Adaptive diversification is driven by selection in ecologically different environments. In absence of geographical barriers to dispersal, this adaptive divergence (AD) may be constrained by gene flow (GF). And yet the reverse may also be true, with AD constraining GF (i.e. ‘ecological speciation’). Both of these causal effects have frequently been inferred from the presence of negative correlations between AD and GF in nature – yet the bi-directional causality warrants caution in such inferences. We discuss how the ability of correlative studies to infer causation might be improved through the simultaneous measurement of multiple ecological and evolutionary variables. On the one hand, inferences about the causal role of GF can be made by examining correlations between AD and the potential for dispersal. On the other hand, inferences about the causal role of AD can be made by examining correlations between GF and environmental differences. Experimental manipulations of dispersal and environmental differences are a particularly promising approach for inferring causation. At present, the best studies find strong evidence that GF constrains AD and some studies also find the reverse. Improvements in empirical approaches promise to eventually allow general inferences about the relative strength of different causal interactions during adaptive diversification.
We here focus on the second and third effects described above – because of the recent flush of work attempting to infer causation from negative association between AD and GF. At the outset, it seems valuable to formally confirm verbal arguments that AD and GF can each negatively influence the other. We use a quantitative genetic model (Appendix S1) to show that (i) variation in dispersal can lead to negative correlations between AD and GF, and (ii) variation in the magnitude of ecological differences can have the same effect (Fig. 1). These results confirm that negative correlations between AD and GF do not, in themselves, allow inferences about which is the cause and which the effect.
Our main goal is to evaluate empirical methods for inferring the arrow of causality between AD and GF in natural populations. Our paper thus forms a bridge between recent reviews that focus primarily on one causal pathway (GF to AD: Lenormand 2002; Garant et al. 2007) or the other (AD to GF: Rundle & Nosil 2005; Hendry et al. 2007). In particular, we illustrate how understanding ecologically driven diversification requires a clear understanding of both causal pathways. We first summarize some of the main ecological and evolutionary forces that influence diversification (Fig. 2), and then discuss how best to reveal the action of these forces in nature. We focus primarily on discrete populations, rather than clinal scenarios, because of the diverse, but diffuse, recent work in the former context. We will argue that the greatest inferential power can be achieved through a simultaneous consideration of multiple ecological and evolutionary forces, as well as through experimental manipulations in nature. We close by considering the inferences drawn from study systems where to date the best inferential methods have been applied. These studies consistently find support for GF constraining AD, and often also for the reverse. More work of an integrated nature is needed, however, before we can ascertain the generality of these initial observations.
A summary of causal effects and interactions
We start by summarizing the main causal pathways that promote or constrain AD (Fig. 2a). On the promoting side, populations that occupy increasingly different environments should experience increasing divergent selection (path 1 in Fig. 2a) and should therefore undergo greater AD (path 2). On the constraining side, an increasing number of individuals dispersing between environments should increase the proportion of immigrants (m, path 3), which should increase GF (path 4) and therefore reduce AD (path 5). Paths 2 and 5 thus represent the classically recognized tension between diversifying selection and homogenizing GF, or the ‘migration-selection balance’ (e.g. Haldane 1948; Mayr 1963; Jain & Bradshaw 1966; Ehrlich & Raven 1969; Slatkin 1973; Felsenstein 1976). We next incorporate the idea that AD can reduce GF via the evolution of reproductive isolation (path 6 = ecological speciation, Schluter 2000). Paths 5 and 6 thus represent the bi-directional arrow of causality that makes it hard to infer cause and effect between AD and GF.
This simple framework (Fig. 2a) immediately suggests some interesting feedback loops (Rice & Hostert 1993; Hendry et al. 2001; Crespi 2004; Hendry 2004). For instance, a decrease in dispersal should reduce GF, which should permit an increase in AD, which should further reduce GF (ecological speciation), which should allow more AD, and so on until some potential equilibrium (or quasi-equilibrium). Working the other way, an increase in dispersal should increase GF, which should reduce AD, which might further increase GF, and so on until perhaps some other equilibrium. These feedbacks then suggest the possibility of alternative stable states, such as near-complete adaptation vs. near-complete maladaptation (e.g. Ronce & Kirkpatrick 2001; Holt et al. 2004), species fission vs. fusion, or perhaps a stable tension at some other intermediate point.
Many additional complexities can be layered onto this simple framework, and we take here up some that are especially important to our later discussions (Fig. 2b). The first set of complications arises via the potential evolution of dispersal. When AD reduces the fitness of migrants between environments (Nagy & Rice 1997; Hendry 2004; Nosil 2004; Nosil et al. 2005), increasing AD should favour the evolution of reduced dispersal (path 7 in Fig. 2b; Billiard & Lenormand 2005; Fraser & Bernatchez 2005). An interesting feedback loop emerges here because the evolution of reduced dispersal will decrease GF (Fig. 2b) and thereby allow increased AD. This increase in AD may favour further evolutionary reductions in dispersal, although reduced dispersal decreases the proportion of the population under selection. These complicated effects are ripe for theoretical examination.
Another set of complications emerges through the effects of demography. First, increasing AD may increase population sizes (path 8) when better-adapted populations grow faster or have higher equilibrium abundances (Kirkpatrick & Barton 1997; Tufto 2001; Lenormand 2002). This effect is most likely when density dependence is weak (Gomulkiewicz et al. 1999; Saccheri & Hanski 2006; Kinnison & Hairston 2007; Kokko & López-Sepulcre 2007). Second, increasing dispersal may directly increase the size of recipient populations (path 9) and may thereby reduce the negative effects of small population size (Holt et al. 2004). If these demographic effects lead to changes in the relative size of interacting populations (i.e. asymmetries in population size), then GF may change – because relatively larger populations will experience relatively lower immigration rates (m) for a given number of immigrants (path 10; Holt & Gomulkiewicz 1997; Tufto 2001; Hendry 2004). This change may result in lower GF (path 4), increased AD (path 5) and further feedback loops. For example, increasing AD that increases local population size may reduce GF – and thereby further increase AD. It is worth noting here that asymmetries in dispersal and population size between environments can lead to an asymmetric equilibrium, whereby adaptation is primarily to only one of the two environments (e.g. Holt & Gaines 1992; Kawecki 2000; Ronce & Kirkpatrick 2001; Kawecki & Holt 2002; Kisdi 2002).
We have sketched only some of the major forces influencing diversification, with the goal to summarize the major effects, generate interesting hypotheses and the following exploration of empirical methods. It is important to recognize, however, that we have not included many other effects, such as evolutionary changes in genetic variation (Guillaume & Whitlock 2007), frequency-dependent selection and co-evolutionary dynamics (Nuismer et al. 2000; Thompson 2005). It is also possible that environmental differences directly impact dispersal even in the absence of AD – e.g. if individuals within populations ‘imprint’ on their local conditions. We hope that theoretical models can ultimately be used to examine these and other interactions in an integrated framework.
Most studies of interactions between AD and GF in nature have used simple correlative approaches. We therefore begin our discussion with such simple designs before turning to alternatives. Note that we avoid lengthy lists of citations to studies that have employed the least effective designs – instead reserving space for particularly informative and robust methods. Note also that all of the studies focus on systems where AD is expected, and may therefore ignore cryptic genetic divergence and cryptic reproductive isolation, which may well be very important in nature.
Some studies focus on a single population with a phenotype unexpected for its environment, inferring a constraining role for GF simply for this reason. This inference is obviously improved by confirming that the observed deviation is in the direction of nearby populations experiencing a different environment, and further by determining whether GF may be sufficiently high to cause the inferred constraint. It is also important to confirm, however, that the unexpected phenotype is not simply the result of unaccounted selection. Ideally, all of this information would then be incorporated into theoretical models designed to test whether the observed phenotypic deviation is consistent with the measured parameters (King & Lawson 1995; Hendry et al. 2001; Moore et al. 2007).
An improvement to the above single-population approach can be to sample multiple pairs of populations in divergent environments. Studies adopting this approach often find that AD is lower for population pairs that exchange more genes, a pattern used to infer that AD constrains GF (e.g. Smith et al. 1997; Gíslason et al. 1999; Lu & Bernatchez 1999) or that GF constrains AD (e.g. Storfer et al. 1999; Langerhans et al. 2003; Hendry & Taylor 2004). These analyses are strongest when the population pairs are evolutionarily independent and numerous, which has been the case for few studies to date. And, of course, these correlations cannot by themselves illuminate cause and effect.
Despite their limitations, correlative studies are likely to remain common, and so we now consider correlative methods that show the greatest potential for causal inferences. For the sections that follow, it is important to remember that dispersal and GF are not the same thing (Kawecki & Ebert 2004). For example, GF can be higher than dispersal when populations are inbred – because immigrants may have higher fitness than residents (Ingvarsson & Whitlock 2000; Ebert et al. 2002). Alternatively, dispersal can be higher than GF owing to selection against migrants and hybrids (Hendry et al. 2000; Hendry 2004; Nosil et al. 2005). Therefore, estimates of dispersal cannot be used as surrogates for GF nor vice versa.
AD to GF inferred from environmental differences and GF
The problem with drawing causal inferences from correlations between AD and GF is the bi-directional arrow of causality that links them (e.g. paths 5 and 6 in Fig. 2). One solution might therefore be to design a test with at least one uni-directional arrow. For example, GF may rarely cause environmental differences, and so we might infer that AD constrains GF when population pairs that occupy more divergent environments show lower GF. Exceptions to uni-directional causality in such comparisons may occur if environmental differences are determined by (i) competition that changes with dispersal (e.g. density- or frequency-dependence, Gomulkiewicz et al. 1999), or (ii) co-evolution of the population and the ‘environment’ (e.g. predator-prey or host-parasite interactions, Thompson 2005). As long as these effects are not particularly strong, a negative correlation between environmental differences and GF implies that environmental differences drive divergent selection, which influences AD, which influences GF – this last effect being the one we wish to infer.
A typical application of the above approach tests whether GF (estimated from genetic markers) decreases as populations occupy increasingly different ecological environments. A number of studies have found this very result (e.g. Smith et al. 1997; Schneider et al. 1999; Ogden & Thorpe 2002; Rolán-Alvarez et al. 2004; Grahame et al. 2006), whereas others have not (e.g. Hendry & Taylor 2004; Crispo et al. 2006). These conflicting outcomes suggest that divergent selection does not inevitably reduce GF, that factors other than AD more strongly influence variation in GF (e.g. geographical features, arbitrary sexual selection, drift, cryptic divergence or isolation), or that parameter estimates are not always reliable. In hopes of reducing the latter possibility, we now consider issues related to the estimation of environmental differences and GF.
For environmental differences, a key is to examine the specific ecological variables that determine selection on the traits of interest. This is not always straightforward and, as in so many cases, detailed knowledge of the organism’s natural history is critical. One might also formally quantify divergent selection on the traits, although accurate estimates of selection are logistically difficult (Kingsolver et al. 2001; Hereford et al. 2004; Hersch & Phillips 2004). Even with accurate estimates, it is important to remember that divergent selection depends not only on the environmental difference but also on the amount of GF. That is, increasing GF leads to stronger selection – because populations are held farther from their respective optima (García-Ramos & Kirkpatrick 1997; Bolnick & Nosil 2007). In fact, path 2 (Fig. 2) could be redrawn with bi-directional causality, again complicating interpretations of cause and effect. Thus even if selection is quantified, it remains important to assess the important environmental differences. In general, selection and adaptation are a function of both the environment and the phenotypic distribution. Inferences would therefore best be drawn by constructing adaptive landscapes – although this has been exceptionally rare for natural populations (Schluter 2000).
Accurate estimates of GF are equally important but many problems arise here also. For instance, most studies use neutral genetic markers to estimate ‘historical’ GF – but no consensus exists as to the best such method (Slatkin & Barton 1989; Beerli & Felsenstein 1999; Whitlock & McCauley 1999; Abdo et al. 2004). Estimating historical GF from genetic markers also assumes that the populations are at equilibrium, which can take some time to achieve depending on effective population size (Whitlock & McCauley 1999). Moreover, most studies estimate the ‘effective number of migrants’ (Nem), even though it is the rate of immigration (m) that most directly influences AD (Kirkpatrick & Barton 1997; Hendry et al. 2001; Tufto 2001; Lenormand 2002). Unfortunately, estimating m is even more difficult than estimating Nem because the former also requires the estimation of Ne (Wang 2005). Another complication is that historical GF will vary among neutral loci depending on their linkage to loci under selection (e.g. Kelly 2000; Emelianov et al. 2004; Gavrilets 2004, pp. 147–148; Grahame et al. 2006). These difficulties might encourage the use of assignment methods to estimate ‘contemporary’ GF (Hauser et al. 2006; Waples & Gaggiotti 2006). These estimates, although useful, are more relevant to dispersal (Berry et al. 2004) than to GF, and so one is again left with the problem of measuring GF. Somewhere between ‘historical’ and ‘contemporary’ GF, linkage disequilibrium can be used to estimate GF that has occurred in the recent past, as has been demonstrated for clines and hybrids zones (e.g. Mallet et al. 1990). At present, we suggest it is most valuable to measure GF using multiple methods and to look for correspondence among them. In doing so, it is important to recognize that the accuracy is maximized, and bias reduced, by analyzing many loci and carefully assessing outliers. Overall, relationships among dispersal, contemporary GF, and historical GF (as well as their estimators) are still unclear, calling for additional theoretical and empirical work.
One reason for caution when extrapolating from specific reproductive barriers to GF is that effects on different potential barriers may offset each other. One possible example comes from Trinidadian guppies (Poecilia reticulata). In particular, brightly coloured males from low-predation populations above waterfalls that move into less-colourful high-predation populations below waterfalls will have lower survival (increased susceptibility to predators) but possibly higher mating success (increased attractiveness to females) than residents. Effects of environmental differences on one potential barrier (natural selection disfavours migrants) may thus be offset by the effects on another barrier (mate choice favours migrants), potentially leading to no net effect on GF (Crispo et al. 2006). More studies of interactions between AD and GF should therefore examine multiple reproductive barriers (e.g. McGraw & Antonovics 1983; Via et al. 2000; Ramsey et al. 2003; Nosil 2007).
GF to AD inferred from dispersal and AD
Bi-directional causality might here be avoided by testing for a negative correlation between AD and the potential for dispersal (e.g. geographical distance). This should work because the potential for dispersal is unlikely to be influenced by AD, thus achieving uni-directional causality in the statistical test (caveats are discussed below). Indeed, several such studies have found a negative correlation between the potential for dispersal and phenotypic divergence (e.g. Sandoval 1994a,b; Langerhans et al. 2003). This approach depends, however, on appropriate estimates of AD and dispersal, subjects to which we now turn.
Estimating AD might seem straightforward – simply measure phenotypic differences – but inevitable complications arise (Kawecki & Ebert 2004). First, it is important to carefully ascertain which traits are subject to consistent divergent selection – and therefore of interest with respect to a constraint. Second, a simple correlation between phenotypic divergence and GF does not necessarily reveal the extent of trait maladaptation in a given population, because this also requires knowledge about the optimum phenotype (Estes & Arnold 2007; Moore et al. 2007). Third, analyses based on a subset of relevant traits will not reveal the constraint on overall adaptation (i.e. the migration load: García-Ramos & Kirkpatrick 1997; Lenormand 2002), which requires measurements of fitness itself. Fourth, phenotypic differences may reflect phenotypic plasticity rather than genetic differences (Pigliucci 2001), and plasticity may even be favoured by GF (Sultan & Spencer 2002). Moreover, environmental and genetic effects on traits may offset each other in nature, leading to apparent phenotypic similarity among populations despite underlying adaptive genetic divergence (counter-gradient variation, Conover & Schultz 1995). In such cases, selection might, for example, act more strongly against immigrants than would be expected from a comparison of phenotypes among wild populations. For all of these reasons, it is important to isolate adaptive genetic differences from phenotypic effects (Kawecki & Ebert 2004), as has recently been done in some elegant long-term studies of AD and GF in the great tit (Parus major; Garant et al. 2005; Postma & van Noordwijk 2005). Arguments for a GF constraint on AD can be further strengthened by studying variation at specific loci under selection, or linkage disequilibrium between loci, as has been done in some clinal studies (e.g. Mallet et al. 1990; Lenormand & Raymond 2000).
Estimating dispersal is notoriously difficult – because it is usually impossible to survey all potential sites and because point estimates may not reflect long-term patterns (Koenig et al. 1996). Because these problems and their solutions (increased sampling effort and temporal replication) are well known, we will not dwell on them further. Instead, we focus on assessing the potential for dispersal, such as the distance between sites (Langerhans et al. 2003), barriers between sites (Bertness & Gaines 1993; Crispo et al. 2006), relative population sizes (Sandoval 1994a; Dias & Blondel 1996; Nosil 2004) and dispersal vectors (Bohonak 1999). Here, it is important to verify that these variables really do influence dispersal as predicted – and this is not always the case (Moore et al. 2007). It is also important to consider possible covariation between dispersal (or the potential for dispersal) and environmental factors that might influence AD. For example, more distant sites may also be more ecologically different, which could cause a negative correlation between phenotype and distance as a result from selection instead of GF (Moore et al. 2007). Moreover, when distant populations are smaller, distance-based reductions in the number of immigrants may not translate into similar reductions in the rate of immigration (e.g. Antonovics 1976).
GF to AD inferred through other correlations
Several other correlations can help provide evidence that GF constrains AD. First, one can test whether divergent selection is positively correlated with GF – while also controlling for variation in environmental differences. The reason is that increasing GF holds populations farther from their local optima, and therefore maintains stronger selection, as has recently been shown for Timema walking sticks (Bolnick & Nosil 2007). Second, one can test whether the phenotypes of populations are correlated with the relative phenotype and frequency of immigrants, such as has recently been shown for great tits (Garant et al. 2005; Postma & van Noordwijk 2005).
Integrated correlative studies
Given the great number and complexity of interactions, the best way to approach correlative work is to quantify as many of the relevant factors as possible, with those in Fig. 2 as a reasonable starting point. This integrated approach can be illustrated by work on Timema walking sticks adapted to different host plants (e.g. Nosil et al. 2002, 2008; Nosil 2004, 2007; Nosil & Crespi 2004; Bolnick & Nosil 2007). For many populations, data were collected on environmental differences (host plant type), divergent selection (phenotypic changes within a generation), phenotypic divergence (morphology and colouration), premating isolation (mate choice and selection against migrants and hybrids), GF (measured using mtDNA and AFLP markers) and the potential for dispersal (geographic isolation and the relative sizes of adjacent patches of host plants). With these data, it has been possible to show that, for colouration at least, environmental differences promote divergent selection (path 1), divergent selection promotes AD (path 2), an increase in the potential for dispersal increases GF (path 3 to path 4 – dispersal itself has not been measured), GF decreases AD (path 5) and AD leads to the evolution of reproductive barriers that likely limit GF (path 6). All of the above effects were found also for other morphological traits, except that AD in morphology did not seem to influence mating isolation. Genetic evidence further suggests that AD constrains GF, at least to some extent, also in nature (Nosil et al. 2008).
For this and other integrated studies, statistical analyses based on path models and structural equations might prove useful (Shipley 2002). Different causal hypotheses could be specified in alternative models and standardized coefficients from the best models used to infer the strength of each pathway. Such models have already been used, albeit rarely, in studies quantifying the role of divergent selection on AD (e.g. Johnson 2002), but apparently not yet in combination with GF. Effectively implementing such models is not trivial, because it requires the accurate estimation of numerous parameters in many populations. Yet, with rigorous sampling on particularly suitable model systems at least, such integrated statistical models of the various factors influencing adaptive diversification might ultimately allow rigorous insight.
As explained earlier, AD and dispersal can both influence population size, which can then influence GF and feed back on AD. In addition, population size can influence the relative contributions of dispersal and AD to GF. For example, the relative contribution of selection against migrants to reducing GF will decrease as the number of immigrants decreases relative to the number of residents (Hendry 2004). When the proportion of immigrants is small, AD may therefore have little direct effect on GF.
Other demographic effects are also possible. For example, high immigration may hold population size above the carrying capacity, thus causing negative density dependence, reductions in mean population fitness and declines in adaptation (Gomulkiewicz et al. 1999; Kawecki 2000). Alternatively, immigration may facilitate adaptation owing to positive density dependence (i.e. Allee effects, Holt et al. 2004) and by sustaining populations until adaptation can occur (Holt & Gomulkiewicz 1997). Incorporating demography into correlative studies may not be easy, but the above summary suggests that the potential payoffs are high.
Experimental evolution in natural populations
Experimental manipulations of AD and GF is a powerful way to infer causation, as has been shown in many laboratory studies (Endler 1977; Rice & Hostert 1993; Cuevas et al. 2003; Forde et al. 2004; Swindell & Bouzat 2006). If we are to infer causation in nature, however, such studies need to be done on natural populations. At present, we know of only one experimental manipulation of dispersal that was aimed at testing the constraining role of GF (Riechert 1993; Riechert & Hall 2000; Riechert et al. 2001). This experiment was motivated by the observation that spiders (Agelenopsis aperta) from environments with different predation pressures differed in anti-predator behaviour – except in one population. To test whether GF was the cause of the phenotypic deviation, Riechert (1993) constructed drift fences that reduced dispersal between the environments. A single generation later, the formerly deviant population had evolved appropriate antipredator behaviour – thus confirming the original constraint imposed by GF. This rare manipulation convincingly demonstrated the constraining role of GF on AD in nature, but more such studies are clearly needed.
Testing the opposite causal pathway (AD to GF) can involve the experimental manipulation of selection, coupled with the monitoring of AD and GF in subsequent generations. For example, populations can be introduced into new ecological environments, and then periodically examined for evidence that increasing AD reduces GF. For example, native insects adapting to introduced host plants often show substantial reproductive isolation from their ancestors on native plants (e.g. Feder et al. 1990). Other work on introduced populations suggests that GF between ancestral and descendent populations can be reduced after only dozens of generations (Hendry et al. 2000, 2007; Sheldon & Jones 2001). Additional opportunities to examine how contemporary adaptation influences GF (or GF influences AD) are manifold given the large number of organisms introduced to new environments (Reznick & Ghalambor 2001). An important next step would be to design and implement such studies with the expressed intent of measuring the rate at which reproductive isolation evolves and GF decreases through time.
We have mainly focused on ways to improve causal inferences, but it also seems appropriate to attempt some initial conclusions about the causal interaction between AD and GF in nature. For this, we focus on a few study systems that have been examined with the best inferential methods (Table 1) to infer whether AD constrains GF or GF constrains AD. We again focus on studies of discrete populations/environments (except where they are combined with studies of clinal variation and for the classical clinal work on Anthoxantum odoratum adapting to mine tailings). These systems represent only a small subset of all studies on AD and GF, but they have (in our opinion) done the best job of demonstrating the arrow of causality – especially when they have explicitly addressed both causal pathways. As in many clinal studies (Lenormand 2002), we find consistent and clear evidence that GF constrains AD in nature (Table 1). These inferences are often only qualitative, but theoretical models applied to water snakes (King & Lawson 1995; Hendry et al. 2001), lake and stream stickleback (Hendry et al. 2002; Moore et al. 2007) and Timema walking sticks (Bolnick & Nosil 2007) have confirmed that estimated levels of GF and other parameters are indeed consistent with the observed AD. These results therefore belie the old expectation (Ehrlich & Raven 1969) that rates of GF in nature are too low to have much of an effect on AD.
Table 1. A selected set of empirical study systems explicitly examining causal relationships between adaptive divergence (AD) and gene flow (GF) in natural populations
Adaptive trait/ reproductive barriers (yes/n.a.)
Other factors estimated
Type of study
The column ‘Inference’ indicates the support for GF constraining AD or AD constraining GF based on the components measured. Shown are the number of sampling sites or populations (n), the metric by which GF was estimated, the adaptive traits that were examined, whether or not reproductive barriers were found, and the expected selective factor driving the divergence. Other factors that were estimated are indicated: strength of selection (s), the expected difference in optima (opt) determined by comparing divergence in the presence and absence of GF, population sizes (N), actual dispersal rates (dr) and genetic basis (h). Type of study refers to correlations between AD and GF (corr) and actual manipulations of GF and/or selection (exp) to infer causality. (See text for further details).
Evidence for the opposite causal pathway (AD to GF) is also present but is restricted mostly to the demonstration that AD generates particular reproductive barriers (Schluter 2000; Rundle & Nosil 2005). Studies that have directly tested whether AD reduces GF in nature are rare (but see Gow et al. 2007), and we therefore argue that the jury is still out on whether AD can have as large effects on GF as the reverse.
Negative correlations between phenotypic differences and gene flow can arise if adaptive divergence constrains gene flow or if gene flow constrains adaptive divergence. We suggest that the first of these causal pathways might be inferred by testing for negative correlations between environmental differences and gene flow. The second causal pathway might be inferred by testing for negative correlations between the potential for dispersal and adaptive divergence. Overall, the greatest inferential power in correlative studies is gained by simultaneously measuring and integrating multiple ecological and evolutionary factors, including environmental differences, divergent selection, dispersal, gene flow, adaptive divergence, reproductive barriers and population size (Fig. 2b). A good knowledge of the natural history of the study system is clearly essential. Particularly robust inferences may emerge from appropriate experimental manipulations of dispersal (e.g. increasing or decreasing movement between populations) and selection (e.g. introductions or environmental manipulations) in nature. At present, there appears to be more qualitative evidence for the constraining role of gene flow on adaptive divergence than for vice versa. However, much more integrated studies of exemplary model systems are needed before we can make rigorous inferences about generalities in nature.
We thank Nelson Hairston Jr, Thomas Lenormand and anonymous referees for valuable comments on earlier versions of this manuscript. KR was supported by The Swedish Research Council and Swiss National Science Foundation. APH was supported by the Natural Sciences and Engineering Research Council of Canada.