Extreme drought pushes stream invertebrate communities over functional thresholds

Abstract Functional traits are increasingly being used to predict extinction risks and range shifts under long‐term climate change scenarios, but have rarely been used to study vulnerability to extreme climatic events, such as supraseasonal droughts. In streams, drought intensification can cross thresholds of habitat loss, where marginal changes in environmental conditions trigger disproportionate biotic responses. However, these thresholds have been studied only from a structural perspective, and the existence of functional nonlinearity remains unknown. We explored trends in invertebrate community functional traits along a gradient of drought intensity, simulated over 18 months, using mesocosms analogous to lowland headwater streams. We modelled the responses of 16 traits based on a priori predictions of trait filtering by drought, and also examined the responses of trait profile groups (TPGs) identified via hierarchical cluster analysis. As responses to drought intensification were both linear and nonlinear, generalized additive models (GAMs) were chosen to model response curves, with the slopes of fitted splines used to detect functional thresholds during drought. Drought triggered significant responses in 12 (75%) of the a priori‐selected traits. Behavioural traits describing movement (dispersal, locomotion) and diet were sensitive to moderate‐intensity drought, as channels fragmented into isolated pools. By comparison, morphological and physiological traits showed little response until surface water was lost, at which point we observed sudden shifts in body size, respiration mode and thermal tolerance. Responses varied widely among TPGs, ranging from population collapses of non‐aerial dispersers as channels fragmented to irruptions of small, eurythermic dietary generalists upon extreme dewatering. Our study demonstrates for the first time that relatively small changes in drought intensity can trigger disproportionately large functional shifts in stream communities, suggesting that traits‐based approaches could be particularly useful for diagnosing catastrophic ecological responses to global change.


| INTRODUCTION
Vulnerability assessments are increasingly using species' functional traits to explain and infer their sensitivities to long-term climate change (e.g., Domisch et al., 2013;MacLean & Beissinger, 2017;Pacifici et al., 2017;Pearson et al., 2014). Traits have less commonly been used to diagnose ecological responses to climatic extremes, which are projected to become more frequent and intense globally (Dai, 2013;Fischer & Knutti, 2015) and are less likely to offer opportunity for species adaptation (Poff et al., 2018;Thompson, Beardall, Beringer, Grace, & Sardina, 2013;Vázquez, Gianoli, Morris, & Bozinovic, 2017). Extreme events such as drought can push ecological communities beyond critical thresholds (Bailey & van de Pol, 2016), defined here as the point(s) along an environmental gradient where a relatively small change in conditions provokes a disproportionately large biotic response (Capon et al., 2015;Groffman et al., 2006;Kelly et al., 2015). Anticipating the ecological impacts of drought hinges on understanding when and why these thresholds are crossed (Standish et al., 2014). However, gradient-based studies that can detect causal relationships and nonlinearities in the relevant response variables are largely lacking (Kreyling, Jentsch, & Beier, 2014).
In running waters, abrupt ecological responses to drought may be expected as critical habitats are lost, such as when the drying of riffles fragments the channel into isolated pools, or when the streambed dries completely (Boulton, 2003;Chadd et al., 2017).
However, this nonlinearity has predominantly been explored with structural metrics (species richness, community composition), and it remains unclear whether thresholds can also be detected in the functional trait profiles of stream biota. By explicitly linking environmental perturbation to species response, functional traits can provide greater mechanistic understanding of disturbance impacts than taxonomic approaches (Chessman, 2015;Floury, Usseglio-Polatera, Delattre, & Souchon, 2017), and as environment-trait relationships potentially transcend biogeographic boundaries, they should yield more universally relevant findings (Menezes, Baird, & Soares, 2010;Schriever & Lytle, 2016;Walters, 2011). Moreover, traits-based indices, particularly frequency distributions of individual traits, appear to be stronger indicators of ecosystem functioning than taxonomic composition (Gagic et al., 2015). A traits-based approach to threshold detection therefore has the potential to significantly improve our understanding of drought, providing (a) information on the key biological mechanisms driving abrupt community shifts; (b) transferable observations of species' vulnerabilities to critical habitat loss; and (c) insights into when and how community functioning may be most affected (Dézerald, Céréghino, Corbara, Dejean, & Leroy, 2015).
Traits-based studies in freshwaters have primarily focused on macroinvertebrates, reflecting their wide distribution, high diversity and prominent role in ecosystem functioning (Menezes et al., 2010).
Various studies have explored macroinvertebrate trait responses to hydrologic disturbance (e.g., Bêche, Mcelravy, & Resh, 2006;Bonada, Dolédec, & Statzner, 2007;Schriever et al., 2015;Leigh et al., 2016), but these have overwhelmingly investigated seasonal drying events which do not represent true extremes for their locale, and to which species are preadapted with a suite of suitable traits and coping mechanisms (Lytle & Poff, 2004). For instance, in environments with a history of severe drying, the strongest biological changes are typically delayed until surface water is completely lost, reflecting local biotic adaptation to all but the most severe disturbance (Boersma, Bogan, Henrichs, & Lytle, 2014;Bogan, Hwan, Ponce, & Carlson, 2017). Community resistance to extreme drought is typically much lower (Lake, 2003), and such events could therefore trigger marked ecological responses long before the streambed dries. We might expect the timing of any such responses to be trait-specific, with changes in species behaviour as drought initially intensifies giving way to subsequent shifts in morphology and physiology, as survival becomes progressively more difficult without physiological adaptations to drying (Hershkovitz & Gasith, 2013;Stubbington & Datry, 2013).
Despite observed and projected increases in the frequency of extreme droughts, such events are still rare in running waters, creating an urgent need for large-scale experiments which can expose species to novel conditions beyond their evolutionary envelopes (Kayler et al., 2015;Knapp et al., 2017;Ledger & Milner, 2015). Furthermore, most definitions of an ecological threshold relate the rate of change in ecosystem state to that of a specific environmental pressure in isolation (Capon et al., 2015;Groffman et al., 2006). This is difficult or impossible to validate as a causal driver-response relationship in correlational studies, which are often beset by confounding influences beyond the stressor of interest, and instead favours detection in an experimental setting (Kayler et al., 2015;Kreyling et al., 2014). Mesocosms are thus suitable as they can isolate trait responses to stream drought from possible confounding factors (Woodward et al., 2016), such as changing pollutant levels, underlying climatic and hydrological regimes and other site-specific contingencies, including surrounding land use (Ding et al., 2017;Durance & Ormerod, 2009;Floury et al., 2017;Thomson et al., 2012;Yao et al., 2017). Crucially, of all experimental approaches, mesocosms also allow for the greatest compromise between realism and replicability (Stewart et al., 2013).
We therefore tested for thresholds in the responses of macroinvertebrate traits across an experimental gradient of drought intensification that encompassed several critical stages of habitat loss. Here, we use the term threshold in a statistical sense, namely a stage in a relationship where the response variable changes more rapidly than the predictor (Groffman et al., 2006;Kelly et al., 2015;Yin, Leroux, & He, 2017). Statistically robust ecological threshold detection methods are commonly used to gauge maximum permissible levels of habitat fragmentation in terrestrial ecosystems (Swift & Hannon, 2010), but have received relatively little attention in the aquatic realm (King & Baker, 2014). Such detection methods nonetheless offer a potentially powerful tool for freshwater ecologists since, by fragmenting habitat, stream drying broadly mimics the impacts of land-use disturbances. Recognizing that individual traits typically covary, as a product of trait coevolution and fitness trade-offs (Menezes ASPIN ET AL. et al., 2010;Poff et al., 2006), we used two separate approaches.
We firstly analysed 16 individual traits with clear, established linkages to drought, thus minimizing the possibility of observing spurious environment-trait relationships (Pilière et al., 2016;Verberk, Noordwijk, & Hildrew, 2013). We then explicitly accounted for trait intercorrelations by grouping taxa according to their trait profiles and analysing responses of these trait profile groups (TPGs) to drought (following Pilière et al., 2016). Our study thus comprised both readily interpretable observations of community-weighted individual traits and models of complete trait profiles.
For all individual traits analysed, we made a priori predictions of functional responses to drought (see Table 1), which were ancillary to three overarching hypotheses. These were formulated on the basis that trait selection is likely to shift abruptly as drought intensifies and habitats are lost, and were as follows: (1) moderate-intensity droughts (pool habitat fragmentation) would predominantly trigger responses in behavioural traits (e.g., dispersal, locomotion); whereas (2) under high drought intensity (streambed drying), changes in morphology and physiology (e.g., towards dessication resistant forms and aerial respiration) would also be apparent; and (3) individual trait and/or TPG responses to drought would be highly nonlinear, with some thresholds detected before complete surface water loss.

| Study site and experimental design
The research was undertaken over 2 years (February 2013-January 2015) across 21 stainless steel, flow-through stream mesocosms (spring-fed headwater stream analogues, each 15 m × 0.5 m × 0.5 m). These were sited next to a perennial reach of the Candover Brook, a mesotrophic chalk stream in the River Itchen catchment, Hampshire, UK (51°10′21″N, 1°18′70″W). Initially, borehole water was pumped into each mesocosm (to capacity) through an inlet pipe and drained over an outlet weir. Our outdoor, once-through setup thus followed design recommendations for maximizing the physicochemical and biological realism of stream mesocosms (Ledger, Harris, Armitage, & Milner, 2009). Bed material comprised fine and coarse gravel distributed to create alternating sections of deep and shallow habitat typical of lowland, low-energy chalk streams (Sear, Armitage, & Dawson, 1999;Sear, Newson, & Thorne, 2004). In each mesocosm, we created three shallow sections using bed layer depths of 25 cm, and four deep sections using bed layer depths of 15 cm. This necessarily simplified design could not capture the full morphological and hydraulic complexity of natural riffle-pool sequences, but it did Crawlers are vulnerable to predation in shrinking pools and dessication upon water loss; burrowers are better able to access streambed refugia and survive fine sediment deposition Bonada, Rieradevall, and Prat (2007), Díaz et al. (2008), Griswold et al. (2008), Robson et al. (2011), Walters (2011), Vadher, Leigh, Millett, Stubbington, and Wood (2017 Burrowing (↑) Respiration Tegument (↓) Oxygen depletion in shrinking pools and loss of water favour aerial over tegument respiration Bonada, Dolédec, et al. (2007), Bonada, Rieradevall, et al. (2007), Robson et al. (2011) Spiracle (↑) Diet Generalist (↑) Taxa with broad dietary preferences are better adapted to cope with prey loss/resource shortages during drought Williams (1996), Vázquez and Simberloff (2002) Thermal preference Cold: <15°C (↓) Eurythermic taxa are more tolerant of water temperature extremes during drought Chessman (2015Chessman ( , 2018 Eurythermic (↑) Note. Body size classes were assigned based on body mass estimates (mg dry mass).
include a core subset of properties that influence ecosystem responses to drought in field settings (i.e., variability of depth and substrate and associated refugia), thus allowing us to test for ecological responses to the progressive loss of critical stream habitat.
Throughout the manuscript, we use the terms "riffle" and "pool" to denote shallow and deep sections of stream habitat, respectively, to ensure that our terminology is consistent with other studies (e.g., Boulton, 2003). Macrophytes (Ranunculus penicillatus subsp. pseudofluitans (Syme) S.D. Webster), algae and macroinvertebrates were collected from nearby perennial stream reaches to seed the channels with taxa from the regional species pool. The mesocosms were then left to run undisturbed for 6 months to allow for community development. The channels were also accessible to aerial colonists throughout the experiment, during both this pre-disturbance period and the drought phase.
In August 2013, the sluices on the inlet pipes were adjusted to simulate a gradient of drought intensity, with each sluice maintained at a fixed setting throughout the remainder of the experiment (until January 2015) to sustain the gradient. Each channel represented a distinct treatment with a unique wetted area (range 6.5-0.25 m 2 ), water volume (1.9-0.001 m 3 ), flow (2.2-0.001 L/s) and temperature range (6-40°C maximum temperature range; Supporting Information Figure S1). During stream drought, these primary stressors covary to elicit physicochemical (e.g., oxygen availability, conductivity) and biological responses (Lake, 2011). The wide range of conditions we simulated was designed to expose the biota to levels of environmental stress beyond their typical limits, as recommended by Kayler et al. (2015) to infer potential responses to future climate extremes.
Our gradient approach offered several advantages over a more conventional factorial design with true replicates, as it allowed us to rigorously test for thresholds (Kreyling et al., 2014) and conduct analyses with significantly greater statistical power (i.e., regressionbased vs. analysis of variance-based; Cottingham, Lennon, & Brown, 2005).
Although groundwater-fed chalk stream reaches are typically hydrologically stable (Sear et al., 1999) are predicted to become more frequent given projected declines in groundwater recharge and baseflows under climate change (Jackson, Meister, & Prudhomme, 2011). Furthermore, the timing of our drought phase, beginning in summer and ending in winter, was realistic: in a groundwater-dominated stream such an event could be triggered by rainfall deficits over two consecutive winters (Wood & Petts, 1999). Drought termination might plausibly then occur the following winter in response to increased autumn rainfall, reflecting the long hydrological lag times characteristic of chalk systems (Parry, Wilby, Prudhomme, & Wood, 2016).

| Sampling and processing
In January 2015, we used a Surber sampler (0.0225 m 2 , mesh size 300 µm) to collect four benthic macroinvertebrate samples per channel (one sample per pool), which were then preserved in 70% industrial methylated spirit. Each sample comprised the uppermost 3 cm of bed gravel spanning the entire surface area of the Surber frame, allowing us to directly compare flowing and non-flowing channels. In the most drought-affected treatments, samples consisted of both dry and wet gravels: surface water was largely absent, but in the upper layer of substrate (<3 cm depth) interstitial refugia persisted and supported macroinvertebrates. Samples were taken only from pools as our focus was to compare aquatic habitats across the drought gradient: the riffle sections of over half of the treatments consisted of exposed, dry gravels. Moreover, our simplified riffle and pool habitats did not differ markedly in either flow profile (broadly uniform) or substrate type (clean gravel), and thus supported similar faunal assemblages. In the laboratory, we used a microscope to separate macroinvertebrates from detritus and identify specimens to genus (except Oligochaeta, which were recorded as such). Taxa were counted and abundance data from each of the four technical replicate samples were pooled and converted to a measure of density (individuals per m 2 ).
We recorded water temperature at 15-min intervals using Tinytag loggers (Gemini Data Loggers Ltd, Chichester, UK) placed in the terminal pool of each channel. Since oxygen depletion can be a critical stressor during stream drought (Lake, 2011), we also recorded dissolved oxygen (DO) levels in each stream at 5-min intervals over one 24-hr period each month using MiniDOT loggers (PME Inc., Vista, CA, USA) suspended midway through the water column. Temperature data were used to calculate the maximum recorded water temperature range, and oxygen data the mean daily minimum DO level, as environmental extremes are typically a stronger predictor of species' responses than means (Vasseur et al., 2014;Vázquez et al., 2017).

| Abiotic variables
We used the axis one scores of a centred, covariance principal component analysis (PCA, explained variance = 94%) to integrate measurements of the four primary drought stressors (wetted area, water volume, flow, maximum recorded temperature range) into a compound index of drought intensity (DI; Supporting Information Table S1). The index was rescaled to vary from 0 (no drought disturbance) to 1 (most severe drought). Low DI (<0.2) was characteristic of channels that  Figure S2b; Boulton, 2003). Consistent with these trends, we observed a broadly linear decline in minimum DO levels across the gradient (Supporting Information Figure S2e).

| Traits
Trait values were assigned at the genus level, using fuzzy-coded information from the European trait databases published by Serra, Cobo, Grac, and Feio (2016) Table S2). Body mass data from all channels were then aggregated to obtain size-frequency distributions for each genus.
To test our prediction that drought would increase the proportion of generalists in the community (see Table 1

| Trait profile groups
To delineate TPGs, we used the same nine grouping features, but this time incorporated a greater number of traits (n = 30 vs. 16; Table 2) to group taxa based on comprehensive trait profiles, thus ensuring that the core traits of each genus were represented within each grouping feature. We applied Gower's distance-based hierarchical cluster analysis (Pavoine, Vallet, Dufour, Gachet, & Hervé, 2009) using Ward's method to the normalized trait-by-genus matrix to identify clusters of taxa with similar trait profiles. Gower's distance was used in conjunction with Ward's method as a double-centering of the Gower dissimilarity matrix indicated that the dissimilarities closely resembled Euclidean distances (after Bruno, Gutiérrez-Cánovas, Sánchez-Fernández, Velasco, & Nilsson, 2016). An iterative procedure was used to select the optimal number of clusters, distinguished by the highest analysis of similarities (ANOSIM) R value, which would indicate maximum dissimilarity among clusters (for parsimony, and to avoid overfitting, we set an upper limit of 10 clusters as a starting condition). Random forest analysis was used to identify the most important traits and grouping features in TPG selection. Importance was calculated using Gini impurity, which describes the impurity (i.e., classification contamination) produced by splitting a particular trait in two (e.g., high ovoviviparity vs. low ovoviviparity) at each node within a decision tree (Liaw & Wiener, 2015). We measured the importance of each trait for each TPG as the mean decrease in Gini impurity (hereafter Gini value), which computes the overall (forest-wide) decrease in Gini impurity attributable to each trait (i.e., the higher the Gini value the more influential the trait).

| Statistical modelling
As trait responses to drought were highly nonlinear, we used generalized additive models (GAMs) to analyse the relationships between drought intensity and (a) trait occurrence (i.e., the standardized abundance-weighted occurrence of a particular trait in the community) and (b) TPG abundance (i.e., the untransformed abundance of taxa belonging to a particular TPG, expressed as individuals per m 2 ).
Cross-validation was used to guide the optimal level of smoothing (Wood, 2008) with minor modifications to avoid over-smoothing, as recommended by Zuur, Ieno, Walker, Saveliev, and Smith (2009).
GAMs were applied to rescaled data (see below), with diagnostic tests validating the choice of basis dimension for each smooth.

| Individual traits
Of the 16 individual traits analysed, 12 (75%) responded significantly to drought intensification, with shifts towards smaller body sizes, aerial dispersal and respiration, burrowing habitat, generalist feeding, dessication resistance and broad thermal tolerance largely corroborating our a priori predictions of trait filtering (Table 3; Figure 1). Note. The traits "susceptible" and "specialist" were calculated by subtracting the standardized "resistant" and "generalist" values from one.
Moderate-and high-intensity droughts were associated with distinct changes in community trait profiles. Our findings suggest that drought-driven habitat losses represent nested trait filters, with channel fragmentation and streambed drying both selecting for suitable behavioural traits but only the latter invoking high physiological resistance. It should be noted that these results could be conservative, as the communities of higher-energy streams with greater numbers of specialist riffle-dwellers (torrenticoles and rheophiles) might also display functional responses before the fragmentation stage (Boulton, 2003;Boulton & Lake, 2008). Here, some of the traits that became more prevalent as channels fragmented could have been a response to escalating biotic stress (e.g., burrowing as a predator avoidance mechanism, generalist feeding to cope with resource depletion), reflecting the potential for species' interactions to intensify as wetted habitat shrinks (Boulton, 2003;Lake, 2003;McIntosh et al., 2017). The abrupt shifts in morphology/physiology at the more extreme end of the gradient are more likely to reflect environmental filters sensu stricto (Kraft et al., 2015). Such shifts are consistent with the results of a separate analysis, where functional turnover patterns indicated that severe dewatering gave rise to resistance strategies uncompetitive at lower levels of disturbance (Aspin et al., 2018). Few studies to date have analysed how trait selection evolves along a continuous stress gradient, hampering our ability to formulate general predictions regarding species' sensitivities to intensifying extremes. Although continua of stressors are increasingly being T A B L E 3 GAM output for significant relationships between drought intensity and both relative occurrence of individual traits and abundances of TPGs Note. Response type is linear (L) or threshold (T). The number in brackets after response type denotes the portion of the drought gradient where the slope of the fitted GAM is >1 or <−1. "Deviance explained" provides a measure of model performance, comparable to the R 2 value in ordinary regression. Significance value denotation is as follows: ns = non-significant (p > 0.05); *p < 0.05; **p < 0.01; ***p < 0.001. All asterisked F-values are significant (p < 0.05) following the Benjamini and Hochberg (1995) correction for controlling the false discovery rate. For complete results see Supporting Information Table S3. described across natural streams (e.g., Ligeiro et al., 2013;Poff et al., 2018), the need to determine cause and effect in environment-trait linkages (Poff et al., 2006)  The ability to disperse to more favourable habitats may partly explain why the resistance of stream invertebrates to severe drying (ability to endure drought stress) is typically much lower than their resilience (resistance plus capacity to recover following flow resumption, sensu Hodgson, McDonald, & Hosken, 2015;Acuña et al., 2005;Boersma et al., 2014;Datry et al., 2014). However, recent studies of intermittent streams (Stubbington & Datry, 2013;Stubbington, Gunn, Little, Worrall, & Wood, 2016) have revealed viable life stages in dry bed sediments, indicating higher resistance than previously thought. The responses of TPGs D and F suggest that such resistance may extend to perennial stream communities. However, the success of these groups was not attributable solely to physiological resistance mechanisms: active aerial dispersal and burrowing habit were most prevalent in dewatered channels, indicating that regular recolonization from external lentic and semi-aquatic source habitats adjacent to our mesocosms (e.g., ponds, drainage ditches, wet soils)

Type N (no trend)
Relative trait occurrence (  and access to subsurface refugia may also have been important for survival. Drought extent is therefore likely to be a critical factor determining community persistence, as without sources of recolonists even the best-adapted taxa could be vulnerable on supraseasonal timescales (Stubbington et al., 2016). Here, as with other experimental studies, the proximity of mesocosms to one another (20-80 cm), and thus the distance between drought-affected habitats and recolonist sources, reflected the physical constraints of our site.
This necessary simplification of metacommunity dynamics implies that our observations of drought impacts are, again, likely to be con- water abstraction effects and flow buffering by the aquifer can give rise to patchy drying patterns (Kendon et al., 2013;Westwood et al., 2006), dispersal between disturbed and undisturbed habitats could plausibly occur over short distances.
The taxa most adapted to drought are often small and r-selected, as high reproductive rate and rapid maturation offer resilience to disturbance (Bonada, Dolédec, et al., 2007;Chessman, 2015;Ledger et al., 2012;Ledger, Edwards, Brown, Milner, & Woodward, 2011;Patrick & Yuan, 2017 These changes in TPG abundance could be considered analogous to the trait abundance shifts described by Boersma et al. (2016), whereby a decrease in the abundance, but not extirpation, of a particular trait combination (or here functional group) can provide an early warning signal of forthcoming functional extinctions (Säterberg, Sellman, & Ebenman, 2013). We therefore recommend that future traits-based Hydroptila (T)

Micropsectra (D)
Procladius (D) Note. The third and fourth columns list the five traits with which each group has the highest and lowest association, respectively. The numbers in brackets are measures of the decrease in Gini impurity resulting from taking the trait into account (the higher the number, the more influential the trait in delineating the TPG). The final column gives the genera belonging to each TPG, as well as the order to which the genus belongs (A = Amphipoda, C = Coleoptera, D = Diptera, E = Ephemeroptera, G = Gastropoda, H = Hirudinea, I = Isopoda, M = Megaloptera, P = Plecoptera, T = Trichoptera, Tc = Tricladida).
studies of drought look beyond community-averaged response variables (e.g., individual trait occurrences), to ensure that potentially catastrophic functional impacts do not go undetected.
Ecological responses to extreme climatic events are typically highly idiosyncratic (van de Pol, Jenouvrier, Cornelissen, & Visser, 2017), so our ability to predict the ecological impacts of severe droughts will largely hinge on the mechanistic insights offered by controlled, manipulative experiments and traits-based approaches.
Understanding which traits confer resistance (and vulnerability) to extreme drought should allow for more targeted conservation efforts during water deficits. For instance, the tendency for most taxa with high physiological resistance to drying to be aerial dispersers underscores the importance of maintaining a network of refugia to act as sources of recolonists. More generally, the high sensitivity of many traits to drought intensification highlights their value as functional biomarkers for resistance and resilience at both species and community level, potentially supplementing existing taxonomy-based biomonitoring metrics (e.g., DELHI index; Chadd et al., 2017).

ACKNOWLEDG EMENTS
TA was supported by a studentship from the NERC CENTA doctoral