Climate change would lead to a sharp acceleration of Central African forests dynamics by the end of the century

Impacts of climate change on the future dynamics of Central African forests are still largely unknown, despite the acuteness of the expected climate changes and the extent of these forests. The high diversity of species and the potentially equivalent diversity of responses to climate modifications are major difficulties encountered when using predictive models to evaluate these impacts. In this study, we applied a mixture of inhomogeneous matrix models to a long-term experimental site located in M’Baïki forests, in the Central African Republic. This model allows the clustering of tree species into processes-based groups while simultaneously selecting explanatory climate and stand variables at the group-level. Using downscaled outputs of 10 general circulation models (Gcm), we projected the future forest dynamics up to the end of the century, under constant climate and Representative Concentration Pathways 4.5 and 8.5. Through comparative analyses across Gcm versions, we identified tree species meta-groups, which are more adapted than ecological guilds to describe the diversity of tree species dynamics and their responses to climate change. Projections under constant climate were consistent with a forest ageing phenomenon, with a slowdown in tree growth and a reduction of the relative abundance of short-lived pioneers. Projections under climate change showed a general increase in growth, mortality and recruitment. This acceleration in forest dynamics led to a strong natural thinning effect, with different magnitudes across species. These differences caused a compositional shift in favour of long-lived pioneers, at the detriment of shade-bearers. Consistent with other field studies and projections, our results show the importance of elucidating the diversity of tree species responses when considering the general sensitivity of Central African forests dynamics to climate change.


Introduction
Among the multiple increasing anthropogenic factors that will affect tropical forests (Lewis et al 2015), climate change appears to be both the most global and inevitable one. Through the alteration of temperature and rainfall, climate change could alter the structure, composition and functioning of tropical forests. Knowing the climate response of these forests is a major challenge in view of their crucial role as large-scale carbon stocks and sinks to contain global warming well below 2°C (Griscom et al 2017, Rogelj et al 2018).
Existing studies have largely focused on Amazonia. Simulations of climate change impact have suggested that Amazonian forests could rapidly decline as a result of enhanced droughts and accelerated temperature increase: the Amazonian die-back (Cox et al 2004). Inversely, some models have projected a greening-up of Amazon forests with an overall gain in productivity as the fertilisation effect of increasing CO 2 concentration overcompensates for the losses due to changes in temperature and precipitation (Huntingford et al 2013). Considering Amazon forests as a carbon sink may however be excessively optimistic and is still being debated (Lewis et al 2015).
As the second largest continuously forested area in the world, Central Africa is a key region for both climate change mitigation and adaptation (Niang et al 2014). Its importance for mitigation is largely recognised, with forests sheltering 80% of the above-ground African carbon stock (Mayaux et al 2013). In this region, depending on the climate scenarios, temperature is projected to increase in a range between 2 and 4°C above the baseline observed in the second half of the 20th century reference (Aloysius et al 2016). Compared to temperature projections, precipitation projections from general circulation models (GCM) or from regional circulation models (RCM) are more heterogeneous, partly because of the difficulty in resolving climate mechanisms driving the variability of precipitation across Central Africa The response of Central African forests to climate change has remained largely understudied until recently (Sonwa et al 2013), partly due to the lack of regional climate information (Lennard et al 2018). Tree species composition is strongly correlated with rainfall and temperature gradients (Fayolle et al 2014). The global study of Zelazowski et al (2011) on the role of climate in the distribution of humid tropical forests and the first regional impact assessment of climate change on the Congo Basin by Ludwig et al (2013) concluded that, in contrast to Amazonia, a die-back in Central African forests would be unlikely.
Relationships between climate variables and forest dynamic processes, i.e. growth, mortality and recruitment, are complex (Swann et al 2016). They are widely variable from one species to another resulting in species uniqueness in their possible response to climate change (Uriarte et al 2018). The same changes in climate could be beneficial to some species but detrimental to others, leading to shifts in floristic composition.
Tropical tree species responses to climate can be approximated through their shade-tolerance, generally by distinguishing ecological guilds (Hawthorne 1995) such as pioneers (P), non-pioneer light-demanders (NPLD) and shade-bearers (SB). In the Central African Republic (CAR), Ouédraogo et al (2013) showed that slow-growing SB species were the least sensitive to drought. On the contrary, in Ghana, drought might lead to a shift from SB species to dry-adapted P and NPLD species (Fauset et al 2012).
From a modelling perspective, tree species grouping, and more generally, plant functional classification (Lavorel et al 2007) is a tool for simplifying floristic complexity. There are various methods for grouping species, including groups based on the subjective ecological knowledge of species, such as widely-used Hawthorne (1995)ʼs guilds, or groups based on species dynamic processes, namely growth, mortality and recruitment (Gourlet-Fleury et al 2005). While subjective ecological groups facilitate the ecological interpretation of results, when it comes to simulating forest dynamics, it is better to build groups based on their own processes in order to improve the estimation of dynamics parameters and the projection performances (Mortier et al 2013, Picard et al 2010. In this study, we aimed at disentangling the longterm effects of different climate variables on the structure and dynamics of a tropical rainforest in Central Africa. We hypothesised that climate change will have differential impacts on species dynamics with varying responses driven by different climate and structure variables within and across ecological guilds. We used a modelling approach based on Mortier et al (2015)ʼs mixture of inhomogeneous matrix models (MIMM), a methodology particularly well adapted to simulate the dynamic behaviours of species-rich ecosystems. We applied it to forest inventory data from a unique 30-year-long experiment established at M'Baïki, in CAR. To integrate the variability of climate projections, we produced 10different versions of MIMM, one for each GCM. After comparing observed and simulated forest stands, we examined the impact of climate change on forest dynamics, in terms of both structure and composition, over the 21st century.

Material and methods
2.1. Data 2.1.1. Forest inventory data Annual forest inventories have been performed since 1982 at the M'Baïki experimental site (figure S1 available online at stacks.iop.org/ERL/14/044002/ mmedia), established in CAR (3°54′N, 17°56′E). The vegetation is a Guineo-Congolian semi-deciduous moist forest, with canopy dominated by Cannabacae, Myristicacae and Meliacae families (Boulvert 1986). Ten permanent sample plots, four ha each, were established in two neighbouring forest reserves. Between 1982 and 2012, all trees 10 cm diameter at breast height (DBH) have been monitored. The floristic description of the trees is complete since 1992. Three disturbance treatments were implemented: three plots were left as controls, seven plots were selectively logged of which four were subsequently thinned. Selective logging occurred between 1984 and 1985 and consisted of harvesting commercial trees with DBH 80 cm. Thinning occurred two years after logging and consisted in poison girdling all non commercial trees with DBH 50 cm and cutting all lianas. A complete description of the experimental site can be found in Gourlet-Fleury et al (2013).

Climate data
To predict the climate at M'Baïki over the next century, we used outputs of regional climate simulations over 1992-2100 conducted over the Coordinated Regional Climate Downscaling Experiment (COR-DEX)-Africa domain at a resolution of ≈50 km We used 10GCM outputs (table S1), downscaled by the Rossby Centre regional atmospheric model (Samuelsson et al 2011). For each GCM, we used three sets of climate projections: the historical runs over the period 1992-2005 extended over 2006-2099, which we refer to as baseline scenario under constant climate; and the projected runs under two Representative Concentration Pathways (RCP), RCP 4.5 and RCP8.5 (Moss et al 2010), over the period 2006-2100. The baseline scenario allowed, by comparison with the other two scenarios, the identification of forest dynamics that would or not be affected by climate change.

Forest dynamics modelling methodology
We used the MIMM methodology (Mortier et al 2015) to model the forest dynamics based on each of the 10 GCM used. MIMM is a statistical correlative model that identifies differential growth, mortality and recruitment processes, and selects their associated subsets of explanatory variables. It combines ideas of mixture models, generalised linear models (GLM), variable selection and matrix models. The MIMM method consists of two steps: (i) for each dynamic process, penalised finite mixtures of GLM are used to classify species into groups, select the best explanatory variables, and estimate the model parameters; (ii) a Usher (1966,1969) matrix model is then calibrated for each non-empty identified group and used to project forest dynamics. Further details are provided in the supplementary section S1.3.

Simulation protocol 2.3.1. Fitting MIMM on GCM versions for M'Baïki
We used data collected on all the M'Baïki plots, regardless of the disturbance treatment applied, and data from CORDEX-Africa historical runs, to fit MIMM to the various GCM and to define the composition and the parameters associated to species groups, over the 1992-2005 period. This enabled us to quantify the dynamic processes over a large range of stand conditions, using ≈120 000 diameter increments measures, 3600 death and 6100 recruitment events.
We processed the climate data after extracting the 30 arc s (≈1 km 2 ) pixel corresponding to the M'Baïki site from raster maps of climate variables. For this pixel we averaged, on an annual basis, monthly time series of 13 climate variables. We excluded six highly correlated variables, using a variance inflation factor (VIF) threshold of 5 (Dormann et al 2013). We thus kept seven climate variables to calibrate MIMM (table 1). Climate variables were complemented by four structure variables calculated from the M'Baïki forest inventories and directly linked to the development of trees and forest stands: DBH (in cm) and its logarithmic transformation, log-DBH, to deal with the nonlinear association between DBH and the growth and mortality processes; stand density (stems per hectare, in ha −1 ) and stand basal area (in m 2 ha −1 ) to quantify the tree competition effect on growth, mortality and recruitment processes.
We assessed the observed and projected stand structures based on four output variables: stand density (ha −1 ), stand basal area (m 2 ha −1 ), quadratic diameter (cm) and carbon biomass (t eq CO 2 ha −1 ). We calculated the first three variables directly from the MIMM outputs at each time step while we estimated carbon biomass by combining the MIMM outputs with a pantropical allometric equation (  Although the calibration and validation procedures could be successfully completed, the simulations over the projection period systematically failed to converge when including CO 2 . Forest dynamics reached unrealistic levels, with infinite values in stand characteristics and processes. This failure was due to the perfect correlation of the annual average atmospheric CO 2 concentration with time. Thus, CO 2 effect was confounded with linear temporal trend, a situation that prevents any valid projection. In the following, only results obtained without CO 2 concentration are presented.

Summarising MIMM outputs
Because the number of species combined to the 10GCM versions led to an impractical number of information, results were summarised by re-aggregating MIMMʼs outputs according to the two following approaches. The first, exogenous to MIMM, is based on ecological guilds, the second, endogenous to MIMM, is based on the similarities of species groupings between GCM versions. For the latter, we used hierarchical clustering to define new meta-groups. The number of meta-groups was estimated using the gap statistics (Kaufman and Rousseeuw 2005). In comparison to guilds, meta-groups are more consistent with the information from the data as they follow the same logic of species grouping based solely on the similarity of their dynamic processes (Picard et al 2010). Further details are provided in the supplementary section S2.1. The contributions of variables to projected changes in dynamic processes were quantified using weighted averages across GCM versions of the correlations between the linear predictor associated to a variable and the projected values (see supplementary section S1.3).

Species groups and explanatory variables
Among the 10different GCM versions used for applying MIMM to the M'Baïki data, the number of groups identified ranged between 5and 8 for the growth process, and 2and 4 for the recruitment process. For the mortality process, all versions agreed on 3groups. Allocating species across growth, mortality and recruitment groups resulted in 11to 17 non-empty growth-mortality-recruitment combined groups. In these groups, depending on the GCM version, the number of species was variable with the largest groups containing between 35and 67species, and the average group between 12.2 and 18.9species. The number of species groups obtained by model fitting and the species composition of these groups showed similarities and differences across GCM versions ( figure 1(a)). The cluster allocation of species and the parameter estimates are provided for each GCM version in supplementary section S3.
To refine comparisons, we disaggregated model projections from the group to the species level, and then re-aggregated them at the guild level and at the meta-group level. The differences between groups, guilds and meta-groups are illustrated in figure 1 and detailed in figure S2. Some species groups, especially some P-dominated or SB-dominated groups were quite well discriminated on the basis of their guild composition. The majority of groups mixed species from two or all three guilds. In particular, no group was exclusively composed of NPLD.
Ninemeta-groups were identified with a hierarchical cluster analysis and gap statistics (figure S2 and table 2). Meta-groups1, 5 and6 were highly dominated by SB. Meta-groups2 and3 were dominated by NPLD, meta-group4 by long-lived P and meta-group9 by short-lived P. The two remaining meta-groups 7 and 8 are dominated by NPLD and SB, in similar proportions.
The climate and structure variables selected to be associated with the dynamic processes varied across the GCM versions, guilds and meta-groups (tables S2 and S3). At the guild level, all structure variables were generally selected to explain dynamic processes with the exception of stand density for growth. All climate variables are generally selected to explain growth, but were not systematically selected for mortality and recruitment. The selection of the variables varied within guilds, reflecting the heterogeneity of specific responses within guilds. For example, climate variables had a large variability in their selection for mortality among P. At the meta-group level, the frequency of selection for the variables had narrower range, indicating less heterogeneity of species responses within meta-groups than within guilds, the best example being meta-group5. However, contrasts in terms of variable selection can be observed between meta-groups, for example on growth (stand density is less selected in meta-groups 1 and 2) and on recruitment (climate variables are less selected in meta-groups 2 and 3).
Over the validation period (2006-2012), simulated forest structure and dynamics remained close to the observed data for all climate change scenarios and disturbance treatments, except a slight overestimation of growth and recruitment in treated plots. In particular, some trends such as the decrease of stand density and recruitment, as well as the increase of basal area, quadratic diameter and carbon biomass were well captured by MIMM.   1(b)). For each point, the size is proportional to the number of trees clustered in the group or meta-group, and the colour and stand position correspond to the guild composition of these trees: the more trees from P (respectively, NPLD and SB) in the group or in the metagroup, the closer the point to the P (respectively, NPLD and SB) summit of the simplex with green (respectively, red and blue). Table 2. Ecological description of the meta-groups. From left to right, columns indicate the most abundant species inside each meta-group, the number of trees, the number of species, the maximum DBH (in cm), the diameter increment (in cm yr −1 ) and the proportions , expressed in reference to the number of trees, of each regeneration guild (in %). At the guild level, some differences were also found between projections in 2099 and the observed values in 2005. Projected quadratic diameter and carbon biomass of NPLD and SB species in 2099 were all significantly higher than the observed values in 2005 (p-values <10 −3 ). The basal area of SB species was also significantly higher than the one observed in 2005 (p-value <10 −3 ). Conversely, the basal area and quadratic diameter of P species projected in 2099 were significantly lower than the observed values in 2005 (p-values <0.05). Moreover, projected growth in 2099 was significantly lower for all guilds than the levels observed in 2005 (p-values <0.01). Projected mortality in 2099 was significantly higher than observations for NPLD species in 2005 (p-value <10 −3 ). We found no difference between projected and observed recruitment for all guilds. At the meta-groups level, the composition remained almost unchanged between observations in 2005 and projections in 2099, and we found no significant difference in stand density (figure 4). Projected quadratic diameters in 2099 were significantly higher compared to 2005 observed values for all metagroups except 4, 5 and9 (significantly lower for metagroup 9). Growth was projected lower in 2099 compared to the observed values in 2005 except for meta-groups2 and8. Projected mortality in 2099 was significantly higher than the observed values in 2005 for meta-groups 2, 3 and9. Recruitment was projected significantly lower for meta-group 1 and significantly higher for meta-groups 5 and8 in 2099 compared to the 2005 observed levels. These results were also consistent with forest ageing, with a general decline in growth, and a decreased abundance of pioneers in favour of shade-tolerant species.

Guild composition (%)
3.4. Impacts of climate change on forest structure and dynamics RCP 4.5 and 8.5 scenarios projected a strong impact of climate change on the structure and the dynamics of forest stands, compared to the projections under the Figure 2. Change over time of the mean observed and simulated output variables describing forest structure for each disturbance treatment under different climate scenarios: stand density (in ha −1 , subfigure 2(a)), stand basal area (in m 2 ha −1 , subfigure 2(b)), quadratic diameter (in cm, subfigure 2(c)), above-ground carbon biomass (in t eq CO 2 ha −1 , subfigure 2(d)), growth (in cm yr −1 , subfigure 2(e)), mortality (in yr −1 , subfigure 2(f)), recruitment (in ha −1 yr −1 , subfigure 2(g)) and turnover (in yr −1 , subfigure 2(h)).  . Average values and standard errors of output variables describing the structure and dynamics of observed and projected forest stands, at the stand and at the guild levels. Structure variables are stand density (in ha −1 , subfigure 3(a)), tree basal area (in m 2 ha −1 , subfigure 3(b)), quadratic diameter (in cm, subfigure 3(c)) and above-ground carbon biomass (in t eq CO 2 ha −1 , subfigure 3(d)). Dynamics processes are growth (in cm yr −1 , subfigure 3(e)), mortality (in yr −1 , subfigure 3(f)) and recruitment (in ha −1 yr −1 , subfigure 3(g)) while turnover rate (in yr −1 , subfigure 3(h)) is also given. For each variable or process, values are indicated at the whole-stand level (first set, in grey) from left to right, and at the guild level (green, red, blue for P, NPLD and SB, respectively). Each set of barplot is composed of four bars, with more or less saturated colour, representing from left to right: M'Baïki control plot observations between 1992 and 2005 (first bar, the most saturated), projected stands under constant climate (second bar, the least saturated), projected states under RCP4.5 and RCP8.5 in 2099 (3rd and 4th bars, respectively, with lower and upper intermediary saturations). baseline scenario ( figure 3). The same trend, namely an acceleration of forest dynamics due to an enhancement of growth, recruitment and mortality processes, was observed in the great majority of the simulations under all scenarios, but with a higher variability for RCP8.5 than for RCP4.5.
Growth increased by a factor of 3.4±1 under RCP4.5 and a factor of 6.8±4 under RCP8.5 compared to the baseline scenario. Mortality was multiplied by 3.7±2 under RCP4.5 and 5.7±3 under RCP8.5, while recruitment, divided by stand density, was multiplied by 3.9±2 under RCP4.5 and 5.7±4 under RCP8.5. For the baseline scenario, the turnover (calculated as the average of mortality and recruitment rates) was found to equal 1.1±0.05% while for the RCP 4.5 and 8.5, it equalled 4.7±3% and 7.1±5% respectively. Considering the contribution of each variable to each dynamic process (figure 5), projected changes of growth appear to be mainly driven by stand basal area and sunshine duration under RCP4.5 and by temperature under RCP 8.5. Projected changes of recruitment appear to be mainly driven by stand basal area and stand density, while projected changes of mortality appear to be driven by temperature and, to a lesser extent, by total runoff.
Climate change significantly impacted the stand density as well as the structural characteristics of the forest: a thinning phenomenon was observed with stand density falling from 588±25ha −1 to 473±48ha −1 under RCP4.5 and to 474±81ha −1 under RCP8.5, paralleled with an increase of the quadratic diameter from 31±1 cm to approximately 33±3 cm under RCP4.5 and 37±4 cm under RCP8.5.
The impacts of climate change differed from one guild to another. It was found to favour P over SB (no significant effect on NPLD). The proportion of SB fell from 62±0.5% to 56±4% under RCP4.5 and to 55±5% under RCP8.5, while the proportion of P increased from 10±0.4% to 16±5% under RCP4.5 and to 17±5% under RCP8.5. Meanwhile, the proportion of NPLD remained around 26%. P turnover rose from 2.3±0.1% for the baseline scenario, to 6.8±3% under RCP4.5 and 8.2±5% under RCP8.5. NPLD and SB turnover shared the same pattern of increased turnover: from 0.78±0.03% to 3.7±2% and 7.1±5% (NPLD species); from 1±0.04% to 3.8±2% and 7.2±5% (SB species). Dynamic processes appeared to be driven by the same climate variables at the guild-level than at the standlevel (figure S3, same plotting method as in figure 5). NPLD and SB dynamic processes appeared to be driven by a larger number of climate variables than P dynamic processes, suggesting a higher sensitivity of SB and NPLD to climate variables.
Forest dynamics acceleration, albeit with much more variability under RCP8.5 compared to RCP4.5, was a general trend projected for all processes and all meta-groups (figure 4), with the sole exception of recruitment for meta-group1 that decreased with time. The magnitude of increase varied greatly from one meta-group to the other leading to changes in the Figure 5. Trivariate representation of the contribution of explanatory variables, except tree variables DBH and log-DBH, to the change of growth, mortality and recruitment over time, for RCP4.5 and8.5. Segments represent the variability (standard errors) across the different GCM versions. A value of 1 means that the projected contribution of the variable is perfectly correlated to the projection of the dynamic process based on all the explanatory variables while a null contribution means that the variable does not contribute to the projection of the dynamic process.
The climate drivers of dynamic processes were variable across meta-groups ( figure S4 and S5, same plotting method as in figure 5). Under RCP4.5, growth was driven by duration of sunshine for all meta-groups except 9. Under RCP8.5, growth was driven by temperature for all SB-and NPLD-dominated meta-groups, but no driver was found for the growth of P-dominated meta-groups4 and9. For mortality, the temperature was the most relevant variable for all meta-groups and scenarios. Under RCP4.5, recruitment was driven by duration of sunshine for all meta-groups, by temperature for SBdominated groups and also by total runoff for meta-group9. Under RCP8.5, temperature was found relevant for the recruitment of P-and SB-dominated meta-groups.

Discussion
We used MIMM, a method allowing the simultaneous clustering of species into groups according to dynamic processes (growth, mortality and recruitment) and the identification of group-specific explanatory climate and structure variables. We projected the dynamics of M'Baïki forest stands over the 21st century under a constant climate scenario and under two scenarios of climate change, RCP4.5 and8.5. We showed that climate change would lead to a sharp acceleration of forest dynamics, with an increase of growth, mortality and recruitment, leading to changes in forest structure and composition, namely a natural thinning effect with fewer but larger trees, and an increase in the relative proportion of pioneers at the detriment of shade-bearers. Our simulations projected a slight increase of carbon biomass. This result is also consistent with previous projections in tropical forests which suggested that climate change induces a shift from low-biomass forests to high-biomass forests (Kim et al 2017). The increase of carbon biomass projected in our baseline scenario, without climate change, is consistent with observations of forest ageing. It can be explained by a recovery from past human-induced disturbance (Morin-Rivat et al 2014).

Climate change would affect forest composition by favouring long-lived pioneers
The magnitude of the projected changes in dynamics varies from one group of species to another and this differential response leads to a change in the relative species composition of forest stands, resulting in an increase of the relative proportion of pioneers, at the detriment of shade-bearers. In particular, we projected a rise of long-lived pioneers (meta-group 4) at the detriment of shade-bearers (meta-groups 1, 6 and 7). Compared to shade-bearers, pioneers have higher light requirements, faster growth and are more disturbance-adapted. Some studies have already reported shifts in tree composition favouring fastgrowing species over slow-growing ones (Wright 2005) and have hypothesised a response to climate change. The expected higher dominance of pioneers may also raise questions about the future capacity of Central African forests to cope with droughts. Indeed, although pioneers are mostly deciduous and therefore well adapted to seasonal drought, the increase in their dominance may have a negative effect on the global ecosystem resistance to water stress because pioneers that allocate their resources to fast growth are less resistant to water stress (Aubry-Kientz et al 2015).

Climate drivers and particularly temperature have significant effects on the dynamic processes
Projected growth appeared to be mainly driven by stand basal area and sunshine duration under RCP4.5 and by temperature under RCP8.5. Projected recruitment of forests is found to be mainly driven by stand basal area and stand density, while projected mortality appeared to be driven by temperature and, to a lesser extent, by total runoff. These results highlight the sensitivity of forests to climate (Anderson-Teixeira et al 2013) with sunshine duration (a proxy of the energy reaching the forest system), water availability and temperature as the main drivers. The positive influence of sunshine duration on growth has already been stressed by Dong et al (2012). The influence of total runoff as a proxy for water availability, is in line with other studies showing the importance of drought in tree mortality (Phillips et al 2010, eg).
The positive correlation between mortality and temperature is consistent with observations from field studies (Clark et al 2010). However the positive correlation between growth and temperature contradicts most of the results from the Amazon (Aubry -Kientz et al 2015). The temperature-sensitivity of growth observed in Amazon forests may not be valid in African forests, as the latter are presumably more resilient to climate change than Amazon forests (Malhi et al 2014). This contradiction could also be resolved considering the nonlinearity of relationships between temperature and photosynthesis (Hüve et al 2011): tropical tree species may be near a high temperature threshold that, if exceeded, will greatly reduce growth (Doughty and Goulden 2008). Amazon forests would already have crossed this threshold whereas Central African forests may still be below it. In this case, our projections might not have captured this threshold effect.
4.4. Meta-groups sharpen differences in species response to climate change that are less evident with guilds The elaboration of optimal species aggregation techniques for modelling the dynamics of species-rich ecosystems remains an active research area (Kazmierczak et al 2014). The comparison of several climate calibrations shows a great similarity in the way species are grouped, making it possible to identify meaningful ecological meta-groups. The identified meta-groups differ in many ways from Hawthorne (1995)ʼs guilds: species with different dynamic profiles can be distinguished within the same guild, notably among pioneers, while other species are grouped together although they are considered to belong to distinct guilds, notably SB and NPLD. Compared to exogenous guilds, our meta-groups have the advantage to be an endogenous result from the model and to reflect the species-specific dynamic responses to climate variables in a more accurate way than guilds. More generally, in the continuation of the work of Gourlet-Fleury et al (2005), Picard et al (2010) and Mortier et al (2015), our results show the value of using dynamics data-driven classifications rather than classifications based on ecological knowledge to simulate long-term forest dynamics.
To refine the grouping of species based on their response to climate change, further developments can be considered, for example through the ongoing CMIP6 (Eyring et al 2016) and advances in the use of extended matrix of GCM-RCM experiments (Wilcke and Bärring 2016), or through the incorporation of ecophysiological processes (Feng et al 2018), especially to deal with the effects of increased atmospheric CO 2 . These would offer new opportunities to enhance the future projections of climate change and to improve the understanding of its impact on tropical forest dynamics (Malhi et al 2014).

Acknowledgments
We thank the Forestry Research Support (ARF, Appui à la recherche forestière) and its seven partners: the French Development Agency (AFD, Agence Française de Développement), the Centre for International Cooperation in Agricultural Research for Development (CIRAD, Centre de coopération internationale en recherche agronomique pour le développement), the CAR Institute of Agricultural Research (ICRA, Institut Centrafricain de Recherche Agronomique), the Ministry of Waters, Forests, Hunting and Fisheries (MEFCP, Ministère des Eaux, des Forêts, de la Chasse et de la Pêche) of Central African Republic, the Service of Cooperation and Cultural Action (SCAC, Service de coopération et d'action culturelle) of the French Ministry of Foreign Affairs (MAE, Ministère des Affaires Étrangères), the University of Bangui and the Central African Republic Company of Wood Peeling (SCAD, Société centrafricaine de déroulage) for providing access to the site and to the database of M'Baïki. We are grateful to volunteers appointed by the SCAC of MAE and the fieldworkers who participated in the project management, data collection and data capture. We thank the two anonymous reviewers for the clarity and relevance of their comments, which significantly contributed to improving the quality of the paper.
The Bureau for Economic Theory and Applications contributes to the LABEX ARBRE (ANR-11-LABX-0002-01). Frédéric MORTIER and Mahlet TADESSE were partially supported by a Georgetown Environment Initiative award. Sylvie GOURLET-FLEURY was partially supported by the DYNAFFOR project (supported by the French Fund for the Global Environment (FFEM), grant number no. CZZ1636.01D) which also supported the publication.
The authors declare no conflict of interest.