A long-standing goal in benthic stream ecology has been to explain and predict species distributions according to key features of the environment (Theinemann 1954, Hynes 1970, Townsend and Hildrew 1994, Statzner et al. 2001b). This goal has proven challenging, not only because of taxonomic uncertainties, but also because environmental features in streams are embedded in a landscape context (Hynes 1975, Malmqvist 2002) that is both multiscaled (Frissell et al. 1986) and dynamic (Resh et al. 1988). Given the difficulties with predicting species composition in lotic systems, researchers have advanced functional classifications of species into groups with similar biological and ecological traits that are expected to respond consistently along specific environmental gradients. Cummins (1973) was the first to articulate fully the functional approach in stream biology by assigning insect species to feeding groups intended to reflect dominant modes of food acquisition. The River Continuum Concept (Vannote et al. 1980) built upon that work by presenting a theoretical construct for relating the trophic characteristics of species to the dominant longitudinal physical and energy gradients of rivers on a global scale. Over the past few decades, many species traits have been collated to facilitate further development of traits-based analyses of insect communities in stream benthic ecology (reviewed in Poff 1997, Statzner et al. 2001b).

The habitat template (templet) model of Southwood (1977, 1988) is the primary theoretical underpinning of traits-based approaches in stream ecology. This model predicts that where environmental conditions (drivers) are similar, species trait composition should converge, even across biogeographic boundaries where species pools differ (cf. Orians 1980, Schluter 1986). This model has been adapted to streams in several manifestations (e.g., Schlosser 1987, Poff and Ward 1990), but most extensively by Townsend and Hildrew (1994), who proposed that benthic communities should consist of species possessing traits well suited to the spatial and temporal heterogeneity of their habitat. The metaphor of environmental filtering (e.g., Tonn et al. 1990, Keddy 1991) entered the aquatic literature as community ecologists began to understand that local community composition reflects regional-scale processes (Ricklefs 1987, Ricklefs and Schluter 1993) and that multiscaled processes can be characterized in hierarchical terms (e.g., Allen and Starr 1982, Frissell et al. 1986). Building on this metaphor, Poff (1997) considered how species traits may be filtered at multiple landscape scales to influence the functional (trait) composition of local stream communities. The template model in benthic community ecology has been supported by numerous studies that explain trait composition of populations or invertebrate communities in terms of environmental drivers (e.g., Scarsbrook and Townsend 1993, Resh et al. 1994, Statzner et al. 1997, Townsend et al. 1997a, b, Robinson and Minshall 1998, Chessman and Royal 2004, Finn and Poff 2005, Heino 2005, Bêche et al. 2006). A handful of traits-based analyses conducted at multiple spatial scales (Richards et al. 1997, Parsons 2001, Weigel et al. 2003, Statzner et al. 2004, Lamouroux et al. 2004, NLP, unpublished data) also give credence to the multiscale filtering concept in lotic systems.

The mechanistic basis for understanding and predicting trait–environment relationships underlying the template model has encouraged some to argue that traits-based approaches could be used in the burgeoning field of stream biomonitoring (Dolédec et al. 1999, Hawkins et al. 2000, Statzner et al. 2001a, Bady et al. 2005). The strong theoretical framework for linking benthic community responses to environmental alteration (especially by anthropogenic stressors) could be used to create ecologically meaningful biometrics that are broadly applicable across multiple spatial scales and that span biogeographic boundaries. Furthermore, for benthic invertebrates, traits may be more time- and cost-efficient than taxonomic methods of defining assemblages. Taxonomic identification at coarser levels than species (e.g., genera and families) can provide an accurate description of the trait diversity at a locality (Dolédec et al. 2000, Gayraud et al. 2003) and, therefore, taxonomic expertise and time-consuming lab work are minimized. The application of traits-based biomonitoring approaches has been best developed in Europe (Dolédec et al. 1999, 2000, Charvet et al. 2000, Usseglio-Polatera et al. 2000b, Statzner et al. 2001a, Gayraud et al. 2003, Bady et al. 2005, Ducrot et al. 2005), although there are also examples from North America (e.g., Merritt et al. 2002, Vieira et al. 2004), Australia (Chessman and Royal 2004), and New Zealand (Dolédec et al. 2006). Trait applications in biomonitoring thus far include defining assemblage types, determining expected trait diversity under reference conditions, and developing metrics to detect environmental gradients.

The traits-based perspective presents a powerful paradigm for broad-scale stream community ecology (both basic and applied) because it offers a framework for mechanistically linking traits in a community to major environmental drivers that influence the fitness of the species possessing those attributes. However, even where such mechanistic linkages can be made easily, fundamental challenges confront the development of a more predictive traits-based approach to community ecology. The overarching need is to develop a more robust multivariate framework, in which the responses of multiple, independent traits can be related to multiple environmental gradients characteristic of most landscapes. Univariate approaches consider single trait responses along individual environmental gradients (e.g., Lamouroux et al. 2004), whereas most multivariate approaches aggregate multiple traits into groups (e.g., Poff and Allan 1995) or ordinate them to reduce dimensionality (e.g., Finn and Poff 2005) and then examine how they vary along environmental gradients. However, these multivariate approaches do not account for statistical or phylogenetic relationships among the traits, and neither do more formal multivariate-based techniques that test for statistical relationships between environmental variables and individual trait responses (e.g., Dolédec et al. 1996, Legendre et al. 1997).

The implicit challenge in adopting such a multifaceted approach is to account quantitatively for the fact that traits are often linked together by evolution and, therefore, cannot necessarily be treated as independent entities. That traits can be linked together into syndromes has been known for some time, e.g., the classic r–K strategies in life-history traits (Pianka 1970). Stream ecologists also have recognized such linkages and pointed out that these linkages may cause individual traits to appear decoupled from environmental conditions (see Resh et al. 1994, Townsend and Hildrew 1994, Statzner et al. 2004). This linkage is analogous to the concept of pleiotropy in genetics, in which selection for a particular gene may alter the frequency of other, linked genes. However, to date, virtually no effort has been expended to disentangle trait linkages quantitatively to reveal the unique information available in each trait and, thus, to allow development of more-informed models relating responses of multiple traits to environmental condition(s).

Our primary objective in this paper is to explore the issue of inter-trait correlations for lotic insects and to identify opportunities and challenges for advancing the theory and application of traits-based approaches in stream community ecology. We pursue this objective in 4 sections. First, we analogize a taxon's specific combination of traits to the ecological niche to conceptualize how linked traits may dictate the response of whole communities to the environment. Second, we quantify the statistical correlations of traits for a large collection of families and genera of North American lotic insect taxa. This analysis allows us to identify highly correlated traits that may influence how entire benthic communities respond along environmental gradients. Third, we use an inferred phylogeny of all taxa based on molecular and morphological data to identify traits that are least constrained by phylogenetic history and, therefore, that can be used to test hypotheses regarding community responses to environmental change. As a subset of this analysis, we focus on the Ephemeroptera, Plecoptera, Trichoptera (EPT taxa), orders often used as water-quality indicators (e.g., Resh and Jackson 1993), to show how a traits-based approach can allow a priori expectations of how these particular taxa may respond differentially to specific environmental gradients. Fourth, we conclude with a brief discussion of some of the analytical limitations to advancing the multivariate traits-based approach, and we recommend future research directions.

The Functional Trait Niche—Linkage among Traits within Taxa

Traits are characteristics of species that are often used to define some biological feature of the organism or its direct relation to the environment. As such, the combination of traits an organism possesses is thought to be the product of natural selection by the environmental conditions in which a species or population has evolved (Southwood 1977). Any individual organism itself possesses multiple traits, and traits are not necessarily concordant with phylogeny (e.g., not all members of the Plecoptera are predators, and not all predators are Plecoptera). Because traits are not necessarily dependent on phylogeny, a traits-based approach can, in principle, be applied across biogeographic boundaries to examine issues such as community convergence under similar selection regimes (e.g., Lamouroux et al. 2002). From this point forward, we use the term trait to indicate some attribute, such as voltinism, that has multiple potential states (e.g., multivoltine, univoltine, or semivoltine) or, in some cases, that can be characterized in a continuous fashion (e.g., absolute body size). The delineation of states for each trait may be arbitrary and reflect either the ability to characterize a discrete state or the ecological question of interest.

The specific states making up some given set of traits can be identified for any species (or higher taxon). We define this specific combination of trait states as a taxon's functional trait niche (FTN), which is analogous to the ecological niche of a species, i.e., the totality of attributes defining its unique relation to its environment and other species. (See Chase and Leibold 2003 for a cogent discussion of the ecological niche and Rosenfeld 2002 for the related concept of the functional niche. Also see McGill et al. 2006 for a discussion of the relationship between functional traits and the realized and ecological niches of species.) Generally speaking, FTNs are nonrandom because they have been structured by evolution and, accordingly, closely related taxa are more likely to have similar FTNs than are distantly related taxa (discussed further below).

The FTN is an important concept because trait linkages at the individual species level represent a specific, evolutionarily constrained structure within the aggregate community trait composition. This structure has some important implications for predictive traits-based community ecology. First, the trait states comprising a community cannot be treated as independent statistical entities, a fact well understood by both ecologists (e.g., Townsend and Hildrew 1994) and statisticians (e.g., Legendre et al. 1997). Second, and more subtly, the response of a benthic community along some environmental gradient(s) may depend on the specific collection of FTNs represented by members of the community. As a consequence, composite measures of community trait composition that do not explicitly account for the evolutionary linkage of trait states (e.g., functional diversity; Charvet et al. 2000, Bady et al. 2005) are unlikely to be adequate indicators of traits-based community response to environmental change.

The simple hypothetical example presented in Fig. 1 illustrates the utility of the FTN concept. In this example, 2 communities each consist of 3 taxa. Each taxon is described by a specific combination of 3 traits (a, b, and c), each of which has 3 states (e.g., trait a has states a₁, a₂, and a₃). Note that the 6 taxa making up the 2 communities are, in fact, unique in terms of their FTNs, i.e., no 2 taxa share an identical combination of trait states. However, each community has the same functional diversity, i.e., number of unique trait states. If we now impose on both communities a strong environmental filter that specifically selects against trait states a₁ and b₂, we see a differential response between the communities as a function of their constituent FTNs. In the 1^st community, trait states a₁ and b₂ occur in separate taxa, both of which are subsequently lost from the community. By contrast, in the 2^nd community, trait states a₁ and b₂ occur in only 1 taxon, and only that taxon is eliminated. Thus, under the same environmental selection regime, the specific combinations of trait states (i.e., the FTNs of the constituent species) can directly influence the resultant functional diversity of the whole community. Of course, a different selection event that, say, selects for a₁ and b₁ would have the opposite consequence regarding the functional diversity of these hypothetical communities.

We present this simple example to make 2 important points. First, it is possible to achieve the same functional diversity with different sets of contributing taxa, an observation that underscores the point that functional and taxonomic diversity need not be equivalent. Second, linkage among traits and trait states (as represented by FTNs) raises the possibility that the overall functional diversity of a community may change in unanticipated ways in response to some environmental selective force acting on a specific trait or trait state. These observations emphasize the importance of considering multiple traits and their linkages in developing a traits-based predictive ecology. The most informative traits and trait states in broad-scale analyses of community responses along environmental gradients would be those that are relatively independent, i.e., weakly correlated and not strongly linked by evolution. Such traits might be particularly useful for biomonitoring applications, a theme we return to below. Another insight afforded by these observations is that an FTN represents a multidimensional response potential for a species along some environmental gradient(s). Therefore, the FTN concept may help ecologists improve our presently limited understanding of how combinations of trait states may interact with the environment by providing an evolutionarily informed, multivariate framework for exploring how particular trait combinations are associated with different selection regimes.

Correlations among Traits across Taxa

We characterized 311 taxa in 75 families of North American lotic insect taxa (from Merritt and Cummins 1996, with nomenclature updates) according to 20 traits coded in 59 states that are broadly categorized as life-history, mobility, morphological, and ecological traits (Table 1). We selected all traits that were readily available in the literature or that could be evaluated confidently by expert opinion and applied to the entire set of North American lotic insect taxa. Our trait matrix was cross-referenced with a comprehensive meta-database of traits for North American invertebrates that we developed for the US Geological Survey (USGS; see Vieira et al. 2006 for database and extensive documentation). The USGS meta-database can be used to express traits quantitatively or as multistate, fuzzy categories (e.g., Resh et al. 1994). However, we assigned each taxonomic unit to only one trait state (i.e., a binary approach) using literature and expert opinion for several reasons. First, detailed trait information typically is not available for invertebrate taxa at the continental scale in North America (unlike in Europe). Second, our goal was to use the trait matrix to explore some dominant relationships among taxa that could guide future research, rather than to provide an exhaustive and authoritative listing of lotic insect traits. Third, this approach greatly simplifies our phylogenetic analysis (see below), which is a novel approach for examining the evolutionary relationships among traits and trait states for a continental invertebrate fauna. Furthermore, the highest taxonomic resolution we considered was the genus level (except for some Chironomidae where tribe was used) because of great taxonomic uncertainty at the species level (for aquatic juveniles) and because congeneric species are often functionally equivalent and arguably can be viewed as ecologically equivalent (Wiggins and Mackay 1978). We also completed analyses at the family level, where we used a majority rule to assign the most common genus-level trait state to the family (regardless of species diversity within genera). The assignments to specific trait states for each taxon are provided in the Appendix. Hereafter, we will refer to the various taxonomic levels as “genus” or “genera” only.

There are >350,000 theoretical combinations of the 59 various states (based on multiplying number of possible trait states for each of the 20 traits) in our trait matrix. However, by definition, a maximum of only 311 unique sequences has been realized during evolution among the 311 extant taxa we considered (at the genus/tribe level of resolution). Of this number, 233 unique combinations of trait states actually occur, i.e., some genera have identical FTNs at the level of trait resolution used in the analysis. In this section, we explore in more detail how these trait states are associated across the North American lotic insect taxa to identify which trait states are only weakly correlated across taxa and, thus, make particularly good candidate traits for use in predictive community ecology.

We used multivariate ordination and cluster analyses to explore relationships among the 20 biological traits at 2 taxonomic resolutions: the genus (or tribe) level (n = 311) and the family level (n = 75). For both levels of taxonomic resolution, we conducted 2 analyses using Gower's similarity coefficient (Gower 1971), which calculates similarity between 2 taxa according to categorical traits. First, we examined the similarity of the taxa in terms of the multistate biological traits. We used Principal Coordinates Analysis (PCoA) on the genus and family similarity matrices to reduce dimensionality and to summarize the dominant patterns of variation among the biological traits (Legendre and Legendre 1998). We retained the first 2 PCoA dimensions to facilitate visual interpretation of the resulting plots. Second, we used Gower's similarity index to construct a dendogram showing similarity in traits across all taxa.

The multivariate summary of the 311 genera (Fig. 2A) shows strong separation in 2-dimensional space, where the 1^st and 2^nd PCoA axes explained 17.8% and 13.6% of the variation in trait space, respectively (both components were statistically significant based on the broken-stick model; Jackson 1993). Genera were distributed in this ordination space with discernible separation among orders in most cases, although some orders (Coleoptera, Plecoptera) contain genera that span much of the ordination space and overlap in their trait composition with other orders (Fig. 2B). These patterns suggest that most genera can be separated according to 3 major trait gradients based on combinations of traits that explain dominant levels of variation among the genera in this 2-dimensional space (Fig. 2C).

The 1^st trait gradient (positive axis I and negative axis II scores) primarily represents genera of Hemiptera and Odonata, although some Coleoptera and Plectoptera genera are included. The primary trait states are high crawling rate (Crwl3), long adult life span (Life3), semivoltinism (Volt1), strong adult flight (Flgt2), and predatory feeding mode (Trop4). Two other traits important to some genera in this group are lack of attachment to the benthos (Atch1) and large adult body size (Size3). The 2^nd trait gradient (negative axis I scores and negative axis II scores) primarily represents Ephemeroptera and some Plecoptera genera. The primary trait states are fast seasonal life cycles (Devl1), no body armoring (Armr1), very short adult life span (Life1), and small adult size (Size1). These genera are common or abundant in the drift (high negative loading on Drft1; see Table 1). The 3^rd trait gradient (positive axis II scores) mainly describes Trichoptera and some Diptera and Coleoptera genera that are characterized primarily by no swimming ability (Swim1), some attachment (Atch2), high rheophily (Rheo3), body armoring (Armr3), and short adult life span (Life2).

The trait gradients shown in Fig. 2 can be viewed essentially as trait syndromes in which specific trait states co-occur across the range of North American lotic insect taxa. Specific sets of trait states are strongly associated with particular groups of genera and families, a finding similar to that reported for European lotic insect taxa using different traits (see Usseglio-Polatera et al. 2000a). We note that these syndromes are interpreted in only a 2-dimensional space that explains <⅓ of the variation in trait states across all the taxa. Adding additional axes would probably better define particular groups (e.g., some of the Plecoptera that cluster with the Trichoptera), but the major components of the FTNs of the insect taxa are reasonably depicted by these 2 dimensions, and our goal here is simply to show general patterns.

We found similar patterns for the 75 families of North American lotic insects. The first 2 PCoA components explained 18.0% and 13.8% of the variation in the original traits, respectively (both statistically significant), and the families were distinguished basically according to the same 3 major trait gradients identified in the analysis of genera (results not shown). This finding, which is consistent with other studies showing similarities in trait analyses between genus and family levels (Dolédec et al. 2000, Gayraud et al. 2003) suggests that a substantial amount of variation in trait diversity within benthic communities can be captured by a coarse taxonomic (family-level) resolution. However, use of family-level could result in the loss of some potentially important resolution, especially for the many families that possess a diversity of trait characteristics among the constituent genera. For example, 17 of the 20 limnephild genera in the database (Appendix) have unique FTNs, i.e., they differ in their unique combinations of trait states. Diversity is reasonably high among baetids (10 of 13 genera have unique FTNs), heptageneiids (9 of 12), and perlids (12 of 14), whereas diversity is less pronounced among groups such as hydropsychids (4 of 11) and gomphids (6 of 13). Thus, when attempting to explore community responses to complex and multiple environmental gradients, use of genus-level information is expected to provide more sensitivity to detection of subtle biological and ecological responses.

In Fig. 3 we show the quantitative similarity among the 59 trait states (and 20 average trait values–inset) as defined by the 311 taxa used in the analysis. Trait states that have high similarity in the dendogram do not necessarily show close association in Fig. 2C, which is only a 2-dimensional representation of relationships among trait states (as compared to the 311-dimensional representation of Fig. 3). Several sets of trait states show high similarity (>80%) across the dendogram. For example, at the bottom of the of the dendogram, a suite of trait states describing semivoltinism (Volt1) with a preference for depositional habitats (Rheo1) and long-lived (Life3), strong-flying adults (Flight2) occurs. This syndrome is represented mostly by members of the Odonata and corresponds closely to the trait gradients observed in Fig. 2. Immediately above this cluster are some groups of trait states that have similarity >90%. One consists of air breathing (Resp3) taxa, with skating habit (Habi6) and ability of the adult to exit the stream (Exit2), i.e., mainly hemipterans. Another consists of the trait states high juvenile drift occurrence (Drft3), ability to both attach and range freely (Atch3), and nonseasonal development (Devl3). Both of these clusters are closely allied with the warm eurythermal trait state (Ther3) as well as with the trait states of swimming habit (Habi5), bi/multivoltinism (Volt3), and collector-filterer feeding mode (Trop2). If the similarity threshold is dropped to 80% the additional trait states of desiccation tolerance (Desi2), poorly synchronized emergence (Sync1), and habits of climbing (Habi2) and burrowing (Habi1) are included. The trait states in the remainder of the dendrogram (i.e., above the trait state Desi2) generally show low similarity; however, scattered throughout this region are sets of trait states with >80% similarity across the 311 taxa. These sets include the following associations: sprawling (Habi3) and shredder/detritivore (Trop5); strong crawler (Crwl3) and predator (Trop4) along with large size (Size3) and strong swimmer (Swim3); no swimming (Swim1) with some attachment (Atch2); strong armoring (Armr3) and herbivore (Trop3); short adult life (Life1) with fast seasonal development (Devl1); univoltine (Volt2) with well-synchronized emergence (Sync2); low tolerance to desiccation (Desi1) with inability of adult to exit water (Exit1); weak adult flyer (Flgt1) with short aerial dispersal (Disp1); and rare in drift (Drft1) with slow-seasonal development (Devl2).

The analysis of trait similarity across all genera also revealed a number of interesting patterns for average trait correlations (Fig. 3 inset). Traits associated with larval dispersal and potential resilience to disturbance occur together (Drft, Desi, Exit, Flgt, Disp), and life-history traits have some strong affinities (Life, Devl and Sync, Volt). Some traits within our 4 general trait groups (Table 1) show low similarity across the 311 taxa. For example, the morphological traits Size, Atch, Armr, and Shpe are only about 80% correlated. Trophic (Trop) and habit (Habi) traits are only weakly correlated with other traits, a pattern that may reflect, in part, the fact that these traits have a relatively large number of trait states or that these traits are more evolutionarily labile than other traits (see below).

Evolutionary Lability among Traits across Taxa

Given the fact that many trait states occur together as tightly-linked syndromes that have apparently strong taxonomic affinities, the question arises as to how one might go about identifying traits that could be used generally for any benthic community. Ideally, we would like to use traits and associated states that are relatively unconstrained by phylogeny. A focus on such evolutionarily labile traits could help avoid the problems raised in Fig. 1. Specifically, convergent evolution is expected among highly labile traits because they are responsive to local selection; therefore, similarity among taxa with respect to these traits often will be largely independent of their phylogenetic relationships (e.g., Cavender-Bares et al. 2004). Indeed, environmental filtering has been shown to act more predictably on evolutionarily labile traits than on other traits in plants (Patterson and Givnish 2002, Cavender-Bares et al. 2004).

Speculation among ecologists about lability of traits through evolutionary time has a long history (e.g., Gause 1947, Mayr 1974, Diamond 1986), but this topic has received only brief attention in the stream ecology literature. Townsend and Hildrew (1994) noted it as an issue for predictive traits-based community ecology, and Usseglio-Polatera et al. (2000a) noted qualitatively that biological traits (e.g., morphology, physiology, life history) appear to be much more constrained by phylogeny than ecological traits (e.g., behavior, habitat preference). However, to date, no one has formally mapped aquatic insect traits onto an inferred phylogeny of North American lotic insects to quantify evolutionary lability and, thus, identify which traits may be applied most robustly to broad-scale community analyses. Furthermore, if phylogenetically unconstrained traits can be related directly to important environmental gradients (filters), then they could provide a good starting point for mechanistic hypothesis testing of a priori expectations about community-level responses to certain types of environmental change. In this section, we describe a formal approach to quantifying the phylogenetic constraint of trait states of North American lotic insect taxa, with the intent of identifying independent (more labile) traits that can be used in predictive community ecology.

New Insights and Continuing Challenges

Traits-based approaches have a rich (though relatively recent) history in both basic and applied stream ecology and are attractive because they represent evolutionary responses to environmental selective forces that often can be quantified across broad geographic extents. The theoretical underpinnings of how various benthic invertebrate traits will respond to environmental gradients have been well established (e.g., Townsend and Hildrew 1994, Poff 1997, Statzner et al. 2001b). However, we remain limited in our ability to predict community functional composition empirically. One problem has been that investigators have focused on univariate responses of particular traits to single environmental gradients or, in cases using multivariate approaches, have ignored the FTNs of individual taxa, which often consist of a group of trait states that are closely linked evolutionarily. We have offered the beginnings of an approach for disentangling the series of traits making up individual FTNs as a way to identify which traits have the greatest potential for a predictive ecology, i.e., those with relatively low statistical correlation to other traits and with higher evolutionary lability across taxonomic groups.

Selecting what might be considered the most robust traits for use in predictive community ecology seems to require balancing 3 important constraints: 1) Ideally, traits with high evolutionary lability should be selected because they can be used for a variety of taxa. Several such traits were identified in Table 2. 2) Low statistical and phylogenetic correlations (which are inherently related) among traits are desirable to ensure statistical independence among traits and to maximize information content when multiple traits are to be used. We identified several labile traits (e.g., voltinism and trophic habit) that were statistically and phylogenetically uncorrelated. For our categorical trait states, statistical correlations provided an adequate approximation of phylogenetic constraint among trait states. The next obvious step would be to explore lability and phylogenetic/statistical correlations among trait states with more informative coding, such as quantitative values or fuzzy coding. 3) The ability to link the trait mechanistically to a specific environmental gradient (preferably one that is quantifiable) also is desirable. This quality may not be feasible for all traits with high evolutionary lability where little is known about trait response to environmental change, e.g., maximum crawling rate is not obviously related to any quantifiable environmental gradient. However, it may be achievable for such traits as voltinism, which has been reported to vary among sites having different disturbance regimes (e.g., Richards et al. 1997) in conformance to theoretical expectations (Townsend and Hildrew 1994).

Figure 5 presents a simple conceptual model that places phylogenetically labile traits into a mechanistic framework given several important environmental gradients. This model provides a way to visualize how individual traits, or combinations of traits, may change under different environmental conditions. In principle, taxa in any regional species pool could be positioned in this 3-dimensional space to provide a priori expectations of community composition in multivariate space.

The preceding discussion offers a way forward in predictive community ecology, but we raise the point that our ability to identify robust traits is currently limited by our incomplete understanding of how traits may interact in a given environmental context. For example, the degree to which single vs multiple traits may be involved in species replacements along the hypothetical disturbance gradient in Fig. 5 is unknown. Disturbance traits may interact with other, less labile traits to confer some resistance or resilience to environmental change, but we have limited understanding of these relationships in nature. Given this knowledge gap, we view the FTN framework as providing a basis for examining, on a taxon-by-taxon basis, how combinations of trait states may be assorted along different environmental gradients. This area is in much need of empirical research. Compiling this type of information also may help inform contemporary theoretical questions about the relationships between functional diversity and taxonomic diversity (e.g., Díaz and Cabido 2001, Petchy and Gaston 2002, Mouillot et al. 2005) or ecosystem function (e.g., Naeem and Wright 2003, Petchey et al. 2004) in biotic communities, and how these relationships may change along environmental gradients (Heino 2005, Stevens et al. 2003). Indeed, the systematic analysis of how functional traits vary along environmental gradients is a promising approach to defining the quantitative relationship between species' fundamental and realized niches, thereby allowing community ecology to become a more predictive science (McGill et al. 2006).

To achieve the full potential of a functional, traits-based ecology, we believe that several challenges must yet be met. A fundamental need exists for a more comprehensive characterization of organismal traits and a more-mechanistic understanding of organism–environment relationships (Poff 1997). For example, many of the traits available to benthic ecologists are based simply on convenience of characterization, not necessarily on some underlying functional relationship with demonstrated or inferred adaptation to the environment. Even for those traits that are most robust for predictive modeling, uncertainties arise from ontogenetic shifts in trait states, such as trophic habit, changes in mobility, etc. An unstated assumption of traits-based analyses at this time is that characterizing traits for late instars is adequate to examine organism–environment relations, but this may not always be the case.

Another assumption of current traits-based analyses is that biotic traits, i.e., those that reflect intensity of ecological interactions among organisms, are typically unimportant. For example, subtle differences in terms of resource acquisition can exist among species in the same trophic guild (e.g., Kohler 1992), and these subtleties can vary along environmental gradients (Poff et al. 2003). Likewise, foodweb structure and ecosystem energy flow can be mediated by behavioral traits of predators and prey (e.g., McIntosh and Peckarsky 1999, Power and Dietrich 2002, Schmitz et al. 2004). Certainly, this kind of trait information is difficult to develop in general, as it typically is derived from more experimental studies; however, development of such traits could add increased resolution to predictive community ecology.

Last, even assuming that all of these biological issues can be resolved, a basic statistical challenge remains. Current multivariate analyses seeking to describe community trait composition such as RLQ analysis (Dolédec et al. 1996) and the 4^th-corner method (Legendre et al. 1997) do not explicitly account for evolutionary linkage of traits; therefore, metrics derived from such analyses may not be adequate indicators of traits-based community response to environmental change. Furthermore, these approaches consider only a single trait at a time, and some can analyze only binary species data (e.g., Legendre et al. 1997). In short, new multivariate statistical methods that account for inter-trait correlations and that can be applied to multiple environmental gradients also are needed to advance a predictive traits-based ecology.

In conclusion, we hope that the approach presented here will encourage more mechanistic, traits-based analyses to advance the predictive ability of trait applications. We believe our approach for defining traits relevant to a specific environmental gradient and for reducing the number of traits (variables) to a more tractable and robust subset for analyses (by filtering out noninformative and intercorrelated/redundant traits) will be useful in many practical applications, as well as in basic stream community ecology. Furthermore, by examining how trait states are linked within various taxa (the FTN), we may gain insight into trait interactions that influence species' responses along multiple environmental gradients.