E. Cam, USGS, Patuxent Wildlife Research Center. 11510 American Holly Drive. Laurel, MD 20708–4019, USA
1. Many studies have provided evidence that first-time breeders have a lower survival, a lower probability of success, or of breeding, in the following year. Hypotheses based on reproductive costs have often been proposed to explain this. However, because of the intrinsic relationship between age and experience, the apparent inferiority of first-time breeders at the population level may result from selection, and experience may not influence performance within each individual. In this paper we address the question of phenotypic correlations between fitness components. This addresses differences in individual quality, a prerequisite for a selection process to occur. We also test the hypothesis of an influence of experience on these components while taking age and reproductive success into account: two factors likely to play a key role in a selection process.
2. Using data from a long-term study on the kittiwake, we found that first-time breeders have a lower probability of success, a lower survival and a lower probability of breeding in the next year than experienced breeders. However, neither experienced nor inexperienced breeders have a lower survival or a lower probability of breeding in the following year than birds that skipped a breeding opportunity. This suggests heterogeneity in quality among individuals.
3. Failed birds have a lower survival and a lower probability of breeding in the following year regardless of experience. This can be interpreted in the light of the selection hypothesis. The inferiority of inexperienced breeders may be linked to a higher proportion of lower-quality individuals in younger age classes. When age and breeding success are controlled for, there is no evidence of an influence of experience on survival or future breeding probability.
4. Using data from individuals whose reproductive life lasted the same number of years, we investigated the influence of experience on reproductive performance within individuals. There is no strong evidence that a process operating within individuals explains the improvement in performance observed at the population level.
The processes underlying these hypotheses are clearly different. The selection process operates at the population level: it is based on a progressive change in proportions of individuals with different survival rates. Such a process can be invoked when comparing demographic parameters at different ages using data from individuals, some of which die earlier than others. We will refer to that level as the ‘population level’. In contrast, an optimization process or a process based on cumulative experience operates at the ‘individual level’. This can be observed within each individual, or formally tested by comparing performance at different ages of individuals with the same longevity. From a methodological viewpoint, a major difference between these processes is the level at which they can be addressed. The optimization/cumulative experience hypotheses permit making predictions at the individual level, whereas the selection hypothesis does not.
Practically, investigating processes potentially operating at the individual level can raise difficulties: it is possible to address reproductive performance (individuals attempt to breed several times in their reproductive life), but obviously not survival (mortality occurs only once). In contrast, when addressing variation in demographic parameters throughout reproductive life at the population level, processes operating at the population and the individual level are both relevant. The purpose of this paper is to present analysis of factors rarely examined simultaneously: not only the influence of experience (first reproduction vs. subsequent breeding attempts) on survival and future breeding probability, but also phenotypic correlations between fitness components. The sign of these correlations will indicate whether a selection process is likely to occur. We also address variation in breeding performance at the individual level and compare the patterns observed within individuals to that observed at the population level. A given pattern of variation in demographic parameters within individuals does not necessarily translate into the same pattern at the population level (Vaupel & Yashin 1985).
Our first objective is to test the hypothesis of an effect of experience on breeding success, survival and the probability of breeding in the following occasion in the Kittiwake [Rissa tridactyla (L.)], a long-lived seabird. This step will provide results that can be compared to previous studies of the influence of the first breeding attempt on demographic parameters. The second objective is to address phenotypic correlations between components of fitness while taking experience into account. Investigation of reproductive success and lifetime production in the Kittiwake led Coulson (1968), Coulson & Porter (1985), Coulson (1988) and Fairweather (1994) to put forward the hypothesis of differences in individual quality. Similarly, in a previous study on the same species, Cam et al. (1998) showed that there are positive phenotypic correlations between current reproduction and survival, as well as between current and future reproduction. In this framework, our first prediction is that breeders (regardless of experience) exhibit a higher survival and probability of breeding in the following year than birds that skipped a breeding opportunity (i.e. nonbreeders). We compare the demographic features of birds that do not breed to those of experienced breeders on the one hand, and of first-time breeders on the other hand. This addresses the sign of the phenotypic correlation between reproduction and survival. We also expect lower survival and breeding probabilities in inexperienced compared to experienced breeders.
Our third objective is to examine in more detail the question of experience by controlling for age (i.e. using a multivariate approach). This analysis addresses whether experience still influences demographic parameters after accounting for the influence of age, which forms the basis of the selection hypothesis. We also control for reproductive success. Assuming that individual quality influences success, successful individuals should exhibit the highest survival and lowest probability of nonbreeding in the following year. A higher proportion of those individuals in older age classes (as assumed by the selection hypothesis) could explain an apparent inferiority of first-time breeders. Controlling simultaneously for age and success permits testing for the influence of experience after accounting for two potential confounding factors linked to selection. These analyses were performed using data from a decreasing number of individuals as they age, which is required to draw inference about the selection hypothesis.
We also address whether a process other than selection could account for the influence of experience on demographic parameters observed at the population level. This requires investigation of variation in performance in individuals with similar reproductive longevity. As we suspect that individuals that die earlier have different features from birds that survive longer, pooling data from birds whose reproductive life lasted at least n years is likely to lead to heterogeneous subsets. This could mask the phenomenon of interest if variations in performances throughout life differ according to quality. Consequently, we performed analyses using data from birds whose reproductive life lasted exactly 2, 3,…, n years. In other words, we used reproductive longevity as a surrogate of quality. We tested for an effect of the number of years elapsed since the start of reproductive life on success probability, and the probability of breeding in the following year. We also address the question of the influence of first-reproduction on reproductive success and future breeding activity using the same approach. Improvement of breeding performance within individuals would support the hypothesis of a positive effect of cumulative experience on reproduction, including reduced reproductive costs, or of a long-term optimization of reproductive investment throughout reproductive life.
Data collection and selection
Data were collected in Brittany (France) from 1979 to 1994. Individuals were marked using a unique colour combination of plastic bands. Only data from individuals marked as chicks and that recruited in one of the study colonies were retained. The data used here were mostly collected between April and August. Each year, individuals were classified as follows: (i) nonbreeders that never previously bred (prebreeders); (ii) first-time (inexperienced) breeders; (iii) experienced breeders; and (iv) nonbreeders that previously bred (nonbreeders). Data from prebreeders were excluded. Individuals were considered as breeders if they built a nest that reached the completion criterion (Maunder & Threlfall 1972). As our objective was to investigate the specific influence of first reproduction on fitness components, we defined a variable experience with two modalities: first-time breeder and experienced breeder. Birds were considered as nonbreeders only if they were resighted in a given year and were known to have skipped the breeding occasion.
Individuals were categorized as successful if at least one of the young fledged. Only data collected from 1985 onwards were retained for analyses requiring knowledge of reproductive success (improvement of field methods led to higher precision in determination of breeding state). Reproductive activity (i.e. breeding/nonbreeding) and reproductive success (i.e. in breeders: fledged at least one chick/fledged no chick), were used as measures of components of fitness. The two components used are: (i) adult survival; and (ii) production of at least one chick (i.e. raised to independence) vs. failure to produce any young. We believe that variation in these components reflects differences in fitness.
Analyses were carried out using two methods, in two steps. Proportions based on counts of individuals observed in year t and returning in year t+ 1 depend on both survival and recapture probability. Unless the latter is equal to one, return rates provide biased estimates of survival probability (Lebreton et al. 1992; Clobert 1995). Proportions based on counts of individuals that return and breed also provide biased estimates of future breeding probability. Differences in return rates among individuals belonging to different categories can reflect variation in recapture probability (Nichols et al. 1994). Breeding activity and success can influence individual behaviour, and thus recapture probability (e.g. individuals that fail may leave breeding places earlier than others, which may influence the probability of observing them). To account for variation in recapture probability we used multistate capture–recapture models (Nichols & Kendall 1995), which incorporate state-specific detection probabilities. They permit estimation of parameters—survival or breeding probability—specific to yearly reproductive state (Nichols et al. 1994; Nichols & Kendall 1995).
Multistate capture–recapture models (Brownie et al. 1993; program mssurviv, Hines 1994) were primarily used to test the hypothesis of an effect of state on recapture probability. They also permitted addressing the influence of experience on survival and the probability of breeding in the following year, and exploration of correlations between components of fitness while taking experience into account. The variable state had three modalities: (i) experienced breeder; (ii) nonbreeder; and (iii) first-time breeder. This method permits estimation of time- and state-specific recapture and survival probabilities and time-specific transition probabilities between states (Nichols et al. 1994; Nichols & Kendall 1995). Some 794 individual histories have been analysed using this approach. The most general model included time- and state-specific recapture and survival probability, and time-specific transition probability between states (Hestbeck et al. 1991). Model notation has been defined in Nichols et al. (1994).
(recapture probability) is the probability that a bird is resighted at time t in state r, given that it is alive and present at time t.
(survival probability) is the probability that a bird in state r at time t survives until occasion t+ 1.
(transition probability) is the probability that a bird in state r in occasion t is in state s at occasion t+ 1 given that the individual survived from occasion t to occasion t+ 1. We used the following notation for states: 1 = experienced breeder; 2 = nonbreeder; and 3 = inexperienced breeder.
Preliminary analysis using this approach indicated that the estimated recapture probability of experienced breeders and nonbreeders was extremely close to one (0.982 ± 0.128*10− 1 for these strata; there is no ‘recapture probability’ for first-time breeders). This high resighting probability permitted us to use return rates to estimate survival probabilities. An individual was considered as dead (or permanently emigrated) if it was not resighted in a given occasion. Similarly we estimated transition probabilities between breeding states r and s using the proportion of individuals released in year t in state r and resighted in year t+ 1 in state s. Survival and transition probabilities were modelled as linear combinations of explanatory variables using a logit link function (Agresti 1990). The same approach was used to address the question of the effect of age and experience on the probability of success in a given year. In addition, as the probability of observing breeders and birds skipping reproduction is very close to one, it is very unlikely that investigators missed first-time breeders established in the study area. Consequently, the age at which birds were observed as recruits can be considered as a reliable measure of age of first local reproduction.
Software and model notation. Logit models were built using proc genmod, proc logistic and the macro glimmix of sas (logistic regression: SAS Institute Inc. 1995; Stokes, Davis & Koch 1995; Littell et al. 1996). Model notation is defined in the STAT User's Guide (SAS Institute Inc. 1988) for generalized linear models. Main effects are represented using capital letters (E for experience, A for age and S for reproductive success), and interactions using an asterisk (E*A for example). Only pairwise interactions between explanatory variables were included in initial models.
Accounting for extradispersion and nonindependence between observations. We estimated an extradispersion parameter and used QAIC for model selection when that parameter indicated over- or underdispersion (see also Model selection). As there is no objective criterion to interpret the extradispersion parameter (Littell et al. 1996), we chose the arbitrary range of values of 0.90–1.10. We used scaled deviance if the value was lower than 0.90 or higher than 1.10 (or very close to these thresholds). In addition, as we used several records from the same individual, nonindependence between observations could bias the analyses. According to Allison (1995), this is not relevant in analyses of survival as a function of the time elapsed since the individual entered the sample (i.e. the first reproduction in our case). We used the repeated statement of the macro glimmix (LittelL et al. 1996) in analyses of the probability of breeding, and the probability of success. We used two correlation structures corresponding to two possible scenarios. In all cases we used a compound symmetric structure, which corresponds to the idea that individuals differ in success probability for example, and that they keep that individual value for their entire reproductive life. We also used an autoregressive structure of order 1 in models that did not include previous reproductive state. This structure corresponds to a situation where observations from the same individual spaced at long time intervals are not correlated in the same way as observations close in time. Models including previous reproductive state already accounted for a possible relationship between consecutive breeding attempts.
Choice of initial logit models. Our objective was primarily to test a priori hypotheses on the effect of experience, age and reproductive success. We did not include all possible factors known to influence reproductive performance, survival or transition probability. In particular, when models contained the variable age, sample sizes corresponding to some modalities of the response variables were sometimes small. This occurred when data were stratified according to the independent variables age, experience, reproductive success and time in models requiring data from two consecutive years (for breeding state transition probabilities). Consequently, we limited the number of explanatory variables: we excluded the effect of time when too many parameters could not be estimated.
Effect of age. In initial models including the variables age, experience and reproductive success, age classes were defined a priori so that experienced, inexperienced, successful and failed birds are represented in each class: [2-, 3-, 4-year olds], [5-year olds] and [6-, 7-year olds]. Only one individual bred for the first time at age two during the study and 12 at age seven, with known breeding success. Data from birds older than seven were excluded from analyses.
Model selection. Model selection was based on Akaike's information criterion (AIC or QAIC; Akaike 1973; Burnham & Anderson 1992; Lebreton et al. 1992; Burnham & Anderson 1998), likelihood ratio (LR) tests and F-tests (Littell et al. 1996). NP corresponds to the number of estimated parameters. Model selection based on these criteria requires that the most general models fit the data adequately, which was assessed using a G-statistic (White 1983) for capture–recapture models and a Hosmer and Lemeshow statistic (Glanz & Slinker 1990) for logit models. For logit models, we systematically built every possible model corresponding to simplifications of the most general model, which always included main effects and pairwise interactions. We only present the general model and candidate constrained models from that initial step. The lowest-AIC model was selected. When the AIC values of two nested models differed by less than two with a higher value for the constrained model we used a LR test or F-test (Lebreton et al. 1992). After selecting a model new constraints were used to test specific a posteriori hypotheses.
Effect of experience and breeding activity
Recapture probabilities. Recapture probability varied with state, but not time (Table 1).
Table 1. . Likelihood ratio tests for an effect of time and state on recapture, survival and transition probability. AIC values for corresponding models
Survival probability. Survival probability varied with time and state (Table 1). There was no significant difference between the survival probabilities of first-time breeders and nonbreeders (Table 2). Experienced breeders had a higher mean survival probability than first-time breeders and nonbreeders
Table 2. . Influence of age (A), breeding success (S) and experience (E) on survival probability
Description of the models
Goodness-of-fit test for M0: Hosmer and Lemeshow statistic = 3.09, d.f. = 8; P= 0.93.
1597 observations from 656 individuals.
Selected model in the initial selection procedure: M1.
State transition probability. Transition probabilities were time-specific (Table 1). We found no evidence of a difference between breeding transition probabilities of first-time breeders and nonbreeders. Inexperienced birds and nonbreeders had a lower probability of breeding in the following year than experienced breeders .
Interaction between state and time. We built a model where temporal variations of survival probabilities of experienced breeders on the one hand and of a category composed of first-time breeders and nonbreeders on the other hand, were parallel on a logit scale. This model was not selected (AIC = 377.7 vs. M3, Table 1). However, estimates made under model M3 (i.e. with the interaction) indicated that the survival probability of experienced breeders was higher or equal to that of inexperienced birds and nonbreeders, in all years (Fig. 1). We also built a model with an additive effect of state and time for breeding transition probabilities. This model had a higher AIC (664.0) than model M3 (AIC = 344.6, Table 1). Estimates of the probability of skipping the next breeding opportunity made under the model with interaction (M3) showed that experienced birds had a lower probability of nonbreeding than first-time breeders and nonbreeders, in all years (Fig. 2).
Success probability. The general model included the effect of experience, time and the interaction between them. We used models with a correlation term between observations from the same individual. We first used an exchangeable correlation structure. The extradispersion parameter estimated using this model indicated underdispersion (0.89). Consequently, we used QAIC for model selection. General model (E T E*T), QAIC = 3289.30, NP = 20, Hosmer and Lemeshow statistic = 1.85, d.f. = 8, P= 0.98. We found no evidence of an effect of the interaction (additive model: QAIC = 3277.34, NP = 11). Models with main effects only were not retained (‘experience’: QAIC = 3346.59, NP = 2; ‘time’: QAIC = 3424.78, NP = 10). The model with no effect of time or experience on success probability was not retained either (QAIC = 3424.43, NP = 1). Experienced breeders had a higher probability of success than first-time breeders, in all years (Fig. 3). Results were similar using a first-order autoregressive correlation structure.
Effect of age, experience and breeding success
Survival probability. Survival probability was modelled as a function of experience, age, reproductive success and the three corresponding pairwise interactions (Table 2). The extradispersion parameter was very close to 1.00 The lowest-AIC model (M1) included the effect of age and breeding success. We tested several a posteriori hypotheses using the model M1 (Table 2). The model with a linear trend in age was not selected (M3). The model with equal survival in 5-year-old birds and younger individuals (M4) was not retained either. We pooled 5-year-old birds with older individuals (M5). The AIC value of this model was slightly higher than the lowest value (1497.26 vs. 1496.59), but the LR test between these models was not significant (χ2= 4.67, d.f. = 2, P= 0.10). Consequently, we selected the most parsimonious model (with only two age classes: M5). Successful breeders had a higher survival probability than failed ones when age was controlled for (Fig. 4a). In addition, the youngest successful breeders had the highest survival probability (Fig. 4a).
Probability of breeding in the following occasion. As above, the initial model comprised the effect of experience, age, breeding success and pairwise interactions (Table 3). The extradispersion parameter was very close to 1.00 . Consequently, we used AIC for model selection. Three models had very close AIC values (M1, M2 and M3). These models were considered as candidate models to test a posteriori hypotheses (Table 3).
Table 3. . Influence of age (A), reproductive success (S) and experience (E) on the probability of nonbreeding in the following occasion
Description of the models
Goodness-of-fit test for M0: Hosmer and Lemeshow statistic = 11.76, d.f. = 8, P= 0.16.
1130 observations from 521 individuals.
M1, M2 and M3: lowest-AIC models from the initial selection procedure.
Μ′1, Μ′2 and Μ′3: constrained models corresponding to M1, M2 and M3, respectively.
We first built models with a linear trend in age, but they were not selected [Μ′1 (1), Μ′2 (1) and Μ′3 (1)]. Neither was the model where 5-year-old and younger birds have equal probability of breeding in the following occasion [Μ′1 (2), Μ′2 (2) and Μ′3 (2)]. We also built models where 5-year-old and older birds had equal probabilities of breeding in the following year [Μ′1 (3), Μ′2 (3) and Μ′3 (3)]. The constrained model with an additive effect of age and reproductive success [Μ′3 (3)] was the lowest-AIC model (857.19). We estimated breeding probabilities using this model: failed breeders had a higher probability of nonbreeding in the following year when age was controlled for (Fig. 4b). In addition, the youngest breeders had a higher probability of nonbreeding in the following occasion when reproductive success was controlled for (Fig. 4b). Another model with only two age classes could also be considered as a candidate model [Μ′2 (3)]. However, the LR tests between this model and the lowest-AIC model [Μ′3 (3)] was not significant (Table 3) (χ2= 0.47, d.f. = 1, P= 0.49). Consequently, we conclude that there is no strong evidence that incorporation of the variable experience improves the fit of the models.
Success probability. The initial model included age, experience and the interaction (Table 4). The extradispersion parameter estimated using an exchangeable correlation structure indicated slight underdispersion . The lowest QAIC model in the initial selection step was the general model (M0): we used this model to test a posteriori hypotheses. The model with a linear trend in age was not retained (M1). We found evidence that the probability of success of 5-year-old birds differs from that of younger individuals (M2 was not retained). We tried to pool 5-year-old birds with older birds: that model was selected (M3). We built a model where there was no effect of age in experienced breeders (M4) or in first-time breeders (M5), successively. The former was selected (M4). Similar results were obtained using a first-order, autoregressive correlation structure. The youngest inexperienced birds have the lowest probability of success. Estimated probability of success: experienced breeders (95% C.I.: 0.42–0.49); 5-,6- and 7-year-old inexperienced breeders (95% C.I.: 0.32–0.47); 2-,3- and 4-year-old inexperienced breeders (95% C.I.: 0.22–0.30).
Table 4. . Test for an effect of age (A) and experience (E) on success probability
Description of the models
Goodness-of-fit test for M0: Hosmer and Lemeshow statistic = 9.91, d.f. = 8, P= 0.27.
Success probability. We first modelled the probability of success as a function of the number of years Y elapsed since first reproduction (included) (Table 5). This variable was retained in one case: birds with a reproductive life of 4 years. There is an important initial increase in success probability (between the first reproduction and the following attempts), but this is not observed in birds with a longer reproductive life (Table 5). We then compared the first reproduction to subsequent attempts (Table 5): this contrast was significant in birds with a reproductive life of 4 and 5 years. Estimates were consistent with those previously obtained: success probability was lower in the first breeding attempt. When Y was considered as a continuous variable (i.e. a linear trend in breeding probability with Y), none of the analyses provided evidence of an influence of that variable (these analyses are not presented here). Results were consistent when using an autoregressive correlation structure of order 1.
Table 5. . Test for an effect of the number of years Y elapsed since first reproduction (included) and of experience (first vs. subsequent attempts), on the probability of success (φ) in individuals whose reproductive life lasted n years. AIC or QAIC values for corresponding models
Influence of the number of years elapsed since first reproduction
Contrast between the first and subsequent breeding attempts
Reproductive life span (n)
test χ2 or F
LR test χ2 or F
H0 = no influence of experience on φ; Ha: experience influences φ.
Information criterion in bold: selected model.
Correlation structure: compound symmetry.
Number of individual histories used: 96, 76, 42, 41, 33, 14 and 13 for y= 2, 3, 4, 5, 6, 7 and 8, respectively.
A: estimated probability of success (φ) and 95% confidence interval, for birds in the yth year of reproductive life (φy): φ¯^= 0.14 (0.05–0.32); φ¯^2= 0.58 (0.41–0.73); φ¯^3= 0.23 (0.12–0.40); φ¯^4= 0.35 (0.26–0.46).
B: estimated probability of success (φ) and 95% confidence interval: φ¯^(first reproduction)= 0.14 (0.05–0.32); φ¯^(subsequent attempts)= 0.38 (0.28–0.49).
C: estimated probability of success (φ) and 95% confidence interval: φ¯^(first reproduction)= 0.13 (0.05–0.33); φ¯^(subsequent attempts)= 0.35 (0.26–0.46).
Breeding transition probability. The probability of nonbreeding in the following year was modelled as a function of Y (Table 6). We did not find evidence of such an effect. The contrast between the first breeding attempt and subsequent attempts was significant in birds whose reproductive life lasted 5 years: the probability of nonbreeding in the following year was higher in first-time breeders (Table 6). As above, we did not find evidence of an influence of Y on transition probability when Y was considered as a continuous variable (these analyses are not presented here). In addition, results were similar using a first-order, autoregressive correlation structure.
Table 6. . Test for an effect of the number of years Y elapsed since first reproduction (included)) and of experience (first vs. subsequent attempts), on the probability of breeding in the following year (ω) in individuals whose reproductive life lasted n years. AIC or QAIC values for corresponding models
Influence of the number of years elapsed since first reproduction
Contrast between the first and subsequent breeding attempts
Reproductive life span (n)
test χ2 or F
LR test χ2 or F
H0 = no influence of Y on ω; Ha: Y influences ω.
Information criterion in bold: selected model.
Correlation structure: compound symmetry.
Number of individual histories used: 68, 40, 40, 33 and 13 for y = 3, 4, 5, 6 and 7, respectively.
A: estimated probability of breeding in the following year (ω) and 95% confidence interval: ω¯^(first reproduction)= 0.66 (0.46 − 0.82); ω¯^(subsequent attempts) = 0.84 (0.74–0.91).
A third hypothesis can be put forward: the apparent influence of experience on survival and future breeding probability, observed when age and success are not controlled for, could reflect an underlying age-related change in the proportions of individuals of different quality in the population. A set of elements suggests that selection is likely to occur in that population. First, we did not find evidence of negative phenotypic relationships between reproduction and survival or between current and future reproduction. This supports the hypothesis of heterogeneity in individual quality, which is a prerequisite for selection to occur. This is consistent with previous studies on the same species (Coulson & Porter 1985; Coulson & Thomas 1985; Coulson 1988; Aebischer & Coulson 1990; Cam et al. 1998). Second, failed breeders have a lower survival and a lower probability of breeding in the following occasion, regardless of experience. Last, we found evidence that first-time breeders have a lower success probability. The higher proportion of failed breeders in first-time breeders might reflect a higher proportion of lower-quality individuals in this category.
Selection is not the only potential process explaining the improvement of performance in experienced breeders compared to first-time breeders. There is an initial increase in performance between the first breeding attempt and the following attempt in some subsets of individuals with similar reproductive longevity. This provides slight evidence that a process operating at the individual level contributes to explain the influence of experience on breeding activity and performance observed at the population level. However, this pattern is not observed in subsequent breeding attempts. In addition, results vary drastically among subsets of individuals with similar reproductive longevity. The majority of our analyses indicated that experience does not influence breeding performance within individuals. This process alone can probably not account for the results obtained at the population level.
Our results are consistent with the scenario described by Vaupel & Yashin (1985): patterns of variation within homogeneous categories can be very different from patterns observed at the population level. An increase in survival (and parameters positively correlated with survival, such as performance) can be observed in populations while there is no variation within homogeneous categories of individuals. This can also be observed when performance varies within categories without any systematic trend, or even deteriorates. The hypothesis of earlier disappearance of individuals with lower survival and performance and the hypothesis of improvement of performance within individuals are not alternative: results supporting one of them do not allow rejection of the other. Forslund & Pärt (1995) suggested that several processes operate simultaneously in populations. In the present case, patterns observed at the individual level cannot explain the pattern at the population level (but see Rockwell et al. 1993).
Considering what Vaupel & Yashin (1985) called ‘heterogeneity's ruses’ can be critical when testing the hypothesis of differential reproductive costs using distinct samples that might include different proportions of individuals of different quality; for example when conducting experiments (e.g. clutch or brood size manipulation) using data from different individuals (inexperienced and experienced individuals or individuals of different age). The selection process should also be considered when addressing variation in survival or reproductive performance at the population level. Heterogeneity in survival among individuals can lead to age- or experience-related patterns of variation in demographic parameters at the population level, apparently consistent with the predictions of various hypotheses based on the notion of reproductive costs or cumulative experience. Heterogeneity can also mask a decrease in survival in older individuals (Johnson, Burnham & Nichols 1986; Nichols, Hines & Blums 1997), which is an assumption of the hypothesis of long-term optimization in reproductive effort.
Data from individuals with similar reproductive longevity permit access to processes other than selection. Such an approach should be used in manipulative as well as observational studies to address processes operating within individuals. For fitness components that cannot be addressed that way, such as survival, we recommend investigation of phenotypic correlations between reproductive success and survival using an unmanipulated reference sample (e.g. Rockwell et al. 1993; Boyd et al. 1995). If results support the hypothesis of heterogeneity in the population, only development of statistical methods accounting for individual differences or use of quality as a covariate (i.e. an individual characteristic determined a priori) would provide means of disentangling the various processes responsible for improvement of fitness throughout reproductive life. Otherwise, several nonexclusive hypotheses have to be considered as possible explanations for observed patterns in demographic parameters. One of the major results of syntheses on long-term studies in birds is the very important disparity among individuals in lifetime reproductive success (Newton 1989). Although environmental stochasticity partly explains this disparity, heterogeneity in individual quality is likely to be a common feature in long-lived species (Newton 1989; McNamara & Houston 1992, 1996; in the kittiwake: Coulson 1968). Thus, selection is a process likely to be operating in most populations of long-lived species.
Many models developed to investigate age-specific reproductive strategies are based on the theory of optimization (Charlesworth 1980; Nur 1988; Stearns 1992; Seger & Stubblefield 1998). Heterogeneity in quality has long been recognized as one of the main obstacles to testing the predictions of these models using data from observational studies (Nur 1988; McNamara & Houston 1996). Incorporation of sources of heterogeneity in models is viewed as a promising means of progressing in the study of the evolution of life histories (McNamara & Houston 1992, 1996). Age-specific reproductive strategies might be optimal within quality groups (Nur 1988; McNamara & Houston 1992, 1996; Cichoń, Oljniczak & Gustafsson 1998; Pettifor, Perrins & McCleery 1998), while they appear suboptimal in heterogeneous groups such as populations. As an illustration, we could speculate that the higher probability of nonbreeding in first-time breeders observed in the present study supports the restraint hypothesis. Indeed, this group probably includes a higher proportion of lower-quality birds likely to incur higher costs for a given investment. In lower-quality birds, nonbreeding could be associated with higher survival and future breeding probabilities, but at the population level nonbreeding is associated with the lowest values for these parameters. The inference that nonbreeding reflects reproductive costs or is beneficial in terms of fitness would have to be restricted to lower-quality individuals and cannot be drawn at the level of a heterogeneous group.
Understanding age-related variation in life history traits requires access to patterns corresponding to homogeneous groups that can be considered as reliable descriptions of patterns at the individual level. In addition, developing state-dependant life-history models (McNamara & Houston 1992, 1996) and testing corresponding predictions using data collected in the wild requires increased stratification of data according to criteria influencing individual performance. This requires estimation of fitness components specific to quality categories (Cichońet al. 1998). From that perspective, it is necessary to increase the efforts devoted to identification of the variables influencing quality (McNamara & Houston 1992, 1996). The influence of heterogeneity on age-related variation in survival has long been recognized in human populations (e.g. in medical surveys) and has led to important methodological developments (Manton & Stallard 1981; Manton, Stallard & Vaupel 1981; Hougaard 1984; Manton & Stallard 1984). There is an urgent need for development of approaches to estimation of survival and breeding probabilities accounting for heterogeneity among individuals in animal populations (Rexstad & Anderson 1992; Burnham & Rexstad 1993) in situations where not all the individuals are resighted during sampling efforts.
We are grateful to the Conseil Général du Finistère and to the Société pour l’Etude et la Protection de la Nature en Bretagne for allowing us to work in the Nature Reserve of Goulien Cap-Sizun (Brittany, France) since 1979. We thank Bill Link and Jim Nichols for advising us for statistical analyses and Jim Hines for his help at various steps. We also thank Evan Cooch, Etienne Danchin and Roger Pradel for providing helpful comments on earlier versions of this manuscript and many other people for assistance with the fieldwork. Last, we thank a referee who provided constructive comments, but whose signature remained a mystery; we regret that we are unable to acknowledge the contribution of this person properly.
Received 17 March 1998;revisionreceived 20 July 1999