The repeatability of behaviour: a meta-analysis

Abstract

There is increasing interest in individual differences in animal behaviour. Recent research now suggests that an individual’s behaviour, once considered to be plastic, may be more predictable than previously thought. Here, we take advantage of the large number of studies that have estimated the repeatability of various behaviours to evaluate whether there is good evidence for consistent individual differences in behaviour and to answer some outstanding questions about possible factors that can influence repeatability. Specifically, we use meta-analysis to ask whether different types of behaviours were more repeatable than others, and if repeatability estimates depended on taxa, sex, age, field versus laboratory, the number of measures and the interval between measures. Some of the overall patterns that were revealed by this analysis were that repeatability estimates were higher in the field compared to the laboratory and repeatability was higher when the interval between observations was short. Mate preference behaviour was one of the best studied but least repeatable behaviours. Our findings prompt new insights into the relative flexibility of different types of behaviour and offer suggestions for the design and analysis of future research.

Within the field of animal behaviour, there is growing interest in consistent individual differences in behaviour (Dall et al. 2004; Sih et al. 2004a, b; Dingemanse & Reale 2005; Bell 2007; Reale et al. 2007). Accumulating evidence from a wide variety of species suggests that some individuals are consistently more aggressive, more exploratory, or more bold than other individuals and that these consistent individual differences in behaviour are often heritable (Boake 1994; Stirling et al. 2002; Kolliker 2005; van Oers et al. 2005) and related to fitness (Dingemanse & Reale 2005; Smith & Blumstein 2008). However, to date, the published data have not been summarized in a way that allows us to assess the evidence for consistent individual differences in behaviour and to explain why the magnitude of individual differences is greater in some studies compared to others.

Many studies over the past several decades have already quantified consistent individual differences in behaviour by measuring the behaviour of individuals on more than one occasion. A variety of statistics have been used to estimate behavioural consistency such as the product moment correlation or the Spearman rank correlation, but the most widely used statistic is the intraclass correlation coefficient, which estimates repeatability (Hayes & Jenkins 1997). Repeatability is the fraction of behavioural variation that is due to differences between individuals. Formally, repeatability is $r=s_A^2{s}^2+s_A^2$ where $s_A^2$ is the variance among individuals and s² is the variance within individuals over time. Behaviours that show relatively low within-individual variance compared to high among-individual variance are more repeatable. In other words, when individuals behave consistently through time and when individuals behave differently from each other, then the behaviour is repeatable. In the past, most studies measured repeatability as a first step towards studying the genetic basis for a behaviour in order to set an upper bound to heritability (Boake 1989; Dohm 2002). A different rationale for estimating repeatability is to assess interobserver reliability and the internal consistency of an instrument, (Hoffmann 2000).

From a different perspective, the large collection of repeatability estimates provides an opportunity to evaluate whether there is good evidence for consistent individual differences in behaviour and to determine whether there are systematic factors that can explain variation in behavioural consistency. Therefore summarizing this literature (previously reviewed in part in: Boake 1989; Hayes & Jenkins 1997; Forstmeier & Birkhead 2004) will provide a strong foundation for moving the study of animal personality forward.

Here, we perform a meta-analysis of the large number of published estimates of repeatability that are based on observations of a single behaviour measured on the same individuals on more than one occasion. Although closely allied with concepts of behavioural syndromes (Sih et al. 2004a), temperament (Reale et al. 2007), personality (Gosling 2001) and coping styles (Koolhaas et al. 1999), all of which generally refer to behavioural consistency through time and across situations, repeatability is more restrictive than these concepts because it ideally refers to consistency of a particular behaviour through time, not necessarily behavioural consistency across situations or contexts. However, in many cases, the specific environmental situations in which a behaviour is being measured are not known. As a result, repeatability estimates reflect both consistency through time and consistency across unmeasured situations (Martin & Reale 2008). Obviously, using a similar framework to assess the evidence for behavioural correlations across contexts is a promising subject for future meta-analyses.

In addition to assessing the claim that individual differences are common, we wish to know whether there are generalizations that can be made about the factors influencing repeatability. We perform an exploratory analysis to address the following questions.

Are Certain Types of Behaviour More Repeatable Than Others?

Studies have estimated the repeatability of behaviours ranging from mate preference to exploratory behaviour to parental behaviour. Therefore, we have an opportunity to ask whether certain types of behaviour are more repeatable than others. One prediction is that behaviours that are more sensitive to the environment (more plastic) are less repeatable. For example, we might assume that behaviours under morphological or physiological constraint should be relatively stable compared to behaviours influenced by energetic needs or the immediate social environment (Castellano et al. 2002; Smith & Hunter 2005). However, if all individuals respond to the environment in a similar way, the behaviour can still be repeatable despite this plasticity. Instead, repeatability estimates are especially affected by individual*environment interactions, or when individuals respond differently to the environment (Nussey et al. 2007; Martin & Reale 2008). Therefore comparing the repeatability of different types of behaviour has the potential to reveal new insights about the flexibility or canalization of different types of behaviour.

Are Certain Taxa More Repeatable Than Others?

Reviews of heritability estimates have found strong taxonomic differences (Mousseau & Roff 1987). Among vertebrates, for example, the heritability of morphological traits is significantly lower for ectotherms than it is for endotherms (Mousseau & Roff 1987), perhaps because ectotherms are more influenced by their environment. Here, we follow Mousseau & Roff’s lead and test whether the same pattern applies to repeatability. We compared patterns of repeatability variation within four major phylogenetic groupings: invertebrates versus vertebrates and endothermic vertebrates versus ectothermic vertebrates. The invertebrate–vertebrate comparison allows us to evaluate the suggestion that the behaviour of taxa with less flexible nervous systems is less plastic, leading to higher repeatability estimates for invertebrates.

Does Repeatability Decrease with the Interval Between Observations?

From a genetic perspective, repeatability might decrease with the interval between measurements because the ‘same’ phenotypic trait may be influenced by different sets of genes at different ages. Therefore increasing the interval between measurements should decrease repeatability of the phenotypic traits because the two measures do not represent exactly the same trait at the genetic level.

Environmental effects might also cause repeatability to decrease with the interval between observations. For example, when the interval between observations of behaviour is short, it is likely that the animals are of similar state (hunger, size, age, condition, dominance, etc.) during both observations and are experiencing similar environments. For example, we might expect reproductive effort in birds to be more repeatable within broods rather than across seasons (Potti et al. 1999; Moreno et al. 2002). In contrast, when the interval between observations is long, there is more opportunity for developmental change; individuals are more likely to undergo dramatic change such as sexual maturity or a niche shift over a longer period of time. Indeed, consistency decreases with the interval between observations in humans (Roberts & DelVecchio 2000) and great tits, Parus major (Dingemanse et al. 2002). Published estimates of repeatability have used a wide variety of intervals between measurements; therefore, they provide an opportunity to test this intuitive suggestion.

Does Repeatability Increase with the Number of Observations Per Individual?

Several studies have measured the same individuals more than two times to calculate repeatability. On one hand, increasing the number of measurements per individual can decrease the measurement error associated with each observation, and therefore might increase repeatability (Hoffmann 2000). On the other hand, when individuals are measured repeatedly, they might habituate to the behavioural assay and become less responsive, or alternatively, become sensitized (Martin & Reale 2008). It is of practical importance to evaluate the relationship between the number of measures and repeatability for the design and analysis of future experiments (i.e. if there is much to be gained by measuring individuals more than twice; Adolph & Hardin 2007).

Does Repeatability Vary Among Age Groups?

In humans, behavioural consistency increases with maturity (Roberts & DelVecchio 2000); older people behave more consistently than younger ones, perhaps because the cumulative experience of the environment leads to increasing consistency with age. Other mechanisms that could cause repeatability to increase with age are the process of consolidated identity or reputation (Roberts & DelVecchio 2000; Dall et al. 2004). Because some studies have estimated the repeatability of behaviours in juveniles while others have measured adults, here, we have an opportunity to ask whether the same trend applies to nonhuman animals.

Do Repeatability Estimates Differ Between the Field and the Laboratory?

Presumably, environmental variance is greater in the field compared to the stable conditions in the laboratory. To the extent that a changing environment is associated with behavioural plasticity, we might expect repeatability to be lower in the field, as has been found for estimates of heritability in Drosophila (Hoffmann 2000).

Do Males and Females Differ in Repeatability?

Two lines of thought in the literature suggest that males might be more repeatable than females. First, the older literature on the persistence of aggression (e.g. Andrew 1972; Wingfield 1994) suggests that testosterone can cause males to be more predictable than females. Second, honest indicator models of sexual selection predict that the behaviours indicated by a sexually selected trait are predictable because females use the trait as a reliable cue for how her mate will behave, for example, as a father (Kokko 1998; Garamszegi et al. 2006a). Here, we assess whether there are sex differences in repeatability generally.

Measuring the behaviour of individually marked animals on several occasions is laborious and therefore it is not surprising that few studies have systematically attempted to compare repeatability across ages (Bakker 1986; Masters et al. 1995; Battley 2006 Missoweit et al. 2007), sexes (Dingemanse et al. 2002; Schwagmeyer & Mock 2003; Nakagawa et al. 2007), locations (Howard & Young 1998; Kolluru 1999), intervals (Allen 1998) or treatments (Spencer & Thompson 2003; Magellan & Magurran 2007). Indeed, simultaneously addressing all of the questions listed above is beyond the scope of any project. However, meta-analysis allows us to address these questions using the growing body of literature concerning repeatability. By drawing on the available data, we can test our hypotheses regarding repeatability as it relates to behaviour, taxa, developmental stage, sex, and so on, to gain insight into how and why repeatability varies. This broad, exploratory analysis is also useful for stimulating new hypotheses and identifying particularly unexplored research directions.

METHODS

We compiled the data set by searching for published estimates of repeatability using the Web of Science search engine with combinations of the following topic terms in July 2008: repeatability, behav*, repeatab*, intraclass correlation coefficient, mate choice, preference, migration, predator. We also searched the reference list of each paper to identify studies that were missed in the initial search.

We used the following criteria when compiling the data set. (1) Studies had to measure the repeatability of individual behaviour, as opposed to the repeatability of the behaviour of a group, pair, chorus, colony, etc. (2) Studies on domesticated animals or animals in a zoo setting were excluded. (3) To facilitate comparisons across studies, the study needed to estimate repeatability as the intraclass correlation coefficient. This criterion excluded studies that measured binary behaviours (e.g. Preziosi & Fairbairn 1996), or that estimated repeatability using Pearson or Spearman correlation coefficients. (4) Studies that measured the repeatability of physiological (e.g. metabolic rate, hormone titre), performance-related (e.g. sprint speed, etc.), morphological (e.g. sperm characters such as the number of sperm per ejaculate or feather length), or life history (e.g. timing of breeding, laying date, clutch size, hatch date, arrival date, growth rate) traits were excluded. The final list of studies is given in Table 1 and the entire data set is given in Supplementary Table S1.

Table 1.

Studies included in meta-analysis organized by behavioural class and species

Author	Species	Taxonomic class
Activity
Benesh et al. 2008	Asellus aquaticus	Malacostraca
Nemiroff & Despland 2007	Malacosoma disstna	Insecta
Smith & Doupnik 2005	Rana catesbeiana	Amphibia
Kralj-Fiser et al. 2007	Anser anser	Aves
Affiliation
Kralj-Fiser et al. 2007	Anser anser	Aves
Aggression
Brown et al. 2006	Acheta domesticus	Insecta
Clark & Moore 1995	Gromphadorhina portentosa	Insecta
Bakker 1986	Gasterosteus aculeatus	Pisces
Riddell & Swain 1991	Oncorhynchus kisutch	Pisces
Kralj-Fiser et al. 2007	Anser anser	Aves
Garamszegi et al. 2006b	Ficedula albicollis	Aves
Pavlova et al. 2007	Sturnus vulgaris	Aves
Antipredator
Johnson & Sih 2007	Dolomedes triton	Arachnida
Bonte et al. 2003	Erigone atra	Arachnida
Fuiman & Cowan 2003	Sciaenops ocellatus	Pisces
Brodie & Russell 1999	Thamnophis sirtalis sirtalis	Reptilia
Kralj-Fiser et al. 2007	Anser anser	Aves
van Oers et al. 2004	Parus major	Aves
Courtship
Thornhill 1983	Harpobittacus nigriceps	Insecta
Hager & Teale 1994	Ips pini	Insecta
Aspi & Hoikkala 1993	Drosophila montana, D. littoralis	Insecta
Beeler et al. 1999	Nicrophorus orbicollis	Insecta
Brandt et al. 2005	Achroia grisella	Insecta
Hoffmann 1999	Drosophila melanogaster	Insecta
Kolluru 1999	Teleogryllus oceanicus	Insecta
Meffert & Hagenbuch 2005	Musca domestica	Insecta
Sattman & Cocroft 2003	Enchenopa binotata	Insecta
Allen 1998	Sciarasaga quadrata	Insecta
Butlin & Hewitt 1986	Chorthippus brunneus	Insecta
Gillham & Devrijer 1995	Chloriona spp.	Insecta
Higgins & Waugaman 2004	Gryllus texensis & G. rubens	Insecta
Du et al. 1987	Yponomeuta rorellus	Insecta
Rivero et al. 2000	Hygrolycosa rubrofasciata	Arachnida
Bee & Gerhardt, 2001	Rana catesbeiana	Amphibia
Gerhardt et al. 1996	Hyla versicolor	Amphibia
Runkle et al. 1994	Hyla versicolor	Amphibia
Howard & Young 1998	Bufo americanus	Amphibia
Sullivan 1992	Bufo americanus	Amphibia
Ryan & Rand 2003	Physalaemus pustulosus	Amphibia
Smith & Hunter 2005	Litoria booroolongensis	Amphibia
Sullivan & Hinshaw 1992	Hyla versicolor	Amphibia
Tarano 2001	Physalaemus enesefae	Amphibia
Wagner & Sullivan 1995	Bufo valliceps	Amphibia
Michalak 1996	Triturus montandoni	Amphibia
Malmos et al. 2001	Bufo microscaphus	Amphibia
Gamble et al. 2003	Poecilia reticulata	Pisces
Travis & Woodward 1989	Poecilia latipinna	Pisces
Rushbrook et al. 2008	Gasterosteus aculeatus	Pisces
Garamszegi et al. 2006a	Hirundo rustica	Aves
Gil & Slater 2000	Phylloscopus trochilus	Aves
Birkhead & Fletcher 1995	Taeniopygia guttata	Aves
Forstmeier & Birkhead 2004	Taeniopygia guttata	Aves
Helfenstein et al. 2003	Rissa tridactyla	Aves
Sanvito & Galimberti 2003	Mirounga leonina, M. angustirostris	Mammalia
Exploratory behaviour
Dingemanse et al. 2002	Parus major	Aves
Quinn & Cresswell 2005	Fringilla coelebs	Aves
Foraging
Missoweit et al. 2007	Panorpa vulgaris	Insecta
Martins et al. 2005	Clarias gariepinus	Pisces
Koteja et al. 2003	Mus domesticus	Mammalia
Habitat selection
Blanckenhorn & Perner 1994	Aquarius remigis	Insecta
Dohm et al. 2001	Bufo marinus	Amphibia
Smith & Doupnik 2005	Rana catesbeiana	Amphibia
Sheldahl & Martins 2000	Sceloporus occidentalis	Pisces
Kamel & Mrosovsky 2005	Eretmochelys imbricata	Reptilia
Kamel & Mrosovsky 2004	Dermochelys coriacea	Reptilia
Janzen & Morjan 2001	Chrysemys picta	Reptilia
Spencer & Thompson 2003	Emydura macquarii	Reptilia
Serrano et al. 2005	Falco naumanni	Aves
Mate preference
Brandt et al. 2005	Achroia grisella	Insecta
Greenfield et al. 2004	Ephippiger ephippiger	Insecta
Hager & Teale 1994	Ips pini	Insecta
Isoherranen et al. 1999	Drosophila virilis	Insecta
Reinhold et al. 2002	Chorthippus biguttulus	Insecta
Wagner et al. 1995	Gryllus integer	Insecta
Wilkinson et al. 1998	Cyrtodiopsis whitei, C. dalmanni	Insecta
Verburgt et al. 2008	Gryllus bimaculatus	Insecta
Bosch & Marquez 2002	Alytes cisternasii	Amphibia
Bosch et al. 2003	Alytes cisternasii, A. obstetricans	Amphibia
Gerhardt et al. 2000	Hyla versicolor	Amphibia
Howard & Young 1998	Bufo americanus	Amphibia
Jennions et al. 1995	Hyperolius marmoratus	Amphibia
Michalak 1996	Triturus montandoni	Amphibia
Murphy & Gerhardt 2000	Hyla gatiosa	Amphibia
Archard et al. 2006	Poecilia reticulata	Pisces
Aspbury & Basolo 2002	Heterandria formosa	Pisces
Brooks, 2002	Poecilia reticulata	Pisces
Brooks & Endler 2001	Poecilia reticulata	Pisces
Kodric-Brown & Nicoletto 1997	Poecilia reticulata	Pisces
Cummings & Mollaghan 2006	Xiphophorus nigrensis	Pisces
Godin & Dugatkin 1995	Poecilia reticulata	Pisces
Howard et al. 1998	Oryzias latipes	Pisces
Kodric-Brown & Nicoletto 1997	Poecilia reticulata	Pisces
Morris et al. 2003	Xiphophorus nigrensis	Pisces
Hoysak & Godin 2007	Gambusia holbrooki	Pisces
Gabor 2008	Poecilia latipinna	Pisces
Lehtonen & Lindström 2008	Pomatoschistus minutus	Pisces
Banbura 1992	Hirundo rustica	Aves
Møller 1994	Hirundo rustica	Aves
Roulin, 1999	Tyto alba	Aves
Forstmeier & Birkhead 2004	Taeniopygia guttata	Aves
Holveck & Riebel 2007	Taeniopygia guttata	Aves
Johnsen & Zuk 1996	Gallus gallus	Aves
Riebel 2000	Taeniopygia guttata	Aves
Mating
Edvardsson & Arnqvist 2006	Tribolium castaneum	Insecta
Radwan 1998	Rhizoglyphus robini	Insecta
Tallamy et al. 2003	Diabrotica undecimpunctata howardi	Insecta
Otronen 1997	Dryomyza anilis	Insecta
Michalak 1996	Triturus montandoni	Amphibia
Magellan & Magurran 2007	Poecilia reticulata	Pisces
Travis & Woodward 1989	Poecilia latipinna	Pisces
Whittingham et al. 2006	Tachycineta bicolor	Aves
Migration
Kent & Rankin 2001	Melanoplus sanguinipes fabricius	Insecta
Semlitsch et al. 1993	Ambystoma talpoideum	Amphibia
Battley 2006	Limosa limosa baueri	Aves
Bety et al. 2004	Anser caerulescens atlanticus	Aves
Petersen 1992	Chen canagicus	Aves
Møller 2001	Hirundo rustica	Aves
Potti 1998	Ficedula hypoleuca	Aves
Other
Brandt & Allen 2004	Uta stansburiana	Reptilia
Kralj-Fiser et al. 2007	Anser anser	Aves
McDonald et al. 2007	Manorina melanophrys	Aves
Masters et al. 1995	Eptesicus fuscus	Mammalia
Parental behaviour
Honza et al. 2007	Sylvia atricapilla	Aves
Gray et al. 2005	Puffinus puffinus	Aves
Nakagawa et al. 2007	Passer domesticus	Aves
Schwagmeyer & Mock 2003	Passer domesticus	Aves
Freeman-Gallant & Rothstein 1999	Passerculus sandwichensis	Aves
MacColl & Hatchwell 2003	Aegithalos caudatus	Aves
Potti et al. 1999	Ficedula hypoleuca	Aves
Lemon 1993	Taeniopygia guttata	Aves

We used the following grouping variables to characterize each repeatability estimate: developmental stage (adult, juvenile or both), sex (male, female or both), whether the study was conducted in the field or laboratory, if the interval between observations was greater or less than 1 year, the minimum number of times that individuals were measured in the study, the functional class of behaviour and taxonomic group. Following Mousseau & Roff (1987), we looked for differences between vertebrates versus invertebrates and ectothermic versus endothermic vertebrates. We categorized behaviours into the following functional classes: courtship, mate preference, activity, affiliation, aggression, antipredator (including risk taking), exploratory behaviour, foraging, habitat selection and territoriality (including nest site selection and thermal preference), migration, mating (any behaviours performed during mating and including extrapair copulations), parental care and other.

In an important paper, Lessells & Boag (1987) pointed out that MS_a (the mean square among individuals) depends on n₀, the coefficient representing the number of observations per individual. When the number of observations per individuals is unequal, n̄ is greater than n₀. Estimates that do not correct for different numbers of observations per individuals systematically underestimate repeatability; the difference between n̄ and n₀ increases with increasing spread in the number of measures per individual. Therefore, we compared repeatability estimates that either did or did not correct for different numbers of measures per individual, as suggested by Lessells & Boag (1987).

An advantage of meta-analytic techniques is that it scales the weight given to the results of each study based on its power and precision. This is done through the conversion on the original test statistic (here, repeatability) to an effect size. The effect size of each repeatability estimate was calculated in MetaWin 2.1 (Rosenberg et al. 2000). The average effect size was computed as a weighted mean, whereby the weights were equal to the inverse variance of each study’s effect estimator. Larger studies and studies with less random variation were given greater weight than smaller studies. Analysis of effect sizes rather than raw repeatability estimates is preferable because more weight should be given to more powerful studies. Therefore, all subsequent analyses were performed on estimates of effect size, rather than the raw repeatability score.

To understand the causes of variation in repeatability estimates, we used fixed effects categorical or continuous models in MetaWin. For comparisons between groups of studies, we report Q_b, the between-groups homogeneity. This statistic is analogous to the between-groups component of variance in conventional analysis of variance, and it is χ² distributed with n groups minus one degree of freedom. We also report effect sizes and their 95% confidence intervals as CL1 ≤ effect size ≤ CL2.

Limitations of the data set and statistical options available for meta-analysis precluded us from formally testing statistical interactions between the grouping variables. We explored patterns in the data set by analysing subsets of the data according to different levels of the factor of interest. For example, after testing for a difference in effect size between males and females using all the data, we then performed the same analysis when field studies were excluded. We repeated the analysis when laboratory studies were excluded, and so forth. We infer that patterns that were common to several subsets of the total data set are robust and do not depend on other grouping variables (see Table 2). If the effect of a grouping variable was significant for one level of a different grouping variable but not for the other level, then we infer that there might be an interaction between the two grouping variables. We also pay particular attention to effect sizes because when a subset of data was eliminated from the analysis, our power to detect a significant effect was reduced. Therefore, in addition to asking whether comparisons are statistically significant for certain subsets of the data, we also report whether effect sizes changed. We view this exploratory analysis as a mechanism for generating hypotheses and to suggest promising areas for future study. We ranked the P values in each column in Table 2 and used the sequential Bonferroni procedure to account for multiple comparisons (Rice 1989).

Table 2.

Summary of meta-analysis results for repeatability and behaviour

Variables included in analysis	N	Invert/vert	Ectotherm/endotherm*	Test interval (short/long)†	Developmental stage (juvenile/adult)‡	Location (field/lab)	Sex (male/female)
Entire data set	759	NS	Ec<En	NS	NS	L<F	F<M
Males	388	I<V	Ec<En	L<S	NS	L<F	—
Females	275	V<I	Ec<En	NS	A<J	L<F	—
Invertebrates	266	—	—	?	NS	L<F	M<F
Vertebrates	493	—	Ec<En	NS	NS	L<F	F<M
Ectotherms*	267	—	—	L<S	A<J	L<F	F<M
Endotherms*	226	—	—	NS	NS	L<F	F<M
Short interval	690	NS	NS	—	NS	L<F	F<M
Long interval	69	?	Ec<En	—	?	L<F	NS
Juveniles	50	NS	NS	?	—	L<F	NS
Adults	706	NS	Ec<En	NS	—	L<F	F<M
Field	293	V<I	NS	L<S	A<J§	—	NS
Lab	466	V<I	Ec<En	L<S	NS	—	F<M
Excluding courtship	432	V<I	Ec<En	NS	—	L<F	F<M
Excluding mate preference	611	I<V	En<Ec§	NS	—	L<F	M<F
Excluding: Serrano et al. 2005; Hoffmann 1999	756	V<I	NS	L<S	NS	L<F	F<M

All comparisons except those with NS were significant at P < 0.05 according to the heterogeneity test. Sample sizes reflect the number of estimates considered in all comparisons except sex and developmental stage, where estimates that were based on measures of both males and females or both juveniles and adults were excluded.‘?’: not enough estimates were available for the comparison.

^*Only vertebrates were included in this analysis.

^†Interval between observations was either less than 1 year ('short') or greater than 1 year (‘long’).

^‡Adult-specific behaviours (mate preference, mating, courtship and parental behaviours) were excluded from these comparisons.

^§Nonsignificant following a sequential Bonferroni correction for multiple comparisons.

Many papers reported more than one repeatability estimate, introducing the possibility of pseudoreplication if multiple estimates from the same study are nonindependent of each other. For example, studies of calling behaviour in frogs often measure more than one attribute of a male’s call on multiple occasions, including amplitude, duration, frequency, and so forth. If the attributes are correlated with each other (e.g. fundamental frequency is positively correlated with dominant frequency; Bee & Gerhardt 2001), then repeatability estimates for the different attributes are not independent. There is no clear consensus about how to handle multiple estimates reported from the same study in meta-analysis (Rosenberg et al. 2000). On one hand, we want to avoid nonindependence among effect sizes, but on the other hand, we do not want to lose biologically meaningful information by using only one estimate per study (e.g. the study’s mean). The loss of information caused by omission of such effects may lead to more serious distortions of the results than those caused by their nonindependence (Gurevitch et al. 1992).

Therefore, we took multiple strategies to address possible bias caused by the nonindependence of multiple estimates per study. First, in cases where studies reported separate repeatability estimates on behaviours measured on more than two occasions, we did not include estimates that provided potentially redundant information (Bakker 1986; Hager & Teale 1994; Archard et al. 2006). For example, a study that measured individuals on three occasions could potentially report repeatability for the comparison between measures one and two, measures two and three, and measures one and three. In this case, we excluded the estimate of repeatability between measures two and three, as it would not provide additional information (for the purposes of our analysis) compared to the repeatability reported in times one and two. We did include the repeatability estimate between times one and three, however, as this represents a different interval between measures, one of the factors in which we were interested. Similarly, when studies reported repeatability for both separate and pooled groups (e.g. males, females, and males and females), we did not include the pooled estimate (Gil & Slater 2000; Archard et al. 2006; Battley 2006).

Second, we compared studies that reported different numbers of repeatability estimates (as in Nespolo & Franco 2007). We found no relationship between the number of estimates reported and the value of those estimates (slope = −0.002, Q_regression = 1.19, P = 0.28). This suggests that the number of estimates reported by a study does not systematically change the effect size reported. Third, we removed, one at a time, studies that contributed the greatest number of estimates to the data set to evaluate whether they were primarily responsible for the observed patterns. Removing studies that reported the highest numbers of estimates did not change any of the main effects (results not shown).

Finally, because a large proportion of estimates were based on just two behaviours (courtship and mate preference, see Results), we reanalysed the data set when either courtship behaviours or mate preference behaviours were excluded. We paid particular attention to whether effect sizes changed in the reduced data sets to determine whether these widely studied behaviours disproportionately influenced the results.

Two studies (Hoffmann 1999; Serrano et al. 2005) in our data set measured a much larger number of individuals (N = 972 and N = 1318, respectively) to estimate repeatability and were therefore weighted much more heavily in the meta-analysis. For comparison, the average sample size of the remaining data set was 39. Serrano et al. (2005) measured habitat preference across years in adult kestrels in the field and found relatively high repeatability for this behaviour. Hoffmann (1999) measured two courtship behaviours of male Drosophila in the laboratory and estimated relatively low repeatabilities.

On one hand, the purpose of meta-analysis is to take differences in power into consideration when evaluating across studies; therefore, it follows that these two studies should be weighted more heavily in our analysis. On the other hand, these two studies are not representative of most studies on repeatability (the next highest sample size after Serrano et al. 2005 in the data set is N = 496) and therefore they might bias our interpretation. For example, the repeatability estimate in the Serrano et al. (2005) was relatively high (R = 0.58) and was measured in the field. Therefore, this heavily weighted result might cause it to appear that repeatability is higher in the field than in the laboratory. To address the possibility that these particularly powerful studies were driving our results, we ran our analyses when the three estimates from these two studies were excluded.

To determine whether our data set was biased towards studies that found significant repeatability estimates (the ‘file drawer effect’), we constructed funnel plots (Light & Pillemer 1984) and calculated Rosenthal’s (1979) ‘fail-safe numbers’ in MetaWin. Funnel plots are useful for visualizing the distribution of effect sizes of sample sizes in the study. Funnel plots with wide openings at smaller sample sizes and with few gaps generally indicate less publication bias (Rosenberg et al. 2000). Fail-safe numbers represent the number of nonsignificant, missing or unpublished studies that would need to be added to the analysis to change the results from significant to nonsignificant (Rosenberg et al. 2000). If these numbers are high, relative to the number of observed studies, the results are probably representative of the true effects, even in the face of some publication bias (Rosenberg et al. 2000).

RESULTS

Summarizing the Data Set

We identified 759 estimates of repeatability that met our criteria (Fig. 1). The estimates are from 114 studies, representing 98 species (Table 1). The sample size (number of individuals measured) ranged from 5 to 1318. Most studies measured the subjects twice, although some studies measured individuals as many as 60 times, with a mean of 4.41 measures per individual. The majority of repeatability estimates (708 of 759) considered in this meta-analysis were calculated as suggested by Lessells & Boag (1987). As predicted, estimates that did not correct for different numbers of observations per individual (mean effect size = 0.47, 95% confidence limits = 0.43, 0.52; hereafter reported as 0.43 ≤ 0.47 ≤ 0.52) were higher than those that did correct for different numbers of observations per individual (0.35 ≤ 0.37 ≤ 0.38, Q_b = 23.0, N = 759, P < 0.001) (Lessells & Boag 1987). However, we found no evidence that this confounded our overall results.

Figure 1.

The distribution of repeatability from published studies. Estimates less than zero occur when there are large standard error bars surrounding estimates of between-individual variance components.

Studies measured the repeatability of a wide variety of behaviours; courtship (327 estimates from 40 studies) and mate preference (148 estimates from 34 studies) were particularly well studied (Table 1, Fig. 2a). Most estimates came from studies of vertebrates (493 versus 266 estimates for invertebrates), with 201 estimates from birds alone (Fig. 2b). The majority of behaviours were studied in adults (706 versus 50 estimates on juveniles, 3 estimates on both adults and juveniles), and more estimates came from studies of males than females (388 versus 275; 95 estimates for both). Most studies measured individuals repeatedly within 1 year, although 69 estimates were based on an interval between observations that was greater than 1 year. Fewer estimates were made in the field (293 estimates) compared to the laboratory (466 estimates).

Figure 2.

Average effect sizes by (a) behaviour and (b) taxonomic class. Numbers in parentheses indicate sample sizes and error bars denote 95% confidence intervals.

Altogether the data overwhelmingly support the hypothesis that behaviour is repeatable (Fig. 1). The average repeatability across all estimates was 0.37, and the weighted effect size across all estimates was significantly greater than zero (0.36 ≤ 0.37 ≤ 0.38, Q_total = 3860.9, N = 759, P < 0.001).

Evaluating Hypotheses

Are certain types of behaviour more repeatable than others?

Repeatability estimates varied widely across different classes of behaviour. The most repeatable classes of behaviour were mating, habitat selection and aggression, while the least repeatable behaviours were activity, mate preference and migration (Fig. 2a). The two best-studied behaviours, mate preference and courtship, had very different repeatabilities; courtship was more repeatable than mate preference.

Are certain taxa more repeatable than others?

There was not a clear difference in the repeatability of the behaviour of invertebrates compared to vertebrates (Q_b = 2.79, N = 759, P = 0.0951; Figs 2b, 3a), but further analyses suggested that the difference between invertebrates versus vertebrates might depend on the behaviour under consideration. On behaviours other than courtship, for example, invertebrates were more repeatable than vertebrates (0.41 ≤ 0.45 ≤ 0.48 versus 0.32 ≤ 0.33 ≤ 0.33; Q_b = 33.6, N = 432, P < 0.001; Table 2). For behaviours other than mate preference, on the other hand, vertebrates were more repeatable than invertebrates (0.42 ≤ 0.43 ≤ 0.45 versus 0.37 ≤ 0.39 ≤ 0.41; Q_b = 13.7, N = 633, P < 0.001; Table 2). It is likely that the interaction between taxonomic grouping and behaviour was influenced by the fact that mate preference behaviours, which generally had low repeatability, were typically measured on vertebrates.

Figure 3.

Average effect sizes for each of the grouping variables. The bars show means and 95% confidence intervals and numbers in parentheses indicate sample sizes. An inset graph is included if the overall result changed when a subset of the data set was excluded. (a) Vertebrates versus invertebrates; (b) ectotherms versus endotherms; (c) interval; (d) developmental stage; (e) location; (f) sex.

As with heritability (Mousseau & Roff 1987), we found suggestive evidence that endothermic vertebrates were more repeatable than ectothermic vertebrates (Q_b = 14.7, N = 493, P = 0.001; Fig. 3b). This pattern depended on whether the animals were measured in the field or the laboratory: in the field, there was no difference (Table 2), but in the laboratory, endotherms were more repeatable (0.32 ≤ 0.36 ≤ 0.40 versus 0.22 ≤ 0.24 ≤ 0.27; Q_b = 15.1, N = 186, P = 0.001; Table 2). The large estimate reported in Serrano et al. (2005), which was measured in an endotherm in the field, may have been driving the overall difference between endotherms and ectotherms, but it seems unlikely that it was solely responsible for the difference because endotherms were more repeatable than ectotherms in laboratory studies only (Table 2).

Does repeatability decrease with the interval between observations?

Initially, it appeared that there was no difference in repeatability based on short versus long intervals between observations (Q_b = 0.87, N = 759, P < 0.350; Fig. 3c). However, closer analysis showed that this surprising result was probably caused by two particularly powerful and therefore heavily weighted studies in the meta-analysis: Hoffmann (1999) lowered effect sizes for short intervals, and Serrano et al. (2005) raised effect sizes for long intervals. When these studies were removed, repeatability estimates were higher for behaviours measured close together in time (Q_b = 43.1, N = 755, P < 0.001; Fig. 3c). This significant effect was robust to several other subsets of the data (Table 2).

Does repeatability increase with the number of observations per individual?

We found no evidence that repeatability estimates were affected by the number of observations per individual (slope = 0.008; Q_regression = 0.42, N = 759, P = 0.516; Fig. 4).

Figure 4.

Relationship (nonsignificant) between effect size and the number of times that individuals were measured.

Does repeatability vary among age groups?

For this comparison, we did not consider adult-specific behaviours such as mate preference, mating, courtship and parental behaviour. Overall, there was no difference in the repeatability of behaviour in juveniles or adults (Q_b = 0.6166, N = 220, P = 0.4323; Fig. 3d). However, certain subsets of the data set suggest that there might be important differences in the repeatability of behaviour of juveniles and adults. Among the subsets of the data set for which there was a statistically significant difference, the behaviour of juveniles was consistently more repeatable than the behaviour of adults. For example, among ectotherms, juvenile behaviour was more repeatable than adult behaviour (Q_b = 13.19, N = 72, P = 0.0003; Table 2).

Do repeatability estimates differ between the field and the laboratory?

Overall, we found that behaviours measured in the field were more repeatable than behaviours measured in the laboratory (Fig. 3e). This pattern was robust across all subsets of the data set.

Do males and females differ in repeatability?

Overall, males were more repeatable in their behaviour than females (Table 2, Fig. 3f). The sex difference was observed in adults, but not in juveniles, and was true for all vertebrates (Table 2). However, there was an interaction between sex and the type of behaviour measured. When mate preference was omitted from the data set, the pattern was reversed and females were more repeatable than males, as judged both by the P value and by effect sizes (0.38 ≤ 0.40 ≤ 0.41 versus 0.43 ≤ 0.47 ≤ 0.51; Q_b = 12.3, N = 538, P < 0.001; Table 2, Fig. 3f). Therefore, it is likely that the very low repeatability of mate preference behaviours, which were typically measured on females (9 estimates of the repeatability of mate preference were for males versus 139 estimates for females), shifted the female average downwards.

Testing for Publication Bias

We found no evidence for publication bias based on either a visual inspection of our funnel plot (Fig. 5) or based on Rosenthal’s fail-safe numbers. Our fail-safe numbers were very large relative to our observed sample sizes, with Rosenthal’s numbers ranging from 100 to over 900 times the number of results in our analysis. Even when only the means from studies reporting multiple repeatability estimates were used, our Rosenthal’s number was over 190 times as large as the number of included studies.

Figure 5.

Funnel plot showing effect size as a function of the number of individuals measured (sample size). Note that the X axis is log-transformed.

DISCUSSION

Our analysis provides strong support for consistent individual differences in behaviour. We found that the repeatability of behaviour was significantly greater than zero, and that roughly 35% of the variation among individuals in behaviour could be attributed to individual differences. Despite the heterogeneous nature of the data set, our analysis also uncovered some intriguing patterns.

We found strong evidence that not all types of behaviours were equally repeatable. Overall it is difficult to make inferences about the causes of variation in repeatability of behaviours (i.e. if some behaviours are more repeatable than others because they are the ones that are least influenced by the environment or the most canalized). However, one pattern that was robust among almost all subsets of the data was that individuals (typically females) were not consistent in their mate preferences; just because a female preferred a certain type of male on one occasion did not necessarily mean that she retained that preference on subsequent occasions. This result is consistent with a growing number of studies showing that what a female prefers in a mate is subject to change depending on her age, condition and the environment (reviewed in: Jennions & Petrie 1997; Cotton et al. 2006). In addition, by their very nature, repeatability studies allow the test subject to have more information about the distribution of quality of mates in the local environment in the second testing situation compared to the first (Janetos 1980; Parker 1983; Real 1990; Dombrovsky & Perrin 1994). After receiving additional information in the first test, females might ‘fine-tune’ their preference in subsequent tests, therefore lowering repeatability.

Our analysis also suggests that not all types of taxa are equally repeatable. For example, endotherms were generally more repeatable than ectotherms, as has also been found for heritability estimates (Mousseau & Roff 1987). One interpretation of this pattern is that ectotherms are more sensitive to the environment, and, therefore, individuals are more likely to change their behaviour according to the environment (but not equally). It is intriguing that when we compared endotherms to ectotherms only in field studies, in which we presume there was more environmental variation compared to in laboratory studies, the difference between endotherms and ectotherms disappeared, contrary to our first interpretation. One possible explanation for this is that mate preference studies (which had very low repeatabilities) were typically conducted in the laboratory, and for mate preference behaviours, endotherms were more repeatable than ectotherms (0.24 ≤ 0.28 ≤ 0.33 versus 0.15 ≤ 0.18 ≤ 0.21, N = 112). In contrast, we found little evidence in support of the popular notion that invertebrates are more rigid in their behaviour than vertebrates.

We found strong support for the intuitive hypothesis that individuals are more consistent over short intervals compared to long intervals, at least when Hoffmann (1999) and Serrano et al. (2005) were excluded. Repeatability was significantly higher when the same individuals were measured for a second time within a year of the first measurement. Granted, greater than or less than 1 year is a fairly coarse measure, and one which does not take differences in life span into consideration. That is, a day in the life of a cricket that lives for only a few weeks (Kolluru 1999) represents a considerably longer fraction of its total life span compared to a long-lived organism such as an elephant seal (Sanvito & Galimberti 2003). This rough measure could therefore lead to bias if taxonomic differences were confounded with interval (i.e. short-lived organisms such as invertebrates are relatively repeatable and were also measured over relatively short intervals). However, we found no difference in the repeatability of behaviour of invertebrates versus vertebrate animals, and, therefore, do not consider taxonomic group to be a confounding variable. In addition, when we looked for relationships between repeatability and the interval between measurements while controlling for life span (and age at maturity), the effect of interval did not change (results not shown). As more data become available, it will be useful to carry out this type of broad comparison in the correct phylogenetic framework.

We found suggestive evidence that there might be systematic differences in the repeatability of behaviour of juveniles versus adults. At first glance, it appeared that there was no difference in the repeatability of behaviour of adults or juveniles. Unfortunately, there are only a few examples in the data set of repeatability estimates of juveniles and adults of the same species and they do not suggest a strong pattern (sticklebacks, Gasterosteus aculeatus: 0.68 juveniles versus 0.78 adults; Bakker 1986; big brown bat, Eptesicus fuscus: 0.51 juveniles versus 0.60 adults; Masters et al. 1995; godwit, Limosa limosa baueri: 0.4 juveniles versus 1.19 adults; Battley 2006; scorpionfly, Panorpa vulgaris: 0.30 juveniles versus 0.21 adults; Missoweit et al. 2007). Comparing the repeatability of behaviour of juveniles versus adults within the same species is an important, interesting and relatively unexplored question with no clear predictions about the direction of the effects. On one hand, we might expect juveniles to be undergoing dramatic developmental change and therefore not show repeatable behaviour. On the other hand, we might expect juveniles to be more repeatable because the costs of straying from a developmental trajectory are higher for juveniles (Biro & Stamps 2008).

Changes in repeatability with age might also reflect the action of selection on phenotypic variance. If there is directional or stabilizing selection on a particular behaviour, then phenotypic variance will decrease after selection. This could cause repeatability to decrease with age (if there is less variation among adults compared to juveniles). Alternatively, if traits expressed early in life are subject to stronger selection pressures than traits expressed later in life, then overall repeatability might increase with age (because there is more variation among adults compared to juveniles).

Contrary to our prediction, we found that behaviour was generally more repeatable in the field than the laboratory. Initially, we reasoned that greater environmental variance in the field would increase within-individual variation (s²) and thereby decrease repeatability. Alternatively, greater environmental variance in the field might allow the expression of more behavioural variation among individuals ($s_A^2$), by creating micro-niches, and thereby increase repeatability. Provided there are advantages of behaving consistently (Dall et al. 2004; McElreath & Strimling 2006), this could explain why consistent individual differences in behaviour were greater in the field.

Our results suggest that there might be important sex differences in repeatability, but the direction of the difference probably depends on the behaviour under consideration. Overall, we found that males were more repeatable than females, which was consistent with our predictions (e.g. Andrew 1972; Wingfield 1994). For example, male house sparrows, Passer domesticus, were more repeatable with respect to parental behaviour compared to female house sparrows (Schwagmeyer & Mock 2003; Nakagawa et al. 2007). The lower repeatability of female parental care might be due to the greater responsiveness of females to the needs of the brood. Alternatively, males might be more repeatable than females because of selection favouring honest signalling (Nakagawa et al. 2007).

Closer examination revealed that the overall sex difference was driven by the extremely low repeatability of mate preference, which was typically measured on females (see discussion of mate preference above). Therefore, our evidence for sex differences in repeatability is inconclusive. This is in agreement with the data on the consistency of personality in humans as well, where sex differences in consistency are rarely observed (Robins et al. 2001). This lack of a consistent pattern across such a wide variety of classes of behaviours may, in fact, be expected if there are differences between the sexes and if the best overall behavioural strategy varies by gender.

Although the large collection of published estimates of repeatability offers an opportunity to look for patterns to explain variation in consistent individual differences in behaviour, there are limitations of the data set and several questions remain unanswered because of the heterogeneous and possibly biased nature of the collection of estimates. For example, one possible source of bias in the data set comes from the types of behaviours that were originally studied. Many studies estimated repeatability as a first step towards studying the genetic basis of the behaviour. Insofar as researchers only commenced studies of the genetic basis of behaviour on behaviours that they already suspected were heritable, then published estimates of repeatability might be biased upwards.

Another source of bias is measurement error. Most studies included in the data set did not distinguish measurement repeatability from true ‘trait’ repeatability (Falconer & Mackay 1996; Hoffmann 2000). Here, we are interested in using the published data to infer patterns about the underlying causes of deviation from consistency. However, one reason why it could appear that individuals do not behave consistently different from each other is because there is measurement error associated with each behavioural observation. This introduces a possible source of bias in the data set if certain studies or certain types of behaviour have more measurement error associated with them than others.

An additional factor that can potentially reduce repeatability is mean-level change between measurements (e.g. on average, the population is more aggressive the second time it is observed compared to the first time; Hayes & Jenkins 1997). If mean-level change causes more within-individual than between-individual change between observations, then repeatability will be low. Mean-level change might have contributed to our finding that repeatability declines as a function of test–retest intervals because mean-level changes in behaviour are more likely to occur over longer periods than over shorter periods. In general, however, mean-level change does not preclude the possibility that repeatability will be significantly different from zero. So long as between-individual differences are large relative to within-individual differences, a behaviour can still be repeatable despite mean-level change.

Our results offer several suggestions for the design and analysis of future research. First, repeatability does not appear to depend on the number of times that individuals are measured. Indeed, it seems likely that increasing the number of observations per individual will decrease the error around the estimate, rather than the repeatability estimate itself. This result suggests that if researchers want to estimate repeatability of a behaviour, they have more to gain by measuring more individuals on fewer occasions rather than fewer individuals on more occasions. Second, to facilitate comparisons across studies, it would be helpful if future papers report statistics such as n_o, whether there was mean-level change between observations, and whether variance among individuals was the same at each measurement (an assumption of the intraclass correlation coefficient statistic but rarely reported).

On a related note, repeatability statistics say little about whether individuals themselves are repeatable; the statistic is a property of the population of individuals. It is likely that in most studies, individuals differ in how much their behaviour changes between observations. That is, even when a repeatability statistic is significantly different from zero, it does not necessarily mean that all of the individuals within the population behaved equally consistently; some individuals were probably more consistent than others. Indeed, the literature on coping styles has emphasized that consistency is a trait that varies among individuals; the behaviour of proactive individuals, which tend to be rigid and routinized, is more repeatable than the behaviour of reactive individuals, which tend to be more responsive to cues in the environment (e.g. Benus et al. 1990, 1991; Marchetti & Drent 2000). An important direction for future studies in this area is to define conditions responsible for individual differences in plasticity (Nussey et al. 2007).

This study reveals some interesting, and sometimes surprising, results when repeatability was assessed across a wide variety of behaviours, species and experiments. For example, while higher repeatability over short intervals might be expected, higher repeatability in the field versus the laboratory prompts us to take a more nuanced look at the specific behaviours being measured in these different settings. Variation in repeatability among classes of behaviour could also influence how we interpret these behaviours. The patterns we found not only show what is known to date about repeatability and behaviour, but also serve to stimulate questions about why repeatability varies across behaviours, ages, sexes or taxa. Our results suggest that particularly interesting but relatively unexplored questions include comparing the repeatability of behaviour between different age classes, between males and females and under different ecological conditions.