Shrub growth and expansion in the Arctic tundra: an assessment of controlling factors using an evidence-based approach

Woody shrubs have increased in biomass and expanded into new areas throughout the Pan-Arctic tundra biome in recent decades, which has been linked to a biome-wide observed increase in productivity. Experimental, observational, and socio-ecological research suggests that air temperature—and to a lesser degree precipitation—trends have been the predominant drivers of this change. However, a progressive decoupling of these drivers from Arctic vegetation productivity has been reported, and since 2010, vegetation productivity has also been declining. We created a protocol to (a) identify the suite of controls that may be operating on shrub growth and expansion, and (b) characterise the evidence base for controls on Arctic shrub growth and expansion. We found evidence for a suite of 23 proximal controls that operate directly on shrub growth and expansion; the evidence base focused predominantly on just four controls (air temperature, soil moisture, herbivory, and snow dynamics). 65% of evidence was generated in the warmest tundra climes, while 24% was from only one of 28 floristic sectors. Temporal limitations beyond 10 years existed for most controls, while the use of space-for-time approaches was high, with 14% of the evidence derived via experimental approaches. The findings suggest the current evidence base is not sufficiently robust or comprehensive at present to answer key questions of Pan-Arctic shrub change. We suggest future directions that could strengthen the evidence, and lead to an understanding of the key mechanisms driving changes in Arctic shrub environments.


Introduction
The Arctic tundra biome provides essential regulatory effects to global climate, in particular albedo (Juszak et al 2014), storage of organic carbon in its living biomass (Nauta et al 2014), and permafrost dynamics (Blok et al 2010). Over at least the last three decades, changes in vegetation composition have occurred that have significant consequences for the regulatory capability of tundra environments. Specifically, the ability of woody shrub species to produce biomass has increased, leading to shrubs of greater maximum height (Epstein et al 2012). Spatial expansion has also occurred: latitudinal 'shrublines' have advanced (Myers-Smith and Hick 2017), and new recruitment has enabled progressive filling of patchy landscapes (Tape et al 2006, Myers-Smith et al 2011, both at the expense of mosses and lichens (Elmendorf et al 2012b). Such 'shrubification' has been a Pan-Arctic trend since the 1980s, supported by data from experimental plots (Elmendorf et al 2012a), remote sensing and repeat photography (Sturm et  Air temperature and growing season lengths have increased in tundra ecosystems more than at lower latitudes, due to positive feedbacks that snow and ice (both on land and at sea) have with climate (Serreze and Barry 2011). Shrubification can be attributed primarily to air temperature changes (Myers-Smith et al 2015), and to a lesser extent soil moisture (Myers-Smith et al 2015, Ackerman et al 2017, although shrub responses are heterogeneous. Data from the International Tundra Experiment (ITEX) long-term plot network demonstrates regional differences in the responses of tundra vegetation to summer air temperatures (Elmendorf et al 2012a). Similarly, shrub ring chronologies indicate heterogeneous long-term responses to mean summer temperature, with maximum sensitivity in warmer and wetter tundra sites (Myers-Smith et al 2015).
The observed heterogeneity suggests that other processes are important in controlling shrubification trends.
Shrubification has been linked to satellite-derived observations of widespread 'greening' (increases in vegetation productivity, as measured by the Normalised Difference Vegetation Index-NDVI). Recently, the NDVI index has shown widespread negative trends across the Arctic tundra for the first time in decades (Epstein et al 2015, Ju andMasek 2016). NDVI has been demonstrated as a correlative proxy for shrubification (e.g. Forbes et al 2010), but predictions based on NDVI assume that (a) correlations between plot-scale productivity and NDVI holds across Arctic regions, despite local-scale factors introducing uncertainty , and (b) the relationship holds under future conditions (e.g. increased landscape shrub biomass). Recognising these uncertainties, the recent negative NDVI trends could be driven by complex environmental controls on shrubs beyond simple temperature metrics, such climatic extremes, and/or discrete disturbance events (Phoenix and Bjerke 2016). A progressive decline in the relationship between air temperature and NDVI since 1982 (Piao et al 2014, Kremers et al 2015 further supports the role of controls beyond air temperature.
Rapidly increasing air temperatures or increased growing season lengths appear responsible for shrubification trends, but with significant roles for other controls that contribute to heterogeneity in shrub-temperature responses. Without a robust assessment of these controls, one cannot ascertain their relative importance, the adequacy of current study designs, or the evidence required to reveal mechanisms driving shrubification processes. We conducted an evaluation of the current evidence base to answer the following questions.
1. What are the suite of controls that may act upon shrub growth and expansion in the Arctic tundra?
2. Do study designs take account of controls to shrubification and the mechanisms that may drive them, and are there spatial gaps in the evidence base that may limit our ability to detect their significance?
3. Do study designs take account of temporal characteristics sufficiently comprehensively to enable inferences to be drawn about likely mechanisms?

Protocol
To establish the controls that may be operating on Arctic shrub growth and expansion, the quantitative evidence base for each control, and gaps in current research directions, we systematically mapped recently published literature (full protocol in supplementary appendix A, available online at stacks.iop.org/ERL/12/ 085007/mmedia). Briefly, we searched the online database Web of Science Core Collection for 'topic = Arctic AND Shrub Ã ' , limited to publication years January 2012-January 2017. The following inclusion criteria were then applied.
1. Shrub response. The study carried out statistical analysis within which at least one direct measure of shrub growth or expansion was used as a response variable (see supplementary appendix A).

2.
Control. Within the statistical test(s), an environmental control external to the shrub was used as a predictor to test against shrub response(s) identified in 1.

3.
Location. At least one site for which the statistical test was completed must occur within the Arctic tundra. We defined the Arctic tundra as any land north of the Arctic treeline (Walker et al 2002) and 'Oro-Arctic' areas (Virtanen et al 2016).
For each included source, we identified every environmental control used as a predictor, at every independent site. The many-to-many relationship between sources, controls, and sites was multiplied out to form source-control-site data points, hereafter referred to as evidence points.

Delineation of methodologies
Methodology was characterised for each evidence point as non-experimental or experimental, then into subclasses depending on temporal characteristics. Following best practice in evidence synthesis (Collaboration for Environmental Evidence 2013), we characterised the data used within statistical analyses and not the data collected. For non-experimental evidence, observational Environ. Res. Lett. 12 (2017) 085007 controls had measurements taken through time to form a time-series of two or more time points. Spatial gradients used multiple measurements across space to substitute for time, while chronosequences attributed such variation across space to specific previous times to form a retrospective time-sequence. For environmental controls that had been manipulated, we defined four broadly distinct forms of experimental design based on the temporal nature of the data used within statistical analysis. i. A time-series factorial was defined as an experiment in which measurements of both the environmental control and shrub response(s) were taken through time, and included in statistical analysis.
ii. A response-only factorial only included timeseries for the response variable, with no predictor time-series.
iii. A non-temporal factorial contrasted the effect of a manipulation with a control plot, but no time series was present. For example, a nutrient addition experiment that tests for an effect on budding date after 18 years, with no 'before' point, and using differences between control and manipulation plots as a substitution for time, would fit this category.
iv. An experimental chronosequence used multiple plots through space with varying treatment lengths to assess the role of treatment on shrub response(s).

Classification of controls
We classified the environmental controls found in the evidence base into two major categories-ultimate and proximal-to provide scope and rigour to the systematic analysis via this underlying framework (figure 1). Proximal controls are defined as environ-mental state parameters that directly impact the ability of a shrub individual to increase in biomass, reproduce or establish, without the need for any intermediate environmental properties (e.g. soil moisture, fire). Proximal controls provide the minimal degree of complexity from which to characterise the underlying mechanisms controlling shrub growth and expansion. Where a proxy measure was used that could be directly attributed to a proximal control (e.g. thaw degree days, for ice and frost), this was included as an evidence point for the proximal control (all proxy measures listed in table A1). Proximal controls are driven ultimately by further environmental properties that influence their occurrence in space and time (ultimate controls), such as the role of sea ice on local air temperature, but without support for any direct mechanistic relationship to shrub performance. Shrub traits (e.g. leaf size and properties, reproductive strategy, wood and vessel structure, metabolic adaptations, growth form, species-related symbiotic relations, etc.) are significant determinants of plantenvironment interactions, and can vary between genera, species, populations, ecotypes, and functional type (Chapin et al 1996). As we did not consider effect sizes in this analysis, we do not formally characterise internal controls here, and leave this for discussion and as a future avenue for research.

Analysis of spatial characteristics
To assess the degree to which the evidence points were spatially clustered or dispersed, we computed spatial autocorrelation using the Global Moran's I statistic (using an inverse distance spatial relationship over Euclidean distance). This approach was additionally utilised to identify spatial clustering for control categories, controls, and experimental designs. To identify specific hotspots of evidence production, we calculated the Getis-Ord Gi Ã statistic (Getis and Ord 1992).

Temporal Integration
Shrub Growth

Shrub Expansion Shrub Traits
Phenotypic / Genetic Figure 1. Conceptual overview of the framework used for the analysis. Proximal controls are state parameters that directly influence the performance of shrub individuals, without any intermediary role of other environmental parameters. These controls may be resources that can become limiting (e.g. soil moisture, nutrients), or disturbance drivers that can cause damage (e.g. gall mites, storm damage). The effectiveness of proximal controls is mediated by shrub traits (leaf size and properties, reproductive strategy, wood and vessel structure, metabolic adaptations, growth form, species-related symbiotic relations). The occurrence of proximal controls depends on additional environmental parameters-ultimate controls (Env a , Env b . . . Env x ).
Environ. Res. Lett. 12 (2017) 085007 To identify research gaps in terms of broad environmental/ecological variability, we computed intersections between available Pan-Arctic layers and all evidence points, calculating Getis-Ord Gi Ã for each resultant landscape component. c. for biodiversity, we used Arctic floristic groups and sectors (Elvebakk et al 1999, Yurtsev 1994, which represent broad patterns of plant species diversity (occurring due to regional differences in glacial and landscape history).

Results
135 of the 432 sources identified met the inclusion criteria and were included in the final analysis. We found 1140 source-control-site evidence points reported during the period January 2012-January 2017 (inclusive), derived from the 135 sources.

Suite of controls
We identified 23 proximal controls (1029 evidence points), presented in table 1. Despite the range of potential proximal controls, there was predominant focus on just five: air temperature (including mean, maximum, minimum, above-freezing mean, growing degree days, and diurnal regional temperature proxies-429 evidence points, or 41.69% of all proximal evidence points), soil moisture (including precipitation mean and sum, groundwater level, water track presence, and soil drainage proxies-263, 25.56%), active layer depth (124, 12.05%), and to a lesser extent herbivory (66, 6.41%), and snow depth/cover (including snow-free date-37, 3.56%). We also identified analysis of 24 ultimate controls within the evidence base, outlined in table 2.

Spatial characteristics of the evidence base
Spatial analysis revealed areas of research focus, and spatial gaps (full results in appendix A.2.3). Analysis of the spatial structure of all evidence points revealed global clustering (Moran's index = 0.237, z = 2.13, p = 0.033). Hotspot analysis indicated six significant (p < 0.05) hotspots of evidence production, centred in Alaska (Toolik Lake, Barrow, and Atqasuk), Alexandra Fiord (Canada), Endalen (Svalbard), and Abisko (Sweden). Patterns of spatial clustering were significantly different between study designs, and controls (figures 2

Methodological and temporal characteristics of the evidence base
In total, 86% of evidence points were derived from observation, with 14% derived from experimental data. For proximal controls, we found the greatest use of spatial gradient approaches for air temperature (14%), herbivory (5%), and soil belowground conditions (soil moisture (40%), and active layer depth (28%) predictors ( figure 4(A)). Spatial gradient evidence points constituted 40% of the total. Soil belowground resources were assessed for a median timespan of eight years, biotic interactions for ten years, air-plant interface controls for 29 years, soil belowground conditions for 50 years, and soil surface conditions for 240 years. Evidence was generally limited to no more than 25 years, aside from certain controls and study designs where long-term observational data could be obtained ( figure 4(B)). Decadal to centennial evidence was dominated by weather-station-derived proxy measures (coupled with dendroecological and repeat photography response variables): gridded, interpolated data products enabled numerous long-term studies of air temperature (proxy: regional air temperature), and soil moisture (proxy: regional precipitation). Space-for-time substitution was used widely, specifically for soil moisture, air temperature, and to a lesser extent herbivory, and snow dynamics. While observational evidence was used for all proximal controls aside from atmospheric CO 2 and insolation (including UVb), manipulations were limited to 13 out of the 23: air temperature, snow dynamics, herbivory, nutrient availability, ice formation, insolation, CO 2 , and soil abiotic conditions. Experimental design and the resulting evidence was weighted towards the use of non-temporal analyses ( figure 4(B)). This was especially pronounced for certain controls: for soil 4. Discussion 4.1. Current evidence base 4.1.1. Suite of controls A predominant focus on air temperature, soil moisture, and herbivory controls suggests that other proximal controls-that may explain recent shrubification trends-are being overlooked. A scoping exercise (appendix A) identified additional proximal controls that were not included in the compiled evidence base: abrasion by snow and ice crystals (Sonesson and Callaghan 1991), wind damage, microbial (Sedlacek et al 2014) and mycorrhizal (Deslippe and Simard 2011) associations, pollinators (Rich et al 2013), allelopathy (Bråthen et al 2010), soil micronutrients, and soil texture . These proximal controls, alongside those that made a low proportion of the evidence base (atmospheric CO 2 , insolation, cryoturbation, erosion (including aeolian and thermo-erosion), and fungal infection), may have been overlooked.

Spatial gaps
Strong spatial clustering of the evidence base towards Alaska and Fennoscandia (figure 2), as well as spatial gaps in the Eurasian Arctic (figure 5), indicate that full spatial variability may not be captured for each proximal control. 65% of the evidence was generated within the warmest parts of the Arctic tundra biome, where summer (July) temperatures average above 9°C. Consequently, any controls and their mechanisms occurring exclusively, or with greater strength, in colder regions may be missed. Dominant processes driving shrubification vary between warmer tall shrub-dominated tundra (spatial infilling), and northernmost shrublines (increasing height and northward expansion). As these processes differ by biological mechanism, responses to controlling factors are likely different. The elevation gradient at Brooks Range has been used as a proxy for bioclimatic subzone, with elevation as a proxy for latitudinal space; however, the non-carbonate bedrock and acidic soils of the range do not account for the variability of plant functional forms and environmental conditions within higher latitude bioclimatic subzones.
Although shrubification trends appear to be driven by key species with Pan-Arctic distributions (Betula nana, Salix sp.), there are indications of Environ. Res. Lett. 12 (2017) 085007 regional genotypic variation in these, and other, shrub species (Abbott and Brochmann 2003, Eidesen et al 2007, Eidesen et al 2013, Jørgensen et al 2012. Similarly, there is evidence for significant phenotypic plasticity within shrub species in response to some proximal controls (Edwards et al 2005, Berner et al 2015, such as within-species spatial gradients from prostrate to erect growth forms. Significant focus of evidence in the 'Alaskan Tundra' floristic sector (Tkach et al 2010), and the Fennoscandian Oro-Arctic, may limit coverage of unique Eurasian ecotypes, species, and thus adaptations, resilience and/or vulnerabilities (figure 5). Spatial focus on long-term ITEX plots at Toolik Lake and Daring Lake (Alaska, USA) has provided comprehensive evidence for moist, low shrub tundra environments; however, this habitat does not account for the breadth of tundra physiognomies (aside from tussock sedge, dwarf shrub, moss tundra'), where other mechanisms may be significant.

Temporal limitations
We noted temporal limitations to soil controls, where the extent of temporal evidence (aside from chronosequence) was generally limited to below 25 years ( figure 4(b)), while hypothesised drivers (e.g. changes in carbon and nitrogen cycling) may occur over decadal to centennial timescales. The mechanisms through which controls may operate vary by their timescales, from diurnal to centennial timescales. The lower temporal resolutions for soil-based controls (appendix figure A8) also limits inference of withinseason and inter-annual control variability, such as how seasonal variability may impact different lifestages (budding, flowering).
Without time series, one can establish the directionality of response, but not the functional form (linear, non-linear) of the mechanism(s) at work. As 42% of experimental evidence utilised nontemporal approaches, these evidence points cannot be used independently to ascertain temporal dynamics, but may only be useful when combined in metaanalyses (e.g. Elmendorf et al 2012b), assuming methodologies can be compared. Similarly, climatic gradients, and the Finland-Norway herbivory gradient, provided a large fraction of evidence. Such spacefor-time substitution approaches do mask the rate and order of temporal processes, and have been empirically proven to overestimate the effects of air temperature on tundra shrub growth compared to experimental and observational data (Elmendorf et al 2015).

Applications and limitations of our approach
One or more mechanism(s) may be responsible for the aggregate effects of a proximal control on shrub growth and expansion through space and time. In the context of global change, these mechanisms need characterisation if we aim at predicting future changes in shrub performance, habitat, and distribution. Our methods of metadata collection can be used as a tool to assess the suitability of the evidence base to support or refute possible mechanistic hypotheses. This approach is demonstrated for soil nutrients in section 4.2.1.
We acknowledge limitations in our approach. First, the evidence gap between the Eurasian and North-American Arctic represents a publication bias; our search strategy does not cover Russian-language or other non-English scientific literature: spatial gaps in Eurasia may therefore have been accentuated. Second, we did not attempt to characterise the importance and strength of proximal controlling factors (resource limitations and discrete events) in space and time, but only the nature of the recent evidence base. Third, as our aim was to characterise the current trends in, and direction, of research, our analysis only represents the most recent five years of research, while older research may display different research quantities and priorities. We extended our search protocol to cover past research, finding that our study analysed 37.5% of all research captured by the search criteria (appendix A.2.1).

Soil nitrogen and shrubification
Soil macronutrients-including nitrogen (N), phosphorus (P), and potassium (K)-are essential resources for plant survival and fitness. Their availability is spatially heterogeneous at all scales (Walker et al 2005, DeMarco et al 2011, as a result of geology, glacial history, landscape processes, abiotic microhabitats, and plant community composition. Nitrogen is one of the most limiting macronutrients to growth in high latitudes (Bobbink et al 2010). There are multiple hypotheses for trajectories of tundra N availability, including: (1) increasing N availability as elevated soil temperatures increase the efficiency of N-mineralising microbes (Sturm et al 2005); (2) sequestration of N into long-lived woody biomass, reducing plantavailable N in soils over decadal to centennial timescales (Progressive Nitrogen Limitation-PNL) (Luo et al 2004);and (3) increasing anthropogenic N deposition (Bobbink et al 2010).
Elevated N increases shrub aboveground biomass and shrub cover, with combined N-P limitation occurring in certain locations (Zamin and Grogan 2012). Evidence was limited to 25 years, which is not long enough to support or refute some shrub-N interactions such as PNL: short-term mechanisms can distort long-term (decadal to centennial) processes (Johnson 2006). Exclusive use of non-temporal experimental approaches (figure 4(b)) limits our understanding of rates of change, providing only single measures of 'length and strength of manipulation' to elevated response. The predictors do not quantify soil bioavailable N, essential to infer starting conditions and stressing and limiting levels of N, nor its forms, essential for understanding mechanisms of uptake and their variability between taxa and environments (i.e. organic versus inorganic forms). Manipulations often do not reflect the rates of change Environ. Res. Lett. 12 (2017) 085007 hypothesised for bioavailable N, fertilising at levels beyond expected quantities and rates of change (Bouskill et al 2014).
Past and future trajectories of N, thus N-shrub interactions, may be determined with alternative methodologies. Spatial variability in N or shrub traits (mycorrhizal associates, N-use efficiency) may explain the differences in observed N limitation across space, requiring measures of N and shrubification beyond ITEX plots. Temporal data could allow partitioning of short-and long-term responses that are difficult to differentiate using non-temporal approaches. Ideally, time-series measurements of bioavailable N on the same timescales as shrub responses would enable researchers to characterise rates of change within and between years whilst accounting for background N variability. Such time series could be interrogated using statistical modelling techniques, to infer the model and parameters of N-dependent growth.

Mechanisms driving recent and future shrubification trends
To reduce uncertainty and increase predictive capability of future shrubification trends, we require mechanistic rather than correlative understandings of the underlying processes. We suggest three key knowledge gaps that must be reduced to gain such an understanding.
1. Spatio-temporal trends of shrubification. Properties beyond biomass and cover that receive lesser attention, such as phenology (Prevéy et al 2017), and advancing shrublines (Myers-Smith and Hik 2017), could be measured for enhanced clarity over Pan-Arctic shrubification trends.
2. Effectiveness of proximal controls. Study designs may be sought that can assess the effectiveness per-unit variability within controlling factors on the identified mechanisms of shrubification, within the present range of environmental variability.
3. Past and future variability of control(s). Each proximal control will vary through time due to a suite of underlying ultimate controls. Establishment of variability for the recent period, over which shrubification has occurred, and linking this to effect sizes, could enable establishment of (a) controls that are varying over the recent period, and (b) controls that may be responsible for observed changes.
We suggest four methodological directions through which tundra ecologists could enhance their study designs to address the above knowledge gaps.
1. Incorporation of time series, to establish the directionality and functional forms of shrub responses to environmental controls.
2. Direct measurement of proximal controls. Many factorial studies did not measure the environmental control being studied, but rather measured the size and rate of perturbation. These methods assume that there is a direct link between perturbation and control (e.g. addition of 5 g nitrogen fertiliser raises bioavailable nitrogen by a linear quantity). Inference of mechanisms could be enhanced by measuring the proximal control(s) directly, for example using automatic continuous loggers rather than gridded climate products. For time series, this will require creative solutions to overcome control-specific difficulties. Soil belowground resources, for example nutrients, require measurements by field researchers, but new technologies should be sought to increase automatic data collection capabilities.

Conclusions
Whereas there is significant evidence for an important role of air temperature and precipitation as drivers of Arctic shrubification, our systematic approach identified 23 proximal controls (those operating directly on the individual shrub and potentially affecting its growth and/or expansion) reported between January 2012-January 2017, spanning soil properties, biotic interactions, and the plant-atmosphere interface. The focus of shrubification research has prominently been on air temperature and precipitation, while evidence suggesting a progressively declining role of climate requires us to consider other potential controls. We found spatial gaps in the evidence for all proximal controls, with research concentrated in the warmest bioclimatic zones of the tundra, and spatial gaps in Western and Central Arctic Siberia. These regions of research concentration already have a high percentage of tall Environ. Res. Lett. 12 (2017) 085007 shrub cover, while regions in the intermediatelatitude tundra (bioclimatic subzones B-D) were sparsely covered. There is a basic mechanistic understanding of manyof the controls on tundra shrubification, mostly derived from experiments conducted in acidic, low shrub, low latitude tundra, where shrubs are already a major component of the vegetation. In comparison, there is little focus on the mechanisms of range expansion and northward dispersal, operating at the northernmost range limit. In the studies included here, we found limitations in the temporal extent and resolution of evidence used, although this varied considerably depending on the proximal control considered. Study designs were in general found to be insufficient for investigating the mechanistic relationship between controls and shrubification, due to frequent use of non-temporal approaches. Reliance on space-for-time and non-temporal approaches risks not accurately reflecting the true rate and order of processes operating within the system.
We identify three knowledge gaps and four recommendations that tundra ecologists can consider to enhance the value of their data and future research. If progress is to be made toward predicting future spatial-temporal shrubification trends, more emphasis must be placed on the mechanisms underpinning shrubification.