Correlational selection does not explain the evolution of a behavioural syndrome


  • C. S. Han,

    Corresponding author
    1. Evolution and Ecology Research Centre, School of Biological, Earth and Environmental Sciences, The University of New South Wales, Sydney, NSW, Australia
    • Correspondence: Chang S. Han, Evolution and Ecology Research Centre, School of Biological, Earth and Environmental Sciences, The University of New South Wales, Sydney, 2052 NSW, Australia.

      Tel.: +61 2 9385 2587; fax: +61 2 9385 1558; e-mail:

    Search for more papers by this author
  • R. C. Brooks

    1. Evolution and Ecology Research Centre, School of Biological, Earth and Environmental Sciences, The University of New South Wales, Sydney, NSW, Australia
    Search for more papers by this author

  • Data deposited at Dryad: doi:10.5061/dryad.624gb


Correlated suites of behaviours, or behavioural syndromes, appear to be widespread, and yet few studies have explored how they arise and are maintained. One possibility holds that correlational selection can generate and maintain behavioural syndrome if certain behavioural combinations enjoy greater fitness than other combinations. Here we test this correlational selection hypothesis by comparing behavioural syndrome structure with a multivariate fitness surface based on reproductive success of male water striders. We measured the structure of a behavioural syndrome including dispersal ability, exploration behaviour, latency to remount and sex recognition sensitivity in males. We then measured the relationship between these behaviours and mating success in a range of sex ratio environments. Despite the presence of some significant correlational selection, behavioural syndrome structure was not associated with correlational selection on behaviours. Although we cannot conclusively reject the correlational selection hypothesis, our evidence suggests that correlational selection and resulting linkage disequilibrium might not be responsible for maintaining the strong correlations between behaviours. Instead, we suggest alternative ways in which this behavioural syndrome may have arisen and outline the need for physiological and quantitative genetic tests of these suggestions.


Numerous studies have demonstrated consistent differences between individuals in behavioural traits such as aggressiveness, antipredatory behaviour, activity or other social behaviour (Dall et al., 2004; Réale et al., 2007; Biro & Stamps, 2008; Bell et al., 2009; Réale & Dingemanse, 2010; Garamszegi et al., 2012). Recently there has been a surge of interest in these differences as well as the fact that the pattern of covariance among these behaviours can persist across different environmental situations (Sih et al., 2004a; Réale et al., 2007). These correlated suites of behaviours – or behavioural syndromes – have been reported in many animal taxa including mammals, bird, fishes, spiders and insects (Dingemanse et al., 2007; Dochtermann & Jenkins, 2007; Johnson & Sih, 2007; Kortet & Hedrick, 2007; Pruitt et al., 2008; Evans et al., 2010; Gabriel & Black, 2010; Adriaenssens & Johnsson, 2012).

The widespread presence of behavioural syndromes raises the question of how correlations between behaviours arise and are maintained. One possibility holds that correlational selection can generate and maintain correlations between behaviours if certain behavioural combinations enjoy greater fitness than other combinations (‘correlational selection hypothesis’, see review of Sih et al., 2004b; van Oers et al., 2005; Bell, 2007). Correlational selection is thought to be widespread and an important determinant of genetic covariance patterns (Zhang & Hill, 2003; Brooks et al., 2005; Johnson & Barton, 2005). Thus, correlational selection can cause the emergence of a behavioural syndrome that is inherited via linkage disequilibrium when there are no shared physiological or genetic mechanisms among the traits involved (see review of Sih et al., 2004b; van Oers et al., 2005; Bell, 2007).

Correlational selection builds linkage disequilibrium between traits that are affected by genes at different loci, driving the emergence of correlations between traits (Price & Langen, 1992; Falconer & Mackay, 1996; Lynch & Walsh, 1998). As a result, when selection subsequently causes a change in the mean of one trait, other correlated traits can shift in correlated response to selection (Sinervo & Svensson, 2002; McGlothlin et al., 2005).

Despite a widespread acceptance of correlational selection (Kingsolver et al., 2001; Blows & Brooks, 2003; Garant et al., 2007), relatively little attention has been paid to the ways in which selection shapes suites of behavioural traits (Kingsolver et al., 2001), and a few studies have examined the effect of correlational selection on the origin and maintenance of behavioural syndromes (Bell & Sih, 2007; Adriaenssens & Johnsson, 2012).

Selection on behavioural traits is expected to vary depending on the environment (Alvarez & Bell, 2007; Dingemanse et al., 2007; Dochtermann et al., 2012). Environmental variation, such as variation in biotic factors (e.g. density, predation risk or sex ratio) or abiotic factors (e.g. temperature, refuge), complicates the link between behavioural traits and fitness, because a behavioural trait that is successful in one environment may not be successful under a different environment (Wilson & Yoshimura, 1994; Roff & Fairbairn, 2007). For example, individuals that are cautious against predators and less exploratory may enjoy higher fitness benefit in risky environments compared with bold individuals, whereas the bold phenotype may do better in the absence of risk. As the relationship between phenotype and fitness depends on prevailing environmental conditions, fluctuating environmental conditions that favour different suites of behavioural traits over time can contribute to the maintenance of variation in behaviours and in the combinations of behaviours that we know as syndromes.

In this study, we test the role of correlational selection in the emergence of behavioural syndrome in a water strider. First, we measured behavioural syndrome structure among dispersal ability, exploration in a novel environment, latency to remount and sex recognition sensitivity in water strider males. As local sex ratio can alter mating behaviour of water striders (Arnqvist, 1992; Rowe, 1992; Krupa & Sih, 1993; Jablonski & Vepsalainen, 1995; Vepsalainen & Savolainen, 1995; Han et al., 2012), we measured which behaviours of water strider males are favoured in various sex ratio environments. In addition, to test whether behavioural syndrome structure results from correlational selection on behavioural traits, we compared behavioural syndrome structure with a multivariate fitness surface based on their reproductive success.

Materials and methods

Study species and rearing condition

Tenagogerris euphrosyne is the commonest water strider species in eastern Australia, but behavioural study of this species has been scarce. The first generation (spring generation) was collected in Hacking River, Otford, New South Wales on 6 November 2010, and the second generation (summer generation) was collected on 8 January 2011. All the individuals were separated by their sex to maintain similar levels of mating experience in males and similar sexual receptivity of females. Males were reared individually in a small tank (home tank, 10 × 15 cm) during the test to prevent the effect of intrasexual interactions on behaviour assays. Females were placed 30 individuals per tank (20 × 15 cm). They were fed ad libitum with surplus frozen crickets (Teleogryllus commodus) and neriid flies (Telostylinus angusticollis and Telostylinus lineolatus) every day. The pieces of floating styrofoam were provided along the margin of the tank as the resting sites. All individuals were released to the original habitat after the experiment was completed.

Mating assays under varying sex ratios

In general, reproductive water strider males and females in male-biased sex ratios mate with multiple mates (Arnqvist, 1997). However, the local sex ratio can vary dramatically over time or between stretches of water divided by water flow or physical barriers. As local sex ratio affects male mating behaviour (Arnqvist, 1992; Rowe, 1992; Krupa & Sih, 1993; Jablonski & Vepsalainen, 1995; Vepsalainen & Savolainen, 1995; Han et al., 2012), we have to consider various sex ratio conditions to measure the effect of behaviour types on mating success.

Using 90 males, we drew 6 groups of 15 males each (3 groups for each generation) and measured their mating success under 4 different sex ratio conditions (15, 10, 5 and 3 females = M : F 1 : 1, 1.5 : 1, 3 : 1, 5 : 1) with 72-h interval between the 4 measures. To prevent mating assays from being biased by the carry-over effect of behaviour assays, mating assays were conducted before behaviour assays. For the first time, each group was tested at a sex ratio of 1 : 1 and then we used each of the six possible treatment orders. At the initiation of each mating assay, both sex individuals were separated by a barrier for 5 min as an acclimation period. Then, after the barrier between the sexes was removed, we recorded males’ mating status after 1 h. For logistical reasons, we reused every female once. To maintain similar levels of mating experience in males and similar sexual receptivity of females, males that failed to copulate in mating assays were also allowed to copulate with females after the assay. A male was returned to his home tank after copulating once.

The initiation of mating has an important effect on male reproductive success, especially in species in which post-copulation guarding lasts for several days (Han et al., 2010). At the initiation of males’ mating attempt, only males that try to grasp the female and overcome female resistance can copulate successfully with a female. Tenagogerris euphrosyne males guard females for up to 10 days post-copula (Han & Brooks, 2013), which is much longer than most other water strider species (Arnqvist, 1997). This represents up to ten percentage of maximum adult male longevity (3–4 months). Thus, the successful initiation of a single mating attempt can represent a large portion of a male's lifetime reproductive success (Fig. S1).

Behaviour assays

In group-living water striders, reproductive males compete for females and harass females to force them to copulate and mate, even when they have enough sperm from previous matings (i.e. scramble competition polygyny). As the sex ratio of wild T. euphrosyne is usually male biased and mate availability is severely limited (Han & Brooks, 2013), males able to distinguish single females from single males and males that persist in attempts to remount females after being dislodged are more successful in mating. Also, water striders encounter microhabitats along the creek where availability of potential mates and prey varies. So tendencies to disperse into unfamiliar environments and to thrive there might strongly influence fitness.

Approximately 3 days after the completion of mating success assays, we commenced measurements of the four behaviours. We measured the following four behaviours in each of 90 individual males: (i) dispersal ability, (ii) exploration in a novel environment, (iii) sensitivity to distinguish the correct sex (sex recognition sensitivity) and (iv) latency to remount (remounting attempts of the male after artificial dislodgement from a female prior to copulation). Each behaviour assay was separately conducted to control the habituation effect and carry-over effect (‘effects of a previous treatment on subsequent treatment’, see Dochtermann, 2010), and behavioural variables from (i) to (iv) were tested in a Williams design (Diaz-Uriarte, 2001) with 48-h interval between each assay. Males were randomly assigned to one of the sequences in a balanced manner. The last behavioural measurement was repeated again after 48 h for estimating repeatability of the behaviour (17 males for dispersal ability, 20 males for exploration and 20 males for latency to remount). Repeatability of three types of behaviours (except sex recognition sensitivity) was measured because the repeatability of sex recognition sensitivity was tested in our previous research (Han & Brooks, 2013). Experiments for the measurement of ‘sex recognition sensitivity’ and ‘latency to remount’ were performed in each male's home tank. The water in the home tank was changed 12 h before the experiment conducted.

Dispersal ability

Dispersal ability was assayed using a 3 × 3 maze, which consists of nine testing tanks (30 cm × 45 cm; each tank, 10 cm × 15 cm; water depth, 0.5 cm, Fig. S4), where the side of each testing tank (except the side facing the margin of the maze) had a hole (2 cm × 2 cm; water depth, 0.5 cm) which males could use to cross. Tanks connected with a small tunnel mimics small pools connected with a narrow stream. In this artificial system, less dispersive males do not tend to move into other tanks through the tunnel, but stay in the original tank where they were placed. More dispersive males, however, actively move into other tanks with less hesitation (shorter duration of pauses) at the front of the tunnel. So this assay does not only simply reflect males’ activity but also measure their decision whether they move into other habitat or not. A male was transferred to the centre of the testing tank of the maze and left undisturbed for 16 min. We measured the number of crossings between the tanks for 15 min, ignoring the first minute to allow for acclimation.

Exploration in a novel environment

We placed males in an unfamiliar environment and measured the time they spent exploring and grooming. Grooming requires a male to stand still, and because grooming and exploration are mutually exclusive, they are negatively correlated. We measured both behaviours in a novel environment and extracted a single principle component as a composite measure of the grooming vs. exploring behaviour. The novel environment consisted of a tank (50 cm × 50 cm × 50 cm; water depth, 10 cm) illuminated from above. The test male was transferred to one corner of the tank and released (males released in the centre of the tank move immediately to the margins for reasons unrelated to exploration; C. S. Han, personal observation). The male was left undisturbed for a 30-s acclimation period and then for a 5-min assay. The male could move freely in the open novel environment, and the behaviour was filmed with a video camcorder. We measured two behaviours of males for 5 min: (i) the number of strokes (movement of middle legs to move forward = ‘exploring’) and (ii) grooming duration (cleaning legs or body during staying on the water).

Actively exploring males spent less time grooming, but inactive males spent more time grooming. We extracted a single principle component (eigenvalue = 1.62, explained variance = 81.1%, component loadings, stroke = 0.55, grooming = −0.55) and component scores for each male. Males that had higher PC scores spent less time grooming, but explored a larger area. Males that had lower scores on PC spent more time grooming, but they were inactive.

Latency to remount

Male water striders mate by grasping and then mounting females. In response to male grasping, females struggle and attempt to dislodge the male, often succeeding despite male persistence (Arnqvist, 1997; Han et al., 2010). Less eager males give up attempts to remount after being dislodged by resistant females, whereas more eager males attempt to remount resistant females more often despite the dislodgement. To measure the level of a male's eagerness to remount reluctant females, we put three females in a male's testing tank and waited until the male mounted one of the females. After mounting, we carefully dislodged the mounting male by catching the pair, gently lifting the male off the female and then placing them next to each other on the water. Then we measured the latency until the male remount after artificial dislodgement.

Sex recognition sensitivity

Reproductively active water strider males indiscriminately attempt to mount other individuals and then dismount if their target is either another male or a mating pair. The speed with which a mounting male dismounts is defined as ‘sex recognition sensitivity’. A male's sex recognition sensitivity is consistent through time, but shows between-individual variation (Han & Brooks, 2013). To measure a male's sex recognition sensitivity, we followed the method outlined by Han & Brooks (2013). Briefly, we presented the focal male with three other males in the focal male's home tank and recorded the behaviour of the focal male for 5 min. We recorded the duration from the moment the focal male mounts another individual until the moment he dismounts (duration of the first mount). Insensitive males stay longer on the back of other males, whereas sensitive males would spend less time on them.

Statistical analysis


We calculated the repeatability of behavioural traits and the standard error of the repeatability following Nakagawa & Schielzeth (2010). With the exception of dispersal ability, normality of residuals was tested using Shapiro–Wilk test, and ‘latency to remount’ and ‘the number of stroke’ were log-transformed to improve normality. To calculate the repeatability, linear mixed model approach was used for ‘grooming duration’, ‘latency to remount’ and ‘the number of stroke’, and a generalized linear mixed model approach with the log link was used for ‘dispersal ability’ (Nakagawa & Schielzeth, 2010). We used multiplicative overdispersion models because they allowed for underdispersion. The calculation was conducted using r package rptR (Nakagawa & Schielzeth, 2010).

Behavioural syndrome

Correlations between behavioural traits were assessed using Spearman's rank nonparametric correlations. All behavioural measures were standardized before calculating the correlation, and behaviours in exploration assay were summarized and standardized using principal components analysis. Statistical significance was inferred if P values remain significant after the procedures for controlling false discovery rate (Benjamini & Hochberg, 1995; Storey & Tibshirani, 2003). Because Bonferroni adjustment has been shown to be overly conservative (Benjamini et al., 2001; Nakagawa, 2004; Narum, 2006), we used a false discovery rate (Benjamini & Hochberg, 1995; Benjamini & Yekutieli, 2001; Storey & Tibshirani, 2003; Dochtermann, 2010) to account for six correlations (α < 0.0204 for six correlations).

Instead of focusing on pairwise behavioural correlations with their inherent problems (for a full discussion, see Dingemanse et al., 2010), it often helps to compare the fit of a priori considered structural equation models (SEM; Dochtermann & Jenkins, 2007; Dingemanse et al., 2010; Bókony et al., 2012). This alleviates the need to control for multiple testing because the test is conducted on the whole matrix. However, as this study measured nonlinear selection acting on the combination of two behaviours, the measurement of pairwise behavioural correlations is also appropriate to compare the pattern of behavioural correlation with correlational selection. Thus, we present pairwise behavioural correlations for behavioural syndrome structure of male water striders, but also provide SEM models of syndrome structures as a supporting information (Fig. S2, Table S1). Statistical analyses for SEMs were performed using amos 21.0, SPSS Inc (Arbuckle, 2006).

To compare the syndrome structures across generations, we used common principal components analysis (CPC; Phillips & Arnold, 1999). Matrix comparison in CPC proceeds by evaluating the similarity of matrices in a hierarchical fashion, from unrelated matrices to partially shared principal components to equal covariance structure of matrices (Phillips, 1998; Phillips & Arnold, 1999).

The effect of behavioural syndrome on mating success

Although Lande & Arnold's (1983) phenotypic selection approach has been applied to study how selection operates on single traits, selection is not likely to act on only single traits. Phenotypic selection approach can also detect selection pressures of multiple traits and combination of traits (i.e. correlational selection) by multivariate selection approach as well as single traits (Lande & Arnold, 1983; Blows & Brooks, 2003; Blows, 2007).

We used linear mixed model to estimate the effect of behavioural traits on mating success in different sex ratio conditions and calculate linear and nonlinear selection gradients. Mating success was significantly repeatable across sex ratio treatments for individual males (repeatability, = 0.18, 95% CI = 0.03–0.34), but the proportion of mating success decreased as treatment sex ratio increased [M : F (1 : 1) = 0.86; M : F (1.5 : 1) = 0.65; M : F (3 : 1) = 0.33; M : F (5 : 1) = 0.18]. We base all selection analyses on within-treatment relative mating success, which we calculated by fitting the sex ratio treatment as a fixed factor in a model with mating success as either 0 (failure) or 1 (success) and a binomial error distribution. We then used this residual as the fitness measure in our selection analysis, thus ensuring that selection gradients are scaled relative to the variation in fitness within each sex ratio. The linear and nonlinear terms of standardized behavioural traits were added to the model as the covariates. Male ID, generations and experimental subgroups were added as random factors. First, we measured standardized linear selection gradients β (i.e. directional selection) by adding all four linear terms. Then we estimated the standardized nonlinear selection gradients γ (i.e. stabilizing/disruptive/correlational selection) using second-order polynomial, quadratic and cross-product regression coefficients. All selection analyses were performed on standardized trait values (mean of 0 and standard deviation of 1). As required for quadratic selection analysis, the values of quadratic regression coefficients and their standard errors were doubled (Stinchcombe et al., 2008). The mixed models were performed using spss (SPSS Inc., Chicago, IL, USA). Lastly, we used nonparametric thin-plate splines (Green & Silverman, 1994) to visualize fitness landscape using the Tsp function in the fields package in r.

The importance of nonlinear selection on a suite of traits is often underappreciated because of the large number of coefficients of quadratic selection and correlational selection (Blows & Brooks, 2003). A canonic rotation of the nonlinear response surface can expose the major axes of nonlinear selection (Phillips & Arnold, 1989). We performed canonical rotation of the matrix (γ) of quadratic selection gradients presented in Table 1 after the method of Phillips & Arnold (1989) to test for multivariate nonlinear selection.

Table 1. Vectors of standardized linear (β) and matrices of quadratic and correlational (γ) selection gradients for four behavioural traits: dispersal ability (D), latency to remount (R), exploration in a novel environment (E) and sensitivity to distinguish the correct sex (S)
  1. The significant terms are indicated in bold. Values in parentheses are standard errors.

Dispersal−0.005 (0.025)−0.036 (0.064)   
Remount latency0.063 (0.026)0.022 (0.044)−0.238 (0.162)  
Exploration0.021 (0.024)−0.008 (0.038)0.060 (0.028)−0.060 (0.046) 
Sex sensitivity0.042 (0.025)0.012 (0.031)0.017 (0.031)−0.017 (0.034)−0.008 (0.060)



Nonzero repeatability simply requires nonzero among-individual variance. In addition, when individuals behave more consistently through time period (low within-individual variance), then the repeatability of behaviour shows a higher value. We found strong support for the presence of consistent individual variation in the number of strokes (= 0.75, = 20, SE = 0.11, = 0.001), grooming duration (= 0.55, = 20, SE = 0.16, = 0.007), latency to remount (= 0.61, = 20, SE = 0.19, = 0.013) and dispersal ability (= 0.58, = 17, SE = 0.26, = 0.04). Thus, individual differences in all behavioural traits were highly consistent over a time period of 48 h.

Behavioural syndrome

The structure of the behavioural syndrome was similar across generations (Table 2, CPC, test of ‘equal matrices’ vs. ‘unrelated matrices’: χ2 = 1.84, d.f. = 10, = 0.99). In both generations, exploration behaviour was not significantly correlated with any of the other behaviours (Table 2). Latency to remount was negatively correlated with sex recognition sensitivity and, marginally, with dispersal ability (Table 2). However, the cross-generation analysis revealed a significantly negative correlation between latency to remount and dispersal ability (Table 2). Thus, males that rapidly attempt to remount on resistant females had poor sensitivity to distinguish single females from single males, and rapidly remounting males also dispersed further.

Table 2. The structure of behavioural syndrome between behaviours, dispersal ability (dispersal), exploration in a novel environment (exploration), sensitivity to distinguish the correct sex (sex sensitivity) and latency to remount (remount) in two different generations, generation 1 (spring generation) and generation 2 (summer generation)
 Generation 1 + 2Generation 1Generation 2
Rs n P Rs n P Rs n P
  1. Bold Rs values (Spearman's rank nonparametric correlations) indicate a statistically significant correlation between behavioural traits after a false discovery rate B–Y adjustment (α < 0.02).

Dispersal – Remount0.29830.008−0.29400.06−0.30430.05
Dispersal – Exploration0.01830.940.10400.54−0.08430.59
Dispersal – Sex sensitivity0.21830.060.14400.400.24430.11
Remount – Exploration0.02830.84−0.001400.990.08430.61
Remount – Sex sensitivity0.4583<0.0010.41400.0090.45430.002
Exploration – Sex sensitivity−0.04830.74−0.16400.33−0.01430.97

Structural equation model analyses also supported the existence of the correlation between latency to remount, sex recognition sensitivity and dispersal ability (Table S1). In the model that best explained the data, latency to remount negatively covaried with dispersal ability and sex recognition sensitivity (Fig. S3, Table S1).

The effect of behavioural syndrome on mating success

Directional selection on males favoured shorter latency to remount (β = −0.063, SE = 0.026, F1,327 = 6.03, = 0.02, Table 1). That is, males that attempted to grasp a single female rapidly even after the dislodgement from females due to the resistance enjoyed greater initial mating success than others. Also poor sex recognition was favoured (β = 0.042, SE = 0.025, F1,327 = 1.67, = 0.10, Table 1). There were no differences in linear selection between sex ratio treatments (Models 2 and 3, ΔAIC = +3.4, χ2 = 4.54, d.f. = 4, = 0.34, Table 3). There was also no significant evidence of nonlinear selection on the suite of four behavioural traits (Models 2 and 4, ΔAIC = +9.6, χ2 = 11.35, d.f. = 10, = 0.33, Table 3) or of differences between sex ratio treatments in the nonlinear components of the fitness surface (Models 4 and 6, ΔAIC = +13.7, χ2 = 14.28, d.f. = 14, = 0.43, Table 3).

Table 3. Model selection procedure. Model selection was based on the Akaike's information criterion (AIC) and the likelihood ratio test, starting with the model with sex ratio (SR) treatment only (model 1) and ending with model 6, which includes all terms, linear terms, nonlinear terms and their interactions with sex ratio treatments
 Model structure (Mating success as a dependent variable)d.f.AIClog Lχ2 P
  1. The best model is indicated in bold. Male ID, generation and experimental subgroups were included as random factors in all model structures. SR: sex ratio; nonlinear terms: quadratic terms and interactions between linear terms of behavioural traits.

2SR + linear terms9 388.4 −
3SR + linear terms + SR*linear terms13391.8−
4SR + linear terms + nonlinear terms19397.0−
5SR + SR*linear terms + nonlinear terms23400.3−
6SR + SR*(linear terms + nonlinear terms)33410.7−

We conducted a canonical rotation of the γ matrix of nonlinear selection gradients to extract the main dimensions along which nonlinear selection occurs (Blows & Brooks, 2003 for a full rationale of this procedure). This canonical rotation resulted in a matrix (M) of four major axes (m1–m4) of nonlinear selection. Only one of these eigenvectors, m4, proved to be under significant nonlinear selection (λ4 = −0.26, SE = 0.08, F1,323 = 10.685, = 0.001, Table 4). As m4 was heavily influenced by latency to remount with a smaller contribution from exploration ability, stabilizing selection on m4 indicated selection acting on the combination of latency to remount and exploration ability. However, the peak of convex selection was biased towards smaller values of m4, and negative directional selection operated on m44 = −0.06, SE = 0.03, F1,327 = 4.819, = 0.03, Table 4). This effect of fast remounting and greater exploration ability on mating success was largely driven by negative correlation selection between remounting and exploration (γ = −0.06, SE = 0.03, F1,317 = 4.50, = 0.04, Table 1, Fig. 1).

Table 4. The M matrix of eigenvectors from the canonical rotation of the nonlinear response surface, and gradients of nonlinear (λi) and linear (θi) selection on each eigenvector
mi M Selection
DispersalRemountExplorationSex sensitivityλiθi
  1. λi: The quadratic selection gradient; θi: the linear selection gradient; mi: eigenvector. The quadratic selection and linear selection were only significant along one eigenvector, m4 (indicated in bold). Values in parentheses are standard errors.

m10.3740.186−0.4020.8150.010 (0.034)0.012 (0.025)
m20.8250.112−0.215−0.510−0.038 (0.054)−0.038 (0.027)
m30.416−0.2020.8440.272−0.055 (0.036)0.040 (0.022)
m4−0.0820.9550.282−0.0420.258 (0.080)0.056 (0.025)
Figure 1.

The fitness surface of the two major axes of nonlinear selection. Numbers on contours are in units of the mating success. All behavioural traits were standardized to mean zero with unit variance. Smaller values of sex recognition sensitivity indicate that males more rapidly dismount after mounting on the back of another male. Smaller values of remounting attempts indicate that males more rapidly attempt to remount females even after they are dislodged in premating struggling.


We found that water strider males showed similar pattern of behavioural syndrome across two cohorts within a summer. This replication of the estimated behavioural syndrome provides added confidence that the syndrome is a real and consistent phenomenon. Despite the presence of significant linear and nonlinear selections, however, the behavioural syndrome structure was not significantly associated with the pattern of correlational selection on behaviours.

The syndrome also appears unlikely to have been shaped by the pattern of multivariate linear selection. The most successful males were those with poor ability to recognize the correct sex (long latencies in sex sensitivity scoring) and those which rapidly attempted to remount females when dislodged (low latency to remount). One would expect selection to fix alleles that conferred poor sex recognition (long latencies in sex sensitivity scoring) and fast remounting (low remount scores), eliminating alleles with the opposite combinations of effects, and thus leave segregating only alleles that have positive effects on both or negative effects on both. This would leave the two behaviours positively correlated, and yet the strongest correlation in the syndrome is a negative one between these two traits.

Taken together, our results suggest that the strong correlations underpinning the behavioural syndrome we document may not have arisen due to correlational selection and linkage disequilibrium. Although our results suggest that correlational selection, at least as measured under a variety of conditions in the laboratory, is not associated with the pattern of the syndrome, they constitute only a weak refutation of the correlational selection hypothesis as set out by previous research (Sih et al., 2004b; van Oers et al., 2005; Bell, 2007). It remains possible that correlational selection over previous generations in the field may have given rise to the observed behavioural syndromes. If that is the case, we would predict that linkage disequilibrium will decay rapidly if animals are kept for several generations in the laboratory.

One reason why selections that water striders experience in the wild might differ from the ones in the laboratory is that artificial laboratory studies do not mimic all the kinds of environmental conditions. Despite a variety of artificial conditions (e.g. sex ratio or population density) in this laboratory-based experiment, we could not include the change of other abiotic environmental conditions (e.g. predation risk, food level and shelter) in the measurement of individual fitness. Thus, it remains unknown whether selection similarly acts in wild populations (Smith & Blumstein, 2008; Adriaenssens & Johnsson, 2009). In consequence, it remains possible that correlational selection over previous generations in the field may not be detected under laboratory condition.

Similarly, the syndrome may arise due to correlational selection at other life stages. Even in the absence of nonlinear selection at the current environment, fluctuating directional selection at different life-cycle stages can produce net lifetime nonlinear selection (McGlothlin, 2010). Thus, the snapshot of correlational selection in adulthood may not fully capture the true fitness surface or the forces shaping the behavioural syndrome.

We used phenotypic correlations to obtain behavioural syndromes. However, statistically, behavioural syndromes do not simply indicate significant phenotypic correlation (Dingemanse et al., 2012; Dingemanse & Dochtermann, 2013). Recent debate over the statistical definition of behavioural syndrome (Dingemanse et al., 2012; Garamszegi & Herczeg, 2012; Brommer, 2013) indicates that behavioural syndromes are between-individual correlations rather than phenotypic correlations including both between-individual correlations and within-individual correlations. That is, phenotypic correlations can occur even in the absence of between-individual correlations when the change in one behaviour strongly correlates with the change in other behaviours within an individual (within-individual correlations; Dingemanse & Dochtermann, 2013). To distinguish between-individual correlations from within-individual correlations, it would be appropriate to apply a multivariate mixed effect model to estimate correlation structure at the between- and within-individual levels (Dingemanse & Dochtermann, 2013). However, as repeated measures used to calculate repeatability were separately conducted from assays to calculate behavioural syndromes and selections in this study, we were unable to calculate between-individual correlations.

Nonetheless, the evidence we present that the syndrome is not tightly shaped by contemporary correlational selection or by the combination of linear selection gradients also suggests that linkage disequilibrium may not be maintaining the nonrandom associations between traits. For now, we must consider pleiotropy a more likely basis for the behavioural correlations. The genetic basis of the correlations between behaviours that we document remains to be formally dissected, but it might be fruitful to test for pleiotropic genetic variation and shared physiological mechanisms that might generate the observed behavioural syndromes. In this case, the behavioural syndrome might arise via regulatory hormones coordinating the expression of multiple traits (i.e. hormonal pleiotropy; Ketterson & Nolan, 1999; Flatt et al., 2005; Lessells, 2008).

As multiple phenotypes in insects can be controlled by a single endocrine regulator (Dingle & Winchell, 1997; Gade et al., 1997; Hartfelder, 2000), one of the stress hormones might induce a male to grasp and court other individuals rapidly and indiscriminately irrespective of sex. If so, males with poor sex recognition sensitivity simultaneously have a quick response to remount on resistant females. Also, a mechanism to increase activity of males may affect both tendencies to disperse and remount on females. Thus, pleiotropic genetic effects and common physiological pathways might underpin the observed behavioural syndrome (Sih et al., 2004b; van Oers et al., 2005; Bell, 2007), regardless of the action of any correlational selection.

Another possibility that merits future testing is condition dependence (state dependence) (Luttbeg & Sih, 2010; Wolf & Weissing, 2010). If all traits depend on an underlying quality such as condition, then the large mutational variance in that quality can maintain considerable variation in and covariation among the traits (Rowe & Houle, 1996; Tomkins et al., 2004). It also has been shown that many individual behavioural traits are condition dependent (Houston & McNamara, 1999; Clark & Mangel, 2000; Luttbeg & Sih, 2010; Wolf & Weissing, 2010). In the same way, if individual behaviours in different contexts are influenced by the same inherent condition, individual condition can also play a role of producing behavioural syndrome. The strong negative covariation between sex sensitivity (for which long latencies in sex sensitivity scoring confer high fitness) and remounting (short latencies equate to high fitness) suggests there may be a general axis of quality underpinning these two traits. High state males might spend their energy to attempt to mate more randomly (poor sex recognition) and to remount more quickly. Whether the expression of behavioural correlations depends on individual condition remains to be tested.


We have demonstrated that the structure of behavioural syndrome did not match with the pattern of correlational selection on the fitness landscape. This provides the first empirical test of the idea that a behavioural syndrome might be generated by correlational selection (and resulting linkage disequilibrium). We found no evidence in support of this idea. However, to more strongly reject the correlational selection hypothesis, we would need to clarify lifetime selection on behaviours of water strider males. Thus, in addition to the genetic estimates of animal personality traits (Weiss et al., 2002; van Oers et al., 2004, 2005; Bell, 2005), our findings emphasize the need for research on the genetic architecture of behavioural syndromes (Dingemanse et al., 2012; Garamszegi & Herczeg, 2012; Brommer, 2013; Dingemanse & Dochtermann, 2013), especially within appropriate ecological contexts (Boake et al., 2002). We reiterate Dochtermann & Dingemanse's (2013) exhortation that genetic studies of behavioural syndromes may provide important insights into how behavioural syndromes constrain evolutionary responses (Dochtermann & Dingemanse, 2013).


This research was supported by a Rosemary Grant Award from the Society for the Study of Evolution to CH and an ARC Fellowship grant to RB. We thank Michael Jennions, Niels Dingemanse and anonymous reviewers who provided helpful comments on this manuscript. We thank Heather Try for her assistance with maintaining the animals. CH thanks Tom Weir for providing collecting information of water striders in NSW.