1. Recent studies have documented shifts in habitat associations of single tropical tree species from one life stage to the next. However, the community-level consequences of such shifts have not been investigated, and it is not clear whether they would amplify, neutralize or completely alter habitat structuring during the transitions to the adult community.
2. We compared habitat-driven species assemblages at three life stages (i.e. recruitment, juvenile and reproductive stages) and six censuses for tree and shrub species in a fully censused 50-ha plot of Panamanian lowland forest. Habitat types were determined using multivariate regression trees that group areas with similar species composition (i.e. species assemblages) according to their topographical characteristics.
3. Three topographical variables (a topographical wetness index, slope and elevation) were major determinants of species assemblages. When analysing individuals of all life stages together, we found a distinct and temporally consistent structuring of the plot into four dominant habitat types (low and high plateaus, slope and swamp) which was consistent with previous classifications. Basically, the same habitat structuring emerged for the juvenile communities of individual censuses. However, recruits showed a weak and temporally inconsistent habitat structuring.
4. A notable homogenization in species assemblages occurred during the transition from juvenile to reproductive, through both a reduction in the number of species assemblages (in 3 censuses, one large reproductive assemblage covered 93% of the plot, and in others, an additional slope habitat emerged) and a reduction in the classification error. Overall, habitat structuring became noisier and weaker over the 25 years of the study.
5.Synthesis. Our results suggest that mortality processes during the transition from recruits to juveniles must enhance the signal of habitat structuring. However, during the transition to the reproductive stage, species may have lost the advantage of being in the habitat with which they had become associated, or the quality of habitat changed during their life span because of larger climatic changes. The homogeneous assemblages of the reproductive stage could be interpreted as support for neutral theories, but further research is required to unravel the mechanisms behind these intriguing observations.
Niche differentiation has been put forth as an explanation for the maintenance of local diversity in multispecies communities (Ashton 1969; Tilman 1982). According to niche theory, species can coexist if they perform best under different abiotic conditions. If abiotic conditions, such as soil attributes or topography, are spatially structured, their structure will be reflected in species distributions through species–habitat associations at the individual species level (Whittaker 1956). As a consequence, different species assemblages should form at the community level within habitat types. However, habitat partitioning in many species-rich (plant) communities may not provide separate niches for hundreds of plant species as required by classical coexistence theory (Valencia et al. 2004), because plants depend on and compete for the same few of resources and acquire them in similar ways (Daws et al. 2002; Silvertown 2004; Comita, Condit & Hubbell 2007). One hypothesis to solve this conundrum is the unified neutral theory (Hubbell 2001), which focuses on adult plants and assumes no niche differences. An alternative hypothesis in the niche context is that coexistence is possible through the partitioning of the ‘regeneration niche’ (Grubb 1977; Tilman 1982). During early life stages, trees may experience their environment as more heterogeneous, e.g. in terms of light availability related to canopy gaps (Ricklefs 1977) or other physical factors, such as soil moisture (Daws et al. 2002; Engelbrecht et al. 2007; Comita & Engelbrecht 2009) or nutrients (John et al. 2007).
In addition, previous studies have found that habitat associations of single tropical tree species typically do not form at early life stages (i.e. seed germination and seedling establishment) and that species often show different ecological habitat associations across life stages (Webb & Peart 2000; Paoli, Curran & Zak 2006; Comita, Condit & Hubbell 2007; Lai et al. 2009). Thus, it is not clear whether adult plants conserve a signal of habitat association that they acquired during early life stages or whether shifts in a species’ ecological preferences occur from one life stage to the next. On the community level, the latter may neutralize habitat structuring for adults or generate a habitat structuring different from that of the recruits and juveniles. However, this issue is complicated by the fact that habitat suitability for a given species at a given location may not be static, but could shift in response to long-term variations in climatic conditions or extreme disturbance events such as severe El Niño droughts (Condit, Hubbell & Foster 1995). Thus, when looking from the community level, emerging species assemblages associated with certain habitat types may not only shift with life stage but also in time because of supra-annual variability in climatic conditions.
An important driver of habitat diversification is topography, which is a first-order control on spatial variation of hydrological conditions that affect the spatial distribution of soil moisture (Harms et al. 2001; Daws et al. 2002; Sørensen, Zinko & Seibert 2006) and nutrients (John et al. 2007), which are crucial for plants. Previous studies conducted at large fully censused tropical forest plots have investigated topographical habitat associations at the level of individual species (e.g. Harms et al. 2001; Gunatilleke et al. 2006; Svenning, Normand & Skov 2006; Comita, Condit & Hubbell 2007; Lai et al. 2009). However, because of sample size limitations, these studies were restricted to the most common species in the community. Here we take a different approach to look at habitat associations. Instead of focusing on individual species–habitat associations, we focus on detection of habitat-driven species assemblages and how they might change with age or size class and through time. The advantage of this approach over individual species associations is that it focuses on the broader picture and allows inclusion of data from all species and not only from more abundant species. This may be particularly important for understanding habitat partitioning in diverse communities, because rare species, which are typically excluded from individual species–habitat association tests, may be rare because they are habitat specialists.
In this study, we assess habitat associations at the community level using multivariate regression tree analysis that groups areas with similar species composition according to their topographical characteristics, thereby defining different habitat types (De’ath 2002; Legendre et al. 2009). We ask whether emerging habitat types and topographical determinants were consistent in time (i.e. census) and with life stage (i.e. recruitment, juvenile and reproductive stages). For this purpose, we use data on tree and shrub species from six censuses conducted in the Barro Colorado Island (BCI) 50-ha Forest Dynamics Plot (FDP), Panama (Fig. 1). Previous species-level studies of seedling and tree stages in the BCI plot found that many species were positively associated with one or more habitats and appeared to exhibit different ecological habitat preferences at the smaller and larger stages (Harms et al. 2001; Comita, Condit & Hubbell 2007). Here we ask whether and how such developmental changes in habitat associations translate into the emergence of local species assemblages at different life stages, and which environmental variables structure the habitats that host different species assemblages. In addition, we use the data from repeated censuses of the BCI plot to assess the consistency of these patterns over time in the light of interannual climatic variation.
Given that topographical variation at BCI is not very large, we expect a noisy answer, but the questions of whether different development stages produce the same habitat structuring and whether that is temporally consistent are intriguing. For example, under the regeneration niche hypothesis, we would expect recruitment and juvenile stages to show strong habitat structuring. If neutral theory were true, the adult community should basically form one large species assemblage. Our special interest is to identify the critical life stage transition where the emergent species assemblages change and to explore the role of long-term changes in precipitation regime or disturbances, such as El Niño Southern Oscillation (ENSO) events.
Materials and methods
The forest site studied is the 50-ha FDP at BCI, Panama (9°9′N, 79°51′W), which receives a mean of 2600 mm of rain per year, most of which falls during the 8-month wet season from May to December (Windsor 1990). Particularly, severe El Niño droughts occurred during the 1982–83 and the 1997–98 dry seasons (Wright 2005). Detailed descriptions of the climate, geology, flora and fauna of BCI can be found in the study by Croat (1978), Leigh, Rand & Windsor (1982), Leigh (1999) and Losos & Leigh (2004). The plot consists of mainly old-growth lowland moist forest, with the exception of a 2-ha area of secondary forest along the north-eastern edge of the plot (Fig. 1; young). Elevation on the FDP ranges from 120 to 155 m a.s.l. (Hubbell & Foster 1983). Despite being mostly flat, the edges of the plot include slopes with up to 20° inclination (Fig. 1; slope). Slopes are wetter than plateaus (Daws et al. 2002; Leigh et al. 2004) and experience a shorter duration of drought during the annual 4-month dry season (Becker et al. 1988; Daws et al. 2002). The high plateau (Fig. 1) is the driest area, having lower dry season soil water availability than slopes or the low plateau (Comita, Condit & Hubbell 2007). In addition to the topographically defined habitat types, a 1.5-ha seasonal swamp area is inundated during the 8-month wet season and streamside habitats comprise steeply sloped areas adjacent to seasonal streams that tend to contain water into the dry season (Fig. 1; Hubbell & Foster 1986; Daws et al. 2002).
The BCI FDP plot was divided into 1250 20 × 20 m quadrats, and abundance of each species was calculated in each quadrat. For each census, we classified a living plant as recruit, juvenile or reproductive and conducted separate analyses for these three life stages. Recruits of census t were defined as plants which were absent in census t-1 (i.e. d.b.h. < 1 cm) but present and alive in the following census t (i.e. d.b.h. ≥ 1 cm). We thus obtained recruit data for five censuses, but not for the first census. Classification of all nonrecruits into juveniles and reproductive individuals was done for each census based on the actual size of the individual and thresholds for reproductive size that were provided along with the census data (Hubbell, Condit & Foster 2005). Note that a recruit will typically be classified in the next census as juvenile, but may also pass directly into the reproductive class. Our classification was done with respect to a functional population dynamics point of view that takes different growth forms into account. For example, shrubs reach the reproductive stage at much smaller sizes than canopy trees. However, we also used an alternative classification based purely on size that included: small: 1–10 cm, intermediate: 10–20 cm and large: >20 cm individuals.
To describe the species assemblages of the entire community, we conducted an analysis for each census based on all living plants recorded in the census (i.e. we pooled the recruits, juveniles and reproductive classes). While the data from individual censuses allow us to explore temporal differences in habitat types (and associated species assemblages) for a given life stage, we also conducted analyses where we pooled the data for each life stage from all censuses. This allowed us to reveal average long-term habitat types to be compared to that of individual censuses. Because recruitment data from different censuses were nonoverlapping, we simply joined the data from five individual censuses to obtain the overall recruits data set. However, for pooling data for juveniles and reproductive individuals, we eliminated repeated counts of the same individual stemming from different censuses. Additionally, we created a data set of all individuals that were alive during the first census and still alive during the last census (in the following called ‘survivors’). We assumed that these trees, which survived all censuses, were located in favourable habitats and used these survivors as point of reference for comparison with recruits, juveniles and reproductive individuals. Our analysis also included rare species, and it was not clear a priori if this influenced our analyses. We therefore repeated the analyses, excluding rare species with <13 individuals in the given class. We selected a threshold as low as 13 individuals to not remove more than c. 3% of all recruits at any given census. Sample sizes for all analyses are reported in Tables S1 and S2 in Supporting Information.
Multivariate Regression Tree Analysis
Multivariate regression trees (MRT; De’ath 2002; Larsen & Speckman 2004) were used to group the 20 × 20 m quadrats with similar species composition according to their topographical characteristics. In this way, we can discriminate habitat types that are occupied by different species assemblages. In general, MRT can be used to explore relationships between multispecies data and environmental characteristics (De’ath 2002). Multivariate regression trees analysis is based on a recursive algorithm. In our case, the root node consists of all quadrats. The algorithm systematically searches the environmental variable (and an associated threshold value) that partitions the parent node (i.e. all quadrats within the plot) into two ‘child’ nodes (i.e. subareas of the plot) that are internally most similar in terms of species composition. This procedure is then recursively repeated for each node until a certain stopping condition is satisfied (Larsen & Speckman 2004). Once an environmental variable is chosen for a given node, it is not excluded from playing a role in defining subsequent nodes. A given variable may therefore define several nodes. We used the Bray–Curtis dissimilarity index:
to compare the species abundances between two quadrats j and k, where xij and xik are the abundances of the ith species in quadrats j and k, respectively, and s is the total number of species. The Bray–Curtis measure of dissimilarity is generally regarded as a good measure of ecological distance when dealing with species abundance, because it allows for nonlinear responses to environmental gradients (Faith, Minchin & Belbin 1987; De’ath 2002).
Regression trees are summarized by their size (i.e. resulting number of nodes) and overall fit. The final tree selection is done by detecting the tree size that has the lowest cross-validated relative error (CVRE). This error measure varies from zero for a perfect predictor to close to one for a poor predictor. The tree fit could also be defined using relative error, as it gets lower as more variables are included into the model. However, the CVRE estimates the predictive accuracy of the resulting classification better (De’ath 2002): after decreasing to its minimum, the CVRE curve slightly increases to a plateau and the predictive power of the model (the number of ‘end branches’) does not increase further. The classification accuracy is then given as being the value that is 1 SE unit higher than the lowest mean value of the CVRE (the 1-SE rule; Breiman et al. 1998). To avoid overfitting the data, we pruned back the fitted trees and selected a tree size that minimized the cross-validated error in all cases. To facilitate comparison of the results between different life stages, we kept a fixed maximal size of five habitat types in all our MRT analyses when possible. Preliminary analyses showed that allowing more habitat types did not improve the fit.
Indicator species analysis (Dufrêne & Legendre 1997) was conducted to identify the species that were statistically significant indicators of the habitat types that are defined by the nodes of the tree. An indicator species index is defined as the product of relative abundance and relative frequency of occurrence of the species within a habitat type. If there are no occurrences of the species within a habitat type, the index takes the value of zero and it increases to a maximal value of 1 if the species occurs at all 20 × 20 m quadrats within the habitat type but does not occur at any other quadrat. The index can be calculated for each habitat type, and species with high index values for a habitat type are regarded as indicator species.
Six topographical variables were considered in the analyses. Elevation (Elv) values were provided along with the census data at every 5 m within the plot. We derived terrain convexity (Con), slope (S) and aspect (Asp) based on the elevation data with the original 5-m resolution using the Spatial Analyst Tools in ArcGIS 9.3 (ESRI, Redlands, CA, USA: Environmental Systems Research Institute). All variables were then converted to the 20-m resolution of the quadrats by taking the mean of all 5-m cells within a 20-m quadrat. This reduced potential edge effects. Additionally, we used SAGA GIS (http://www.saga-gis.org) to calculate two indices that are commonly used to quantify topographical control on hydrological processes. The topographical wetness index (TWI) represents the ratio of the area upslope of each quadrat to the local slope for that quadrat. Because the amount of water on slopes depends on the size of the catchment area above (Daws et al. 2002), we expect that this index should capture important information on wetness. We calculated TWI using Tarboton’s Deterministic Infinity method (Tarboton 1997; Sørensen, Zinko & Seibert 2006). The second hydrological index is the vertical distance from the channel network (Chn). Ideally, we would include elevation data outside the 50-ha plot for the calculation of TWI and Chn, because the true catchments of edge quadrats may extend beyond the borders of that area. We therefore critically inspected results involving these variables for potential bias.
Because we expected an influence of the biotic conditions on recruits, we also used the census data to calculate biomass Bm, maximal height Hmax, basal area Ba and a light index I for each 20 × 20 m quadrat (see Appendix S1 in Supporting Information for details on calculation of these variables). We related these variables to the recruitment data of current and subsequent censuses. Maps of the environmental variables are shown in Fig. S1.
The pooled data of recruits, juveniles and reproductive individuals and survivors resulted in more than 100 000 individuals in each class (Table 1) with juveniles being the largest class and survivors being the smallest class. Species richness was highest at the reproductive stage (306 species for all censuses together and 289 for the first census) and lowest at the juvenile stage (253 species for all censuses together and 246 for the first census; Table 1; Table S5). In analyses where we eliminated rare species (<13 individuals in a given class), recruits lost 41–52% of the species, juveniles 14–17% and reproductive individuals 20–21% (Table S1b).
Table 1. Results of multivariate regression tree analysis for survivors (individuals alive during all censuses) and pooled data of all censuses of the category of survivors, recruits, juveniles and reproductive individuals in the Barro Colorado Island 50-ha Forest Dynamics Plot, Panama
No. of plants
No. of species
No. of 20 × 20 m cells
No. of indicator species
1: (high plateau): Elv ≥ 150.2, S < 6.5, TWI < 12.4
2: (low + intermediate plateau): Elv < 150.2, S < 6.5, TWI < 12.4
3: (slope) S ≥ 6.5, TWI < 12.4
4: (swamp) TWI ≥ 12.4
1: (high plateau): Elv ≥ 151.6, S < 6.5, TWI < 12.8
2: (intermediate plateau): Elv ≥ 144.1, Elv < 151.6, S < 6.5, TWI < 12.8
3: (low plateau): Elv < 144.1, Elv < 151.6, S < 6.5, TWI < 12.8
4: (slope) S ≥ 6.5, TWI < 12.8
5: (swamp) TWI ≥ 12.8
CV error, cross-validated error. The environmental variables: Elv, elevation; Asp, aspect; S, slope; Con, convexity; TWI, topographical wetness index and Chn, vertical distance from the channel network. The labels in parentheses are the authors’ verbal interpretations with respect to the classification proposed by Harms et al. (2001)
1: (high plateau): Elv ≥ 151, S < 6.5
2: (intermediate plateau): Elv ≥ 132.7, TWI < 12.3, Elv < 151, S < 6.5
3: (low plateau) Elv < 132.7, TWI < 12.3, Elv < 151, S < 6.5
4: (swamp) TWI ≥ 12.3, Elv < 151, S < 6.5
5: (slope) S ≥ 6.5
1: (intermediate plateau) S < 8.2, Elv ≥ 126.3, TWI < 12.4
The BCI plot showed a high dynamic between censuses. Between 56% and 63% of all recruits had entered the juvenile class in the next census, between 20% and 25% had entered the reproductive class and between 17% and 19% had died (Table S3). After 20 years, 35% of the recruits were juveniles, 16% were reproductive and 49% had died (Table S3). Between 82% and 87% of all juveniles remained in the juvenile class in the next census, 3–4% had become reproductive, and 10–15% had died (Table S4). After 25 years, 44% of the juveniles were still juveniles, 11% reproductive and 45% had died (Table S4).
Pooled Data of Different Censuses
The MRT analyses indicated that only three topographical variables (i.e. TWI, slope and elevation) were important drivers of species assemblage in the BCI plot, when data of the different censuses were pooled (Table 1; Fig. S3). The other variables were never selected (Table S5). The TWI appeared consistently in each regression tree with value of TWI = 12.4 or TWI = 12.8 separating the swamp area from the rest of the plot for all life stages and survivors (Table 1). For reproductive individuals, four cells of the swamp were separated as high-elevation swamp (Elv > 141.7; Table 1). Because the habitat class defined by the TWI threshold approximates the swamp area (which was previously identified using field observations as reported in the study by Harms et al. 2001), we can assume that the approximation in TWI index did not bias our results. For survivors, recruits and juveniles, the slope variable (S) consistently separated a habitat with steep slopes from the rest of the area at a value of S = 6.5°. For reproductive individuals, the area designated as slope was reduced to steeper areas with a threshold of S = 8.2° (Table 1). Finally, for survivors, recruits and juveniles, elevation (Elv) divided the plateau habitat into low and high plateaus at breakpoint of approximately Elv = 150 m (Table 1). For recruits, elevation also distinguished an intermediate plateau between 144 and 152 m from the low plateau (Fig. 2). The same was true for juveniles, but here the low plateau was only limited to the eastern part of the plot with Elv < 133 m. For reproductive individuals, the low and high plateaus were not separated (Table 1; Fig. 2), but a small area with elevation <126 m appeared at the south-eastern corner of the plot (Fig. 2).
The CVRE was lowest for the reproductive category (0.79; Table 1), intermediate for survivors and juveniles (0.83) and largest for recruits (0.92). The number of habitat types produced by MRT analyses was five in all categories except in the survivors category, where low and intermediate plateaus were not differentiated. Interestingly, the two major classes of reproductive individuals occupied 95% of the area of the plot area (76% intermediate plateau and 19% slope).
The habitat classifications for individual censuses differed substantially from the classification where recruits of all censuses were pooled (Fig. 2; Table S5). Overall, the number of habitat types formed by MRT analysis for recruits varied from two in 1995, 2000, and 2005 to four in 1990 (Fig. 2), and the fit was poor with CVRE ranging between 0.97 in 1985 and 0.999 in 2005. The classification of 1985 separated only the swamp area and a small low-elevation area in the south-easterner corner of the plot, but the rest yielded one large assemblage covering 94% of the plot (Fig. 2), and it showed the lowest error. The classification for the 1990 census was similar to the classification of the pooled data, but did not differentiate between low and intermediate plateau. The subsequent censuses maintained the slope habitat of the 1990 census, but yielded only two habitat types (plateau and slope). Thus, species assemblages of individual recruitment generations showed high noise and relatively little habitat structuring.
The biotic variables biomass, maximal height, basal area and light index, calculated from the data of the previous census and of the same year, did not improve the classification and did not enter the regression tree for any analysis.
Multivariate regression tree analysis yielded five habitat types, when juveniles of all censuses were pooled together (Table 1; Fig. 2). It produced basically the same habitat types as recruits of all censuses, driven by the same environmental variables and similar breakpoints, but for juveniles, the low plateaus were restricted to the eastern part of the plot with elevation thresholds of 132.7 m (Fig. 2; Table 1). The CVRE of this regression tree was 0.83, smaller than that of recruits (0.922).
The number of habitat types formed by MRT analysis for individual censuses varied from 4 to 5 with CVRE ranging from 0.836 to 0.873 (Table S5; Fig. 3). In contrast to recruits, the juveniles showed consistent habitat classifications across censuses that agreed well with that of the pooled data (Fig. 2). Small differences occurred in the 1981–83 census, which did not show separation of low and intermediate plateau habitat (Fig. 2; Table S5). Overall, juveniles showed relatively clear and temporally consistent species assemblages at the different topographically defined habitats.
For reproductive individuals, MRT analysis of the pooled data revealed a habitat structure with relatively low error (CVRE = 0.79) that differed substantially from that of recruits and juveniles (Fig. 2), both in break points (except for TWI) and environmental variables (Table 1). The swamp area with TWI > 12.4 was maintained, and the slope habitat was somewhat reduced (242 quadrats compared to 344 and 348 for juveniles and recruits, respectively). However, the intermediate and high plateaus were joined into one large habitat type (occupying 76% of the plot), and low plateau with elevation <126 m was reduced to the lower right corner of the plot (Table 1; Fig. 2).
The number of habitat types for individual censuses revealed by MRT analysis varied from four in 1985, 1995, 2000 and 2005 to six in 1983 (Table S5), and CVRE ranged between 0.776 for the first census and 0.857 for the last census (Table S1). However, most of these habitat types were limited to a few quadrats: more than 94% of the plot was made up by one habitat type (in 1985, 1990 and 1995) or two habitat types (1981–83, 2000, 2005) (Fig. 2). The 1981–83, 2000 and 2005 classifications were similar to those obtained when all censuses were pooled, but in the 1985, 1990 and 1995 censuses, slope habitat was not differentiated (in 1985 and 1995 censuses, the slope variable appeared in the classifications but was removed after pruning), producing one large assemblage occupying more than 94% of the plot. Thus, reproductive individuals showed the lowest noise and the spatially most homogeneous species assemblages among all life stages, and slope habitat disappeared in 1985 and re-appeared in 2000.
Survivors and Analysis of the Entire Community
Multivariate regression trees analyses for plants that survived all censuses produced four habitat types with CVRE = 0.827 (Table 1), which was comparable to the types emerged from the pooled data sets of recruits and juveniles, but low and intermediate plateaus were not differentiated (Fig. 2). Fifty-four per cent (60 113 individuals) of survivors were juveniles and the remaining were reproductive individuals.
Application of MRT analyses of all living plants, taken separately for individual censuses, revealed five habitat types (except 2000 with four habitat types) with CVRE values ranging between 0.772 in the first census and 0.856 in the last census (Tables S1a and S5). The classifications were highly consistent among censuses and coincided well with the classification for plants which survived from first to last census. Only the first three censuses included an additional habitat type surrounding the swamp.
Approximately 92% of all individuals were small (i.e. <10 cm), 5% intermediate (i.e. 10–20 cm) and 3% large (i.e. >20 cm; Table S2). Unsurprisingly, the classification of small trees was the same as that of all trees (cf. Fig. 2 and Fig. S2). Probably due to the low number of intermediate and large trees (<15 000; Table S2), the MRT analysis yielded very high classification errors for intermediate and large trees (>0.95). The classifications for small trees and juveniles were very similar and temporally consistent, but the low plateau of juveniles was somewhat reduced (Fig. S2, Table S6). The classification for intermediate-sized trees was similar to that of small trees, but the swamp habitat disappeared and instead a high-elevation habitat (Elv > 157.2), never present in the functional classification, appeared (Fig. S2, Table S6). This area has been cleared c. 100 years ago (Harms et al. 2001) and is occupied by young forest with a notably high density of the midstorey species Gustavia superba (with most individuals >10 cm). Large trees differentiated in the first four censuses only into plateau and slope habitat (and the young forest) and in the last two censuses additionally into a high plateau.
Figure 4 shows the number of indicator species for each habitat type for the pooled data, Table S7 lists the most important indicator species and a complete listing of all indicator species and indicator values can be found in Table S8.
For the pooled data sets, the indicator values reached up to 0.260, 0.454, 0.412 and 0.438 for recruit, juvenile, reproductive and survivors, respectively (Table S7). Slope habitat showed 17 indicator species for recruits (the most important ones being Hirtella triandra, Beilschmiedia pendula and Drypetes standleyi with an index >0.25), and 25 indicator species at the juvenile stage (with Ocotea whitei, Poulsenia armata, Beilschmiedia pendula and Unonopsis pittieri with an index >0.35) Ten slope indicator species were also indicator species for recruits. The three species Chrysochlamys eclipse, Ocotea whitei and Trophis caucana, which were indentified as indicator species for reproductive individuals, were also indicator species for juveniles and two of them also for recruits (Table S7).
Of 51 indicator species of recruits, 24 were found also in the juvenile stage, but only seven species were found both in juvenile and in reproductive stages with Bactris major, Elaeis oleifera and Cassipourea elliptica having the highest indicator value (>0.25; Table S7). Of the 34 indicator species of juveniles, 23 were also indicator species of recruits.
Low plateau was the largest habitat type, but was split into western and eastern low plateaus for juveniles, and merged with the high plateau for reproductive individuals (Fig. 2). For juveniles, 20 and 39 indicator species were identified for western and eastern low plateaus, respectively. In recruits, the western low plateau was divided into two plateaus and the eastern low plateau was merged with one of these plateaus. Six indicator species were identified for the low plateau that contained the eastern low plateau and also surrounded the swamp in the middle of the plot. Nineteen species were identified for the remaining western low plateau region. The high plateau comprised five indicator species for recruits, 13 for juveniles, and the pooled plateau habitat of reproductive individuals yielded three indicator species of which one species (Desmopsis panamensis) was also an indicator for the high plateau region for juveniles.
In this study, we adopted a community perspective for revealing species assemblages that emerged in response to local habitat differences rather than analysing the habitat association of individual species, as has been done in most previous studies (e.g. Harms et al. 2001; Gunatilleke et al. 2006; Svenning, Normand & Skov 2006; Comita, Condit & Hubbell 2007; Lai et al. 2009). This allowed us to concentrate on emergent, higher-level structures that consider the entire community, including rare species. Using multivariate regression tree analysis, we assessed whether the BCI plot exhibited topographical habitat types (defined by its associated species assemblages) that were consistent across life stages and among censuses. When analysing the entire communities (i.e. individuals of all life stages together), we found a distinct and temporally consistent structuring of the plot into four dominant habitat types (swamp, slope, low plateau and high plateau). Similar habitat structuring emerged for the juvenile communities of individual censuses and the 1990 recruit census. However, recruits showed a weak and temporally inconsistent habitat structuring and yielded models with low predictive accuracy. Thus, recruit species assemblages were statistically detectable but not very much pronounced.
Surprisingly, the distinct species assemblages observed for juveniles disappeared almost completely for reproductive individuals. Instead, for half of the censuses reproductive individuals formed one large assemblage that covered 93% of the plot, and in the other censuses, a second major habitat type (occupying <30% of the plot) was differentiated at the slopes. Even more surprisingly, this spatial homogenization was accompanied by a homogenization in the emerging species assemblages as indicated by a reduction in the classification error. In general, the classification error was large for recruits and declined with progressing life stage, but also increased for all life stages in time.
Comparison with Previous Habitat Classifications
Interestingly, we found that the habitat types that emerged for all life stages, except reproductive individuals, agreed well with the topographical classification proposed by Harms et al. (2001) that was based on biological arguments and a priori assumptions, but not statistical analysis (cf. Figs 1 and 2). Overall, we confirmed that the BCI plot can be divided basically into four major habitat types ‘low plateau’, ‘high plateau’, ‘slope habitat’ and ‘swamp’. Four of the habitat types defined by Harms et al. (2001) can be represented as a tree of the same structure as the regression tree for recruits (Table 1): the swamp vs. no swamp criterion coincides more or less with the TWI breakpoint, the elevation breakpoint of 152 m in the study by Harms et al. (2001) is very similar to the 150 m breakpoint in our analysis, and the slope breakpoint is similar as well (7 vs. 5.9). However, the swamp area is smaller in the MRT analysis, and the stream and mixed habitats around the high plateau designated by Harms et al. (2001) were not differentiated in our analysis. In some cases, the low plateau was split in our analysis into intermediate and low plateau (Fig. 2).
Separation of swamp is not surprising because swamps are often floristically distinct from the surrounding vegetation (Harms et al. 2001). The separation of slope habitats from plateau, especially for juveniles, can be explained by hydrological conditions at BCI, where slopes are wetter than plateaus (Daws et al. 2002; Leigh et al. 2004) and experienced a shorter drought during the dry season (Daws et al. 2002). This results in low survival of seedlings of slope specialist species in plateaus (Comita & Engelbrecht 2009), thereby creating a distinct juvenile assemblage on the slopes.
Habitat Structuring of Recruits and Juveniles
The MTR classifications for recruits did not exhibit the strong emergent species assemblages that we observed for juveniles and showed high classification errors, indicating that the species assemblages formed by recruits were statistically detectable but not especially homogeneous in terms of species composition and abundance. Emerging species assemblages changed over time with a particular sequence, starting in 1985 with one uniform assemblage (except the swamp and a small low plateau area); in 1990, a high number of recruits emerged and high plateau and slope were differentiated, and after 1990, when recruit numbers were much lower, only slope and nonslope remained. The differentiated recruit assemblages in 1990 may be due to a lag in the effect of the strong 1982–83 El Niño. Seedlings of drought-sensitive species would have likely had higher mortality in the drier habitats, especially the high plateau, relative to the wetter slope habitat. Interestingly, the appearance of the slope habitat for recruits in 1990 was paralleled by the appearance of slope habitat for the reproductive class 10 years later. Ten per cent of the 1990 recruits were reproductive individuals in 2000. For recruits, we additionally used the variables biomass, height, basal area and light index to account for potential regeneration niches, but they did not improve the classification of the regression trees.
One possible explanation for the inconsistency in habitat classifications across censuses and lack of clear species assemblages for recruits is that the topographical variables used here as surrogates for soil moisture and nutrients (and the variables related to light availability and vegetation structure) may not reflect actual regeneration niches or may be too coarse (note that we calculated these variables with a spatial resolution of 20 × 20 m). However, an alternative explanation is that other factors may mask existing regeneration niches. For example, seed input of species shows high year-to-year variability and depends on fluctuations in flowering and fruiting, which are driven by El Niño events (Wright et al. 1999; Wright 2005), precipitation (Zimmerman et al. 2007) or abundance and distribution of seed dispersers (Dennis et al. 2005). Additionally, seeds tend to be dispersed primarily near parent trees (Hubbell 1979; Ribbens, Pacala & Silander 1994), which may mask habitat preferences of recruits to some extent.
The arguments of masking mechanisms and low sample sizes are strengthened by the finding that recruits of all censuses together showed an MRT classification similar to that of juveniles (but with an higher CVRE error). This is reasonable because juveniles accumulate several recruitment cohorts. However, to yield this pattern, ‘filter’ mortality must enhance the signal of habitat structuring during the transition from recruits to juveniles (Comita & Engelbrecht 2009). This mechanism would also, as observed, reduce the noise in the emerging juvenile species assemblages.
Habitat Structuring of Reproductive Individuals
The assemblages of reproductive adults were internally more homogeneous than those of juveniles (as indicated by a reduction in the classification error of the regression tree), which suggests that neutralization does not come with the cost of higher variability. The observed homogenization in the emerging species assemblages of reproductive individuals could be interpreted as support for neutral theories (e.g. Hubbell 2001), which assume that species-rich communities of adults are not primarily structured by habitat differences. Our results suggest that a directed mortality filter is working that both neutralizes habitat association and homogenizes the emerging species assemblages.
The intriguing question that emerges is which processes may have caused the observed homogenization in species assemblages during the transition from juvenile to reproductive adult. During the transition, species must have lost the advantage of being in the habitat with which they had become associated. One hypothesis is that habitat-specific mortality resulted in high densities of juveniles in their preferred abiotic habitat, but that negative density dependence at the transition to reproductive individuals (e.g. Schupp 1992; Condit, Hubbell & Foster 1994; Webb & Peart 1999) outweighed the benefits of the optimal seedling habitat and led to adult trees no longer being associated with that habitat (Webb & Peart 2000). Alternatively, ontogenetic shifts in resource requirements, such as light (Lusk 2004; Poorter et al. 2005), or shifts in water use (Donovan & Ehleringer 1992), may be responsible for conflicting habitat requirements between juvenile and reproductive stages that could neutralize the habitat structure that emerged during the juvenile stage if habitats that previously were good for survival came to have higher mortality rates at the transition from recruits to maturity.
The homogenization of the reproductive assemblages may also be a consequence of different time scales. The reproductive class accumulated individuals that recruited over a long period and were therefore subject to many different environmental conditions, whereas juveniles experienced only more recent environmental conditions (after 25 years only 44% of the juveniles were still juveniles; Table S4). Thus, many older adult trees may be present in habitats as a consequence of past favourable periods, but in the absence of favourable conditions in recent years, no recruits or juveniles would be found there.
At the species level, the strength of habitat associations in the BCI plot is relatively weak, meaning few species were restricted to a single habitat (Harms et al. 2001; Comita, Condit & Hubbell 2007). Rather, most species had higher than expected abundances in their associated habitat but were also encountered in other habitats at lower densities. Such effects are considered in our analysis, because the Bray–Curtis dissimilarity index considered species abundances within 20 × 20 m quadrats. The relatively high classification error (CVRE) at all life stages reflects the relatively weak strength of habitat associations at the species level.
The classification error CVRE was lowest for reproductive individuals, which showed the lowest habitat structuring, and highest for recruits, with juveniles having intermediate error (Table 1). When plotting error over time, we found a consistent tendency to increase for all development stages (Fig. 3). Thus, habitat structuring at the BCI has become noisier and somewhat less structured over the 25 observation years. This increase in noise was accompanied by a slight but consistent decrease in the number of species observed in the different development stages (Fig. 3) and was largest for individuals at the reproductive stage (from 289 species in 1983 to 278 in 2005). Additionally, the number of individuals within different life stages declined slightly with time (Fig. 3). One possible explanation for these findings is the occurrence of a dry period at the BCI plot that lasted through the first half of the study period (Condit 1998) with severe El Niño droughts that occurred in 1982–83 and 1997–98 and caused high mortality. For example, of 37 species defined to be moisture demanding, 33 declined after the severe 1982–83 El Niño drought (Condit, Hubbell & Foster 1996a,b). Such disturbance events may interrupt ‘normal’ species behaviour and introduce considerable and unspecific noise that increases the observed dissimilarity in species assemblages within habitat types. However, recent results demonstrated a strong dependence of seedling mortality on water availability, and especially a higher mortality of (wet) slope specialists in the (dry) plateau habitat (Comita & Engelbrecht 2009). Thus, the opposite effect would be expected: for juveniles the noise in habitat-dependent species assemblages should become smaller because drought should make occupancy of dry and wetter habitats more uniform in terms of species composition. Additionally, the 1995–96 dry season marked the beginning of a wetter period, which should then reverse this trend in juveniles. Further research is required to unravel the mechanisms behind this pattern.
Analyses of indicator species suggested that only few species showed consistent associations with a single habitat type across multiple life stages, consistent with previous species-level studies in tropical forests (Webb & Peart 1999; Paoli, Curran & Zak 2006; Comita, Condit & Hubbell 2007; Lai et al. 2009). Moreover, we found that reproductive individuals formed basically one large assemblage with far fewer indicator species than were found for the high and low plateaus at the recruit and juvenile stages (Fig. 4). The large assemblage of reproductive individuals was supplemented by some additional habitat types that contained a larger number of indicator species e.g. the swamp assemblage with >20 indicator species. Nonetheless, the number of indicator species for different habitat types at the recruit and juvenile stages was always higher than that at the reproductive stage (Fig. 4).
Our analysis of habitat-defined species assemblages of tropical tree species suggests that complex changes in species assemblages occur in time and with life stage. Recruits showed a weak and temporally inconsistent habitat structuring, whereas distinct and temporally consistent habitat types (low and high plateaus, slope and swamp) emerged for juveniles. However, a notable homogenization occurred during the transition to the reproductive stage, through both a reduction in the number of species assemblages and a reduction in the classification error of the regression tree analysis.
The homogenization in the emerging species assemblages of reproductive individuals suggests that species may have lost the advantage of being in the habitat with which they had become associated as juveniles, or habitat quality changed during their life span. While the weak habitat structuring of the reproductive community agrees with predictions of neutral theories, further research is required to unravel the possibly non-neutral mechanisms behind the intriguing homogenization at the transition from juvenile to reproductive. Our results re-emphasize the need for studies to examine species’ habitat preferences at multiple life stages (Comita, Condit & Hubbell 2007). Comparative studies in other tropical forest plots are required to find out whether our results reflect site idiosyncrasies of the BCI plot, which shows relatively weak habitat associations, or general trends.
The BCI forest dynamics research project was made possible by National Science Foundation grants to Stephen P. Hubbell, support from the Center for Tropical Forest Science, the Smithsonian Tropical Research Institute, the John D. and Catherine T. MacArthur Foundation, the Mellon Foundation, the Celera Foundation, and numerous private individuals, and through the hard work of over 100 people from 10 countries over the past two decades. The plot project is part the Center for Tropical Forest Science, a global network of large-scale demographic tree plots. R.K and A.H were supported by the ERC advanced grant 233066 to T.W. L.S.C acknowledges the support of a postdoctoral fellowship from the National Center for Ecological Analysis and Synthesis, a Center funded by U.S. NSF (Grant #EF-0553768), the University of California, Santa Barbara, and the State of California. Two anonymous reviewers provided helpful suggestions for improving the manuscript.