1. In freshwater fishes, inter-population variation in male phenotype is often associated with differences in predation intensity, but these effects can be difficult to disentangle from environmental influences.
2. The western rainbowfish Melanotaenia australis exhibits marked sexual dimorphism – females are plain with a slender body, while males have striking coloration and are deeper in the body. Male traits differ in expression among populations, but this has not been described or explored in the literature.
3. This paper describes a study designed to test for geographic structuring of male phenotype in M. australis and to determine whether between-population variation in male phenotype is attributable to variation in predation regime, after accounting for environment.
4. We collected data describing habitat, and the size, activity and abundance of predators at sites containing M. australis populations. We then used photography, spectrometry and geometric morphometrics to describe colour pattern, spectral reflectance and body shape in males from these populations. Finally, we used permutation-based multivariate statistics to partition variance in these traits according to environment and predation regime.
5. Downstream environments posed higher predation risk to M. australis. Furthermore, males from these sites consistently exhibited larger cheek spots and fewer coloured lateral stripes than those from upstream sites. Variation in predation regime accounted for a significant proportion of the total variance in these traits (30·9%), after controlling for the effects of environment.
6. Variation in predation regime did not explain variation in reflectance or shape. Environmental variation, however, explained a significant portion of the total variance in reflectance (74·9%), and there was a strong trend towards it explaining a portion of the total variance in body shape (34·9%).
7. We conclude that natural selection by predators may be an important determinant of the evolution of colour pattern variation in M. australis, but not of that of body shape or colour reflectance.
8. Further study of M. australis will complement existing models, which show complex relationships between predation regime, environment and phenotype. Understanding these relationships is prerequisite to predicting the evolution of phenotypic variation in natural systems.
Populations can adapt to the presence of predators by evolving colour patterns that help them remain undetected (Endler 1978). In the guppy Poecilia reticulata, an increase in the risk of predation is associated with a reduction in the number and size of colour patches in males (Endler 1980). Selection by predators on colour pattern has also been seen in other fishes (for examples see Semler 1971 and Haas 1976) and other taxa including, but not limited to, moths, birds and reptiles (for examples see Tutt 1896; Slagsvold, Dale & Kruszewicz 1995; Forsman & Shine 1995; Götmark & Olsson 1997; and Niskanen & Mappes 2005). High levels of predation may also select for colour patterns that make greater use of ultraviolet reflectivity. Although predators may have vision in the ultraviolet, in water, ultraviolet light is readily scattered over distance because of its short wavelength. Thus, because potential predators are typically further away than conspecifics, ultraviolet-based patterns may be less visible to predators but still useful as a means of signalling among conspecifics (Endler 1991; Reckel et al. 2002).
The persistence of conspicuous colour patterns in a population, despite natural selection against them, indicates the presence of counter-selection for them (Fisher 1958). Male guppies, for example, must exhibit conspicuous coloration despite the risk of attracting the attention of predators, or they will be at a reproductive disadvantage. It has been widely demonstrated that females in many species, including guppies, discriminate among potential mates on the basis of coloration (Endler 1983; Price 2006). The optimal colour pattern for male fishes may therefore be determined by a balance between natural selection for crypsis and sexual selection for conspicuousness (Endler 1978; Price 2006).
Predation may also facilitate the evolution of morphological adaptations beyond coloration (for examples see Touchon & Warkentin 2008 and Møller, Couderc & Nielsen 2009). Mosquitofish, Gambusia affinis, from high-predation areas have a larger caudal region, smaller head and more elongate body than those from low-predation areas (Langerhans et al. 2004). This body shape confers a significantly faster burst swimming response, which is critical for predator evasion (Langerhans et al. 2004). A similar trend is observed for male, but not female, guppies, suggesting that predation is a more important determinant of morphology for males than it is for females (Hendry et al. 2006). This is unsurprising, given that their colouration makes male guppies more vulnerable to predators (Houde 1997).
More recent empirical evidence, however, suggests that Endler’s (1978) theoretical model alone may not be sufficient for predicting the evolution of cryptic or conspicuous traits under different predation regimes. In natural populations of guppies, Weese et al. (2010) found that colour patterns were not under stronger natural selection at sites with high predation risk, and Houde & Endler (1990) and Rodd et al. (2002) found that females do not always appear to favour males with conspicuous colour patterns. Furthermore, although inter-population translocations of guppies have been performed in the field following Endler (1980), the resulting evolution of traits differed from expectations derived from theory (see Karim et al. 2007; Kemp et al. 2009). Thus, although there is an extensive literature investigating the effects of predation regime on phenotype in guppies, the situation appears more complex than originally envisaged, and there is a clear need for additional research with other systems.
In this paper, we explore the associations between predation regime and phenotype in the western rainbowfish Melanotaenia australis (Castelnau 1875), a highly abundant freshwater fish endemic to the Pilbara and Kimberley regions of north-western Australia. Males have bright colouration (incorporating a brightly coloured spot on the operculum – a ‘cheek spot’– and multiple coloured lateral stripes on the body) and elaborate fins, whereas females are relatively plain, suggesting that sexual selection is acting on male phenotype. There is anecdotal evidence from aquarists (e.g. Allen & Cross 1982) and field guides (e.g. Allen, Midgley & Allen 2002) that these male traits vary considerably across the range of the species. Furthermore, the distributions of sympatric fish species suggest that M. australis may be exposed to a variety of predation regimes across their distribution. Downstream locations, with their estuarine influence and larger, more permanent river channels, are host to more species of fish, and therefore possibly more species of predator, than upstream locations (see distributions in Allen, Midgley & Allen 2002).
The aims of this study were to formally document the phenotypic differentiation present across the range of M. australis, to test for spatial structuring of phenotype using multivariate statistical approaches and to partition phenotypic variance according to predation gradients and environmental effects. We hypothesized that downstream locations would be more dangerous habitats for rainbowfish than upstream locations and that phenotypic traits of M. australis would differ accordingly. We sampled across the distribution of the species in Western Australia and used a combination of photography to describe colour patterns, spectrometry to describe reflectance (visible and ultraviolet) and geometric morphometrics to describe body shape. We then used permutation-based multivariate statistics to test for associations between phenotype and site location and between phenotype and predation regime, while controlling for habitat variation.
Materials and methods
Field sites were initially chosen according to a hierarchical design. Six major drainages were chosen in each of the Pilbara and Kimberley regions of Western Australia and, within each, one downstream site and one upstream site were identified. This gave a total of 24 sites, but for logistical reasons, we were only able to access 21 of these (Table S1). Each site consisted of a pool connected to, or within, the main river channel, in which sampling and observations were conducted. Field work took place during May–June 2007.
Predation regime data
At each site, visual observations of fishes were made at c. 09.00 (±1 h) on the day of sampling. This was always the first task to be undertaken. Two observers simultaneously watched different aggregations of M. australis for 15 min and recorded all other species seen, their sizes (estimated by eye, to the nearest cm) and the occurrence of predatory interactions. Typical interactions were ‘chases’, where another species lunged at or chased a rainbowfish, ‘strikes’, where another species made contact with a rainbowfish, and ‘kills’, where a predator captured a rainbowfish. On an ad hoc basis, depending upon site conditions, underwater video cameras were used to monitor the fishes present. If new species were seen in addition to those recorded during the observations, these were added to the species list for that site. Similarly, if during the course of other operations at a site (netting and trapping for rainbowfish, measurement of water chemistry parameters) additional species were observed, these were also added; our predator presence/absence data thus represent the total number of species seen over the sampling duration (6 h) at each site.
Species were identified as predators of M. australis based on our own field observations and based on the literature (references are provided alongside the data; Table S2). Where information about a species was not available, its likely status as a rainbowfish predator was determined based on information about other closely related species, or on information at a family level. Species identified as likely predators of M. australis were then categorized as either low-risk (fishes which would opportunistically eat rainbowfish) or high-risk (piscivores which would actively hunt rainbowfish).
At each site, salinity, pH, conductivity, temperature and turbidity were measured using an electronic meter (Model 611; YEO-KAL Electronics, Brookvale, NSW, Australia), and physical characteristics were assessed visually (pool volume, water surface velocity, riparian vegetation as a percentage of shoreline cover, aquatic vegetation presence on a scale of 0–5, substrate type on a scale of 0–5, woody structure presence on a scale of 0–5; when scales were used, 0 denoted complete absence and 5 denoted extreme/choking abundance in the cases of aquatic vegetation and woody structure, and 0 denoted detritus and 5 denoted large stones >65 mm diameter in the case of substrate type). Water samples were taken for the purpose of assessing spectral absorbance. Each sample was filtered (#2 paper, Whatman), and its absorbance was measured at 10 nm intervals over the range 325–795 nm, using a spectrophotometer (NovaSpec II; Pharmacia-LKB, Strängnäs, Sweden) calibrated with filtered, deionized water.
Rainbowfish phenotype – raw data collection
Different methods were used to capture M. australis at different sites. Sampling methods were chosen based primarily on operator safety, rather than habitat characteristics; thus, there were no obvious associations between habitat type and particular sampling methods. Indeed, at the majority of sites, a 10 m pocket-seine net with 5 mm mesh was used (Table S1). Seining commenced in deeper water and continued by sweeping the net into the shoreline. At some sites, dip nets and stationary fish traps baited with tinned cat food were adopted in addition to, or instead of the seine. After capture, all fish were euthanased using a 10× overdose of AQUI-S® (an isoeugenol-based liquid anaesthetic designed specifically for fishes; AQUI-S New Zealand, Ltd., Lower Hutt, New Zealand) as described by Young (2009). It was not possible to catch fish at all sites; M. australis phenotypic traits were documented from 10 to 13 sites, depending on the category of trait considered (Table S1).
Reflectance was measured in the field using a USB-4000 spectrometer and a UV-VIS DT-MINI light source (Ocean Optics, Dunedin, FL, USA), with a shield fitted over the reflectance probe to exclude ambient light from the surface of the sample (see Young, Simmons & Evans 2010 for exact procedure). One reflectance measure was taken at each of five places on the fish: at the centre of the tail, on the reflective cheek spot, at the beginning of the lateral line, at the end of the lateral line and at the centre of the lower flank (Fig. 1). Fish were then photographed out of the water by placing each individual on its side, on a white plastic card, within a ‘light box’– an upturned plastic box with opaque white sides to minimize interference from ambient light, and with built-in lights (two 6500 K, 11 W ‘daylight’ fluorescent tubes) to give even illumination. A camera lens protruded into the box from a cut-to-size hole in the top, and photographs were taken using a Fuji S9500 digital camera. Fish were stored in Dietrich’s fixative [by vol.: 58% H2O, 30% EtOH (95%), 10% formalin (37% formaldehyde), 2% glacial acetic acid]. This gives minimal distortion of tissues (Fournie, Krol & Hawkins 2000). In the laboratory, the preserved fish were arranged with fins spread and photographed (7·6× magnification) for morphometric analysis.
Rainbowfish phenotypic data – raw data treatment
Photographs of fresh specimens were imported into ImageJ v. 1.39 (National Institutes of Health, USA). The area of the cheek spot and the number of coloured lateral stripes were measured on each fish. These are hereafter referred to as pattern variables. Each fish had five reflectance spectra associated with it (one spectrum for each of five different body locations). Each spectrum consisted of over 3500 data points, necessitating data distillation. This was performed following Young, Simmons & Evans (2010) and reduced each spectrum to 20 measures of reflectance at wavelengths every 20 nm, from 325 to 705 nm inclusive.
Following distillation, spectra were used in principal component analyses (PCAs). There were five sets of data (five body locations where spectra were measured on each fish); within each were 20 variables (20 measures of reflectance, each at a different wavelength) and 202 samples (those 20 measures of reflectance were recorded on each of 202 fish – but see note in statistical analyses section about sample size in the reflectance data set). For each set of spectral data, the PCA simplified reflectance spectra into principal components (PCs) (orthogonal variables) describing mean reflectance and spectral shape (Endler 1990; LeBas & Marshall 2000). Only those PCs with eigenvalues greater than one, and which collectively explained 90% or more of the observed variance in each PCA, were used in subsequent analyses. These are hereafter referred to as reflectance variables. Absorbance spectra of the water samples taken in the field were treated in the same way, and the resulting PCs were incorporated into the environmental data set.
Photographs of preserved specimens were used for the geometric morphometric analyses. They were imported into TPSdig v. 2.10 (F. J. Rohlf, SUNY, Stony Brook, NY, USA), and landmarks (18 fixed and 27 sliding) were assigned to the outline of the body (Fig. 1). A relative warps analysis was conducted using TPSrelw v. 1.46 (F. J. Rohlf, SUNY Stony Brook), which generated warp scores describing variation in shape among individuals. These are hereafter referred to as shape variables. TPSrelw was also used to calculate centroid size for each individual (a measure of overall body size), which was used as a covariate in some of the subsequent analyses.
Paired t-tests were used to test for differences between upstream and downstream sites in the number of predator species (within drainages) and predator maximum size (within species). The strength of the association between the number of predator species and site predation risk (=number of chases + strikes + kills) was assessed with Spearman’s rank order correlation.
To test for spatial structuring of phenotype, we used three nested permutational multivariate analyses of variance (nested Permanovas; Anderson, Gorley & Clarke 2008) – one each for the pattern, reflectance and shape data. permanova uses permutation methods to test for the response of one or more variables to one or more factors in an anova experimental design, on the basis of resemblance measures (Anderson 2001a). permanova was chosen because of the presence of multiple response variables within each analysis, the unbalanced nature of the data (response data not available for all sites; Table S1) and because the data would violate the assumptions of standard anova. The permanova routine calculates a distance-based pseudo-F statistic for each term in the model, based on the expectations of mean squares, and then calculates P-values using a permutation procedure (Anderson 2001b). Although permanova does not require normality in its predictor and response variables, our data were prescreened for outliers, dispersion biases and multicollinearity. Four fish (of the total of 202 – see Table S1) were removed from the reflectance data because of extreme outlying scores for one of the PCs describing spectral shape of tail reflectance. PCAs were re-calculated, accordingly – all subsequent analyses of the reflectance data set were thus based on a total of 198 individuals – and all other data were acceptable. Each permanova design contained three factors – Region, Drainage (Region) and Site (Drainage) – and the Euclidean distance matrix of the phenotypic scores for individual fish was tested for a response to these factors. Centroid size was used as a covariate to control for any effect of variation in overall fish size at different sites. After each permanova, least squares means were used to further investigate any patterns identified. These were derived from general linear models using only those data for which both upstream and downstream observations were present.
To test for an association between phenotype and predation regime, we used distance-based linear modelling (distlm; Anderson, Gorley & Clarke 2008) on a site-by-site basis. distlm partitions the variation in a multivariate data cloud, as described by a resemblance matrix, according to a regression model (Legendre & Anderson 1999). It can be used to analyse data sets that contain a mixture of categorical and continuous predictor variables, and as with permanova, the general null hypothesis of no relationship is tested using a pseudo-F statistic, and P-values are obtained using permutation (Anderson 2001b). Three models were created: one for the shape data, one for the pattern data and one for the reflectance data. A phenotypic data matrix for each analysis was calculated from average trait scores, for each variable of interest, at each site. A Euclidean distance matrix derived from this phenotypic data matrix was then used as the response. Each model partitioned the total variation in the response into portions explained by each of three predictor matrices.
The first predictor was a covariate, containing mean centroid sizes for each site (a univariate matrix). The second contained environmental data for each site. This matrix incorporated four of the 14 measured environmental variables and those used were different for each analysis. They were chosen using the routine best in the Primer-E statistical package (Anderson, Gorley & Clarke 2008), which, when given sets of response and predictor variables, lists subsets of predictors that best explain patterns in the responses. For each of the three environmental predictor matrices to be used in distlm, we used best to permute all possible combinations of environmental predictors and identify which set of four, in each instance, maximized the Spearman rank order product moment correlation between the Euclidean distance matrix of the predictors and that of the relevant response variables (i.e. patterns, reflectance or shape). See Clarke & Gorley (2006) for an overview of best (and see Clarke & Warwick 2001, where the routine is referred to as BIO-ENV). The third predictor described predation regime, using four variables: the number of species of high-risk predator recorded at each site, the number of species of low-risk predator, the size of the largest high-risk predator and the size of the largest low-risk predator.
The environmental and predation regime predictor matrices were restricted to four variables each because the number of predictor variables that can be used by distlm is constrained by the number of samples in the analysis. The numbers were balanced (four per matrix) so that the relative predictive power of either matrix was not biased. Each distlm fitted the centroid size data, environmental data and predation data, in that order. Phenotypic variance associated with variance in predation regime was thus calculated only after accounting for phenotypic variance associated with centroid size and environment. Finally, best was used to calculate the Spearman rank order product moment correlation between the Euclidean distance matrices of the entire environmental and predation regime data sets, to assess the likelihood of a (potentially confounding) correlation between environment and predation regime.
Multivariate analyses were performed using Primer-E v. 6 with the permanova+ add-on (PRIMER-E, Ltd, Ivybridge, UK), following Anderson, Gorley & Clarke (2008). Other analyses were performed using Minitab v. 15 (Minitab, Inc.), and effect sizes were calculated following Nakagawa & Cuthill (2007).
Predation regime and predation risk
Fourteen species of predator were recorded. Six were categorized as high-risk, and eight as low-risk (Table S2). In drainages where both upstream and downstream sites were sampled, the number of species of predator was significantly higher downstream (Fig. 2; paired t-test, t8 =3·42, P = 0·0090; effect size d =1·70, 95% CI = 0·535–2·865; see also Table S2). For those species of predator seen in both upstream and downstream sites, the maximum size of each species across all downstream sites was not significantly different to that across all upstream sites (Fig. 2; paired t-test, t5 =1·66, P = 0·1570; effect size d =0·79, 95% CI = −0·292–1·862; see also Table S2), despite a trend in the data towards predators from downstream sites being larger than their upstream counterparts (Fig. 2). High-risk predators were usually only present at downstream sites (Table S2). There was a positive correlation between the number of species of predator, and the total number of chases, strikes and kills, at sites (Spearman rank order correlation; rs = 0·639, t17 =3·43, P = 0·0032).
From the five PCAs of the spectral data, a combined total of 11 PCs (variables describing the overall reflectance and spectral shape of reflectance) were retained. For each of the five PCAs (one PCA for each of the five measurement locations on the body), the first PC described overall reflectance at the location of interest. The remaining PCs then described different aspects of spectral shape. For each analysis, the variables selected for subsequent use (three PCs for the tail and two PCs for each other location) accounted for 90% or more of the total variation in reflectance (Table S3).
For each fish, the geometric morphometric analysis generated 86 relative warp scores for describing shape, as well as centroid size for describing overall body size. The first 15 warp scores, which collectively explained 95·5% of the total variance in rainbowfish shape, were retained (Fig. S1). The first three relative warps, which collectively explained almost 60% of the total variation in shape, could be interpreted as describing fish body height relative to body length, fish head size relative to body size and the combination of head size and body height relative to body size and body length, respectively (Fig. S1). The remaining warps explained more subtle, and consequently less interpretable, aspects of fish shape.
Our subsequent analyses combined the reflectance and shape variables into multivariate responses; thus, we do not present their individual characteristics here (but see Table S3, Figs S1 and S2 for complete descriptions, and Fig. S3 for an example of raw reflectance data).
Spatial structuring of rainbowfish phenotype
Rainbowfish patterns, reflectance and shape exhibited different degrees of spatial structuring when analysed with permanova. Centroid size, included in the analyses as a covariate, was a statistically significant factor in determining pattern and shape (Fig. 3). Region was a significant predictor of shape and tended to influence pattern, but had no influence on reflectance (Fig. 3). Drainage was not a significant factor in determining any aspect of phenotype (Fig. 3). For all aspects of phenotype, the effect of site was highly significant (Fig. 3). See Table S4 for the full permanova tables.
The site and region effects of the two colour pattern variables, examined using least squares means (to take the effect of centroid size into account), exhibited clear patterns (Fig. 4). Melanotaenia australis from the Pilbara region had a larger cheek spot area and greater number of lateral stripes than those from the Kimberley region, and within drainages, fish from downstream populations had a larger cheek spot and fewer coloured lateral stripes than those from upstream populations (Fig. 4).
The raw population means for each of the 11 reflectance measures showed no clear patterns. The effect of site was readily apparent, but differences between sites were not in consistent directions (Fig. S2). The site and region effects for shape were examined using least squares means for the first relative warp, which explained 30·5% of the total variance in shape (Fig. S1). Fish from the Pilbara region had greater body depth relative to body length than fish from the Kimberley, but there was no consistent pattern of differences between upstream and downstream sites (Fig. S4). An examination of the remaining relative warps revealed a diminished effect of region, and the directions of the differences between sites within drainages remained inconsistent (data not shown).
Rainbowfish phenotype as a response to predation regime
For distlm of the pattern data, the environmental variables selected by best as predictors were the surface flow rate, amount of aquatic vegetation present, amount of woody structure present and spectral characteristics (absorbance of UV and infra-red light vs. absorbance of white light) of the water at each site (rs = 0·253). For the reflectance data, the amount of aquatic vegetation present, and the pH, turbidity and spectral characteristics (absorbance of light below 600 vs. that above 600 nm) of the water at each site, were selected (best: rs = 0·643). For the shape data, water temperature, pool volume, surface flow rate and amount of aquatic vegetation present, were selected (best: rs = 0·431).
Centroid size was a significant predictor of variance in shape, and tended to predict variance in pattern (Fig. 3), accounting for 32·0% and 31·2% of phenotypic variance in these traits, respectively. Centroid size was not included in the distlm for reflectance because of its lack of significance in the permanova of the reflectance data (Fig. 3). After taking centroid size into account, the environmental data matrices did not predict variance in pattern or shape, although there was a trend towards this in the distlm for shape (Fig. 3). The environmental data matrix did, however, explain a significant portion (74·9%) of variation in reflectance (Fig. 3). After accounting for environment and centroid size (where applicable), the predation data matrix was not a predictor of reflectance or shape (Fig. 3). However, predation was a significant predictor of patterns (Fig. 3), where it explained 30·9% of phenotypic variance (see Table S5 for the full distlm tables). There was no statistically significant correlation between the Euclidean distance matrices of the entire environmental and predation data sets (rs = 0·181, t76 =1·60, P = 0·1137).
Variation in predation regime was directly associated with variation in colour patterns across populations of the western rainbowfish M. australis, in northern Western Australia. In contrast, predation regime did not explain variation in the other two trait categories examined – reflectance and body shape. Instead, variation in these traits was associated with variation in environmental conditions (for shape, as a strong trend rather than a statistically significant association). We therefore conclude that in terms of colour patterns, our data are consistent with our hypothesis of direct selection on traits by predators, but such selection may not be acting directly on reflectances or body shape.
Melanotaenia australis from downstream sites had larger cheek spots and fewer coloured lateral stripes than those from upstream sites, and a significant portion of the total variance in these traits was explained by variance in predation regime. The main difference between upstream and downstream sites was the presence, in the latter, of high-risk predators, and we propose that this greater diversity of predators results in increased predation risk for rainbowfish in the downstream environment. This assertion is strongly supported by the positive correlation between predator abundance and the number of predation attempts on rainbowfish.
The role played by predation as a determinant of overall phenotypic variation in M. australis appears to be variable, depending upon the traits considered. The importance of predation regime in predicting colour pattern variation, and the lack of importance of environment, suggests that pattern traits may have some value for reducing conspicuousness, regardless of the environmental context. At the same time, the lack of importance ascribed by our analyses to predation as a predictor of body shape and reflectance may indicate that predation does not have an overarching role in defining variation in these traits in this system. The relationships between colour patterns, reflectance and predation regime observed here are consistent with those in guppies; Kemp et al. (2009) found that among-population differences in predation regime are not associated with differences in the colours themselves (i.e. reflectances), but are associated instead with differences in the arrangement of those colours (i.e. patterns). Nonetheless, before discarding reflectance entirely as a variable of interest, it would be wise to further investigate its variation in M. australis. We suggest that gathering reflectance data from additional points on the body, and assessing the way that reflectance can vary within individual pattern elements – comparing reflectance at the centre and around the periphery of the cheek spot, for example, and contrasting these with the reflectance of the adjacent skin – may be a fruitful approach.
Importantly, we lack information on the predators themselves. These species may differ from each other in their visual acuities, and there may be temporal, spatial and ontogenetic variation in their feeding behaviour (Endler 1978). Predators may impart size-specific predation impacts, which we controlled for here using predator sizes, but additional data regarding predator gape width and height could further refine this. Additional demographic measures, with respect to the predator community, would allow a better description of the relative predation risk that populations of rainbowfish are exposed to (Millar et al. 2006). It is also obvious that if behaviour is taken into account, the patterns of individual fish may function as part of a more complex predator-avoidance system at the level of the shoal. Consequently, the adaptive value of large cheek spots and fewer coloured lateral stripes in high-risk areas has yet to be determined; this may be difficult to do so, however, if an interaction between behaviour and coloration exists (through hormonal mediation, for example).
None of our principal component scores describing reflectance was weighted heavily by ultraviolet reflectance alone. This suggests that although rainbowfish do exhibit ultraviolet reflectance, it is not a major feature defining differences among individuals, and its utility as a covert communication channel may be limited. Nonetheless, this warrants further investigation. UV vision and colouration have been demonstrated in other species, and its prevalence as a covert communications channel in fishes is not well understood (Losey et al. 1999). The importance of controlling for body size, particularly with respect to geometric morphometric analyses, must also be recognized. The significance of centroid size as a covariate when partitioning variance in shape indicates the presence of allometric changes in shape with body size (Bookstein 1991). By extension, because growth in M. australis is indeterminate, as is common in fishes (Sebens 1987), fish will change in shape with age. The nature of ontogenetic shifts in body shape warrants further research. Palma & Steneck (2001), for example, suggest that behavioural and visual crypsis in marine rock crabs may only be important at early stages during the life history of this species.
If phenotypic variation across populations of M. australis is the result of evolutionary divergence (as opposed to the result of recurrent episodes of selection on plastic traits), any discussion of predation as the selective force behind phenotypic variation in M. australis assumes that this variation is heritable and adaptive (Endler 1978, 1980). There is no published evidence for the former, although inter-population trait differences persist when fish are captive-bred in a common environment (M. J. Young, J. P. Evans & L.W. Simmons, unpublished data), suggesting a genetic basis to these traits. Regarding the latter, it could be argued that neutral drift, not selection, has resulted in the observed variation, especially between the Kimberley and Pilbara regions. Published data suggest only that fish from these regions are of the same species (McGuigan et al. 2000); population genetics work incorporating fish from our study locations is currently lacking.
In conclusion, our results suggest that selection by predation may account for some aspects of the phenotypic variation exhibited by M. australis. This study adds to the body of literature demonstrating that the relationship between predation regime, environment and phenotype is complex and that it may not be possible to readily discern direct effects of predation in all systems. It is important to note that our data represent an exploratory search for associations between phenotype, predation and environment in M. australis, and the patterns we present do not imply causation. As an initial assessment of a novel system, our data suggest several directions for future work. We clearly need experimental approaches that manipulate predator abundance and test for concomitant changes in male phenotype over successive generations, to complement our geographical survey (as per Endler 1980; Karim et al. 2007; Kemp et al. 2009). This paper also serves as an example of how, using permutation-based multivariate statistics, phenotypic variance can be partitioned into components explained by multiple sets of predictors, even with an unbalanced, nested design, large numbers of variables and categorical and continuous predictors.
We thank T. Faithfull, S. Williams, and R. Tuckett for field assistance, J. Prince for statistical advice, A. Bucholz, F. Harvey, A. Kerpent, E. Phillimore, and F. Simmons (with funding from Shenton College, Western Australia) for laboratory assistance, and three anonymous referees for comments on previous drafts. Funding was provided by the Australia New Guinea Fishes Association, the Australian Geographic Society, and the University of Western Australia. This research operated under UWA animal ethics approval.