Three data sets were used in this study. The study site is located in Paracou, French Guiana (5∘18′ N, 52∘55′ W). The forest is typical of Guianan rain forests and the dominant tree families are Fabaceae, Chrysobalanaceae, Lecythidaceae, and Sapotaceae. More than 700 species of trees ≥ 10 cm DBH (diameter at breast height) have been described at the site.

Mean annual precipitation averages 2980 mm (30-year period), and the site receives nearly two-thirds of its annual precipitation during the long rainy season between mid-March and mid-June , and less than 100 mm per month from August to November (Fig. ).

The model used in this study consists of a model coupling growth and mortality processes at the whole community scale. The model is build taking advantage of two preliminary studies where the growth and the mortality sub-models were developed. The likelihood is computed using the distribution probability of mortality (Eqs. 3 and 4) and the computed growth rate (Eqs. 5 and 6). A vigor index is added into the mortality process, taking the past growth of the two previous years into account . We added the climate variables into the two processes to highlight the links between some climate drivers and one particular process. Because the final forest dynamic model was not linear, we build a Markov chain Monte Carlo algorithm under a Bayesian framework to infer the parameter posterior distributions. Growth and mortality processes were linked through tree vigor and are parameterized simultaneously. If tree i stays alive, it grows at a growth rate AGRi,s,t, and its diameter DBHi,t-1 becomes DBHi,t. The joint model likelihood is then ∏t=1nf(DBHi,t|DBHi,t-1)×(1-pi,s,t) if tree i stays alive during the length of the studied period, pi,s,k×∏t=1k-1f(DBHi,t|DBHi,t-1)×(1-pi,s,t) if tree i dies between time k-1 and time k, where

f(DBHi,t|DBHi,t-1) is the probability density for a tree with diameter DBHi,t-1 at time t-1 to have a diameter DBHi,t at time t; this quantity is used to compute the vigor estimator.

pi,s,t=logit-1γ1×clim1+θ1×Vigori,s,t+θ2×DBHi,s,t-1DBHmaxs+θ3×DBHi,s,t-1DBHmaxs2+θ4×Hmaxs+θ5×WDs+θ6×Toughs with Vigori,s,t=log⁡AGRi,s,t-1+1AGR^i,s,t-1+1, and log⁡(AGR^i,s,t-1+1)=(γ2×clim2+θ7×DBHmaxs+θ8×WDs+θ9×Hmaxs+θ10×δ13Cs)×exp⁡-12log⁡DBHi,t-2θ11×DBHmaxsθ12×WDs2, and log⁡(AGRi,s,t-1+1)=log⁡(AGR^i,s,t-1+1)+εi, with εi∼N(0,θ13), where pi,s,t is the probability of dying of tree i of species s between time t-1 and t; AGR^i,s,t-1 is the predicted growth between time t-2 and time t-1. AGRi,s,t-1 is the observed growth between time t-2 and time t-1; DBHmaxs, Hmaxs, WDs, Toughs and δ13Cs are functional traits of species s to which tree i belongs (Table ); θ1,θ2,⋯θ13 are parameters to be estimated, and εi is an individual error term following a normal distribution; γ1 and γ2 are the parameter vectors linking the climate predictors with the processes of mortality and growth respectively; clim1 and clim2 are the vectors of climate predictors included in the processes of mortality and growth, respectively.

To identify the different axes of variation of our climate data set and avoid including collinear variables in the model, we realized a principal component analysis (PCA) on the climate variables.

We included all climate variables one by one in each process of the model and computed the partial likelihood for each sub-model of growth or mortality we obtained. This provides a first result about the importance of each climate variable. Depending on these results and on their degree of collinearity from the PCA, we selected some climate variables and included them in the growth model and in the logit function of mortality.

We implemented a Markov chain Monte Carlo algorithm to estimate the model parameters . A random walk was used as a proposal distribution to sample new values of parameters that were or were not selected, using the ratio of Metropolis–Hasting. Only standard deviation was sampled in an inverse-gamma posterior distribution with a Gibbs sampler. The functional traits used as demographical predictors were uncertain because botanical determination was incomplete for the older censuses, and not all values of functional traits were available for all species. We used the method developed in to handle these uncertainties. All the algorithms and statistical treatments were implemented with R software .

Functional traits were introduced in the final model with an interaction term by multiplying a climatic variable with a functional trait. We did not test all possible interactions but, based on results from a literature survey, we investigated biologically meaningful interactions only (Table ). We included in the model an interaction between wood density and the drought estimator Aunder, an interaction between DBH and Aunder, and an interaction between DBH and precipitation Pre.

Species vary over 1 order of magnitude in their wood density (WD), ranging from 0.08 to 1.39 g cm3 , and the encountered range of wood density is particularly large in species-rich tropical rain forests . Wood density is a key functional trait because of its importance for mechanical stability, defense against herbivores, hydraulic conductivity, photosynthetic carbon gain and diameter growth rates of plants . High wood density implies thin and short xylem vessels with small pit pores, which decrease the risk of embolism and cavitation. Trees with high wood density are then expected to be less sensitive to drought. The term Aunder multiplied by (WDmax-WD) accounts for the effect of drought on trees with low wood density. This term is added in growth and mortality to test this effect (Table ).

The variable selection was realized using the literature, the PCA results, and the results of the univariate analysis.

The growth trajectory was adjusted by a size-dependent diameter growth model (Fig. ). Parameters linking the maximal growth to the functional traits DBHmax, WD, Hmax and δ13C have similar values and interpretations to ; that is, maximum growth rates increase with increasing DBHmax, and decreasing WD, Hmax and δ13C (Table ). Maximum growth rate is attained for a tree diameter equal to 0.794 × DBHmax. The parameters linking the probability of mortality to Hmax, WD and Tough converged around negative values, meaning that the probability of dying is lower when the tree is high, has a high wood density and/or high laminar toughness. The drought estimator (Aunder) converged to negative values in the growth and mortality processes; thus growth and mortality computed at our biannual timescale are lower when the drought estimator Aunder is higher. The parameter linking mortality with precipitation (Pre) is positive. This finding implies that mortality rate is higher during 2-year timescale with high precipitation. In our data set, the highest total precipitation was, albeit non-significantly, rather related to the highest proportion of tree-fall deaths (Fig. ). The parameter linking temperature (tmp) and growth takes negative values; thus growth values are lower during the warmest periods.

In the growth process, interaction between (WDmax-WD) and drought is negative (Table ), implying that trees with lower WD are more sensitive to drought and reduce their growth more. Moreover, interactions linking the current diameter and drought are also negative; thus larger trees are more sensitive to drought and reduce their growth more compared to smaller trees. None of the interaction terms included in the mortality process had an effect (Table ).

The water stress during the dry season, estimated with Aunder, negatively impacts the growth and mortality processes. Trees will thus grow less quickly and have a lower probability of dying during 2-year periods with the most intense dry seasons. The reduction of growth is expected, and has many ecophysiological causes. Indeed, water is essential for sap fluxes and for photosynthesis efficiency. The reduction of growth is furthermore linked with the current DBH and the species' wood density (Table ). Big trees are more sensitive to water stress than small trees. This was expected in light of the results obtained after rainfall exclusion . Indeed, maintenance costs are higher for big trees, making these trees more vulnerable to the driest periods. Regarding the wood density, species with high values are more resistant to drought. This is consistent with our hypothesis that high wood density implies thin and short xylem vessels and thus decreases the risk of embolism and cavitation. As the ability of trees to recover from periods of sustained drought is strongly related to their embolism resistance , a tree with high wood density will be more able to maintain growth during dry years. For similar reasons, we expected a positive impact of Aunder on mortality rates. Experimental trough-fall exclusions conducted in Tapajós and Caxiuanã indeed demonstrated that 50 % rainfall exclusion led to very high mortality rates . Our results show no positive effect of drought intensity on mortality rates (Table ) and look contradictory to . However, the natural variability of the drought intensity (total rainfall from 5486 to 6207 mm) in our data set is hardly comparable to the experimental 50 % reduction in total rainfall. Moreover, our modeling framework prevented us from seeing long-term effects induced by repeated drought events because the drought variable values depend only on the last 2-year climate. One may also expect that standing death is more frequent during the driest periods but, when plotting tree mode of death against drought estimator (Aunder), no evidence was observed for a potential trend (Fig. ). To conclude, our results confirmed that the relationship between drought and mortality may be challenging to estimate and to link with their underlying causes at an inter-annual timescale.

Water saturation Pre had a strong effect on mortality; mortality rate varied between 1.5 and 2 % per 2 years with increasing total precipitation. This is consistent with the hypothesis that trees are more vulnerable when the soil is water saturated. In the Paracou forest, about half of tree deaths are due to standing death and half to tree fall. This ratio looks, albeit non-significantly (R2= 0.61, P=0.08 because of the low number of observations; n=9), linked with total precipitation. The highest total precipitation led to the highest proportion of tree-fall deaths (Fig. ). This confirms the observation of and the hypothesis that waterlogged soils in space or in time are risky for trees. Moreover, during the rainy season, strong rainfall events often come with strong winds that may accelerate this process . Studies observing a relationship between tree mortality and excess of water in the soil primarily focus on geographical variation and conclude that excess water in the soil restricts root establishment because productivity of fine roots and rooting depth are generally low in sandy soils and soils with high moisture content. Our results highlight that the time variation in soil water saturation is also very important and should be reassessed.

Temperature is identified as predictor of trees' decreasing growth. As the temperature rises, the velocity of reacting molecules increases, leading to more rapid reaction rates but also to damage of the tertiary structures of the enzymes and reduced enzyme activity and reaction rates . These two processes are responsible for a bell-shaped curve of growth response to temperature . Temperature can affect photosynthesis through modulation of the rates of activity of photosynthetic enzymes and the electron transport chain, and in a more indirect manner, through leaf temperatures defining the magnitude of the leaf-to-air vapor pressure difference, a key factor influencing stomatal conductances . In tropical forests, as temperatures are already high, rising temperatures may imply lower growth, consistent with results from .

This temperature effect may become the most problematic for tropical forest dynamics, considering the rising temperatures that are predicted, with a great degree of certainty, by climate models for the next century . Indeed, as temperature was identified as a strong predictor of growth, all else being equal, averaged community growth and forest productivity may consequently decrease in time. This decline in productivity in time is perhaps what we are starting to see throughout the Amazon . As no consensus has been reached yet, additional studies using regular inventories are urgently needed to explain the conflicting patterns of the temperature effect found in the extant literature . Finally, we need to acknowledge that we do not know much about how forest dynamics will behave in the next century under temperature conditions that will be so different from what is actually observed. In this context, manipulative warming experiments are increasingly vital to better predict the future of tropical forest dynamics .