Integrating animal temperament within ecology and evolution

(a) Variation

We would like to encourage the reader to use specific, unambiguous tests designed clearly to address each of the five temperament traits defined in Section II.2. For instance, following the proposed terminology, a test measuring boldness ideally should avoid any component of novelty or should take into account the effects of the change in novelty on the expression of boldness. Our choice of restricting the terminology to the five traits described in Section II.2, is based on the importance of avoiding possible redundancy between the terms commonly used in studies on temperament (i.e., timid = shy; affable = sociable; see the review by Gosling, 2001), which may have been a constraint on the development of ecological research on temperament.

Measurements of temperament traits should be based on experiments that are designed to specifically exclude non-target behaviours. For instance, fear of novel objects or food should be measured in a familiar environment (e.g., a home cage) to ensure that the measured behaviour does not reflect exploration of the environment but rather behaviour directed to the novel challenge. In order to avoid correlations between behaviours solely caused by the experimental arrangement, standardized methods are required to measure different axes of temperament that may often be correlated (e.g., boldness, exploration, activity). For example, various studies have shown that ‘boldness’ and ‘exploratory behaviour’ are correlated (e.g., Huntingford, 1976b; Verbeek, Drent & Viepkema, 1994) and concluded that temperament is context-general (Kagan et al., 1988; Coleman & Wilson, 1998). Such correlations between different behaviours, however, may reflect side-effects of the experimental set-up, particularly if all behavioural tests were conducted in a novel environment. For instance, the correlation between predator-inspection behaviour, aggressiveness and exploratory behaviour in three-spined stickleback (Huntingford, 1976b) may not reflect inter-trait correlations but rather individual consistency in fear towards an unfamiliar aquarium (Maier, Vandenhoff & Crowne, 1988; Budaev, 1997).

We encourage researchers to consider more than one measure of each temperament category (Table 3), just as an organism’s body size can be measured using several measures. For example, one could measure boldness by using predator or handler tests following a trappability test, and then test for the redundancy between indices by estimating the strength of the correlation between these two measures. For example, early exploratory behaviour, a combined score for boldness and exploration in great tits (Drent et al., 2003), is genetically highly correlated (r= 0.85) with so-called risk-taking behaviour (van Oers et al., 2004a). Van Oers et al. (2004a) interpret this strong genetic correlation as evidence for strong associations between different personality traits, although one could argue that the strong genetic correlation suggests that these different behavioural tests are all measures of the same trait. If it became clear that the two indices are not related, we would have to expand our terminology. Alternatively, when the two indices are strongly related, the use of different tests would validate the measures, and therefore the use of a combined index (e.g., from a principal components analysis) would be preferred (for example Drent et al., 2003). Another issue is absence of variation in response to a test (Table 2). For example, in a novel-object experiment Eastern chipmunks (Tamias striatus) showed no reaction to a small ball (D. Réale, unpublished results). They continued their foraging routine, and foraged without exploring the ball or even pushed the ball away in order to collect the seeds underneath. When tested with a pair of wool socks, however, chipmunks showed longer and variable latencies to approach the food in the presence of the object. Both objects were “new” to the chipmunks, although given its shape and structure, the pair of socks may have carried information representing “predatory” cues for the chipmunks that a spherical, smooth ball does not. Some tests may also create a situation that is too overwhelming for all the individuals tested, thereby not showing any variation among them. This kind of situation may not be exceptional, but may be rarely reported by scientists, who tend to select tests that generate variation. Absence of variation could also be caused by the presence of genotype x environment (GxE) interaction with no variation at the specific ecological situation of the test (Table 2). This would imply that past selection pressures have eroded the genetic and phenotypic variance of the trait within the range of ecological conditions tested. Running the test in different ecological situations may be a way to reveal any GxE for the trait (see below).

(b) Repeatability (r)

Repeatability can be used as a first estimate of individual consistency in a trait (Boake, 1989; Falconer & Mackay, 1996). Repeatability illustrates how strong the individual consistency for the trait is in the population and is the proportion of phenotypic variation explained by the variation among individuals (Falconer & Mackay, 1996). It can be obtained by running an analysis of variance with individuals included as a fixed factor, when a minimum of two measures of the trait for each individual have been taken (Falconer & Mackay, 1996). It can also be obtained from a mixed model where individual is a random effect (Diaz-Uriarte, 2002). Repeatability is a unit-free standardized measure that can readily be compared among samples (e.g., populations, species). Low repeatability for a temperament trait could be observed if, for example, the conditions chosen for the experimental test do not generate behavioural variation (i.e., all individuals react in the same way to the experimental set-up; Table 2) or by high within-individual variation relative to among-individual variation. When we estimate repeatability we assume that the environment in which the replicated measures have been taken is constant. Though seemingly similar to the experimenter, however, successive replications of a test for the same individual may differ due to micro-environmental effects, thereby changing the expression of the behaviour (Henderson, 1990). Furthermore, learning and habituation may affect the novelty aspect of an open-field test replicated several times (Greenberg & Mettke-Hofmann, 2001; Wahlsten, 2001). For example, great tits were quicker to explore a novel environment when confronted with it a second time and this effect is stronger when the animal is tested shortly after its initial test (Dingemanse et al., 2002). In the same way, the latency of an individual to enter a trap may decrease with repeated exposure because the trap loses its novel dimension. Habituation may be circumvented by exposing subjects to different types of novel objects (e.g., Verbeek et al., 1994), although repeated exposures to different novel objects may also lead to habituation, if the animal can habituate to novelty itself, as opposed to particular novel properties. The test will thus no longer represent the same challenge to the subject. Disentangling the effects of novel and non-novel components of an object/habitat on the reaction of an individual implies that studies on temperament should not restrict measures to one condition, but should also integrate environmental variation and temporal dimensions (e.g., with learning; Seferta et al., 2001). A reaction norm approach could be helpful to circumvent such difficulties: it represents the phenotypic variation of genotypes (here as individuals) across an environmental gradient (here as novelty; Fig. 3). By running a mixed-model analysis using several replicated measures of a trait for each individual it is possible to estimate individual variation for that trait (individual is used as a random effect), and the general effect of the environment (for example the rank of the trial is used as a fixed effect). In addition, an interaction between trial and individual would indicate that individuals differ in the way they respond to the experimental arrangement with experience (Fig. 3). The absence of a significant interaction would indicate that all the individuals habituate at the same rate. This absence of interaction, combined with significant individual variation for the temperament trait, also indicates a potential behavioural syndrome where individual experience is the environmental situation (Sih et al., 2004b). The reaction norm approach can be generalized to any other gradient (e.g., importance of the risk; number of competitors; age of the individual; seasonal variation).

(c) Heritability

Repeatability is assumed to set an upper limit to heritability (Boake, 1989). High repeatability and individual consistency can also come from several other non-genetic sources: maternal effects; common environmental effects and individual histories; experience; learning; and the environmental conditions specific to each individual (Falconer & Mackay, 1996). Repeatability provides initial evidence that among-individual variation is caused by factors intrinsic to the individual, but does not allow for separation of the genetic and non-genetic components of the variance in a temperament trait. Heritability (h²) is a standardised index of the proportion of phenotypic variance explained by additive genetic variance of the trait, and represents its evolutionary potential (i.e., the ability of a trait to change its mean and distribution across generations, as a result of selection; Falconer & Mackay, 1996; Roff, 1997; Lynch & Walsh, 1998). Because temperament corresponds to individual behavioural differences that are consistent over time and across situations, heritability should occupy a central position in the study of temperament. Most studies to date have been conducted under laboratory conditions and in many cases were explicitly aimed at excluding rather than studying environmental sources of variation. Other sources of variation of temperament traits, particularly ontogenetic changes or those induced by the environment (Markowitz et al., 1998; Lowe & Bradshaw, 2001; Maestripieri, Tomaszycki & Carroll, 1999) have largely been neglected. Examples of non-genetic sources of between-individual variation are maternal influences on offspring temperament mediated via variable transmission of androgens in the yolk of bird eggs (Schwabl, 1993) or in the effects of intrauterine testosterone on the behaviour and life histories of female rodents (Compaan et al., 1992; Clark & Galef, 1995). Examples of within-individual sources of variation in temperament include seasonality in birds (Dingemanse et al., 2002), ontogenetic changes in behaviour of birds (Carere et al., 2003), laboratory rodents (Koolhaas et al., 1999) and wild marmots (Armitage & Van Vuren, 2003), and learning in fish (Huntingford et al., 1994). If not considered, all these effects may bias the estimate of heritability. For example, individual variation in the rate of habituation will generate uncontrolled GxE interaction, compromising estimation of temperament trait heritability.

Classical breeding experiments and analyses of variance (ANOVAs) (see Falconer & Mackay, 1996; Roff, 1997; Lynch & Walsh, 1998), or artificial selection experiments (Trut, 1999; Drent et al., 2003), provide a way of measuring the genetic basis of a trait in the laboratory. A recent statistical method, the “animal model”, is particularly promising for studies of temperament traits in natural populations (Réale, Festa-Bianchet & Jorgenson, 1999; Kruuk, 2004; see Bell, 2005, for application to temperament) as it is appropriate for complex pedigrees (Lynch & Walsh, 1998). Methods that permit us to estimate which part of the phenotypic variance in a temperament trait is caused by additive genetic, maternal variance, common environmental or permanent environmental effects would provide valuable information on the mechanisms at the origin of temperament differences within a population. For instance, a strong permanent environmental variance of a temperament trait would indicate that a large portion of the individual variation for that trait was generated during the early life of the individual, but has neither a genetic basis, nor is the result of maternal influence. In the same way, a strong common environmental effect would illustrate the possibility that the environment in which an individual is reared influenced its temperament. When the data collected on a temperament trait involve replicated tests, seasonal variation, ontogenetic effects, or environmental differences, a mixed model with genotype as a random effect and with trial, season, age, or environment as a fixed effect will allow for accurate assessments of heritability, phenotypic plasticity, and GxE of that trait. Mixed-model ANOVAs with full-sib or a half-sib design (Fry, 1992), or with recent advances in random regression using the animal model (Schaeffer, 2004) could be used for estimating these parameters.

(a) Biological and ecological validity

For logistical and methodological reasons, temperament traits should be measured in an experimental context as we can rarely measure them by direct observation. Furthermore, the need for large sample sizes for estimation of heritabilities or selection pressures, means tests should be rapid and easy to perform. Direct observation of individuals and their ranking according to a predetermined scale has, however, been used by psychologists (Gosling, 2001) but can be criticised because of the subjectivity of the interpretation of observations. The experimental approach is open to criticism since it places an animal in a situation that is irrelevant to its natural conditions. Nevertheless, in the experimental approach, we are more interested in differences between individuals in a broader context, than the behaviour in the experimental context per se. For instance, we assume that the behaviour of the mouse in an open field reveals its reactions to a new and open environment and thus its exploratory tendencies.

When possible, testing the biological and ecological validity of the test by looking at the relationship between the temperament trait and physiological traits (i.e., biological validation) or ecologically important traits (i.e., ecological validation) is recommended. Biological validation allows determination of the mechanisms responsible for the variation in temperament. There is extensive evidence that behavioural variation in temperament may reflect underlying hormonal and neuro-endocrine variation among individuals (Bohus et al., 1987; Boissy, 1995; Koolhaas et al., 1999; Kagan & Snidman, 2004; Groothuis & Carere, 2005).

Ecological validation is important to temperament research, because it provides a way to integrate temperament traits in the ecological framework, and put temperament at the interface between ecological factors and selection. Results supporting this link will encourage ecologists to consider temperament in their studies. The link between a behavioural measure of temperament and an ecologically important trait can be studied by using phenotypic correlations between the two traits. For example, dispersal distance in the wild was related to experimental tests of exploratory behaviour in great tits (Dingemanse et al., 2003) and killifish Rivulus hartii (Fraser et al., 2001). It is also possible to compare an individual’s behaviour measured in an experimental test (e.g., aggressive reaction to a mirror image) with observations in the same context in a more natural situation (e.g., rate of aggressive interactions). Such validation may yield intriguing insights as shown by a study on great tits where the correlation between dominance in the wild and experimental tests of exploratory behaviour (which predicts dominance in the laboratory: Verbeek et al., 1994, 1999) was positive for territorial adults but negative for non-territorial juveniles (Dingemanse & de Goede, 2004). Experimental validation is also possible by manipulating the phenotype (e.g., through hormonal implants) of some individuals and by comparing their behaviour with a control group (i.e., phenotypic engineering: Ketterson & Nolan, 1999; see below). Alternatively, ecological validation can be obtained by estimating the genetic correlation between the measured traits and the ecologically important trait. An array of methods is available to estimate genetic correlations between temperament traits [breeding designs: Lynch & Walsh (1998); artificial selection: van Oers et al. (2004a); pedigree analyses and the animal model: Kruuk (2004)].

(b) Behavioural syndromes

The concept of genetic correlation is central to defining a trait and whether two measurements can be considered as one unique trait or as two different traits. The genetic correlation between two characters or between measures of a character in two different environmental conditions gives a standardised index of the degree to which the covariance of the two characters is governed by a common set of genes, generally caused by pleiotropy or linkage disequilibrium (Via & Lande, 1985; Roff, 1997; Lynch & Walsh, 1998). From a mechanistic point of view, a strong positive genetic correlation would indicate common genetic and physiological pathways between the two traits (Crusio, 2001; Sih et al., 2004a, b). Alternatively if the correlation is negative, it would reflect the competition for resources allocated to each trait (Roff, 1997). From an evolutionary point of view, a strong genetic correlation illustrates past co-evolution between the traits (Roff, 1997; Lynch & Walsh, 1998), and will prevent independent evolution of both traits (Roff, 1997; Lynch and Walsh, 1998; Sih et al., 2004a, b). An evolutionary and ecological approach to the study of temperament will thus include estimates of phenotypic and genetic correlations between the traits (Sih et al., 2004b).

Genetic and phenotypic correlations are appropriate for testing for the presence of a behavioural syndrome. On the other hand, a behavioural syndrome can still occur even in the presence of GxE effects. For example, a weak GxE interaction corresponds to a strong genetic correlation across the two environments, and a genetic correlation of 1 can still be found with GxE effects if genetic variances differ in the two environments. Evidence of genetic correlation between temperament traits and other life-history or morphological traits (Table 4) indicates the ecological and evolutionary importance of temperament traits.

The extent to which temperament traits are independent or form part of a behavioural syndrome is a controversial and important issue (Wilson et al., 1994; Koolhaas et al., 1997, 1999; Sih et al., 2004a, b). It is critical to the understanding of the phenotypic organisation, genetic structure and evolutionary history of temperament traits and to determining the generality of findings across contexts, populations, and species. On one hand, implicit in the concepts of behavioural syndromes and coping styles (Koolhaas et al., 1997, 1999; Sih et al., 2004b; Groothuis & Carere, 2005) is the existence of a set of related behavioural and physiological traits. On the other hand, this integrated view of temperament has been challenged by the hypothesis that divergent and context-specific selection pressures could favour the evolution of context-specific temperament traits, with weak phenotypic or genetic correlations across contexts (Wilson et al., 1994; Wilson, 1998). Both hypotheses have been supported by empirical studies on various models (presence of a behavioural syndrome: Riechert & Hedrick, 1993; Benus & Rondigs, 1996; Groothuis & Carere, 2005; Sih et al., 2004a; absence of a behavioural syndrome: Spoolder et al., 1996; Coleman & Wilson, 1998; Réale et al., 2000; D’Eath & Burn, 2002; Bell, 2005).

The nature of the outcome may depend on several possible explanations. In support of the behavioural syndrome hypothesis, a common neuroendocrine system is known to mediate a whole suite of behaviour patterns (Boissy, 1995; Bucan & Abel, 2002). Furthermore, alternative coping strategies result from individual differences in reactivity of the sympathetic nervous system (high in active copers) and the hypothalamic-pituitary-adrenal axis (high in passive copers) in response to stress (Koolhaas et al., 1997, 1999), indicating that fundamental differences in neuroendocrinological and behavioural organisation underlie variation in animal personality (Bohus et al., 1987; Boissy, 1995; Groothuis & Carere, 2005). This gene-neurophysiology-behaviour pathway should be illustrated by strong phenotypic and genetic correlations between the temperament traits studied whether these are different traits or a given trait measured in different contexts (Table 4; see also Sih et al., 2004a, b). The reason why behavioural traits and temperament are mediated by a common neuroendocrine system is not well understood. We can advance three reasons: (1) the evolution of a large number of neurophysiological systems would be costly, thereby creating a strong constraint on the diversity of possible behavioural responses and thus a strong integration between temperament traits. Overall, a few neurophysiological systems appear to play an equivalent role in many different invertebrate or vertebrate organisms (e.g., serotonergic and dopaminergic systems; Koolhaas et al., 1999; Gosling 2001; Libersat & Pflueger, 2004); (2) the homeostasis of the organism necessitates feedbacks and controls at the neurophysiological level, and this can only be done with a limited set of neurophysiological systems. The evolution of gene-neurophysiology-behaviour pathways would thus be limited to a small number of well-integrated systems; (3) traits involved in a given function co-evolve, and are thus characterised by strong genetic correlations between them, suggesting the existence of functional, developmental, and evolutionary modularity of organisms (Cheverud, 1996; Wagner, 1996; see also the literature on cognitive modularity: Reader, 2006).

Explanations for the absence of behavioural syndromes are multiple. Such empirical negative findings may result from methodological limits to these studies; the fact that the relationship between the two traits has been cancelled out by their common relationship with another variable (Shipley, 2000), or from a lack of statistical power. As Wilson pointed out (Wilson et al., 1994; Wilson, 1998), selection could lead to the independent evolution of temperament traits according to the context in which they are measured. Genetic correlation is assumed to constrain the independent evolution of the two traits, and a behavioural syndrome should in principle be constant within and across populations (Sih et al., 2004b; Bell, 2005). Positive correlations between temperament traits, however, can be observed within a population but not across populations because of long-term differences in the selection regime of two populations on a combination of temperament traits (Bell, 2005), founder effects that create linkage disequilibrium at the origin of the correlation (Whitlock, Phillips & Fowler, 2002), gene flow, or genetic drift (Armbruster & Schwaegerle, 1996).

(a) Measuring selection

Natural selection can be directly measured by the covariance between traits and fitness, allowing estimation of both the strength (Lande, 1979; Lande & Arnold, 1983) and shape (i.e., directional, disruptive, stabilizing, correlational) of selection (Brodie, Moore & Janzen, 1995). Such estimation models can be used to assess the way selection acts on correlated characters (Lande & Arnold, 1983; Kingsolver et al., 2001), and so are particularly suitable for the study of selection on temperament traits (Réale et al., 2000). To date, only two field studies have used such an approach in wild populations, measuring both heritability (see Réale et al., 2000; Dingemanse et al., 2002, 2004; Dingemanse & Réale, 2005) and components of fitness under different environmental conditions (Réale & Festa-Bianchet, 2003; Dingemanse et al., 2004, Both et al., 2005). In bighorn ewes Ovis canadensis, weaning success increased with boldness (Réale et al., 2000). Furthermore, boldness was related to survival during years of high predation but not during years of low predation by cougars Puma concolor (Réale & Festa-Bianchet, 2003). In great tits Parus major, slow-exploring females had the lowest probability of nest failure and assortative pairs of extreme phenotype (i.e., either slow-slow or fast-fast pairs) produced heaviest offspring in all of four years (Both et al., 2005). Whilst the direction of selection on temperament during the breeding phase did not differ between years (Both et al., 2005), outside the breeding season it fluctuated between years, sexes, and components of fitness (Dingemanse et al., 2004). Selection in this case was coincidental with the occurrence of masting by beeches Fagus sylvaticus, a factor that affects competitive regimes of these birds (Perdeck, Visser & Balen, 2000). In both studies, selection pressures varied according to environmental conditions, although definitive identification of the important environmental changes would require sufficient replication to show statistically the link between selection and environment. These examples emphasize the importance of long-term longitudinal studies to measure selection: individuals encounter many different social and physical environments during their lives. These studies also illustrate that single components of fitness may not accurately predict an individual’s overall fitness, as the direction of selection may differ between various components of fitness (e.g., survival versus reproductive output).

Manipulation of the environment or phenotypic engineering could facilitate the study of selection. This could be done by measuring the survival and reproductive success of individuals whose phenotype has been manipulated; such as using agonists or antagonists of some neurotransmitters involved in temperament variation (e.g., tryptophan and fluoxetine increase and cyproheptadine and fenfluramine decrease 5-hydroxytryptamine (5-HT) activity, respectively, affecting dominance acquisition in male vervet monkeys Cercopithecus aethiops; Raleigh et al., 1991). The path analysis method proposed by Scheiner, Mitchell & Callahan (2000), a method derived from phenotypic selection analysis, has great potential to estimate selection pressures on a set of hierarchically organised temperament traits.

The phenotypic selection approach may offer another way to investigate whether or not selection can act on a whole suite of traits, and lead to the co-evolution of temperament traits. In bighorn ewes Ovis canadensis, selection seems to disfavour the combination of low levels of boldness and of docility, although the fitness consequences of the interaction between the two traits have not been tested (Réale et al., 2000; Réale & Festa-Bianchet, 2003). Furthermore, in three-spine stickleback Gasterosteus aculeatus the phenotypic correlation between aggressiveness and boldness differed between populations (Bell 2005) and is higher for populations exposed to predators (N.J. Dingemanse, J. Wright, A. Kazem, D. Thomas, R. Hickling & N. Dawnay, in preparation), suggesting correlational selection on those traits. Hence, future fitness studies should attempt to address under which conditions natural selection favours the evolution of correlated suites of behaviours. Focusing on the fitness consequences of a single temperament trait without considering interaction among traits may give misleading results.

Selection studies on temperament traits could provide several insights. For example, we still do not know what components of fitness (survival at different stages, mating success, fecundity, reproductive success) are affected by temperament traits, or how and in what conditions temperament phenotypes are advantageous or disadvantageous. Studies in the future may help us understand if selection favours plasticity or canalisation of temperament traits. Can we observe antagonistic selection according to sex, and what are the consequences of such a selection pattern on the maintenance of variance in temperament traits? More information on the role of temperament as a factor in sexual selection is needed.

(b) Adaptive explanations for the maintenance of variance in temperament traits

The relative importance of the different processes for the persistence of within-population variance in temperament traits (i.e., selection-mutation balance, spatiotemporal or frequency-dependent selection, selection-migration balance) is largely unknown (Wilson, 1998). Theoretical modelling suggests that fluctuating selection pressures can help to maintain quantitative genetic variation at higher levels than expected from mutation alone (Burger & Gimelfarb, 2002). Game theory and the evolutionarily stable strategy approach provide a framework to assess the costs and benefits of alternative behaviours (Magurran, 1993; Dall et al., 2004). In the case of temperament traits, both frequency- and density-dependent processes are likely to contribute to the maintenance of high levels of variation (Wilson et al., 1994; Dall et al., 2004), because the social environment is probably an important determinant of the relation between temperament and fitness. For instance, bird studies have shown that individuals differ in social aggression and foraging behaviour (Verbeek et al., 1996; Marchetti & Drent, 2000), with the pay-offs of alternative strategies frequency dependent (Barnard & Sibly, 1981; Koops & Giraldeau, 1996). Such frequency-dependent games could also involve cooperation instead of competition, as shown in white-throated sparrows, Zonotricia albicollis, a species where a chromosomal inversion caused a polymorphism (resulting in ‘tan-striped’ types that are less aggressive than ‘white-striped’ types). Ninety-eight per cent of birds pair disassortatively (Houtman & Falls, 1994), partly because certain pair combinations make better teams during chick-feeding (Knapton & Falls, 1983). The role of social organisation in the maintenance of individual variation could be addressed indirectly by comparing phenotypic variation in behaviour among social and solitary species. Models also predict that temporal variation in selection can maintain a mix of phenotypically flexible and inflexible genotypes (Wilson & Yoshimura, 1994). Many species live in temporally fluctuating environments, and it is an open question as to why some individuals and species are relatively inflexible in their behaviour instead of showing higher plasticity (Wilson et al., 1994). An important question is to what extent variation in temperament and behavioural syndromes, and limits and costs associated with phenotypic plasticity constrain or reflect selection on behavioural flexibility (Wilson et al., 1994; Koolhaas et al., 1997, 1999; Dingemanse et al., 2004). Importantly, note that the adaptive explanations for the maintenance of variation in temperament traits given above only explain the co-existence of different temperament types given that individuals show consistent behaviour. That is, none of the afore-mentioned explanations explain why individuals are consistent over time, contexts, or situations. There is thus a strong need for a theoretical framework that explains at the same time why different temperament types coexist and why individuals have temperament types (Dall et al., 2004).

Comparative studies of the amount of genetic variation between trait categories provide an alternative way to investigate the effect of past selection pressures on temperament traits relative to other trait categories. Strong and persistent directional selection on a trait should erode the additive genetic variance, relative to the dominance variance, and thus decrease the trait’s heritability (Mousseau & Roff, 1989; Crnokrak & Roff, 1995; Merilä & Sheldon, 1999; Stirling, Réale & Roff, 2002). Comparing the average value of standardized indices of variation such as heritability, coefficient of additive genetic and residual variance (Houle, 1992; Merilä & Sheldon, 1999; Stirling et al., 2002), or dominance variance (Crnokrak & Roff, 1995) for temperament traits relative to other types of traits could therefore provide valuable information about the selection regime of temperament traits. Temperament traits generally show significant heritability, averaging 0.31 (Table 5), a value similar to those observed for behaviour or life-history traits (Stirling et al., 2002). Variance is not only affected by selection, however, but also seems to depend on GxE interaction or epistasic effects (Merilä & Sheldon, 1999; Stirling et al., 2002). More estimates of standardized indices of variation for more temperament traits are needed before we can provide any detailed study of the relationship between temperament and fitness. Studies on great tits Parus major provide an example of how the combination of fitness and quantitative genetics studies can provide clues towards the history of selection on temperament traits in natural populations. Studies in the wild showing fluctuating selection on temperament traits (Dingemanse et al., 2004), and back-cross experiments based on selection lines indicating high levels of both additive and dominance variance (Drent et al., 2003; van Oers et al., 2004c) both suggest the absence of strong consistent directional selection on this type of trait (Mousseau & Roff, 1989; Roff, 1997).

(a) Intra-species population comparison

A classical approach to the study of adaptation in the wild is the comparison of populations living in different environments, and the correlation between population phenotypes and environmental characteristics (Mousseau et al., 2000; Reznick & Travis, 2003). This method has provided interesting results on temperament and behavioural adaptation in the wild (Arnold & Bennett, 1984; Magurran, 1993; Huntingford et al., 1994; Clarke & Boinski, 1995). For example, fish populations living in sites with high risks of predation show higher boldness, predator inspection scores, and faster escape response than those living in safer sites (Huntingford, 1982; Magurran, 1993; Huntingford et al., 1994; O’Steen, Cullum & Bennett, 2002). Similarly differences in habitat have been shown to shape variation in boldness and aggressiveness in desert spiders, Agelenopsis aperta (Riechert & Hall, 2000). Island populations of birds generally show lower neophobia and higher exploration than mainland bird populations (Greenberg & Mettke-Hofmann, 2001). Population comparison has a limit: different environmental conditions experienced by the populations may counteract genetic effects on inter-population differences in phenotypic expression (Conover & Schultz, 1995). One way of disentangling environmental from genetic effects is to raise individuals from different populations in similar environmental conditions where all the observed variation among populations is expected to be genetic in origin (the common-garden approach; Mousseau et al., 2000). The common-garden approach can also be extended to study evolutionary constraints or trade-offs; one could do so by looking at the correlation between two phenotypic traits measured in different populations (Roff, Crnokrak & Fairbairn, 2003). An alternative is reciprocal transplants of individuals between the two populations (Conover & Schultz, 1995). Reciprocal transplants offer the possibility to look at the interaction between genes and the environment on a trait’s phenotypic expression (Carroll et al., 2001), and allow a measurement of selection by comparing the fitness of indigenous and transplanted individuals (Riechert & Hall, 2000).

(b) Interspecific comparative studies

The most common use of the comparative approach is to test whether a trait is adapted for a particular function (Martins & Hansen, 1997). The association between a temperament trait and other traits or ecological conditions can be studied from the information provided by the measure of those traits for a set of species (Webster & Lefebvre, 2001; Marples, Roper & Harper, 2003). Using this approach, Greenberg (2003) tested the “Neophobia Threshold Hypothesis” the idea that specialist species are more neophobic than ecologically plastic species, because neophobia reduces the probability of investigating and incorporating a new resource in the repertoire of a species. In support of this hypothesis, the specialized chestnut-sided warbler Dendroica pensylvanica is more neophobic than the more plastic bay-breasted warbler Dendroica castanea (Greenberg, 2003). This hypothesis was confirmed with wild-caught individuals from two other species, although work on naïve captive-reared birds provided contrary results (Greenberg, 2003).

Comparison of more than two species increases the power of such approaches. Comparative methods, however, have rarely been used to study temperament (but see Mettke-Hofmann et al., 2002; Richardson, 2001). In a pioneering study in zoo animals, Glickman & Sroges (1966) placed a diverse set of novel objects in the cage of over 200 species of mammals and reptiles, and used this information to describe patterns of object exploration within and between species. Extending this approach, Mettke-Hofmann et al. (2002) examined the link between ecological factors and neophobia and exploration in 61 parrot species in captivity. Neophobia was higher for insectivorous than for folivorous birds, potentially reflecting the relatively high costs of approaching novel food items for insectivorous species compared to folivorous ones. Exploration was uncorrelated with neophobia, and was higher in species that live on forest edges or on islands, and that eat nuts, suggesting that exploration is primarily influenced by the value of information. Exploration was also more intense in migratory than in resident species (Mettke-Hofmann et al., 2004).

The interest of the comparative approach goes beyond the test of the adaptive hypotheses: one could examine the functional relationships and trade-offs between two traits, the existence of constraints (e.g., the importance of phylogenetic inertia in the variation of a trait), and the ecological and evolutionary implication of temperament (e.g., the role of temperament in the evolutionary diversification in a taxon).

The comparative approach can therefore be successfully applied to a variety of questions on temperament, although it is not exempt from difficulties. A major stumbling block is the quantification of temperament traits in species with different lifestyles: how should results from novel object tests conducted on the house mouse, the three-toed sloth or the Asian elephant be compared? Although one can in principle apply the same method to quantify a temperament trait for a set of species, whether the measure is actually comparable across species is not so straightforward. Novel object experiments should take account of the variability of habitat preference among species; a standardized novel test on the ground, for example, may lead to differences between terrestrial and arboreal species, as the latter species will be more reluctant to approach and explore novel objects on the ground. One solution to this problem is to compare temperament in closely related species, based on their physiological, morphological and ecological similarities. However, this decreases the generality of the results, and potentially removes biologically meaningful variation in the studied trait. Alternatively, one could use a multiple regression approach that includes potentially confounding effects as covariates (e.g., body size or basal metabolic rate). Another option would be to use physiological surrogates of temperament traits (e.g., monoamine levels), provided that among-species variation in these traits is higher than within-species variation. Although these measures are indirect, they are likely to be less subject to observational error and context-dependent biases (Kamil, 1994).

Species may resemble one another in their temperament, not because of independent, convergent evolution in response to specific ecological conditions, but instead because of common ancestry. Modern techniques use phylogenetic information to deal with this problem of non-independence (Felsenstein, 1985; Harvey & Pagel, 1991; Garland et al., 1993; Garland, Bennet & Rezende, 2005). Another potential danger of the phylogenetic-based methods is that, like any correlational approach, the association between two variables may be caused by their common link with a third variable (Bennett & Owens, 2002). Consequently, a proper comparative analysis of temperament traits should previously have identified and controlled for potential confounding factors that could inflate or obscure the studied relationship. Furthermore, it is essential to estimate the within-species consistency in the trait (i.e., the trait has to be species-specific) by running mixed models in which species are coded as a random factor. The problem of establishing a causal link between the variables under study may in part be overcome with methods that allow estimation of ancestral states (Harvey & Pagel, 1991). Alternatively, Hansen’s adaptation method may be used to estimate the impact of a specified environment on the evolution of a given trait (Hansen, 1997) and path analysis may help establish the most likely causal scenarios (Li, 1975). These techniques can potentially be applied to study the causal link between a temperament trait and another trait or an ecological factor (e.g., neophobia and dietary generalism).