospect in the habitat (Johnson 2007). Mechanisms allowing animals to select the best option among habitats of different quality are thus expected to evolve (Cody 1985). Early habitat selection models assumed that individuals have perfect knowledge of habitat quality in different locations. Individuals are free to distribute according to habitat quality, and realized fitness equalizes among habitats of different initial Manuscript received 11 December 2016; revised 17 April 2017; accepted 10 July 2017. Corresponding Editor: John Sauer. 4 E-mail:
[email protected]
quality because of negative density-dependence – the “ideal free” distribution (Fretwell 1972). Alternatively, competition can constrain individuals so that some preempt higher-quality habitats and force others to settle in lower-quality ones – the ideal “despotic” distribution (Fretwell 1972). Although these models provide a useful departure point to investigate animal distributions, they lack consideration of behavioral mechanisms by which individuals assess habitat quality and disperse. To assess breeding habitat quality, individuals can cue on abiotic or biotic factors (e.g., microclimate: Wilson 1998, food supplies: Orians and Wittenberger 1991). However, these factors provide partial information on habitat quality and might be unreliable (Bollmann et al. 1997, Giraldeau et al. 2002). It could be more reliable and parsimonious to cue on conspecific or heterospecific density, or the breeding success of individuals, if they are sufficiently predictable (Switzer 1993, Doligez et al. 2003). To decide whether to disperse, individuals can use personal information, i.e. their own success, which
2684
October 2017
HABITAT SELECTION ACROSS SPATIAL SCALES
results from habitat quality and individual characteristics (Switzer 1997, Schaub and von Hirschheydt 2009). To decide whether and where to disperse, individuals can also use public information, i.e. the success or density of conspecifics (Danchin et al. 1998, Doligez et al. 2002, Fernandez-Chac on et al. 2013) or heterospecifics (Parejo et al. 2005, Fletcher 2007), which provides a greater sampling power to assess habitat quality (Schmidt et al. 2010). However, individuals might not disperse toward the best habitats. If all individuals select the same location, competition may devaluate habitat quality (Lima and Zollner 1996). Furthermore, dispersal costs (reviewed in Bonte et al. 2012) have been widely neglected in habitat selection studies (Morris 2003, Burgess et al. 2012). Costs can be incurred while searching for, or moving to a new habitat (predation, energy and time spent in movement, information gathering and establishment in a competitive context; Stamps et al. 2005). Dispersers may also incur opportunity costs due to loss of familiarity advantages (e.g., knowledge of foraging routes, territorial dominance, pacified neighborhood interactions; Piper 2011). Refinements of habitat selection theory and empirical studies integrating such costs are required (Piper 2011, Burgess et al. 2012). Moreover, the spatial scale of analyses determines our perception of individual decisions (Bowler and Benton 2005). For instance, a bird changing nest site might also change woodland or not: the factors motivating the decision can differ if they concern the nest site, the woodland, or both. It is necessary to consider a hierarchical framework to disentangle the scales over which information is gathered, and understand individual decisions (Kotliar and Wiens 1990, Orians and Wittenberger 1991). Importantly, information at different scales can conflict: breeding success can be high in the neighborhood, but low at the scale of the entire woodland, which induces a habitat selection dilemma. Based on neighborhood success, individuals should keep their nest site (hence remain in the woodland), but based on woodland success, they should leave the woodland (hence the nest site). Dispersal studies have rarely considered multiple spatial scales: how individuals adjust habitat choices across scales is poorly understood (Bowler and Benton 2005, Schmidt et al. 2010, Matthysen 2012). Integrating costs and constraints on habitat selection across spatial scales can help solve the aforementioned dilemma. Indeed, whereas high-quality habitats might maximize fitness anywhere, dispersal costs are expected to increase with distance to the previous habitat (Van der Jeugd 2001, Baker and Rao 2004, Bowler and Benton 2005). Natural selection should favor a strategy balancing dispersal costs and benefits by settling in the closest habitat which maximizes fitness. We suggest a “sequential proximity search” (SPS): a suite of conditional choices of leaving the previous habitat at an increasing spatial scale. Individuals first assess the quality of their own site and decide whether to resettle there. If not, they assess habitat
2685
quality in their closest neighborhood and decide whether to resettle in this neighborhood. If not, they expend their decision to an enlarged neighborhood and so on until a habitat is accepted. For a given species, the spatial scales at which the SPS can hold should lie below the limit over which information gathering and movement are unaffordable for individuals. Dispersal costs can also differ according to individual “state”, which has to be taken in consideration when investigating breeding habitat selection (Matthysen 2012). For instance, breeders face a trade-off between time allocated to habitat selection and parental care, which is relaxed by breeding failure and is not incurred by nonbreeders. The territorial sex incurs higher costs of establishing in an unfamiliar habitat and leaving the defended breeding site (Greenwood 1980). Experienced individuals with higher competitiveness may preempt breeding sites, but they may have more familiarity with a previous location (Greenwood 1980, Matthysen 2012). Static individual differences in competitiveness and other dispersal-related traits may also yield heterogeneity in dispersal motivations (Matthysen 2012). Here we investigated spatial-scale dependency in habitat selection behavior using data from a 32-yr metapopulation study in the black-legged kittiwake (Rissa tridactyla). We addressed whether dispersal behavior is consistent with the SPS strategy. We defined nested spatial units, and breeding dispersal as an ordinal response indexed by the movement scale. We used mixed sequential binary regressions (Agresti 2010) accounting for the hierarchical nature of habitat patch structure. We assessed the influence of public information, individual state (sex, breeding experience, breeding status and performance), location, individual identity and year on dispersal probability at each spatial scale, conditional on departure at lower scales. In this population, several studies have shown that dispersal motivations vary according to personal success, status, experience, sex, and conspecific success in the smallest patch with clear physical boundaries that includes several nest sites (Danchin et al. 1998, Danchin and Cam 2002, Naves et al. 2006). Moreover, Bled et al. (2011) have found a negative quadratic relationship between probability of nest-site persistence (with the same or another owner) and density in the immediate neighborhood. However, no study considered the hierarchy of scales. Here we drew a distinction between movement among contiguous patches or distant patches, and evaluated habitat quality not only at the scale of contiguous patches, but also at that of larger spatial units including distant patches. Under the SPS hypothesis, we expected that each dispersal decision at a given scale be motivated by habitat quality evaluated at that scale but not at larger ones. If the SPS hypothesis does not hold, dispersal decisions should depend on public information evaluated at all spatial scales because individuals will be attracted by the highest-quality habitats anywhere in space.
2686
PAUL ACKER ET AL.
METHODS Data collection The study (meta)population is located in Brittany (France; Appendix S1). We used data from 1982 to 2012. Birds were marked as chicks using color bands (Cam and Monnat 2000). Resighting probability is virtually equal to one after recruitment (0.998 in Cam et al. 1998), thus recruited individuals not returning to colonies were considered dead or permanently emigrated (Cam and Monnat 2000). Birds were categorized as inexperienced at first breeding and experienced afterwards. Nonbreeders are individuals that bred in the past but did not complete nest building in the current year (Cullen 1957, Cam et al. 1998). Unsuccessful breeders completed nest building but did not raise any chick to fledging. Successful breeders raised at least one chick to fledging (Cam and Monnat 2000). Chicks were considered as fledged if they either left the nest site and came back to be fed by parents, or were seen alive in the nest before fledging with folded wings several centimeters longer than tail. Sex was identified through behaviour (Cam et al. 1998, Naves et al. 2006). The annual location (nest site) was known for every breeder. Nonbreeders that built an incomplete nest were assigned to this location. Alternatively, nonbreeders were assigned to the location they attended the most (Cam and Monnat 2000). Data from nonbreeders that evenly attended different sites were excluded (36% of nonbreeder cases). Focusing on breeding dispersal required data from individuals that were seen at least in two consecutive breeding seasons. We excluded data from individuals with unknown state (sex or breeding status, 1% of individual-year observations).
Ecology, Vol. 98, No. 10
for breeding habitat quality. The breeding attempt outcome of an individual settled in a patch and this patch success cannot be considered independent. Consequently, we excluded the reproductive outcome of this individual to calculate patch success. We kept only data from patches including more than 10 nests (we excluded 4% of individual-year observations) where demographic stochasticity may result in a large mismatch between patch quality and success (Danchin et al. 1998, Naves et al. 2006). We excluded data from individuals or nest sites with uncertain breeding success in a given year. We also excluded data when the number of nests used to calculate patch success was lower than 80% of the total number of nests in the patch (0.3% of individual-year observations). We visually assessed spatiotemporal heterogeneity of patch quality at the cliff, social-group and colony scale (Appendix S2). We addressed predictability of patch quality in a time-series analysis framework, by inspecting the sample autocorrelation function at each spatial scale (Appendix S2). Individual dispersal events We treated breeding dispersal (Fig. 1) as a variable (Y) with five modalities depending on spatial scales of
Spatial scales We treated the breeding habitat as nested spatial units: (1) the nest site, (2) the “cliff”: a cliff wall containing nest sites and separated from other cliffs by rocky ridges or coastal segments without nesting birds (Naves et al. 2006), (3) the “social group”: a set of cliffs constitutive of a cove where visual or vocal contact is possible among birds, (4) the colony: a set of social groups separated from other colonies by at least 500 m (max 12 km; Aubry et al. 2009). Over 1982–2012, the study area hosted annually 2–5 colonies (mean 4.5 0.7), 5–18 social groups (14.0 2.8), 20–44 cliffs (31.0 7.2) and 658–1201 nest sites (935.0 118.0; Appendix S1). Hereafter spatial units above the nest site (cliffs, social groups or colonies) will be called “patches”. Breeding habitat quality We used (1) the number of complete nests in a patch (hereafter “density”; Fernandez-Chac on et al. 2013), (2) and the annual proportion of nests with successful reproduction in the patch (hereafter “patch success”) as proxies
FIG. 1. Schematic representation of the dispersal events. Modalities are ordered according to the scale of the movement: (1) keeping the nest site, (2) leaving the site but staying in the cliff, (3) leaving the cliff but staying in the social group, (4) leaving the social group but staying in the colony, (5) leaving the colony.
October 2017
HABITAT SELECTION ACROSS SPATIAL SCALES
departure and indexed by movement magnitude: re-using the previous nest site (Y = 1), leaving the nest site but staying in the cliff (Y = 2), leaving the cliff but staying in the social group (Y = 3), leaving the group but staying in the colony (Y = 4), leaving the colony (Y = 5). We used observations from 10,702 dispersal events (Y = 1: 7,814, Y = 2: 2,059, Y = 3: 293, Y = 4: 131, Y = 5: 405) concerning 2,558 individuals, 2,376 nest sites, 43 cliffs, 21 social groups, 6 colonies and 32 yr. Modeling We addressed the relationship between dispersal decisions and information on breeding habitat quality using a regression model for ordinal variables (sequential or “continuation-ratio” model; Agresti 2010). The probability space is split into a suite of conditional probabilities following the response ordering. The model includes z-1 binary responses (where z is the number of categories in the ordinal variable) that contrast each category (Y = j) with the grouping of higher-order categories {(Y = j + 1),. . .,(Y = z)}. In this framework, dispersal involves several decisions: whether or not to leave the breeding habitat at each spatial scale, conditional on leaving the lower scales. Each dispersal event Yit concerned one individual i in two consecutive years: t (year of departure) and t + 1 (year of arrival). ðjÞ There are four levels of dispersal probability Pit (Eq. 1): the probability of leaving the nest site occupied at t (j = 1), leaving the cliff occupied at t given that the individual left the site occupied at t (j = 2), leaving the social group occupied at t given that the individual left the cliff occupied at t (j = 3), leaving the colony occupied at t given that the individual left the group occupied at t (j = 4).
2687 ðjÞ
For each probability Pit , public information at lower scales than the scale of the focal movement was excluded from the set of predictors. We thus assumed that individuals did not refer anymore to information concerning spatial units once they had left these units. Further, we accounted for spatial heterogeneity in dispersal probability only at the scale of the focal movement (i.e. only nestsite identity was included at the nest-site scale, only cliff identity was included at the cliff scale, etc.). We used the robit link function, a robust alternative to logit or probit links for binary regressions which is less sensitive to outlying observations (Liu 2004, Appendix S3). Our model was the following (Eqs. 2.1 to 2.4): ð1Þ
ð1Þ
ð1Þ
ð1Þ ð1Þ robitðPit Þ ¼ lð1Þ þ að1Þ s þ ae þ ar þ ðbLW þ cLWr ÞLWit ð1Þ
ð1Þ
ð1Þ
ð1Þ
þ ðbLG þ cLGr ÞLGit þ ðbLC þ cLCr ÞLCit ð1Þ
ð1Þ
ð1Þ
ð1Þ
þ ðbDW þ cDWr ÞDWit þ ðbDG þ cDGr ÞDGit ð1Þ
ð1Þ
ð1Þ
ð1Þ
þ ðbDC þ cDCr ÞDCit þ ðbD2 W þ cD2 Wr ÞD2Wit ð1Þ
ð1Þ
ð1Þ
ð1Þ
þ ðbD2 G þ cD2 Gr ÞD2Git þ ðbD2 C þ cD2 Cr ÞD2Cit ð1Þ
ð1Þ
þ ui þ ut þ uð1Þ n (2.1) ð2Þ
ð2Þ
ð2Þ
ð2Þ ð2Þ robitðPit Þ ¼ lð2Þ þ að2Þ s þ ae þ ar þ ðbLW þ cLWr ÞLWit ð2Þ
ð2Þ
ð2Þ
ð2Þ
þ ðbLG þ cLGr ÞLGit þ ðbLC þ cLCr ÞLCit ð2Þ
ð2Þ
ð2Þ
ð2Þ
þ ðbDW þ cDWr ÞDWit þ ðbDG þ cDGr ÞDGit ð2Þ
ð2Þ
ð2Þ
ð2Þ
þ ðbDC þ cDCr ÞDCit þ ðbD2 W þ cD2 Wr ÞD2Wit ð2Þ
ð2Þ
ð2Þ
ð2Þ
þ ðbD2 G þ cD2 Gr ÞD2Git þ ðbD2 C þ cD2 Cr ÞD2Cit ð2Þ
ð2Þ
þ ui þ ut þ uð2Þ w (2.2)
ðjÞ
Pit ¼ PrðY it [ j j Y it jÞ ¼ 1 PrðY it ¼ j j Y it jÞ; (1.1)
ð3Þ
ð3Þ
ð3Þ
ð3Þ ð3Þ robitðPit Þ ¼ lð3Þ þ að3Þ s þ ae þ ar þ ðbLG þ cLGr ÞLGit ð3Þ
ð3Þ
ð3Þ
ð3Þ
ð3Þ
ð3Þ
þ ðbLC þ cLCr ÞLCit þ ðbDG þ cDGr ÞDGit
8 ð1Þ > < PrðY it ¼ 1Þ ¼ 1 Pit ! k1 Q ðjÞ ðkÞ > Pit ð1 Pit Þ k ¼ 2; . . .; 5: : PrðY it ¼ kÞ ¼
ð3Þ
ð3Þ
þ ðbDC þ cDCr ÞDCit þ ðbD2 G þ cD2 Gr ÞD2Git ð3Þ
ð3Þ
ð3Þ
ð3Þ
þ ðbD2 C þ cD2 Cr ÞD2Cit þ ui þ ut þ uð3Þ g
j¼1
(2.3)
(1.2) ðjÞ
Each probability Pit was expressed as a function of variables characterizing the individual state (sex, experience, breeding status and individual identity in year t), the location of origin (nest-site, cliff, social-group or colony identity, patch density and success at the different spatial scales in year t), and year of departure (t). We also considered interactions between (1) breeding status, and (2) patch density and success. Following Fletcher 2007 and Bled et al. 2011, we included the quadratic effect of density at each level of the model. Individual identity, year and patch identity were treated as random effects (except colony identity because there were only 6 colonies). Other variables were treated as fixed effects.
ð4Þ
ð4Þ ð4Þ ð4Þ robitðPit Þ ¼ lð4Þ þ að4Þ s þ ae þ ar þ ac ð4Þ
ð4Þ
ð4Þ
ð4Þ
þ ðbLC þ cLCr ÞLCit þ ðbDC þ cDCr ÞDCit ð4Þ
ð4Þ
ð4Þ
ð4Þ
þ ðbD2 C þ cD2 Cr ÞD2Cit þ ui þ ut
(2.4) where l stands for intercepts, a for fixed effects of categorical variables, b for fixed effects of continuous variables, c for interactions, u for random effects of categorical variables. L stands for patch success and D for density (continuous variables). Capitalized subscripts indicate the spatial scale corresponding to the parameter or variable: W, G, and C, for cliff wall, social group and
2688
PAUL ACKER ET AL.
colony, respectively. Italicized subscripts indicate the object of the parameter or variable: s for sex, e for experience, r for reproductive status, n for nest site, w for cliff wall, g for social group and c for colony. For individuals that left their social group (j = 4), there were too few cases in the successful breeder category (eight observations) to consider a separate breeding status: we grouped them with unsuccessful breeders (478 observations). There were also too few observations in colony 6 (5/536 observations) to consider a separate patch: we grouped them with observations from colony 5 (the closest colony, 153 observations). The random effects account for non-independence in the data induced by individual, spatial and temporal pseudoreplication. They also provide the opportunity to explore how heterogeneity in dispersal probabilities can be partitioned into consistent influences of the individual, spatial and temporal contexts that are not captured by fixed effects. Further, we used a quadrivariate normal distribution with mean 0 and a different variance-covariance matrix for individual identity and year to consider correlations between individual random effects and between year random effects over the four submodels (Appendix S3). This correlation structure was helpful for Bayesian sampling: autocorrelation substantially decreased and the mixing improved. Also, this model feature provides the opportunity to assess whether individuals have a consistent propensity to disperse to close or remote locations. Correlations between year effects would indicate a tendency for short- or long-distance dispersal in some years. This might result from the wide temporal variation in success, density, and true distance among patches (Appendices S1 and S2), because dispersal magnitude in our model is only a rough approximation of dispersal distance. Parameter estimation Inference was based on a Bayesian approach using Gibbs sampling, a Markov Chain Monte Carlo algorithm, with program JAGS 3.4.0 (see BUGS code in Data S1; Plummer 2003) called from R (R Core Team 2016) with the rjags package (Plummer 2015). We ran 20 chains with different initial values. We used an adaptive phase of 100 iterations. We discarded the first 5000 iterations and used the subsequent 40,000 iterations for exploration of posterior distribution samples (8 9 105 samples in total). Chains were not thinned (Link and Eaton 2012). Continuous variables were standardized, making all effect sizes comparable.
Ecology, Vol. 98, No. 10
penalized through shrinking towards zero, unless there is strong signal for non-zero in the data. This approach separates strong signal from noise and avoids overestimating effects (Carvalho et al. 2010). We did not use the horseshoe prior for the intercepts, but weakly informative normal priors with mean 0 and variance 104. We used a weakly informative uniform prior on the range (0,10) for standard deviation of spatial random effects. For individual and year effects, we used the Cholesky decomposition of the variance-covariance matrix of correlated random effects introduced by Chen and Dunson (2003). The priors used reflected reasonable doubt on variance parameters, and shrunk covariance parameters towards zero (Appendix S3). Posterior distributions Post-processing of MCMC chains was performed in R (R Core Team 2016; Appendix S4). Convergence was ^ assessed using the Brooks-Gelman-Rubin diagnostic R for each parameter (Brooks and Gelman 1998). We ^ achieved convergence with all R
