Adaptation and constraint in a stickleback radiation



The evolution of threespine sticklebacks in freshwater lakes constitutes a well-studied example of a phenotypic radiation that has produced numerous instances of parallel evolution, but the exact selective agents that drive these changes are not yet fully understood. We present a comparative study across 74 freshwater populations of threespine stickleback in Norway to test whether evolutionary changes in stickleback morphology are consistent with adaptations to physical parameters such as lake depth, lake area, lake perimeter and shoreline complexity, variables thought to reflect different habitats and feeding niches. Only weak indications of adaptation were found. Instead, populations seem to have diversified in phenotypic directions consistent with allometric scaling relationships. This indicates that evolutionary constraints may have played a role in structuring phenotypic variation across freshwater populations of stickleback. We also tested whether the number of lateral plates evolved in response to lake calcium levels, but found no evidence for this hypothesis.


Phenotypic radiations are useful for investigating the relative roles of selection and evolutionary constraints on the location of species in morphospace (Schluter, 2000). Marine populations of Holarctic threespine stickleback (Gasterosteus aculeatus) have repeatedly colonized freshwater habitats that became available at the end of the last glacial cycle about 15 000 years ago (Bell, 1994, 2001). These freshwater populations provide a well-known example of rapid phenotypic evolution independently giving rise to similar phenotypes in different populations (Bell & Foster, 1994; Hendry et al., 2008). Assuming that all freshwater lineages originated from the same ancestral oceanic meta-population (Bell, 1984; Bell & Foster, 1994), it can be hypothesized that descendant freshwater populations inherited similar genetic architectures and developmental constraints. Shared phenotypic optima across freshwater populations may accordingly have driven trait adaptation to similar phenotypic states in the different lineages. Although variational patterns may have changed during freshwater radiations (Berner et al., 2010), there is also evidence that morphological changes have been influenced by ancestral patterns of variation (Schluter, 1996a; Hansen & Voje, 2011; Leinonen et al., 2011). Here, we present a historically informed comparative study of threespine stickleback sampled in a large number of Norwegian freshwater lakes and marine locations to test the relative influence of local adaptation and variational constraints in explaining phenotypic evolution.

Much of the morphological variation in threespine sticklebacks across lakes is hypothesized to be due to variation in foraging opportunity (e.g. Walker, 1997). Lake populations are usually uniform, but a few lakes in North America possess two sympatric ecomorphs (or species): a limnetic form specialized for feeding on plankton and a benthic form specialized for feeding on bottom-dwelling invertebrates in the littoral zone (McPhail, 1984). Benthic fish have larger and deeper bodies, smaller eyes and mouths, and shorter and fewer gill rakers than limnetic fish (McPhail, 1984, 1992; Schluter & McPhail, 1992; Hart & Gill, 1994; Schluter, 1996b). It has been suggested that monomorphic populations often resemble the morphology of one of these two ecomorphs: populations in deep, oligotrophic lakes tend to be specialized for feeding on plankton in open water, whereas those in small, shallow lakes tend to be specialized for foraging on invertebrates in the littoral zone (e.g. McPhail, 1984; Lavin & McPhail, 1985). The two sympatric ecomorphs may therefore represent the phenotypic extremes along a limnetic–benthic continuum on which most allopatric populations of sticklebacks are found (Foster et al., 1998). Certain lake characteristics are therefore predicted to correlate with morphological traits relevant to foraging across populations. For example, lake depth affects prey abundance and diversity, shoreline development may play a role in habitat complexity, and lake perimeter relates to habitat size.

The number of lateral bony plates is another trait that varies within and among populations. Marine populations consist almost exclusively of completely plated fish with about 30–36 plates on each side, whereas freshwater populations are predominantly low plated and rarely have more than seven plates per side restricted to the anterior part of the fish (Bell & Foster, 1994). Loss of lateral plates, and armour in general, is a rapid and repeatable feature of freshwater adaptation (Bell, 2001; Bell et al., 2004; Lucek et al., 2010). There is also direct experimental evidence that the low-plated morphs, or at least the genes coding for them, have a selective advantage in freshwater lakes (Barrett et al., 2008; Le Rouzic et al., 2011). The hypothesis that variation in plate numbers can be related to changes in predator regimes and predator intensity between habitats has received some support (e.g. Hagen & Gilbertson, 1972; Moodie & Reimchen, 1976; Reimchen, 1983; Kitano et al., 2008). It has also been hypothesized that the abundance of cover and shelter in littoral habitats might select for fewer plates, as acceleration and manoeuvrability may be adaptations to predator regimes where fast-start performance is important (Bergström, 2002; Walker et al., 2005). A hypothesis that has been less thoroughly investigated is Giles' (1983) suggestion that the reduction in nonessential bony tissue in freshwater-resident sticklebacks is an adaptation to reduced calcium availability (but see Marchinko & Schluter, 2007).

In the present study, we use a comparative approach to test whether eleven metric traits and gill-raker number in 74 freshwater populations of stickleback in Norway show indications of being adaptations to lake depth, lake area, lake perimeter and shoreline complexity. We use a model of evolution that allows the traits to evolve towards niche-dependent optima (Hansen, 1997; Butler & King, 2004; Hansen et al., 2008; Labra et al., 2009; see also Hunt et al., 2008), and hypotheses of adaptation are evaluated based on whether the estimated optima are influenced by environmental factors such as lake characteristics. How fast and reliably the traits evolve towards the optimal states are also estimated. We also estimate evolutionary allometric trait relationships across the freshwater populations and compare them with the same relationships in 13 marine (ancestral) populations to investigate whether linear morphological traits follow different evolutionary trajectories dependent on habitat.

Contrary to recent suggestions that allometry is not an important constraint on stickleback evolution (e.g. McGuigan et al., 2010), we show that most trait variation across populations can be explained by allometric scaling. Marine and lacustrine populations follow the same allometric model, which may indicate that developmental and genetic constraints are important contributors shaping current stickleback phenotypic variation. Only the traits that are the least bound by allometric scaling show (weak) indications of adapting to lake characteristics.

Materials and methods

Stickleback samples

Threespine sticklebacks (Gasterosteus aculeatus) were sampled from 74 freshwater lakes and 13 brackish/marine sites along the Norwegian coastline between 58 and 71 degrees north latitude between 1988 and 2011 (see Fig. 1). Lake samples were collected using minnow traps (Breder, 1960; chamber 30 cm long, 15 cm in diameter, with openings of 1 cm in diameter) unbaited or baited with cheese. Marine samples were collected with hand-held dip nets, large minnow traps (Breder, 1960; chamber 100 cm long, 40 cm in diameter, with 1 cm diameter openings), as well as fine-mesh seine nets (0.6–1.2 cm mesh size). Captured sticklebacks were stored directly in 96% ethanol. Fish analysed were 30 mm or larger to ensure fully developed armour (Hagen, 1973), and reproductive fish were not used to avoid the effects of sexual dimorphism, which is only found in reproductively mature stickleback (Kitano et al., 2007). Lakes were selected to achieve good geographical spread (latitude, longitude and altitude) and large variation in lake age. We chose about 10 fish from each location for analysis, aiming to get an even sex ratio and avoiding fish with parasite infestation (e.g. Schistocephalus sp.).

Figure 1.

Geographical distribution of the 74 freshwater lakes (black dots) and 13 marine sites (white dots) from which threespine stickleback were sampled. Areas of increased site density are shown in inset maps.

Morphological measures of metric traits

Each fish was photographed on a common background in right lateral view using a tripod-mounted CANON EOS 350D digital camera fitted with a 90-mm lens (Tamron macro) from the same distance. The fish were photographed in random order. We placed 23 landmarks on each photograph (Fig. 2) and added a scale bar using tpsDig version 2.16 from the Thin Plate Spline (TPS) suite of software (Rohlf, 2005–2007). Raw data from tpsDig were imported into MorphoJ 1.02j (Klingenberg, 2011), which we used to extract coordinates for each of the 23 landmarks.

Figure 2.

The position of the 22 landmarks used to characterize body shape and linear morphological traits in threespine stickleback. Landmarks 1–20 were used to estimate body shape and the eleven linear characters are based on the following pairs of landmarks: 1–11 body length, 1–19 mouth (gape width), 1–4 snout to edge of eye, 1–3 snout to centre of eye, 3–4 eye radius, 5–17 head depth, 1–20 head length, 7–15 body depth, 6–21 length of first spine, 7–22 length of second spine, 10–12 caudal peduncle/tail width.

Landmark coordinates were used to extract 11 linear measures of morphological traits for each fish using R v.2.10.1 (R developmental Core Team, Vienna, Austria): body length, eye radius, mouth length, distance from snout to beginning of eye, distance from snout to centre of eye, head depth, head length, body depth, length of first and second dorsal spine, and caudal peduncle (see Fig. 2 for details). These traits are ecologically important and might therefore show adaptation to lake types and habitats. We removed the within-population allometric effects of body size on each of the traits by regressing the log of each trait on lake-mean-centred log body length in an analysis of covariance (ancova) with lake as a factor. We assumed a common within-species allometric coefficient for each trait to minimize the potential biasing effects that the modest sample sizes might have on the age and the size of the fish sampled within each population. We mean centred the log body length of fish from the same lake around zero to make the intercept in the regression model equal to the trait mean within each lake. We then used the intercepts estimated for each population as data for the comparative analyses.

Morphological measures of meristic traits

The number of right lateral plates and right upper and lower gill rakers was counted for all 870 fish in a randomized order to estimate the mean trait value for each population.

Shape analysis

Multivariate morphological shape analysis of the study populations was assessed using principal component analysis in MorphoJ 1.02j (Klingenberg, 2011) on the landmark coordinates extracted from tpsDig. Principal component analysis assumes independent and identically distributed data points, making it important to account for history when size-correcting traits (e.g. Revell, 2009). The assumptions of principal component analysis are close to being fulfilled by the majority of all linear morphological traits, as they seem independent of evolutionary history (see 'Results'). Procrustes superimposition was therefore created for the entire data set (870 fish) in MorphoJ to remove all variation due to size and orientation. One analysis was run on the full data set to check whether freshwater populations group together and orient differently on the shape axes than the populations from the marine environments. A second analysis was run on freshwater fish only to investigate shape variation within the freshwater environment.

Lake data

ArcGIS (version 10; Environmental Systems Research Institute; Redlands, CA, USA) was used to determine lake area in km2, perimeter (shoreline length) in km, distance to the ocean in metres, lake depth in metres and shoreline complexity (see equation below). The lakes of interest were extracted from a water-surface data set of all of Norway (Statens Kartverk 2009). Area and perimeter were determined using the Calculate Geometry tool within the attribute tables using the polygon feature and choosing metres as unit and scale. Lake elevation was determined by sampling a 30-m Digital Elevation Model of Norway produced by the United States Geological Survey. Distance to the ocean was determined using the manual measure tool and information from Digital Elevation Models and modern-day flowlines (Statens Kartverk 2007). The distance was calculated by summing drawn straight-line segments (unit = metres) from the lake mouth to the ocean following elevation contours (river valleys) and paths of least resistance (i.e. without passing through major elevation obstacles). Methods used to estimate lake depth were similar to those used by Hollister et al. (2011). These methods assume that the geophysical processes that shape a lake basin are the same as the geophysical processes that shape the topography directly surrounding that basin (Hutchinson, 1957). Consequently, we assume that the slope surrounding the lake will continue to the lake bottom. As a result, lake depth should be a function of the distance from shore and the median percentage slope of the surrounding area (Hollister et al., 2011). We created 100-m buffers around each lake using the Analysis toolpack, for which we calculated the median slope from the 30-m Digital Elevation Models. The Euclidian distance from the point in each lake farthest from the shoreline was determined using the Spatial Analyst toolpack. The maximum distance from the shore was multiplied by the median slope to give an estimate of maximum depth for each lake. Shoreline complexity was measured as

display math

where L is the length of the shoreline and A is the lake area (Wetzel, 1975), such that a perfect circle has the minimum value D = 1. A larger shoreline complexity indicates a higher potential for the development of littoral communities and production. The ages of the lakes were estimated using the program Sealevel32 (Møller, 2003), using info on land uplift (isobase) and the current altitude of each lake.

Giles (1983) suggested that the reduction in nonessential bony tissue in freshwater relative to marine populations of sticklebacks could be an adaptation to reduced available calcium. Two-hundred millilitres of water was collected from 46 of the 74 lakes to measure the calcium content in these lakes. Water samples were held in the field at 15–25 °C and then refrigerated at about 4 °C upon return to the lab (1–4 weeks after sampling), until processing. A total of 100 mL of water was preserved and matrix matched by adding 1 mL HNO3. Standard solutions were produced using a certified 1000 ppm Ca standard to cover the expected concentrations. All samples were analysed on a Varian Vista ICP-OES, using Ar 5.0 (instrument Argon) at a speed of 16–17 L min−1. All lake variables can be found in Table S1.

Comparative method

The different ages of the lakes necessitate a ‘phylogenetic’ comparative approach to deal with the correlations that arise from different times of split from the marine ancestors and the different lengths of exposure to the freshwater niches. Although river paths may change due to isostatic rebound, none of the lakes in our data set are currently connected by rivers. We accordingly made the assumptions that each lake was colonized quickly after the formation of the lake and that gene flow between freshwater populations can be ignored. In one set of analyses, we also assumed that the populations originated from the same ancestral oceanic meta-population (Bell, 1984; Bell & Foster, 1994). This assumption is commonly made in studies of threespine sticklebacks (e.g. Schluter, 1996a; Berner et al., 2010) and has some empirical support (Deagle et al., 2013). We represented this historical scenario with a nonultrametric star phylogeny (i.e. a completely unresolved multifurcating tree in which the distance from root to tip may differ across branches) with branch lengths corresponding to lake ages (Fig. 3a). In another set of analyses, we allowed the freshwater populations to originate from an evolving marine lineage by analysing an ultrametric phylogeny in which the 13 marine populations together with the youngest extant freshwater population constitute a monophyletic star phylogeny as illustrated in Fig. 3b. This model was motivated by recent findings of genomic variability and population structure in marine threespine stickleback (DeFaveri & Merilä, 2013; DeFaveri et al., 2013).

Figure 3.

(a) shows a nonultrametric star phylogeny (i.e. an unresolved multifurcating tree) with unequal branch lengths. The symbols represent different freshwater populations, and the length of the branches portrays the time span each population has been separated from the ancestral marine population. (b) shows an ultrametric phylogeny where all extant taxa have the same distance from the root and where freshwater populations split of the marine lineage (stippled line) as they colonize available lakes. The important difference between the comparative analyses using the two phylogenies is that the marine population evolves in B, but not in A.

The comparative method we used is designed to study adaptive evolution in the presence of phylogenetic inertia (Hansen, 1997; Butler & King, 2004; Hansen et al., 2008; Labra et al., 2009; Bartoszek et al., 2012) and was fitted with R v2.10.1 (R Development Core Team, 2010) using the SLOUCH 1.2 program and some independent scripts that are to be incorporated into the SLOUCH package (scripts available in supplementary material). The model of evolution is based on an Ornstein–Uhlenbeck process and assumes that a trait (e.g. body depth) has a tendency to evolve towards a ‘primary’ optimum θ, defined as the average optimal state populations will reach in the given environment after ancestral constraints have disappeared (Hansen, 1997), at a rate proportional to a parameter α. The model additionally includes a stochastic component with standard deviation σ, which can be interpreted as evolutionary changes in the trait due to unmeasured selective forces, genetic drift, etc. We report this as vy2/2α, which can be interpreted as the expected residual variance when adaptation and stochastic changes have come to an equilibrium. The primary optimum is modelled as a linear function of one or more predictor variables based on lake characteristics (e.g. lake depth). We assumed that these predictor variables have not changed since the inception of the lake. This makes the model different from the ‘random-effect’ model of Hansen et al. (2008), which assumes that the predictor variables evolve stochastically; this necessitated the modifications of the SLOUCH code mentioned above. The method uses generalized least squares for the estimation of the regression parameters and maximum likelihood for estimation of α and σ in an iterative procedure (as detailed in Hansen et al., 2008).

The model estimates the regression of the primary optimum on various lake characteristics (Fig. 4). The coefficients of this ‘optimal regression’ measure the influence of the lake characteristics on the primary optimum, and the amount of variation explained by the regression indicates the relative importance of adaptation to the variable(s) in question. Note that this optimal regression may be steeper than an observed ‘evolutionary’ regression. This is because a lag in adaptation (i.e. a finite value of α) introduces a deviation from the predicted optimum resulting in a shallower slope (Fig. 4). We quantify the lag in adaptation by a half-life parameter, t1/2 = ln(2)/α, which is the expected time it takes for the trait mean to evolve halfway from the ancestral state towards the predicted optimal freshwater state. A half-life value of zero signifies immediate adaptation to the freshwater environment, and then the optimal and evolutionary regressions are identical. A half-life value above zero indicates that the marine ancestry has an influence on the current trait state in the extant freshwater populations.

Figure 4.

The black line represents the ‘optimal’ regression line, which describes the relationship between a trait (e.g. log mouth length), and an optimum, which is a linear function of an environmental variable, for example log lake perimeter), when the evolution of the trait towards the optimum has had sufficient time to lose all ancestral constraints. The slope of the optimal regression tells us how a change in the optimum would lead to a change in the trait. Given fast adaptation, the ‘optimal’ slope and the amount of variance explained by this model are accordingly informative of how important a specific optimum is for the evolution of the trait. The dashed line represents the observed or evolutionary regression. The slope of the evolutionary regression equals the slope of the optimal regression slope times a phylogenetic correction factor (p(αt) = 1 – (1– eαt)/αt, see Hansen et al., 2008 for more details). The correction factor is a function of the estimated rate of adaptation (α) in the comparative model. The evolutionary regression is identical to the optimal regression if adaptation is immediate and the trait tracks the optimum without delay. The evolutionary regression will be increasingly shallower as a function of a slower rate of adaptation towards the optimum. The difference in slopes between the optimal and the evolutionary regression exemplifies a situation where there are some constraints on the adaptation towards the optimum.

The general statistical effects of lake age on the traits can be quantified by estimating the phylogenetic half-life in an Ornstein–Uhlenbeck model including only one intercept for the primary optimum when using the nonultrametric phylogeny (Fig. 3a). Such a statistical influence from evolutionary history on the traits can have two different sources. Younger freshwater populations could be more similar due to the fact that it takes time for freshwater populations to adapt to their new environments (a lag in adaptation or ‘phylogenetic inertia’), but they could also be more similar if the lake characteristics were not randomly distributed with respect to history, that is, if the populations are adapted to lake characteristics that correlate with lake age. The difference between the half-life estimates with and without a predictor variable tells us how much of the lake history effect is due to inertia (resistance to adaptation) as opposed to an age effect in the predictor variables (see Labra et al., 2009 for more discussion on this point).

Akaike's information criterion (AIC) was used to compare different models as suggested by Butler & King (2004), see Burnham & Anderson (1998) or Lajeunesse (2009) for justification.

We used this approach to estimate lake age effects in the eleven metric traits and in the two meristic traits (gill-raker number and lateral-plate number). We then tested whether these traits individually showed signs of being adaptations to lake depth, lake area, lake perimeter, lake shoreline complexity and distance from ocean. As proportional differences in lake depth, area and perimeter may be as relevant as their absolute differences, we also tested the logarithmic values of these three variables as predictors. If more than one of the models that included a lake characteristic did better according to AIC than a nonadaptive model (i.e. a model with only a single intercept with the nonultrametric phylogeny, and a marine and a freshwater intercept with the ultrametric phylogeny), we ran multiple regression models to test whether a combination of lake variables could outcompete the models with single explanatory variables. We also tested whether lateral-plate number evolved towards an optimum defined by calcium concentrations in the lakes. Finally, we investigated whether body size had an effect on lateral-plate and gill-raker numbers across lakes by fitting two models for the relationship. One ‘adaptation model’ was based on evolution towards an optimal relationship as for the quantitative traits above, and the other ‘constraint’ model was based on assuming an immediate response of the traits to changes in body size as expected from, for example, an allometric relation (see Hansen & Bartoszek, 2012 for justification). We still allowed an age or phylogenetic effect in the residual variance of this model.

Evolutionary allometries

The evolutionary allometry is the log-log regression of mean trait size on mean body length across populations. We first tested whether the evolutionary allometric slope of each trait differed between marine and freshwater populations by conducting an ancova using marine or freshwater habitat as a factor and including an interaction between habitat and body length, but we did not detect any significant difference between the two habitats in any of the 10 linear traits. We accordingly estimated a common evolutionary allometry for each trait combining the 74 freshwater and 13 marine populations using the ultrametric phylogeny (Fig. 3b). As above, we fitted two models: one in which the evolutionary allometry arises through adaptation towards an optimal trait–body relationship and one in which it arises due to an underlying constraint (e.g. a static allometry) so that changes in body size lead to immediate correlated responses in the trait. Both models allow for and estimate phylogenetic correlations in the residuals.

Accounting for measurement error

The trait and variable means used in the comparative analyses are subject to sampling errors. We controlled for known estimation error variances in response and predictor variables in the comparative analyses as described by Hansen & Bartoszek (2012), except that regression slopes were not corrected for bias. As all trait means are based on rather small sample sizes (= 10), we assumed that the sampling distributions of the different populations were similar and averaged the sample variance for the different traits across populations. The measurement variances included for each population in the comparative method were then estimated by dividing the average sample variance with the sample size of each population.


Morphological variation

Means and standard deviations of the morphological traits within each of the 74 freshwater populations are reported in Table B in the supplementary material. The among-population coefficients of variation (CV) for the linear morphological traits were similar across freshwater populations and ranged from 8.9% for eye radius to 16.6% for the length of the first dorsal spine. The CV was 63.2% for the lateral plates, mainly due to seven populations having very high plate numbers. Gill-raker number shows the least interpopulation variation with a CV of 5.5%.

All metric traits were larger in the marine than in the freshwater habitat (supplementary Table C). The among-population coefficients of variation for the metric traits in marine populations ranged from 9.1% (mouth length) to 17.7% (first spine) and were comparable to those of the freshwater populations except for gill-raker number (CV = 3.5%) and lateral-plate number (CV = 11.0%), having less variation than in the freshwater populations.

The first three principal components explained 22.3%, 14.6% and 13.2% of the morphological variance across all marine and freshwater fish put together. The shape change described by the first principal component (PC1) appears to describe the way the individual fish curves post-mortem (See Fig. 5a) and was not analysed further. The other two components represent population-based variation in shape. Fish from marine environments have on average deeper bodies (PC2) and smaller heads (PC3) than fish from freshwater, but the two groups showed substantial overlap (Fig. 6). The same three axes of variation appeared in the analysis that was run on freshwater fish only and explained similar amounts of variation (Table 1).

Table 1. Principal component analyses. The three principal components that explained more than 10% of the shape variance.
 Variance explainedPositive PC scoreNegative PC scoreOverall variance in PC scoreAverage within-population variance in PC score
PCA1: all fish
PC122.3%Curved downCurved up0.0003530.000236
PC214.6%Big headLittle head0.0002380.000136
PC313.2%Streamlined bodyBluff body0.0002110.000096
PCA2: freshwater fish only
PC123.0%Curved downCurved up0.0003560.000234
PC214.6%Big headLittle head0.0002300.000146
PC311.8%Bluff bodyStreamlined body0.0001840.000083
Figure 5.

The three major axes of shape change revealed by the principal component analysis of the freshwater threespine stickleback. Landmarks are numbered, hollow circles connected by gray lines show the average landmark placement for all fish, and filled circles connected by black lines show landmark shifts associated with the vector values above (+1 and −1; about twice the value of the biological extremes of each shape vector). Warped outline drawings were generated in MorphoJ. (a) Post-Mortem Fish curvature (an artefact not used for analysis), (b) Big head-Little head, (c) Bluff body-Streamlined body.

Figure 6.

Principal component analysis of all threespine stickleback (N = 870) from 74 freshwater populations (grey) and 13 marine/brackish populations (black). The second component represents streamlined fish at one end (negative) and bluff-bodied fish at the other (positive). The third component represents fish with relatively large heads at one end of the axis (positive) and a fish with a relatively small heads at the other (negative). See Table 1 for more details.

Correlations between lake age and traits across freshwater populations

We used the nonultrametric phylogeny (Fig. 3a) to investigate the effects of lake age on the traits (Table 2A). The majority of the metric traits were only weakly affected by lake age. The best estimates indicate that it took on average <540 years to lose half the ancestral effect in eight of the eleven linear traits. This is <3% of the maximum lake age, and half-lives of zero were included in all the support intervals. However, mouth length and the two snout-to-eye distances had half-lives of more than 5000 years (Table 2A). Populations in lakes of similar age therefore have more similar mouth sizes and distance from the eye to the snout than expected by chance.

Table 2. (A) Results of the comparative analyses of the nonultrametric phylogeny. The estimate of the ‘phylogenetic’ (lake history) effect of each trait is reported when a predictor variable is not given. Models including a predictor variable are only reported if they have a better AIC score than the intercept-only model or if they are discussed in the main text. For all the reported models, we show the maximum-likelihood estimates of phylogenetic half-life, t1/2, in units of maximal tree length (maximal branch length = 1 = 18000 years) with its two-unit support interval, the stationary variance, vy, and estimates of the intercept and slope of the optimal regression with standard errors. Model fit is described with a phylogenetically corrected r2, the log likelihood (logL) and the AIC of the model.
Response variablePredictor variablet1/2 (supp. region) v y Intercept ± SESlope ± SE r 2 logLAIC
Log body length (mm)0.01 (0.00 – 0.19)0.00141.60  131.39−254.77
Log size-corrected mouth length (mm)0.61 (0.03 – ∞)0.00120.48  119.37−230.72
Log lake area in sqkm0.37 (0.04 – ∞)0.00070.42 ± 0.01−0.012 ± 0.00610.61%121.71−231.42
Perimeter in km0.34 (0.03 – ∞)0.00070.42 ± 0.01−0.002 ± 0.00110.52%121.67−233.33
Log perimeter0.33 (0.04 – ∞)0.00060.49 ± 0.02−0.022 ± 0.00812.83%122.59−233.17
shore development0.46 (0.05 – ∞)0.00070.56 ± 0.04−0.048 ± 0.01812.86%122.64−235.27
Log size-corrected eye radius (mm)0.02 (0.01 – ∞)0.00010.26  140.90−273.79
Perimeter in km0.01 (0.01 – ∞)0.00000.27 ± 0.01−0.001 ± 0.00011.51%142.93−275.85
Log perimeter0.01 (0.01 – ∞)0.00000.27 ± 0.01−0.007 ± 0.00411.20%142.80−273.59
Shore development0.01 (0.01 – ∞)0.00000.28 ± 0.01−0.012 ± 0.00611.30%142.84−275.67
Log size-corrected snout to centre of eye (mm)0.43 (0.00 – 2.46)0.00160.70  121.95−235.89
Perimeter in km0.43 (0.00 – ∞)0.00150.71 ± 0.02−0.002 ± 0.0013.43%123.12−236.23
Shore development0.42 (0.00 – ∞)0.00140.76 ± 0.03−0.033 ± 0.0175.34%123.84−237.68
Log size-corrected snout to eye (mm)0.33 (0.00 – 2.52)0.00210.51  109.11−210.22
Shore development0.32 (0.00 –  ∞)0.00190.56 ± 0.03−0.029 ± 0.0175.29%110.55−211.09
Log size-corrected head depth (mm)0.01 (0.00 – 0.21)0.00100.88  125.02−240.62
Lake area in sqkm0.01 (0.01 – 0.10)0.00100.89 ± 0.01−0.001 ± 0.0016.52%125.39−240.78
Perimeter in km0.01 (0.01 – 0.11)0.00100.89 ± 0.01−0.001 ± 0.0017.67%125.85−241.70
Shore development0.01 (0.01 – 0.13)0.00100.90 ± 0.01−0.017 ± 0.0177.02%125.59−241.18
Log size-corrected head length (mm)0.01 (0.00 – 3.08)0.00101.06  126.87−244.16
Lake area in sqkm0.01 (0.00 – ∞)0.00101.06 ± 0.01−0.001 ± 0.0014.47%127.46−244.91
Log lake area in sqkm0.01 (0.00 – ∞)0.00101.06 ± 0.01−0.010 ± 0.0066.15%128.11−244.21
Perimeter in km0.01 (0.00 – ∞)0.00101.07 ± 0.01−0.001 ± 0.0016.78%128.35−246.70
Log perimeter0.01 (0.00 – ∞)0.00101.07 ± 0.01−0.001 ± 0.0017.27%128.54−245.08
Shore development0.01 (0.00 – ∞)0.00101.08 ± 0.01−0.013 ± 0.0077.32%128.56−247.12
Log size-corrected body depth (mm)0.01 (0.00 – 0.59)0.00020.92  120.79−233.58
Lake area in sqkm0.01 (0.01 – 0.14)0.00090.92 ± 0.01−0.001 ± 0.0017.25%122.17−234.35
Distance to ocean0.01 (0.01 – 0.13)0.00090.92 ± 0.010.000 ± 0.0006.36%121.83−233.65
Lake depth in m0.01 (0.01 – 0.14)0.00900.92 ± 0.010.000 ± 0.0006.55%121.90−233.80
Perimeter in km0.01 (0.01 – 0.17)0.00900.92 ± 0.010.000 ± 0.0017.22%122.16- 234.32
Shore development0.01 (0.01 – 0.16)0.00900.92 ± 0.02−0.002 ± 0.0086.47%121.87−233.74
Log size-corrected first spine (mm)0.01 (0.00 – ∞)0.00010.44  93.41−178.50
Lake depth in m0.01 (0.00 – ∞)0.00000.43 ± 0.010.000 ± 0.0005.16%94.37−178.74
Log lake depth0.02 (0.00 −2.84)0.00010.41 ± 0.020.014 ± 0.0069.49%96.08−180.16
Log size-corrected second spine (mm)0.01 (0.00 – ∞)0.00000.50  92.65−177.33
Lake depth in m0.02 (0.00 – ∞)0.00000.49 ± 0.010.001 ± 0.0007.74%94.42−178.83
Log lake depth0.02 (0.00 – 0.92)0.00000.47 ± 0.020.014 ± 0.0069.36%95.02−178.04
Log size-corrected tail width (mm)0.02 (0.00 – ∞)0.00040.19  129.91−251.81
Gill rakers0.13 (0.03 – 0.96)0.640018.63  −101.42210.85
Log lake depth in m0.17 (0.03 – 1.29)0.590018.16 ± 0.270.247 ± 0.10212.72%−98.70209.39
Body size†0.01 (0.01 – 1.85)0.47005.54 ± 5.327.65 ± 2.6916.74%−96.79205.59
Immediate effect of body size0.07 (0.00 – 1.22)0.54006.11 ± 6.737.69 ± 4.2110.68%−99.40208.81
Lateral plates (N = 92)0.10 (0.00 – 0.24)15.35004.95  −206.17420.35
Body size†0.07 (0.00 – 0.18)13.810044.73 ± 24.6319.20 ± 14.918.21%−203.65419.30
(B) Results of the comparative analyses of the ultrametric phylogeny. See Table 1 for explanations
Response variablePredictor variablet1/2 (supp. region) v y

Intercepts ± SE

M = Marine

F = Freshwater

Slope ± SEr 2logLAIC
Log body length (mm)0.02 (0.00 – 0.07)0.0015M = 1.65 ± 0.01F = 1.60 ± 0.01 13.01%152.94−297.38
Log size-corrected mouth length (mm)0.00 (0.00 – 0.09)0.0006M = 0.48 ± 0.01F = 0.44 ± 0.01 8.89%140.48−272.46
Log perimeter0.00 (0.00–0.34)0.0004M = 0.48 ± 0.02F = 0.46 ± 0.01−0.026 ± 0.01614.37%143.16−273.59
Shore development0.15 (0.00–∞)0.0004M = 0.48 ± 0.02F = 0.49 ± 0.02−0.030 ± 0.00812.83%143.19−273.63
Log size-corrected eye radius (mm)0.06 (0.00–∞)0.0001M = 0.28 ± 0.01F = 0.26 ± 0.01 2.52%162.40−316.31
Perimeter in km10.00 (0.00–∞)0.0000M = 0.28 ± 0.01F = 0.03 ± 0.19−0.027 ± 0.0117.99%164.56−318.38
Log perimeter10.00 (0.00–∞)0.0000M = 0.28 ± 0.01F = 0.08 ± 0.19−0.054 ± 0.2237.76%164.44−318.14
Log size-corrected snout to centre of eye (mm)0.02 (0.00–0.11)0.0010M = 0.75 ± 0.02F = 0.69 ± 0.01 11.03%142.95−277.40
Log size-corrected snout to eye (mm)0.01 (0.00–0.07)0.0015M = 0.57 ± 0.02F = 0.49 ± 0.01 17.12%128.27−248.05
Log size-corrected head depth (mm)0.03 (0.00–0.12)0.0011M = 0.95 ± 0.02F = 0.88 ± 0.01 12.47%145.99−283.49
Log size-corrected head length (mm)0.02 (0.00–0.14)0.0010M = 1.11 ± 0.01F = 1.06 ± 0.01 11.11%149.05−289.61
Log size-corrected body depth (mm)0.03 (0.00–0.10)0.0011M = 1.00 ± 0.02F = 0.92 ± 0.01 18.90%140.94−273.38
Log size-corrected first spine (mm)0.02 (0.00–0.06)0.0011M = 0.64 ± 0.02F = 0.44 ± 0.01 55.23%107.07−205.66
Log lake depth0.02 (0.00–0.05)0.0009M = 0.64 ± 0.02F = 0.41 ± 0.020.034 ± 0.01558.42%109.46−206.19
Log size-corrected second spine (mm)0.02 (0.00 – 0.06)0.0010M = 0.68 ± 0.02F = 0.50 ± 0.01 53.15%106.30−204.11
Log lake depth0.02 (0.00 – 0.06)0.0008M = 0.69 ± 0.02F = 0.47 ± 0.020.032 ± 0.01656.18%108.33−203.92
Log size-corrected tail width (mm)0.02 (0.00 – 0.06)0.0005M = 0.29 ± 0.01F = 0.19 ± 0.01 38.08%150.37−292.25
Gill rakers0.01 (0.00 – 0.17)0.56

M = 19.29 ± 0.26

F = 18.40 ± 0.11

Log lake depth0.00 (0.00 – 0.05)0.51M = 19.29 ± 0.34F = 17.97 ± 0.280.456 ± 0.25016.23%−115.03242.79
Log body size (mm)0.00 (0.00 – 0.23)0.49M = 8.82 ± 4.90F = 8.27 ± 4.736.338 ± 2.96017.43%−114.40239.53
Immediate effect of body size0.00 (0.00 – 0.25)0.49M = 8.82 ± 4.52F = 8.27 ± 4.366.338 ± 2.72817.43%−114.40239.53
Lateral plates (N = 84)0.00 (0.00 – ∞)13.94M = 28.73 ± 1.07F = 5.38 ± 0.45 82.37%−240.76490.00
Log body size (mm)0.00 (0.00 – ∞)13.24M = −2.46 ± 16.14F = −24.78 ± 15.5818.877 ± 9.69083.11%−238.91488.55
Lateral plates (N = 70)0.03 (0.00 – 0.05)5.01M = 28.73 ± 0.70F = 4.83 ± 0.32 93.20%−161.47331.55
Ca0.03 (0.00 – 0.05)4.93M = 28.72 ± 0.69F = 5.19 ± 0.50−0.142 ± 0.15293.28%−161.03333.00
Log Ca0.03 (0.00 – 0.05)4.83M = 28.72 ± 0.69F = 5.34 ± 0.49−1.668 ± 1.21193.40%−160.61332.15

The meristic traits had weak associations with lake age with a half-life of 1800 years for lateral-plate number and 1440 years for gill-raker number, and in both cases, the support intervals included zero half-life (Table 2A).

Adaptation to lake characteristics

Predictor variables improving the AIC are reported in Table 2A for the analyses using the nonultrametric phylogeny and in Table 2B for the analyses using the ultrametric phylogeny (complete results are reported in Supplementary Tables D and E). The comparative analyses using the two phylogenies were qualitatively similar, but with generally larger uncertainty and therefore less support for the predictor variables when using the ultrametric phylogeny. In the following, we focus on the results from the nonultrametric phylogeny.

Although many ‘significant’ effects were found, these were generally small with no lake variable individually explaining more than 13% of the variance in any trait across lakes and with no support for models including multiple predictors.

Log lake perimeter and shoreline development each explained up to 13% of across-lake variance in log mouth length and log eye radius (Fig. 7a,b). The slopes of the optimal regressions were negative; an increase in shoreline complexity of one unit (e.g. an increase in lake perimeter from 1 to 4.5 km with a constant lake area of 1 km2, or a reduction in lake area from 1 to 0.05 km2 with a constant lake perimeter of 1 km) would decrease optimal mouth length and eye radius by a little more than a millimetre. We also found that log lake depth had a small positive effect on the optimal log lengths of the first and second dorsal spines (Fig. 7c,d). For both spines, the optimal slopes were 0.014 (± 0.006 SE), which means that a 10% increase in lake depth would increase the optimal lengths of the spines by a mere 0.14%. The optimal regression of gill-raker number on log lake depth explained 13% of the variance and had a slope of 0.247 (± 0.102 SE) rakers, meaning that a fourfold increase in lake depth would increase the optimum number of gill rakers by one.

Figure 7.

Optimal (dashed line) and evolutionary regression (black line) of freshwater populations of threespine stickleback: (a) log mouth length on log perimeter (km), (b) log eye radius on log perimeter (km), (c) log first dorsal spine on log lake depth and (d) log second dorsal spine on log lake depth. The difference in slope between the two types of regression indicates whether there is a lag in adaptation to the factors. The optimal regression is what the line would look like if the populations were given sufficient time to escape their ancestral influences.

Given the weak ‘adaptive’ effects, the estimated rates of adaptation are of limited significance, but most half-life estimates from models including predictor variables were close to and statistically indistinguishable from zero indicating that any adaptive effects are rapidly established. The only exceptions concern the adaptation of mouth length to the effects of (log) perimeter and shoreline development, which gave half-life estimates of 6120 and 8280 years with support intervals excluding zero (Fig. 7a), and the evolution of gill rakers in relation to lake depth, which happened with a half-life of 3060 years and immediate adaptation excluded from the support interval. However, the corresponding analyses using the ultrametric phylogeny were indistinguishable from instantaneous adaptation (Table 2B).

Effect of calcium on lateral-plate number

The relationship between calcium concentration and the mean number of lateral plates had a slope not significantly different from zero independent of which phylogeny we analysed, explained very little variance and had a lower AIC value than the nonadaptive model.

Evolutionary allometry of linear traits

There was little difference between the evolutionary allometries across freshwater and marine populations (Supplementary Table F), so we estimated a common evolutionary allometry for all populations combined with both an adaptation-based and a constraint-based model (Table 3). The adaptation-based model allows evolution towards an estimated optimal regression in the same manner as in the above models of adaptation to lake characteristics, and the constraint-based model assumes instantaneous responses of the trait to changes in body size. Because the half-life estimates from the adaptation model were close to zero, indicating instantaneous adaptation, it was usually not possible to distinguish these two models by AIC. The only exceptions were for the dorsal spines where the constraint model showed better fit (2.18 and 2.12 AIC units for first and second spine, respectively). This was because these were the only traits for which immediate adaptation/no phylogenetic effect could be excluded. For all the other traits, the evolutionary allometries could equally well have arisen as instantaneous adaptation to a functional relationship as to an allometric constraint, and they predicted the traits well; explaining between 69% and 90% of the variance (Fig. 8). With the exception of eye radius, which showed a strong negative allometry, these relationships were approximately isometric, and they all showed a rough correspondence between the evolutionary and the (common) static allometric slopes.

Table 3. Regressions of mean log trait length against mean log body length across marine and freshwater populations (evolutionary allometries). Shown are both the adaptive model and the direct-effect model fitted for each trait in SLOUCH. The common static (within-population) allometric slope we estimated in the ancova analyses when we removed the within-population allometric effects of body size on each of the traits prior to the comparative analyses (see 'Comparative method' in 'Materials and methods') is given for comparison with the evolutionary allometric slope.
Evolutionary allometry of:Model typet1/2 (supp. region) v y Intercept ± SESlope ± SE r 2 Common static all. slopeAICc
Mouth lengthAdaptive0.00 (0.00 – 0.03)0.0000−1.06 ± 0.170.94 ± 0.1073.12%0.99 ± 0.03−343.53
Constraint3.82 (0.00  – ∞)0.0000−1.06 ± 0.170.94 ± 0.1073.23%0.99 ± 0.03−343.53
Snout to eyeAdaptive0.00 (0.00 – 0.01)0.0000−1.48 ± 0.171.24 ± 0.1186.40%1.02 ± 0.02−350.95
Constraint0.61 (0.00 – ∞)0.0000−1.49 ± 0.171.24 ± 0.1186.41%1.02 ± 0.02−350.95
Snout to centre of eyeAdaptive0.00 (0.00 – 0.01)0.0000−0.97 ± 0.151.04 ± 0.0989.87%0.96 ± 0.02−385.09
Constraint2.40 (0.00 – ∞)0.0000−0.97 ± 0.141.04 ± 0.0989.89%0.96 ± 0.02−385.09
Eye radiusAdaptive0.00 (0.00 – 0.03)0.0000−0.83 ± 0.140.68 ± 0.0968.73%0.96 ± 0.02−377.92
Constraint0.00 (0.00 – ∞)0.0000−0.83 ± 0.140.68 ± 0.0968.73%0.96 ± 0.02−377.92
Head depthAdaptive0.00 (0.00 – 0.01)0.0000−0.78 ± 0.141.05 ± 0.0989.32%0.93 ± 0.01−393.60
Constraint0.83 (0.00 – ∞)0.0000−0.78 ± 0.141.04 ± 0.0889.28%0.93 ± 0.01−393.60
Head lengthAdaptive0.00 (0.00 – 0.01)0.0000−0.50 ± 0.130.98 ± 0.0890.30%0.93 ± 0.01−405.80
Constraint2.28 (0.00 – ∞)0.0000−0.50 ± 0.130.98 ± 0.0890.32%0.93 ± 0.01−405.80
Body depthAdaptive0.00 (0.00 – 0.02)0.0000−0.82 ± 0.151.09 ± 0.0979.12%0.97 ± 0.02−356.61
Constraint0.21 (0.00 – ∞)0.0000−0.81 ± 0.151.08 ± 0.0978.93%0.97 ± 0.02−356.67
First dorsal spineAdaptive0.17 (0.07 – 0.60)0.0001−2.46 ± 0.291.81 ± 0.2449.46%0.77 ± 0.07−216.89
Constraint0.13 (0.05 – 0.60)0.0014−1.73 ± 0.301.36 ± 0.1949.01%0.77 ± 0.07−218.71
Second dorsal spineAdaptive0.16 (0.05 – 3.32)0.0001−2.30 ± 0.301.75 ± 0.2448.54%0.83 ± 0.08−215.85
Constraint0.13 (0.00 – 0.67)0.0013−1.61 ± 0.311.32 ± 0.1947.94%0.83 ± 0.08−217.73
Tail widthAdaptive0.00 (0.00 – 0.02)0.0000−1.51 ± 0.151.07 ± 0.1077.58%0.87 ± 0.02−352.67
Constraint0.15 (0.00 – ∞)0.0000−1.49 ± 0.151.06 ± 0.0977.46%0.87 ± 0.02−352.97
Figure 8.

Evolutionary allometry of nine linear morphological traits. Estimates of regression parameters, including r2 are reported in Table 3. Black dots = freshwater populations, grey exes = marine populations. Note that a freshwater lake (Lutvann) has the largest individuals in our set of lakes.

The dorsal spines showed a different and somewhat puzzling pattern. Despite the fact that we could reject the adaptation model in favour of the constraint model, the strong positive evolutionary allometries did not fit with the negative static allometries for these traits. The spines showed a less close fit to the evolutionary allometry, however (just below 50% of the variance explained). All this indicates that the spines are somewhat evolutionary decoupled from body size.

Relationship of metric traits to body size

For number of gill rakers, both the constraint and the adaptation models outcompeted a model without body size for both phylogenies (Table 2). For the nonultrametric phylogeny, the adaptation model was 3.22 AIC units better than the constraint model and explained 16.7% of the variation in gill-raker number. The relationship had a slope of 7.65 (± 2.69 SE) gill rakers, which means that an increase in body length of about 13% increases the number of gill rakers by about 1. The constraint and the adaption models gave identical predictions when we used the ultrametric phylogeny and both returned a slope of 6.34 (± 2.96 SE and ± 2.73 SE for the adaption model and constraint models, respectively), but neither did better than a model without body size.


Freshwater sticklebacks have undergone a rapid phenotypic radiation in which independent migrations from a common marine ancestor into different freshwater habitats have given rise to parallel phenotypes in different populations (Klepaker, 1993; Bell & Foster, 1994). Natural selection is often strongly implicated in instances of parallel evolution (Endler, 1986; Schluter, 2000), but convincing evidence for the many potential selective agents that have been proposed to explain evolutionary change in sticklebacks is still rare. Shared genetic and developmental constraints are alternative explanations for why species might develop similar phenotypes in similar environments (e.g. Langerhans & DeWitt, 2004). Allometric scaling of traits may reflect such evolutionary constraints (e.g. Huxley, 1932; Gould, 1971; Lande, 1979, 1985; Savageau, 1979; Stevens, 2009; Voje & Hansen, 2013) and may bias the direction of evolution in phenotype space.

We found that freshwater and marine populations displayed similar evolutionary allometric relationships. The allometric models predicted trait sizes accurately across adult body sizes ranging from an average length of 34 millimetres in Varpvatn to an average of more than 60 mm in Lutvann, with the marine populations falling on the larger side within this range. Evolutionary allometries across species (or in this case, isolated populations) may result from adaptation to optimize trait size in relation to body size or from variational constraints that restrict the traits to change in concert. It is not always possible to separate constrained evolution from rapid adaptation, as these models make similar predictions when the phylogenetic half-lives are short. In most cases, the half-lives indicate that adaptation, if it took place, would have been very fast. Such immediate changes may also reflect phenotypic plasticity, which could be adaptive or not. Due to the rough similarity between the evolutionary allometries and the static (within-population) allometries for most traits, plus the fact that the static allometries were similar between marine and freshwater fish, we find it most plausible that the evolutionary allometries are largely a result of developmental and/or genetic constraints shared among the ancestral marine populations and the descendant freshwater populations.

The dorsal spines provide a clear exception to these patterns, showing strong positive allometry across populations (i.e. positive evolutionary allometry), but negative allometry within populations. There was also more variation around the evolutionary allometry than for the other traits, and we could reject the hypothesis that spine lengths adapt to body size along some functional allometric relationship. These observations are indicative of adaptation away from allometric constraints. We did not find strong evidence of adaptation to lake variables, however, only a weak effect of lake depth, with larger spines in deeper lakes. A possible interpretation is that spine length adapts to some unmeasured lake variable, such as the presence of predatory fish. The positive evolutionary allometry indicates that the benefit of longer spines is larger for larger-sized fish, but the poor fit of the adaptation-to-body size model indicates that this may not follow a simple functional allometry.

If trait evolution is dominated by allometric constraints, we do not expect to find much evidence for adaptation independent of body size. Interestingly, the traits most influenced by lake variables are those for which the evolutionary allometry explains the least among-population variance. About 12% each of the variances in mouth length and gill-raker number were explained by shore development and lake depth, respectively. The phylogenetic half-lives for these models were 9000 and 3000 years, respectively. Given that many of the lakes are much younger than 9000 years, this means that many populations have mouth sizes and gill-raker numbers that are not yet fully optimal in relation to lake characteristics. In contrast, the best models for all other traits indicate that whatever adaptation to lake variables that may have happened took place very fast. As for body size itself, we did not find any evidence for adaptation to lake variables. The evolution of body size also happened with short half-lives making size unrelated to lake age. As all traits are related to size, the rapid evolution of size may be a common factor in the short half-lives of most traits and models. Identifying the selection pressures that shape the evolution of body size in freshwater lakes may therefore explain the majority of the trait variation across threespine stickleback populations.

No obvious benthic–limnetic continuum

North American stickleback populations usually fall somewhere onto a well-described benthic–limnetic axis of variation. The macroinvertebrate diet of the benthic fish is reflected in their deep bodies, deep heads and short snouts, small eyes, large mouths, and few gill rakers; this falls in contrast to the long, streamlined head and body, large eyes and more gill rakers of limnetic fish, which helps them to specialize on plankton (Hart & Gill, 1994; Walker, 1997). The smaller mouths and more numerous gill rakers in larger fish predicted by our evolutionary allometries overlap the expectation for limnetic stickleback, but their deeper bodies and smaller eyes do not. Body-size variation, in itself, is causing phenotypic variation in a direction that is not directly along the benthic–limnetic continuum.

The ‘optimal’ slopes in our models do not fit well with the limnetic–benthic continuum. A high degree of shoreline complexity is often correlated with high productivity due to an average higher input of nutrients from streams and watershed compared with a lake with a less complex shoreline (Dodds & Whiles, 2010). Lakes rich in nutrients may provide good opportunities for fish to forage on invertebrates along the shore, and it has been argued that a longer shoreline and complex shore structure should favour benthic morphs with larger heads, mouths and eyes (e.g. McPhail, 1984; Lavin & McPhail, 1985). We found only weak effects of these lake variables on morphology in our study.

Failure to find support for specific adaptive hypotheses does not mean that the traits in question are unaffected by selection. Our analyses can neither confirm nor reject adaptation in general; they can only reject specific adaptive hypotheses. It may be that we did not measure the most relevant lake variables or that those we measured were too inaccurate to capture the adaptive effects. It is also possible that our sampling methods did not enable us to sample all potentially available ecotypes in every lake. Studies that have claimed morphological shape adaptations in sticklebacks due to lake characteristics have usually based their conclusions on just a few populations (e.g. McPhail, 1984; Lavin & McPhail, 1985) or on the rare cases of sympatric limnetic–benthic ecotypes (e.g. McPhail, 1984, 1992; Schluter & McPhail, 1992). In lakes with sympatric populations, reinforcement and character displacement may induce different strengths and patterns of selection than seen in allopatric populations (Schluter, 1996b, 2000, 2003; Foster et al., 1998). Furthermore, repeated colonization rather than lake attributes may be the most important determinant for the evolution of divergent species pairs (Ormond et al., 2011). The exact environmental variables that drive phenotypic evolution in lakes with a single randomly mating population are therefore still not fully understood. Fish predators are ubiquitous in Norwegian lakes and the presence of fish predators may alter which niches are available, and consequently change selection pressures across lakes. As such, we cannot rule out that an effect caused by predatory fish drowns out potential ecological effects. Experimental studies identifying selective advantages for certain phenotypes in particular lakes or environments should provide useful hints for variables that ought to be included in future comparative studies.

Adaptation to lakewide variables may also be relatively unimportant in relation to specialization to diverse habitats within lakes. Subtle spatial differences in morphology have been shown to be associated with ecological factors in lake and stream stickleback (Reimchen, 1980; Bell, 1982; Baumgartner, 1992). Much of the phenotypic variance across different habitats could also be due to phenotypic plasticity. Wund et al. (2012) showed that oceanic stickleback raised in a complex habitat and fed a typical benthic diet (macro-invertebrates) grew up to resemble derived, benthic freshwater fish. They were also able to produce an adult body shape similar to the derived limnetic phenotype in oceanic stickleback by raising them in a bare environment and feeding them zooplankton. Svanbäck & Schluter (2012) have shown that plasticity is higher in generalist freshwater populations in British Columbia compared with the highly specialized benthic and pelagic freshwater populations and to marine populations. They hypothesize that this is due to a broader niche. If this is the case in the Norwegian lakes we sampled, a high degree of phenotypic plasticity could explain the relative lack of within-lake clustering in body shape.

Gill-raker number may adapt to body size

We did find a positive relationship between body size and gill-raker number, as did Moodie & Reimchen (1976). They hypothesized that the higher gill-raker number in larger fish is an adaptation to faster swimming speeds (faster fish forage on smaller aggregated planktonic organisms) and claimed that the relationships are not explained by allometric growth. Our results partly support Moodie and Reimchen's hypothesis in that the relationship between the number of gill rakers and body size was better explained by an adaptation model than by a constraint model on the nonultrametric phylogeny, although the models were similar on the ultrametric phylogeny. In any case, no more than 18% of the variance in gill rakers was explained by size, and factors other than body size must be involved in gill-raker evolution.

Lateral-plate number is not related to lake calcium levels

The difference in plate number between marine and freshwater populations can be caused by selection for fewer plates in some freshwater habitats (Barrett et al., 2008; Marchinko, 2009; Le Rouzic et al., 2011). Calcium availability has been implicated as one of the selective forces operating on the Eda locus, which accounts for much of the variation in plate diversity in threespine stickleback (Colosimo et al., 2004, 2005). However, we found no support for Giles' (1983) hypothesis that reduced plate number is an adaptation to the calcium concentration in lakes. Bell et al. (1993) found that the effect of reduced calcium on pelvic reduction was contingent on the absence of predatory fishes. Unfortunately, we were not able to investigate this due to incomplete information about predators. Lateral-plate reduction is much more common than pelvic reduction, however, both generally and in our study lakes, and is often present regardless of the presence of piscivorous predators. Many of our lakes were relatively low in calcium, and it is still possible that calcium can influence lateral-plate number as part of an interaction with other environmental variables we did not measure. It is also possible that the level of calcium affects the size, but not the number, of lateral plates, something we did not test in our analysis. However, calcium alone seems not to be a major explanatory variable for variation in lateral-plate numbers in Norwegian sticklebacks. In contrast, Spence et al. (2013) recently reported a relationship between plate number and calcium across Scottish lakes.


We found indications that the 74 freshwater populations of stickleback have diversified largely in phenotypic directions predictable from static allometric scaling relationships. These relationships are maintained between all marine and freshwater populations we tested. Independent trait adaptation to lake ecology seems to have played a less important role. It would be interesting to investigate these relationships in other populations to see whether this is an isolated or a global phenomenon. In either case, the impact of evolutionary constraints on the evolution of threespine sticklebacks cannot be ignored when trying to discern the different factors that have shaped their phenotypic evolution.


We would like to thank Annette Taugbøl and Elisabeth Wiig for help in the field and Anne-Marie Skramstad for measuring the calcium concentrations. We are also thankful to the large number of landowners for allowing us access to the lakes. Thanks to Annette Taugbøl, Juha Merilä, Mats Björklund, Gavin Thomas and an anonymous reviewer for helpful criticisms and comments on the manuscript. This work was supported by the Norwegian Research Council (Grant Number 16639/F20).