Warming, drought, and disturbances lead to shifts in functional composition: A millennial‐scale analysis for Amazonian and Andean sites

Tropical forests are changing in composition and productivity, probably in response to changes in climate and disturbances. The responses to these multiple environmental drivers, and the mechanisms underlying the changes, remain largely unknown. Here, we use a functional trait approach on timescales of 10,000 years to assess how climate and disturbances influence the community‐mean adult height, leaf area, seed mass, and wood density for eight lowland and highland forest landscapes. To do so, we combine data of eight fossil pollen records with functional traits and proxies for climate (temperature, precipitation, and El Niño frequency) and disturbances (fire and general disturbances). We found that temperature and disturbances were the most important drivers of changes in functional composition. Increased water availability (high precipitation and low El Niño frequency) generally led to more acquisitive trait composition (large leaves and soft wood). In lowland forests, warmer climates decreased community‐mean height probably because of increased water stress, whereas in highland forests warmer climates increased height probably because of upslope migration of taller species. Disturbance increased the abundance of acquisitive, disturbance‐adapted taxa with small seeds for quick colonization of disturbed sites, large leaves for light capture, and soft wood to attain fast height growth. Fire had weak effects on lowland forests but led to more stress‐adapted taxa that are tall with fast life cycles and small seeds that can quickly colonize burned sites. Site‐specific analyses were largely in line with cross‐site analyses, except for varying site‐level effects of El Niño frequency and fire activity, possibly because regional patterns in El Niño are not a good predictor of local changes, and charcoal abundances do not reflect fire intensity or severity. With future global changes, tropical Amazonian and Andean forests may transition toward shorter, drought‐ and disturbance‐adapted forests in the lowlands but taller forests in the highlands.


| INTRODUC TI ON
Old-growth tropical forests were long thought to be in a stable state. However, direct observations are accumulating that over the past decades forests are showing changes in productivity (Brienen et al., 2015), species composition (Feeley, Davies, et al., 2011), and functional trait composition (van der Sande et al., 2016).
These changes can be caused by climate change, such as warming, droughts or increased atmospheric CO 2 concentration, or by natural and anthropogenic disturbances such as logging, hurricanes, and fires (Hirota et al., 2021). Understanding forest responses to multiple potential drivers is of critical importance, as it will help to understand forest responses to future changes. It is, however, difficult to understand the responses of tropical forest to multiple drivers, as they can have opposing effects and act at different timescales, and are therefore rarely considered in the same study. Most studies assess changes in tropical forests across a few decades (Brienen et al., 2015;Feeley, Davies, et al., 2011;van der Sande et al., 2016), whereas the recruitment and growth of trees to the canopy can take centuries, and the successional changes to old-growth forest composition even longer (Rozendaal et al., 2019). We therefore need to assess tropical forest responses at temporal scales of centuries to millennia.
Traditionally, long-term (100-1000 of years) palynological studies link changes in pollen taxon associations (González-Carranza et al., 2012;Smith & Mayle, 2018) to their ecological niche space to interpret the underlying environmental drivers of vegetation change.
However, this way to infer the factors driving environmental change can be made more mechanistic by using functional plant trait characteristics. Functional traits are plant characteristics that determine plant establishment, growth, and survival, and thus the responses of species (Poorter et al., 2006;Violle et al., 2007) and ecosystems (Becknell & Powers, 2014;van der Sande et al., 2018) to environmental change. Therefore, using pollen data to analyze how the trait composition of plant communities is affected by human disturbance and climate change over centennial to millennial timescales will allow for a better mechanistic understanding of forest responses to global changes (Sakschewski et al., 2016). Various studies have used such a functional trait approach by determining the community-weighted mean trait values, to assess short-term, decadal forest responses to environmental change (Enquist & Enquist, 2011;Fauset et al., 2012;van der Sande et al., 2016;Zhou et al., 2014).
Recently, we presented a new approach that analyzed millennialscale shifts in functional trait composition, based on species abundance changes from fossil pollen records (van der Sande et al., 2019(van der Sande et al., , 2020. We found that over a time period of 7000 years, the trait composition of a lowland forest in Peru was affected by past climate change and human disturbance (van der Sande et al., 2019).
Using eight different traits, we showed that the community-mean wood density (WD), an indicator of drought and shade tolerance, decreased during times of high precipitation. Also, the communitymean leaf size increased with precipitation, probably because large leaves need more water for transpirational cooling and/or larger leaves are needed to intercept more light in wetter and denser forests. Furthermore, the community-mean tree height and seed mass (SM) increased with fire activity, perhaps because large species can escape crown fire, and seedlings from large seeds grow faster and have a greater chance of surviving a possible next fire. However, to date no studies have been done to assess how these relationships hold up across the diverse forests of the Amazon and Andes, and we do not know to what extent environmental drivers similarly affect trait composition for other forest types and at broader spatial scales.
Here, we assess the functional trait composition of eight South American tropical forest landscapes (ranging from 0 to 3200 m asl) and their millennial-scale responses to past climate change and disturbances. Along elevation, marked decreases occur in water availability, temperature, atmospheric pressure, and insolation, resulting in changes in species and trait composition (Enquist et al., 2017;Llerena-Zambrano et al., 2021;Maharjan et al., 2021). Such differences in environmental conditions and trait composition may lead to different forest responses to climate change and disturbances.
For example, while the distributions of lowland tropical tree taxa | 4777 van der SANDE et al. are strongly influenced by mean annual rainfall (Toledo et al., 2012), the distributions of montane tree taxa, such as Weinmannia and Polylepis, are mostly determined by mean annual temperature (Duque et al., 2015;Fadrique et al., 2018;Groot et al., 2013). From ca. 2300 m, upslope night frost may occur, making resistance to night frost important for taxa occurring in these sites. This could suggest that Amazonian lowland forests respond most strongly to changes in rainfall, whereas highland Andean forests respond most strongly to changes in temperature and occurrence of night frost (Fadrique et al., 2018). In addition, intense droughts, such as those caused by El Niño events, are generally strongest in eastern and northern Amazonian lowland forests (Marengo et al., 2021). Furthermore, during the Holocene, tropical forests have seen varying degrees of human occupation and associated disturbances (McMichael & Bush, 2019). In many wet tropical forest ecosystems, fires do not naturally occur at high intensity or frequency, and are therefore often considered a human disturbance (Gosling et al., 2022). Forests with a longer history of natural or anthropogenic disturbances, such as fire or clearcut for agriculture (Iriarte et al., 2020), may be more resilient and show weaker responses (due to high resistance and/or fast recovery) to an increase in disturbances (Cole et al., 2014). On the other hand, frequent disturbances such as fires can also reduce an ecosystem's resilience and cause drastic shifts in state, for example to a savanna ecosystem (Staver et al., 2011).
We ask two questions. First, how do temporal changes in past climate (precipitation, temperature anomalies, and the frequency of El Niño events) and past disturbances (fire activity and disturbances in general) determine shifts in community-mean trait values across Neotropical forests over at least the last 2000 years (for some of our sites up to 12,000 years). We hypothesized that increased rainfall and decreasing droughts increase the relative abundance of drought-vulnerable taxa that may have a combination of low WD, large adult stature with large hydraulic path lengths, and large leaves (Aleixo et al., 2019;Guillemot et al., 2022). High disturbance intensity (e.g., from fire, hurricanes, or land-use change) can have contrasting effects: on the one hand, it can create more open spaces and increase the relative abundance of pioneer species with low WD and large leaves. On the other hand, when disturbance events occur frequently, it could increase the relative abundance of taxa that can resist disturbance (e.g., with dense wood), can resprout, or that have a shorter life cycle (e.g., short adult stature). Second, how do these responses of trait composition to climate and disturbances depend on elevation? We hypothesized that precipitation and El Niño effects are stronger at low elevation, because water availability is a more limiting resource, while temperature effects are stronger at high elevation, because of its limiting effect on tree growth (Table 1). To address these questions, we use data from eight fossil pollen records, which we match with data of four key functional leaf, stem, and whole-plant traits that capture the global spectra in plant form and function (Díaz et al., 2015). Subsequently, we calculate community-mean trait values across time and link these to disturbance indices (charcoal to indicate fire, and abundance of the pioneer genus Cecropia spp. to indicate disturbances) and independent climate proxies (for precipitation, temperature, and El Niño events).
We test these hypotheses using cross-site analyses and site-specific analyses.

| Sites
We focused this study on eight sites in forest-dominated landscapes in the Amazon and Andes (Table 2; Figure 1). We selected four lowland (5-600 m), one mid-elevation (1250 m), and three highland (2780-3150 m) sites. For each site, we had one fossil pollen record derived from lake sediments to assess changes in pollen taxonomic composition for the last two millennia, while some pollen records reflect a longer period up to the last ca. 12 kyr. We linked the pollen taxa names in these records (at genus level) to leaf and stem trait data to quantify "community-mean traits" changes over time. From the selected lake sediment cores, we also obtained data on charcoal for fire activity (Whitlock & Larsen, 2001) and quantified the abundance of Cecropia pollen as a proxy of forest openness (Zalamea et al., 2012), which was used as indicator of general disturbances.
Furthermore, we used independent, external sources for climatic proxies of precipitation, temperature, and the frequency of El Niño events (see for a complete description below).

| Fossil pollen records
We used data from eight published fossil pollen records that covered at least the most recent 2000 years of the Holocene (mean = 6830 and max = 10,020 cal years BP; Table 2). Pollen grains and charcoal particles washed or blown in from land are well preserved in lake sediments and are therefore used to assess changes in vegetation and fire through time. We used the calibrated years before present (cal years BP) data as published per site (see references in Table 2), which were obtained from age models based on 14 C from the sediment cores.
We selected these eight fossil pollen and charcoal records be-  Table 2. In total across all records, we had pollen data from 1200 pollen samples (i.e., 1200 points in time).
Samples were prepared following standardized methods (Faegri & Iversen, 1989). Pollen grains in the samples were identified and counted, in most cases up to 300 pollen grains per sample. Details TA B L E 1 Hypotheses for the effects of environmental variables (precipitation, temperature, El Niño frequency, fire activity, and general disturbances) on the temporal changes in four community-mean traits (adult height, leaf area, seed mass, and wood density).

Adult height (H)
Leaf area (LA)

Precipitation
High precipitation increases H because less drought allows the establishment of tall taxa that have longer hydraulic pathlengths (Bennett et al., 2015;Olson et al., 2018) High precipitation increases LA because large-leaved species largely rely on transpiration for their heat cooling, which is easier at high precipitation (Greenwood et al., 2017) High precipitation increases SM because large-seeded taxa are more shade-tolerant and can better germinate and persist in wet forests that are generally taller and more shaded (Baraloto et al., 2005;Moles et al., 2005;Poorter & Rose, 2005) High precipitation decreases WD because soft-wooded taxa are less resistant to droughtinduced cavitation (Eller et al., 2018;Greenwood et al., 2017;Markesteijn et al., 2011) Temperature Temperature decreases H at low elevation because of higher evaporative demand and a relative advantage of taxa with short hydraulic path lengths (Olson et al., 2018). However, temperature increases H at high elevation because of lower risk of freezing-induced cavitation and a relative advantage of tall, cavitation-vulnerable taxa (Mao et al., 2018) Temperature decreases LA at low elevation because small leaves have more efficient convective heat cooling, but high temperature increases LA at high elevation because of denser vegetation and more light competition Temperature decreases SM at low elevation because these species germinate better in the more open conditions due to gap formation and shorter trees, but increases SM at high elevation because these species establish better in the denser and taller forest Temperature increases WD at low elevation because of the higher evaporative demand and a relative advantage of drought-tolerant taxa with high WD (Liang et al., 2021). However, temperature decreases WD at high elevation because of lower risk of freezing-induced cavitation and a relative advantage of cavitation-vulnerable taxa with wide vessels (Yang et al., 2020) and, hence, low WD  (Poorter & Rozendaal, 2008)), or fire activity decreases LA because of more efficient convective cooling in more open and warmer conditions Fire activity decreases SM because small but many seeds enhance colonization of burned sites), or fire activity increases SM because large seeds are associated with high resprouting capacity and fast growth of seedlings, which enhances their chances to escape fire (Lahoreau et al., 2006;Westoby, 1998) Fire activity increases WD, because dense wood is associated with less fire damage and with high resprouting capacity probably due to more carbohydrate reserves (Brando et al., 2012;Poorter et al., 2010) Disturbances Note: The cartoon in (a) provides a simplified overview of relative changes in trait composition that we expect with increased precipitation and reduced droughts (upper row), increased temperature (second row), increased fire activity (third row), and increased general disturbances (last rows). Note that these represent relative changes over time in response to different drivers, and do not represent absolute differences between elevations and drivers. The table in (b) provides more detailed hypotheses on each of the environmental effects (rows) on the four community-mean traits (columns).

TA B L E 1 (Continued)
TA B L E 2 List of metadata information for the eight sites used ordered from high to low elevation.   Note: For each site, metadata are provided on elevation, country, coordinates (in decimal degrees), current-day average annual temperature and precipitation, the source of the paleo-precipitation record used, information on the human disturbance history of the site, the timespan of the pollen and environmental records used per site, the average temporal resolution of the pollen record (and, hence, of the environmental records), the sample size of the site, a description of the current vegetation, and the publications with the original data.
of methods for pollen extraction and identification used for each of the sites can be found in earlier publications, see Table 2. Here, we only focus on taxa identified to genus level and occurring in the ), and we want to assess how the tree community is changing. We included pollen abundances from all identified genera, including those identified with some uncertainty (e.g., the taxa identified as "c.f."), as this was the best available identification. The sites varied in their proportion sampling coverage (between 0.37 and 0.97; Appendix S1) but sampling coverage was similar across years within sites, indicating that the coverage differences between sites would have little effect on the analyses (in which sites are included as random factors, see Section 2.8).

| Functional traits
To evaluate past shifts in functional composition in forests, we use four tree traits that likely respond differently to underlying environmental changes: (1) WD, which is part of the wood economics spectrum ) and associated with drought and shade tolerance and with species successional stage; (2) leaf area (LA), which is part of the plant size spectrum (Díaz et al., 2015) and associated with light capture and convective cooling; (3) adult height (H) as a strategy to avoid crown fire, enhanced light exposure, and is linked with drought vulnerability (Rowland et al., 2015) and vulnerability to strong wind (Onoda & Anten, 2011); and (4) SM, which enhances establishment success in the dark forest understory (Grimme & Jeffrey, 1965)

| Community-mean traits
We calculated the community-mean traits for each pollen sample, weighted by ln-transformed tree taxon abundance. We lntransformed abundance of the taxa to reduce the weight of taxa with high pollen production, and we ln-transform the traits to reduce extreme values. Our previous study showed that the relationship between community-mean traits of forests and elevation was accurately captured by tropical pollen assemblages, both at the species level and genus level (van der Sande et al., 2020). This coherence creates opportunities for assessing the effect of environmental gradients (either spatial or temporal) on community-mean traits from pollen assemblages. It is possible that the absolute es- On average per site, from the total tree pollen counts we had data on WD for 73%, on H for 80% on LA for 84% and on SM for 68%. We refer to community-weighted mean traits as "community" traits throughout the manuscript.

| Climate
To understand temporal changes in community traits, we included three climatic proxies, all derived from independent data sources.

| Fire activity
We used charcoal concentration as measure of fire activity, as high charcoal concentration could indicate the occurrence of spatially larger fires, multiple fires within the timespan of the sample, and/ or severe fires with high amounts of biomass consumed Whitlock & Larsen, 2001). Hence, we use the term "fire activity" instead of fire intensity, severity or frequency, as we do not know which of these is the cause of high charcoal abundance.
In many wet tropical forest ecosystems, fires do not naturally occur at high severity or frequency, and are therefore often considered a human disturbance (Gosling et al., 2022). However, discussion remains on the origin of fires in the tropics, and we therefore refer to them as "fire-related disturbance" in general.
Charcoal concentrations were quantified from the same records used for pollen analyses, although often at a higher temporal resolution. In case charcoal data were available at different sample depths than the pollen analyses, we used the charcoal sample that was used at the time equal or directly preceding the pollen sample, to assess short-term forest responses to fire.
Methods of charcoal counting (e.g., charcoal particle-size range) differed between records and values cannot be compared directly. We therefore standardized the charcoal data in two ways. Second, we applied z-transformations within records by using the scale function in R, as also done for the other predictor variables.
Results of these two approaches are very similar, and we therefore present the latter approach to apply a consistent scaling method for all our predictor variables. At the sites of Lake Pata and Lake Kumpak a , although charcoal was counted, almost none was found.
For Lake Pata, this is consistent with an absence of human occupation (Nascimento et al., 2019).

| General disturbance
As a proxy of general disturbances, for example caused by landuse conversion by humans, landslides, or hurricanes, we used the proportion of Cecropia pollen relative to the pollen sum (including pollen that were not identified, identified to family level, and different life forms such as grasses). Although the average Cecropia abundance differs between sites, the genus is present in all sites and likely increases with disturbances in low-and high-elevation sites.
Therefore, the within-site increases in the proportion of Cecropia pollen indicates increasing frequency, intensity, and/or extent of disturbances.
Cecropia is also included as one of the genera in the calculation of community-mean traits. Therefore, the effect of Cecropia abundance on community-mean traits may not be fully independent. To assess whether this influenced our results, we also ran the analyses described below using community-mean traits without Cecropia.
The results are similar (Appendix S9) and we therefore present the community-mean traits based on all tree genera including Cecropia.

| Analyses
To assess the effect of reconstructed climate (precipitation, tem- Nonsignificant interaction effects were removed from the model. Hence, the final models included the main effects of the six fixed predictor variables (precipitation, temperature, El Niño, fire activity, general vegetation disturbance, and elevation) and a maximum of five interaction terms. We ran linear mixed models using the nlme package in R (Pinheiro & Bates, 2016).
To assess whether this overall model provides similar results to site-specific models, we also ran one generalized least square regression model per site per trait, including precipitation, temperature, El Niño frequency, fire activity, and general vegetation disturbance as fixed predictors. Furthermore, to account for potential temporal autocorrelation in the data, we included a temporal autocorrelation structure using the corCAR1 function, which is designed to deal with a continuous time covariate (Pinheiro & Bates, 2016). We standardized all response and predictor variables (by subtracting the mean and dividing by the standard deviation) prior to analyses so that we could compare effect sizes between models and sites. We ran generalized linear models using the gls function and the corCAR1 function from the nmle package in R (Pinheiro & Bates, 2016). For all analyses, R version 3.6.1 was used (R Core Team, 2019). Models based on a similar time window between sites of 2100 y gave similar results to using the maximum available time window per site (Appendix S5).

| RE SULTS
We assessed how the community-mean tree traits (weighted by pollen abundance) changed with past climate (precipitation, temperature, and El Niño frequency) and disturbances (fire activity and general vegetation disturbances), and how these relationships

| Adult tree height
Precipitation increased community tree height, whereas temperature decreased height and El Niño frequency tended to decrease height ( Figure 2a). The temperature effect, however, was positive at high elevations and negative at low elevations (Figure 3b). The effect of fire activity was negative at high elevations and tended to be positive at low elevations (Figure 3d), and the effect of disturbance was positive only at high elevations (Figure 3e).

| Leaf area
Precipitation, temperature, and general disturbance increased community LA (Figure 2b). The El Niño frequency decreased LA at high elevations and increased LA at low elevations (Figure 3h), and the disturbance effect is especially strong at low elevations (Figure 3j).

| Seed mass
General disturbance decreased community SM (Figure 2c). The negative effect of general disturbance (Figure 3o) on SM was strongest at low elevation. Fire activity decreased SM at high elevation but increased SM at low elevation ( Figure 3n).

| Wood density
Temperature and general disturbance decreased WD (Figure 2d).
This negative effect of general disturbance on WD was especially strong at low elevation (Figure 3t).

| DISCUSS ION
We assessed how community tree trait composition across eight lowland and highland Neotropical forest sites relates to past climate change and disturbances. All four traits responded to these environmental variables, but the direction of effects often varied depending on elevation. Overall, the most important predictor was our measure for general vegetation disturbances, which increased tree height and LA and reduced SM and WD across the lowland and highland sites. Temperature was the most important climatic predictor, leading to increased LA and tree height at high elevations, and to decreased WD and tree height at low elevations. Tree height and LA respond strongly to these drivers, whereas SM and WD respond more weakly. Here, we will first discuss the overall responses of tropical lowland and highland forests and then highlight some of the main site-specific differences.

| More favorable climatic growing conditions lead to more acquisitive trait values
We expected that, under benign conditions, there would be an increase in the abundance of taxa with acquisitive trait values that are associated with fast growth. Such benign conditions would occur with increased water availability (i.e., increased precipitation and decreased El Niño frequency), increased temperatures at high elevation because they prolong the growing season and reduce freezing-induced cavitation risk, and decreased temperatures at low elevation because they reduce evaporation and, hence, reduce atmospheric drought stress (Table 1). We indeed found that community-mean adult height and LA (Figure 2a,b) increased with precipitation, indicating that increased water availability allows for higher dominance of taller taxa with longer hydraulic path lengths (Bennett et al., 2015) and larger leaves with higher transpirational demand (Wright et al., 2017). Interestingly, WD was not affected by changes in precipitation, as we had earlier found for one of the sites (van der Sande et al., 2019). Perhaps increasing precipitation leads on the one hand to higher water availability and an advantage of drought-vulnerable, fast-growing species with soft wood, and on Community responses to temperature were stronger than responses to precipitation and El Niño frequency, as three of the four traits were significantly affected by temperature, and temperature had the largest effect sizes (Figure 2). We found that, at high and cold elevation, adult height increased during warmer times (i.e., with high temperature; Figure 3b, blue line). Tree height generally decreases from low to high elevation (Mao et al., 2018;Simard et al., 2011; van der Sande et al., 2020), and tree size is one of the most important traits predicting species' position along elevation gradients (Maharjan et al., 2021). Hence, increased temperature probably results in less growth limitation and a longer growing season (Mao et al., 2018), which allows for an upslope range expansion of taller species. Such an upward movement of lowland forest species has been observed in Holocene pollen records (Flantua et al., 2014;González-Carranza et al., 2012) and also during current global warming (Fadrique et al., 2018;Feeley, Silman, et al., 2011;Woodall et al., 2013). However, at low and warm elevations, adult height decreased with further warming, probably because stronger vapor pressure deficits and associated atmospheric drought lead to increased mortality of large trees (Figure 3b, red line) (Aleixo et al., 2019;, which have longer hydraulic pathlengths and are more drought sensitive.

F I G U R E 3
Visualizations of all relationships between the five environmental variables in columns (precipitation, temperature, El Niño frequency, fire activity, and general disturbance) and the four community-mean functional traits in rows [(a-e) adult height, (f-j) leaf area, (k-o) seed mass, and (p-t) wood density]. The relationships that showed a significant interaction with elevation (see Figure 2) are here visualized as two lines: low elevation (red, predicted for an elevation of 100 m asl) and high elevation (blue, predicted for an elevation of 3000 m asl). Significant relationships that did not show an interaction with elevation are visualized as one gray line. Nonsignificant relationships are shown without line. The axes represent the scaled axes (scaling was done per site, by subtracting the site's mean and dividing by the site's standard deviation). In both lowland and highland forests, LA increased with temperature (Figures 2b and 3g), which may be determined by different mechanisms. In the lowlands, increased temperatures may increase tree mortality, resulting in more canopy gaps and more establishment opportunities of large-leaved pioneer species that can capture more light (Aleixo et al., 2019). In the highlands, however, higher temperatures reduce the risk of freezing, which may lead to a shift from species with small, dense, and coriaceous frost-tolerant leaves (Olson et al., 2018;Yang et al., 2020) to fast-growing species with large leaves.
High WD is associated with narrow vessels that increase tolerance to drought-induced and freezing-induced cavitation (Olson et al., 2018;Pittermann & Sperry, 2006;Yang et al., 2020), and with high resistance to wind (Putz et al., 1983). Dense-wooded species are therefore often found in dry areas and at higher elevations (Bruelheide et al., 2018;Olson et al., 2018). At low elevations, we expected that increasing temperature would lead to higher transpirational demand and atmospheric drought stress, and therefore to a higher abundance of drought-tolerant, dense-wooded species.
We found, however, a negative effect of temperature on WD, both in lowlands and highlands (Figures 2d and 3q). This does not agree with findings from , who show that acquisitive species were more sensitive to negative climate effects on tree growth. This is possibly explained by higher mortality rates and gap formation, which would favor the establishment of fast-growing, soft-wooded, light-demanding species (Aleixo et al., 2019), or by an increased abundance of taxa with high metabolic rates and fast life cycles.
Hence, across these eight forest sites, wetter climates (i.e., higher precipitation, lower El Niño frequency) increased the abundance of more drought-sensitive taxa with long hydraulic path lengths (Koch et al., 2004;Stovall et al., 2019) and large leaves that can capture more light but rely on transpirational cooling (Wright et al., 2017).
However, climatic warming has opposite effects along the elevation gradient, providing at high elevation more favorable conditions for taller taxa and at low elevations more favorable conditions for shorter taxa.

| Fires lead to more stress-adapted species, while general disturbances lead to more acquisitive species
Disturbances can either lead to more open conditions, favoring the establishment of fast-growing, light-demanding species that are tall and have soft wood and large leaves, or lead to more stressful conditions (due to fire damage, topsoil removal by erosion, and supra- optimal temperatures in open conditions), favoring slow-growing, short species with dense wood and small leaves. We found that adult height and SM decreased with fire activity at high elevation (suggesting more stressful conditions), but not at low elevation (Figure 3d,k).
Small trees with a fast life cycle and small-seeded species that can quickly colonize burned areas have a relative advantage at high fire activity. Although fires have been frequent and widespread in lowland tropical forests (Gosling et al., 2022) we found, surprisingly, no clear effect of fire on lowland trait composition. We can think of five possible explanations. First, the long-term fire history resulted in adapted ecosystems that no longer strongly respond to fire in terms of their species and trait composition. This would be especially the case in the drier, eastern part of the Amazon where fires have been more severe and frequent (Gosling et al., 2022). However, it is not very likely for the wet (western) tropical forests where fires have been less frequent and severe despite human intervention (Gosling et al., 2022) and fire-resistance traits such as thick bark are rare (Paine et al., 2010;Poorter et al., 2014;Rosell, 2016

| Adult height and LA are most responsive to climate and disturbances
Adult height and LA are generally more responsive than WD and SM.
Various reasons may explain these differences. The height of trees is strongly limited by water availability (Koch et al., 2004), and tree height is one of the best predictors of drought-induced tree mortality Stovall et al., 2019), explaining its strong response to climate and to the indirect climatic changes caused by F I G U R E 4 The eight sites [(a-h) same order as Figure 1 and Table 2) ordered from highest to lowest elevation, with a picture of the lake and the current vegetation (first column), the temporal changes in community-mean adult height (in meters) (second column), and the standardized regression coefficients of precipitation (Prec; dark blue), temperature (Temp; light blue), El Niño frequency (EN; yellow), fire activity (Fire; dark red) and general disturbance (Gen; gray) on community-mean adult height (third column). We here only show results for height for illustrative purposes.  disturbances. The strong responses of LA are striking, given that LA is less strongly conserved within genera (Appendix S2) and, hence, more uncertainty arises when calculating community-mean traits based on genus-average data. WD, on the other hand, is associated with resistance to multiple environmental stressors, such as shade (Markesteijn et al., 2011), drought (Poorter & Markesteijn, 2008), and pathogens (Augspurger & Kelly, 2010). These multiple drivers of WD may have opposite effects and result in less clear temporal changes. SM is one of the main dimensions of global variation in plant traits (Díaz et al., 2015) as it influences reproduction, germination, and seedling survival, but it has no direct impact on adult trees, which most strongly determine the community trait composition. Our study shows that the long-term effects of climate change and disturbances are best captured by changes in traits related to plant and leaf size, rather than by WD and SM.

| Local disturbance histories determine local responses to climate and disturbances
Overall, 34% of the 156 relationships tested at the site level were significant (Appendix S7), which is similar to the cross-site analyses (35% significant). The strongest site-level drivers were temperature (significant for 47% of the relationships) and general disturbance (significant for 53%), and the directions of these site-level effects were generally similar to the cross-site effects; temperature and disturbance increased community-mean LA and decreased community-mean SM and WD, and temperature also decreased height (Figure 4; Appendices S7 and S8). This is in line with recent increases in abundances of lowland tree taxa across highland forests due to global warming (Fadrique et al., 2018). We found, however, that effects of El Niño frequency and fire activity varied strongly between sites, indicating that sitespecific responses to some drivers are context dependent. Nearly all sites had some effects of El Niño and fire, but on different traits and in different directions (Appendix S8). This indicates that El Niño and fire events do impact the forest, but that effects can be opposite and local differences in, for example, soil conditions and seasonality may determine local-scale responses to these drivers. Specifically, the sites that show stronger El Niño effects (e.g., Cocha, Pata) are the sites in which El Niño probably causes strongest reduction in precipitation (Appendix S3). Our index for El Niño frequency is derived from one record for all sites, but the climatic effects of El Niño events can vary strongly across regions (Marengo et al., 2021). Furthermore, locally the degree of human influence during the last 10,000 years can vary strongly (Flantua et al., 2016;Piperno et al., 2015), and sites may respond at different rates, hence with different lag times, to fire. The lack of effects of El Niño and fire does not seem to be determined by weaker variation in these indicators (the CV of El Niño and fire was mostly not lower than the CV of other drivers). Hence, although temperature is a commonly strong driver of change across forests, certain forest ecosystems simply seem less responsive and more resilient to other types of changes (e.g., climate and disturbance events). Future studies are needed to assess how different disturbance regimes in the past influence forest resilience to future disturbances.

| CON CLUS IONS
During recent decades, forests have shifted in species and trait composition both in lowland (Enquist & Enquist, 2011;Fauset et al., 2012;van der Sande et al., 2016) and highland (Feeley, Silman, et al., 2011)  This study is the first to assess millennial-scale effects of environmental change on functional trait composition across Amazonian and Andean forests. Although we cannot predict absolute changes in functional composition based on pollen composition, future increases in temperature may lead to relative increases in soft-wooded species with large leaves, and relative increase in forest height in the highlands and relative decreases in forest height in the lowlands.
These shifts in functional composition also imply shifts in biodiversity and ecosystem functioning, with potentially less carbon storage and sequestration in the lowlands but potentially more carbon sequestration in the highlands.