Slow-growing species cope best with drought: evidence from long-term measurements in a tropical semi-deciduous moist forest of Central Africa



  1. Understanding how drought affects annual tree growth in tropical forests is of crucial importance to predict their response to climate change. Previous studies, mainly led in the Neotropics and in Southeast Asia, have yielded contradictory results which might be explained by differences in species studied, in the tree development stages considered, or by differences in other environmental factors than water availability.
  2. Here, we described the growth responses of functional groups of tree species to drought in a Central African semi-deciduous moist forest. Species groups were automatically defined using a finite mixture model, which grouped species according to their growth model parameters. The growth model considered the variation in species response to drought, and the effect of competition for resources and of tree development stage on growth. Groups were further characterized by species functional traits. Nine species groups were identified. They differed in their ability to acquire, use and conserve resources, as suggested by their differences in maximum growth rate, regeneration guild, maximum dbh, wood density and leaf habit. The species were organized along a light requirement gradient that here closely matched a broader continuum of plant strategies for resource use, from slow-growing shade-tolerant conservative species to fast-growing pioneer acquisitive species.
  3. Tree growth decreased with drought intensity, and species drought tolerance was found to be related to resource use strategy: slow-growing species using a conservative strategy were the least sensitive to variations in water availability, while fast-growing species using an acquisitive strategy were the most sensitive.
  4. Synthesis. Shade-tolerant species, characterized by a low potential growth rate and thus a conservative strategy of resource use, were found to be the least sensitive to drought. This supports the hypothesis of a single axis summarizing multiple traits that represents a general trade-off between the conservation and rapid acquisition of resources.


Although the scenarios predicting future climate change in Central Africa are controversial (Hulme et al. 2001; Delire, Ngomanda & Jolly 2008), the variability of climate and in particular of rainfall distribution is expected to increase. This change in water regime may both modify the productivity of tropical moist forests (Clark et al. 2003; Schuur 2003; Phillips et al. 2009) and cause a shift in floristic composition (Engelbrecht et al. 2007; Enquist & Enquist 2011; Feeley et al. 2011), as a result of the growth and mortality responses of the trees. For instance, during short-term droughts, large trees generally suffer greater mortality than small understorey trees (Condit, Hubbell & Foster 1995; Van Nieuwstadt & Sheil 2005; Nepstad et al. 2007; Phillips et al. 2010). In contrast, long-term droughts have been found to increase the abundance of canopy trees over small trees (Fauset et al. 2012).

Previous studies about the effects of drought on annual tree growth, mainly conducted in the Neotropics and in Southeast Asia, have yielded contradictory results: (i) dry years combined with high temperatures were associated with a decrease in growth (Clark et al. 2003; Nath et al. 2006), (ii) no link was found between dry years and changes in growth (Nakagawa et al. 2000), or (iii) dry years were associated with an increase in growth (Clark & Clark 1994; Condit et al. 2004). These contrasting results might be explained by differences in the species studied, differences in the tree development stages considered or differences in other environmental factors than water availability that affect growth (e.g. light).

Tropical tree species vary widely in their growth responses to drought (Newbery & Lingenfelder 2009), and species sensitivity to drought explains species distribution along rainfall gradients (Engelbrecht et al. 2007). Three major strategies of species' adaptation to drought have been identified: (i) evergreen species that tolerate drought stress and are able to maintain physiological functions despite low water availability by having high biomass investment in enduring organs, and minimizing cavitation and transpirational water loss, (ii) deciduous species that delay drought stress by shedding leaves in the dry season and that maximize resource capture during a limited growing season, and (iii) drought-intolerant species that maximize both below- and above-ground resource capture to increase competitiveness for light, at the expense of higher risk of drought mortality (Markesteijn & Poorter 2009). The impact of adaptation to drought has been investigated on seasonal variations of growth, showing that tree growth slowed down or completely stopped at the beginning of the dry season (Prévost & Puig 1981; Worbes 1999), for both evergreen and deciduous species (Pélissier & Pascal 2000; Fichtler, Clark & Worbes 2003; Couralet et al. 2010). To our knowledge, no study has focused on the annual time step.

In tropical moist forests, light also is a critical environmental factor that limits tree growth (Baker, Swaine & Burslem 2003). Therefore, species differ in their strategies to capture and use light, resulting in a continuum of shade tolerance (e.g. Maharjan et al. 2011) with one end corresponding to slow-growing, shade-tolerant species that can establish and grow in the low-light conditions of the understorey, and the other end corresponding to fast-growing pioneer species that need large light gaps to establish and grow (Swaine & Whitmore 1988; Kyereh, Swaine & Thompson 1999). This continuum in shade tolerance is associated with a set of species functional traits. Shade-tolerant slow-growing species tend to have high wood density, evergreen leaves and a small stature, whereas light-demanding fast-growing species often have low wood density, deciduous leaves and a large stature (Turner 2001; Baraloto et al. 2010; Maharjan et al. 2011). It is well known that species differ in their shade tolerance, but much less is known about how these strategies interact with strategies to capture and use water. The theory of optimal allocation predicts that shade-tolerant species cannot tolerate both shade and drought, as a consequence of a trade-off between below- and above-ground biomass allocation (Smith & Huston 1989). In a low-resource environment, plants maximize their surface area for intake of the most limiting resource. Therefore, shade tolerance should be achieved through higher biomass investment to the shoot system to enhance light acquisition, while drought tolerance should be achieved through higher biomass investment to the root system to enhance water acquisition. However, in the field, no significant trade-off between species drought and shade tolerance has been found so far (Engelbrecht et al. 2007; Markesteijn & Poorter 2009).

In this study, we used a long-term experimental site set-up in a Central African semi-deciduous moist forest to assess the impact of drought on tree growth. Contrasted silvicultural treatments were performed in this experiment, which advantageously brought about a gradient in resource availability. Tree growth was assumed to depend on drought, competition for resources and tree development stage, in a species-specific way. In species-rich ecosystems such as tropical moist forests, high diversity implies that the sample size for most species is limited, hindering a good fit of species-specific dynamics models. To overcome this problem, species were clustered into functional groups. A variety of methods has been used to group species (Swaine & Whitmore 1988; Favrichon 1994; Steneck & Dethier 1994; Gitay & Noble 1997; Gourlet-Fleury & Houllier 2000; Bellwood & Wainwright 2001; Picard et al. 2010), while not ensuring that the within-group similarity is maximum or that the number of groups is optimal (Dunstan, Foster & Darnell 2011). To solve these statistical shortcomings, the mixture model framework has been proposed (McLachlan & Peel 2000; Dunstan, Foster & Darnell 2011; Mortier et al. 2013). In the present study, we extended this framework by simultaneously (i) grouping species according to their response to the covariates, (ii) quantifying the relationship between growth and these covariates and (iii) selecting for each functional group the relevant covariates. Then, the species growth patterns obtained from the analysis of more than 2 74 000 annual growth observations were associated with particular species functional traits, to address the following questions: (i) Do species showing similar growth patterns have similar functional traits (light requirement, adult stature, wood density, leaf habit)? and (ii) Are differences in species responses to drought explained by differences in species functional traits?

Material and methods

Study Site

The study was conducted in lowland semi-deciduous moist forest at the M'Baïki experimental site (3.50°N and 18.00°E) in the Central African Republic. The climate is Oubanguian Guinean (Sillans 1958) with an average rainfall of 1738 mm and a dry season that lasts from December to February (when monthly rainfall is lower than average monthly evapotranspiration, i.e. 100 mm).

The site contains ten 4-ha permanent sample plots (200 × 200 m), each surrounded by a 50-m-wide buffer zone (Bedel et al. 1998). Three different silvicultural treatments of increasing intensity were implemented in the plots. Three plots have been left untouched (control treatment). The other seven plots were selectively logged in 1984–85 (logging treatment), and four of these were thinned 2 years later (logging + thinning treatment). Because the different silvicultural treatments were not homogeneous across the 4-ha plots, these plots were subdivided into 1-ha subplots that are the statistical units for analysis. All live trees ≥10 cm dbh (diameter at breast height) in the plots were botanically identified (229 tree species or morphospecies, see Appendix S1 in Supporting Information) and geo-referenced, and their girth measured annually between 1982 and 2008 to the nearest 0.5 cm using a metal tape measure.

Five soil categories with different textures and depths were identified on 9 of the 10 4-ha plots by Freytet & Tandeau de Marsac (1992). This study was based on these nine plots. On average, 50% of the trees grow on Haplic Ferralsols (IUSS-Working-Group-WRB 2006) or ‘sols ferralitiques typiques profonds’ (Freytet & Tandeau de Marsac 1992), 21% on Ferralsols (endoskeletic) or ‘sols ferralitiques gravillonnaires de profondeur’, 18% on Ferralsols (episkeletic) or ‘sols ferralitiques gravillonnaires de surface’, 10.5% on Leptosols or ‘sols peu évolués’, and < 0.5% on Plinthosols or ‘sols ferralitiques indurés de surface’.

Growth Quantification

Tree growth was expressed as the annual tree diameter increment (0.26 ± 0.44 cm year−1 on average, n = 2 74 026 observations, 36 ha measured annually between 1982 and 2008), computed from dbh measurements in two successive years only for live trees with no trunk anomalies. Measurement errors (965 observations) were further eliminated by retaining only diameter increments between −0.4 cm (corresponding to stem shrinkage during the dry season, Baker et al. 2002) and 10 cm (the 95th percentile of observed diameter increments in the fastest-growing species, Musanga cecropioides).

Growth Predictors

Annual tree growth was assumed to depend on (i) drought intensity, (ii) resource availability and (iii) tree development stage. The length of the dry season (number of months, lengthDS), the average rainfall during the dry season (mm, rainfallDS) and the average plant-available soil water content (mm, meanSW; hereafter ‘soil water content’) were used to assess drought intensity. Stand basal area (m² ha−1) and stand density (number of trees ha−1) are classical indices to assess the general environment of competition (Biging & Dobbertin 1995). They are easy to measure and were used to assess resource availability. Tree dbh (cm) was used to assess tree development stage (Table 1).

Table 1. Range of variations in tree growth predictors
Growth predictorsMean(min–max)UnitAbbreviation
Drought effect
Length of the dry season4.2(3–6)Number of monthsLengthDS
Average rainfall during the dry season36.1(2.7–60.3)mmRainfallDS
Annual average soil water content67.7(9.5–108.7)mmMeanSW
Resource availability effect
Stand basal area31.9(17.2–41.2)m² ha−1Basal area
Stand density624.9(381–800)stems ha−1Density
Tree size effect

The three drought indices jointly accounted for climatological drought (lengthDS, rainfallDS), which is the deficit in rainfall, and for local drought (meanSW), which is the effective water stress experienced by the tree that partially depends on the soil available water capacity (AWC). LengthDS and rainfallDS were computed from monthly rainfall records available at the Boukoko meteorological station, close to the M'Baïki site (A. Ougou, pers. comm., 1981–1989 and 1994–2008 periods).

MeanSW was computed using a simple water balance model based on a monthly time step. Monthly plant-available soil water content SWt (mm) was the balance between previous soil water content, monthly rainfall (Pt, mm) and evapotranspiration (Et, mm): SWt = SWt−1 + PtEt. Monthly rainfall records were the ones used to calculate lengthDS and rainfallDS, and evapotranspiration for month t was the average evapotranspiration across years for month t between 1945 and 1982 (Franquin et al. 1988). The monthly sequence of average evapotranspiration was identically repeated across years in the water balance model. Plant-available soil water content SWt is AWC when the soil is at the field capacity water content. Assuming that this is the case in the heart of the wet season, the starting month t0 of the water balance model was set to August 1981. Because AWC varies with soil texture, different AWC values were computed for each of the five soil types, using soil characteristics measured in five soil profiles dug in the study site (Appendix S2). AWC also varies with rooting depth that decreased from Haplic Ferralsols to Ferralsols (endoskeletic), then to Ferralsols (episkeletic) and finally to Plinthosols and Leptosols [the shallowest soils: less than one meter deep (Freytet & Tandeau de Marsac 1992)]. Rooting depth was set to 2 m for Ferralsols (episkeletic), a gravelly soil indicating the proximity of the bedrock (V. Freycon, pers. comm.). Following Canadell et al. (1996), we estimated that Haplic Ferralsols were 4 m deep, and Ferralsols (endoskeletic) 3 m deep. Because rooting depth in these two soil types could have been underestimated (Wonkam 2011), a sensitivity analysis to rooting depth was conducted and showed very little sensitivity for depths > 4 m and 3 m in Haplic Ferralsols and Ferralsols (endoskeletic), respectively (Appendix S3). After having run the water balance model for each soil type, meanSW was computed for each year as the average of SWt across the months of this year. We then attributed a meanSW value to each geo-referenced tree. Because annual tree surveys at M'Baïki are conducted in May, May was taken as the starting month when computing the three drought indices (the one-year time interval started in May of year t and ended in April of year t+1).

LengthDS and rainfallDS were positively correlated (r = 0.55, Pearson's test P-value < 0.05), but were both kept because of their complementary description of the dry season. Stand basal area and density were also positively correlated (r = 0.40, Pearson's test P-value < 0.05), but were both kept for their complementary description of the stand structure (Biging & Dobbertin 1995). All other correlations between growth predictors (including meanSW and tree dbh) were weak and not significant.

Growth Model

We modelled the diameter increment in year t of tree i belonging to species s = 1,…,S, where S is the number of species, in subplot p as ΔDistp = Χistp βs + εistp, where Χistp is the design matrix, βs is the vector of coefficients to be estimated for species s, and εistp is the residual error. Predictors Χistp included the intercept associated with species s, the two variables rainfallDS and lengthDS in year t, the variable meanSW in year t for tree i, the basal area and density of subplot p in year t, and the dbh of tree i in year t. We modelled growth as a finite mixture of regressions (DeSarbo & Cron 1988; McLachlan & Peel 2000), allowing us to simultaneously group together those species with a similar growth response to the environment, and quantify their response to the environment (Dunstan, Foster & Darnell 2011; and see discussion of the method in Appendix S4). Finite mixture models are based on the assumption that observations are distributed among unobserved groups. Because groups were defined at the species level and observations were made at the tree level, we extended the finite mixture of regressions for two-level data: when an individual was assigned to a species group, all its conspecifics were simultaneously assigned to the same group. The procedure ultimately yields a probability wsk called the posterior probability that species s belongs to species group k, thus giving an estimate of classification uncertainties. We then classified species by assigning each species to the group with the highest posterior probability. We considered a species as well classified if its maximum posterior probability ≥0.95.

In the mixture modelling framework, selecting for each group the significant covariates using classical methods would have been computationally intensive given the high number of observations (n = 2 74 026). Therefore, we extended the finite mixture of regressions on two-level data in such a way that different predictors could be selected in each group, and that the predictors selection is realized simultaneously with the species grouping process, using a lasso type penalization (Yuan & Lin 2006; Khalili & Chen 2007). Instead of fitting 2p models when there are p predictors, the lasso penalization selects the most meaningful predictors by shrinking the coefficients of non-meaningful predictors to zero. Therefore, all predictors need to be scaled when fitting the growth model. The selection is performed separately for each species group, in such a way that each one ends up with its own set of predictors. Using lasso penalization for variable selection also offers a way to deal with the temporal autocorrelation between growth measurements (see Appendix S4). The finite mixture of regressions on two-level data with lasso penalization was fitted using an expectation-maximization algorithm (Dempster, Laird & Rubin 1977; Khalili & Chen 2007; see Appendix S4). The R and C codes based on the flexmix (Leisch 2004) and glmnet (Friedman, Hastie & Tibshirani 2010) packages are available from the authors.

Species Traits

To determine whether species with similar growth patterns had similar functional traits, we tested whether species groups were related to traits linked with strategies of resource acquisition and tolerance to stress (leaf habit, wood density), light requirement (maximum growth rate, regeneration guild) and life-history strategies (adult stature, assessed through maximum dbh).

Information on leaf habit (deciduous or evergreen) was extracted from local floras and supplemented with field work (J.-L. Doucet, pers. comm.). Mean values for wood density were extracted from the CIRAD data base on wood technological properties (; Gidoin 2010). Maximum growth rate was calculated from the data used to fit the growth model and was expressed as the 95th percentile of the diameter increment distribution for each species (all trees with dbh ≥10 cm) observed at M'Baïki. Following Poorter et al. (2003), we used it as an indicator of species light requirement. Information on regeneration guilds was collected from the literature and from field work (J.-L. Doucet, pers. comm.). The terminology for regeneration guilds followed Hawthorne (1995), with three guilds: pioneers (P), non-pioneer light-demanders (NPLD) and shade-bearers (SB). Finally, maximum dbh was defined as the 95th percentile of the dbh distribution for each species. To improve the precision of this percentile, dbh distributions were taken from forest inventories encompassing large areas in the Central African Republic: 30 182 0.5-ha plots covering about 1 600 000 ha (Réjou-Méchain et al. 2008) were added to the M'Baïki plots.

Welch's one-way analysis of variance (Welch's anova) and Dunnett's modified Tukey–Kramer pairwise multiple post hoc comparison test (Dunnett's post hoc comparison test) adjusted for unequal sample sizes and unequal variances were used to assess whether maximum growth rate and maximum dbh significantly differed across species groups. This test was not performed for wood density because of an insufficient number of observations in most groups. A chi-square test was used to determine whether species were independently distributed across P, NPLD and SB guilds, and species groups. The same was done for the deciduous and evergreen types. To prevent the results from being contaminated by misclassified species, only those species with a posterior probability ≥0.95 of belonging to a particular group were retained. If a trait was missing for a species, it was also excluded from the analysis (the total number of species considered for the analysis is shown in Table 2).

Table 2. Number of well-classified species (probability of belonging to the group ≥0.95) for which ‘maximum growth rate’, ‘regeneration guild’, ‘maximum dbh’, ‘wood density’ and ‘leaf habit’ traits were available
Species groupNumber of well-classified species with information on
Maximum growth rateRegeneration guildMaximum dbhWood densityLeaf habit


Nine species groups with contrasted growth responses to rainfallDS, lengthDS, meanSW, basal area, density and dbh were found by the finite mixture of regressions (Appendix S5). Groups were labelled G1 to G9 based on the rank of their model's intercept, with G1 having the smallest intercept (Fig. 1a). The 9 groups were highly variable in their number of species, which ranged from 3 (for G9) to 40 (for G1; Table 3). The classification method proved to be highly efficient, with 70% of the species being assigned to a particular group with a probability ≥0.95. Expectedly, this probability decreased as the number of growth observations for a species decreased.

Table 3. Species group distribution in the mixture and group characteristics. For the number of species, the number in parentheses gives the number of well-classified species (with a probability of belonging to the group ≥0.95). Average annual growth is shown with its standard deviation. Computation of the group proportions in the mixture is given in Appendix S4
Species group labelNumber of speciesNumber of growth observationsProportion in the mixtureRelative abundance (%) in 2007Average annual growth (cm year−1)
G140 (22)393110.1414.330.09 ± 0.19
G231 (21)670250.2523.340.15 ± 0.25
G335 (22)337500.1211.980.19 ± 0.29
G413 (8)340690.1211.310.26 ± 0.33
G521 (18)276660.19.830.25 ± 0.39
G629 (17)258990.099.990.29 ± 0.36
G734 (28)244540.098.380.41 ± 0.48
G823 (20)142120.056.70.67 ± 0.70
G93 (3)76400.034.091.08 ± 1.20
Figure 1.

Model parameter estimates for each species group: (a) intercept of the model, parameters associated with (b) diameter, (c) stand basal area, (d) stand density, (e) average rainfall during the dry season, (f) length of the dry season and (g) soil water content. A zero estimate means that the predictor was not selected in the species group growth model. Species groups are ordered along their estimate of the intercept coefficient (that is potential growth). Parameter estimates are of the same order of magnitude and have no unit because all predictors were centred and scaled. Panel (h) gives an outline for interpreting the relationship between growth and resource availability, as follows from the model intercept, stand basal area (an index for the availability of all resources) and rainfallDS (an index for water availability): tree growth increases when resources increases, and the slope of the growth response increases with potential growth, with the slowest-growing species group (G1) being the least sensitive to a change in resource availability.

Functional Groups

Maximum growth rate significantly differed between species groups (Welch's anova, F7,52.894 = 154.32, P-value< 0.0001, Fig. 2a). Post hoc comparisons categorized the species groups into seven supergroups: three slow-growing singleton supergroups (G1, G2 and G3; 0.38, 0.62 and 0.76 cm year−1, respectively), one moderate-growing supergroup consisting of G4, G5 and G6 (0.96 cm year−1) and three fast-growing singleton supergroups (G7, G8 and G9; 1.29, 1.92 and 3.12 cm year−1, respectively). Maximum growth rate correlated very closely with the intercept of the growth models (Pearson correlation coefficient, r = 0.998, n = 9 groups). This intercept represents the potential growth rate, that is, the mean growth rate when the effects of drought intensity, resource availability and development stage have been accounted for. Hence, this intercept can be interpreted as a measure of the species’ intrinsic light requirement.

Figure 2.

Relationship between potential annual growth rate (cm, intercept of the growth model) and the species group average of (a) maximum growth rate, (b) maximum diameter and (c) wood density. Grey squares correspond to mean values and whiskers to standard deviation. Letters in (a) and (b) identify species groups that significantly differed for the trait considered (Welch's anovas followed by Dunnett's post hoc comparison tests). Group G9 and wood density were excluded from the analysis because of a small number of species observations. Only well-classified species (probability of belonging to the group ≥0.95) were considered.

Although fewer observations were available for wood density, species in slow-growing groups tended to have high wood densities, whereas species in fast-growing groups tended to have low wood densities (Fig. 2c).

A significant relationship was detected between species groups and leaf habit (Χ2 = 22.35, P-value = 0.001, Fig. 3b): species in the slowest-growing groups (G1, G2, G3) were mainly evergreen, whereas those in the fast-growing groups G7 and G8 were mainly deciduous. These results were unchanged when examining the distribution of observations across leaf habit types instead of the distribution of species (Fig. 3d): evergreen trees made more than 70% of groups G1, G2 or G3; deciduous trees made more than 70% of group G7; and the moderate-growing group G4 mainly contained deciduous trees.

Figure 3.

Species or observation joint frequency distributions across species groups and categorical functional traits at M’Baïki, Central African Republic: (a) number of well-classified species across species groups and regeneration guilds (P pioneer, NPLD non-pioneer light-demander, SB shade-bearer); (b) number of well-classified species across species groups and leaf habit types (deciduous or evergreen); (c) within-group proportions of well-classified growth observations across regeneration guilds; (d) within-group proportions of well-classified growth observations across leaf habit types. Within-group proportions do not sum to 1 because they are computed among the set of all observations (the gap to 1 gives the within-group proportion of ill-classified observations). These proportions were used to interpret the growth response to drought indices. Groups marked with an asterisk in (a) and (b) were significantly related to one regeneration guild (χ2 = 88.02, P-value< 0.001), or one leaf habit type (χ2 = 22.35, P-value = 0.001).

A significant relationship was detected between species groups and regeneration guilds (Χ2 = 88.02, P-value < 0.001, Fig. 3a). Species in the slowest-growing groups (G1, G2) were mainly shade-bearers, species in the fast-growing group G7 were mainly light-demanders, species in the two fastest-growing groups (G8, G9) were mainly pioneers, while a mix of light-demander and shade-bearer species made the bulk of the intermediate groups G4, G5 and G6. These results were unchanged when examining the distribution of observations across regeneration guilds instead of the distribution of species (Fig. 3c): shade-bearer trees made more than 90% of groups G1, G2, pioneer trees made almost 100% of group G9, while groups G4 and G7 mainly contained non-pioneer light-demanding trees.

Finally, maximum dbh significantly differed between species groups (Welch's anova, F7,51.536 = 13.29, P-value< 0.0001, Fig. 2b). Post hoc comparisons categorized the species groups into two supergroups: species in slowest-growing groups G1 and G2 had the lowest maximum dbh (40.7 cm), while species in the moderate- and fastest-growing groups (G4 to G8) had the highest maximum dbh (82.6 cm). Group G3 occupied an intermediate position (61.8 cm). Although group G9 was not included in the analysis because of a small number of species, its average maximum dbh was consistent with the one of the moderate- and fast-growing groups.

Factors Explaining Growth

The linear growth model with specific parameters for the nine groups explained 26% of the total variability in growth rates (Appendix S6). The explanatory power varied between groups, from 3.9% (G1) to 14.8% (G4; Table 4). Growth was mainly explained by stand basal area (all groups but G5), then rainfallDS (G1, G3 and G6 to G9) or dbh (G2, G4).

Table 4. Proportion (%) of the total variability in annual tree diameter increments explained by each growth predictor. One analysis of variance table was constructed for each species group. A dash means the predictor was not selected as a significant variable
Species groups% explained variancePredictor's share (%) in the total sum of squares
Basal areaDiameterRainfallDSDensityLengthDSMeanSW

Drought intensity

Among the two correlated predictors of drought, rainfallDS captured the main effect of drought, while lengthDS acted as a correction variable. Tree diameter increments in all groups increased when rainfallDS increased (all coefficients were positive, Fig. 1e) and, in five of these groups, tree diameter increments further decreased when lengthDS increased (Fig. 1f): for a given amount of dry season rainfall, a longer dry season logically proved to be more detrimental to growth. A surprising reverse effect of lengthDS was observed in two groups (G4, G9), whereas this predictor had no significant additional effect on growth in the two remaining groups (G1, G3). The slopes of the growth response to rainfallDS and lengthDS were arranged along a gradient that paralleled the potential growth rates gradient: the slower growing the species, the less sensitive it was to rainfallDS and lengthDS.

In five of the nine groups, tree diameter increments increased with meanSW (Fig. 1g), whereas the reverse was observed in one group (G7). Conditionally on the other predictors being included in the model, meanSW had no significant effect on growth in the three remaining groups (G1, G2 and G6).

Resource availability

Between the two correlated predictors of resource availability, stand basal area captured the main effect of competition for resources, while stand density acted as a correction variable. Tree diameter increments decreased in all groups when stand basal area increased (i.e. when resource availability decreased, Fig. 1c). As for rainfallDS, the slopes of the growth response to stand basal area were arranged along a gradient that almost perfectly paralleled the potential growth rates gradient (compare Fig. 1a,c): the slower growing the species, the less sensitive it was to stand basal area.

Once the effect of basal area was accounted, stand density did not have any correction effect on growth for species groups G5, G7 and G9. In contrast, for the six other groups, tree diameter increments further decreased when stand density increased (Fig. 1d), evidencing a sensitivity to the number of competitors: for a given stand basal area, the larger the number of trees, the smaller the tree diameter increment. In other words, species from these groups experienced higher growth in a forest plot made of few large trees than in a plot with the same basal area made of many small trees.

Tree development stage

Tree diameter increments increased with increasing tree dbh (Fig. 1b), except for groups G3 and G9 where growth was not sensitive to tree size (null coefficients). Group G5 showed the highest growth response to dbh.


Growth Responses Reveal Functional Groups of Tropical Tree Species

Finite mixture modelling proved to be an efficient method to analyse long-term records of tree growth and to build functional groups of tropical tree species: species showing similar growth patterns had similar values for several functional traits. It is not surprising that functional groups could be defined on the basis of species growth performance, as growth scales physiological and biomass allocation mechanisms at the plant level. Nevertheless, only the size of the M'Baïki data set, which is unique and unprecedented for Africa, allowed us to identify this relationship so clearly.

The nine species groups were ordered along a light requirement gradient (as defined by the potential growth rate) that paralleled a gradient of maximum growth rate, thus confirming that the latter trait can be used as a proxy for the former (Poorter et al. 2003). The light requirement gradient was clearly identified because light is a major growth limiting factor in tropical moist forests (Baker, Swaine & Burslem 2003), and because the M'Baïki site encompassed stands with contrasted light levels due to silvicultural treatments.

The species grouping at M'Baïki was consistent with those established in other moist forests (Lieberman et al. 1985; Swaine & Whitmore 1988; Condit, Hubbell & Foster 1996; Favrichon 1998; Poorter et al. 2003; Nascimento et al. 2005; Chazdon et al. 2010) and conformed to the big picture that species can be ordinated along two nearly independent axes that sum up correlated traits (Turner 2001; Gourlet-Fleury et al. 2005): adult stature, life span and mortality rate for the first axis; light requirement, growth rate and wood density for the second axis. Nevertheless, the correlation between light requirement and wood density at M'Baïki was not as clear as one could expect, probably because heavy wood species were under-represented: no species at M'Baïki had a wood density > 1 g cm−3, and the average wood density (0.57 g cm−3, with an interquartile range of 0.45–0.71 g cm−3) was lower than the one reported for African moist forests (0.60 g cm−3 with an interquartile range of 0.58–0.67 g cm−3; Henry et al. 2010). This under-representation could be due to a sampling bias (African wood densities are more documented for light woods than for dense woods), or it could indicate that species at M'Baïki are overall more light-demanding than in other African moist forests, suggesting that the forest has recently been disturbed.

Finally, the growth–diameter relationships and growth responses to competition for resources were consistent with the presumed ecology of the species. Most of the trees (dbh ≥10 cm, 191 species) showed increasing growth rates as they got larger. Interestingly, the two species groups (G3 and G9) for which growth was independent of diameter had a particular growth response to stand density. G3 had the strongest negative growth response to density, suggesting that trees in this group were strongly sensitive to the competition with their neighbours. Given their adult stature, species in G3 can be identified with the ‘typical subcanopy species’ (Poorter et al. 2005), consisting of overtopped adult trees in the intermediate strata between understorey and canopy. On the contrary, G9 was insensitive to density, which is consistent with its pioneer nature that is able to escape competition from neighbours by growing rapidly into the canopy.

Slow Growth Confers Drought Tolerance

Overall, tree growth decreased when drought increased, regardless of whether drought was measured by average rainfall during the dry season, length of the dry season or soil water content. Nevertheless, soil water content was the least efficient predictor of growth, suggesting that climatic water deficits were buffered on deep soils (Haplic Ferralsols) (Wonkam 2011).

Species were arranged along two parallel gradients of drought tolerance and of tolerance to shortage in all resources (as indicated by basal area): slow-growing species were the least sensitive to variations in water availability or to variations in resource availability. The key functional trait to locate a species along these gradients was the potential growth rate. This convergence of responses supports the hypothesis of a general continuum of plant strategies, from the resource-conservative strategy (associated with slow growth, drought tolerance, low sensibility to basal area, low light requirement) to the strategy of rapid acquisition of resources (Chapin 1980; Grime et al. 1997; Reich et al. 2003; Díaz et al. 2004; Wright et al. 2004). The association between the resource-conservative strategy and both drought and shade tolerance can be explained by slow growth leading to slow tissue turnover, which reduces resource loss and therefore reduces dependency on the environment for the acquisition of new resources (Chapin, Autumn & Pugnaire 1993; Sterck et al. 2011).

The congruency between drought and shade tolerance can also be explained by a trade-off between hydraulic safety (i.e. resistance to cavitation, low hydraulic conductivity) and hydraulic efficiency (high hydraulic conductivity; Markesteijn et al. 2011a), with pioneer and deciduous species having smaller hydraulic safety margins than shade-tolerant and evergreen species (Markesteijn et al. 2011b). Hydraulic conductance indicates the species strategy for water transport and is related to photosynthesis and growth (Reich et al. 2003). Moreover, hydraulic conductance and vulnerability to cavitation are correlated with wood density (Markesteijn et al. 2011b), which is in turn correlated with shade (Valladares & Niinemets 2008) and drought tolerance (Chave et al. 2009; Maharjan et al. 2011). These correlations reinforce the idea of a single axis summarizing multiple traits, which represents a trade-off between the conservation and the rapid acquisition of resources.

Finally, our results refute the idea that, because of a trade-off between below- and above-ground biomass allocation, shade tolerance and drought tolerance would be incompatible (Smith & Huston 1989) or independent (Maharjan et al. 2011). However, this latter mechanism remains more theoretical than based on field evidence (Niinemets & Valladares 2006; Engelbrecht et al. 2007; Markesteijn & Poorter 2009), and we concur with Reich et al. (2003) to argue that shade tolerance is a whole-plant strategy with associated traits that enhance resource conservation rather than resource acquisition (lower leaf, stem and root respiration rates and longer leaf life span).

Drought Tolerance was Indirectly Related to Leaf Habit

A significant relationship between species drought tolerance and leaf habit was found, with slow-growing drought-tolerant groups mainly composed of evergreen species and fast-growing drought-sensitive groups mainly composed of deciduous species. Given that leaf habit is a proxy for leaf life span (in short, evergreen leaves are long-lived and deciduous leaves are short-lived; Givnish 2002), this relationship corresponds to a pattern where slow-growing species have long-lived leaves, and fast-growing species shorter-lived leaves (Poorter & Bongers 2006). It is probable that leaf life span would better predict drought tolerance than leaf habit (Markesteijn & Poorter 2009; Markesteijn et al. 2011b): first, because species showed a gradient of drought tolerance rather than a dichotomic response, and secondly, because any species group with a clear dominance of evergreen or deciduous species encompassed divergent responses to the drought indices. Thus, among the groups dominated by evergreen trees (G1, G2, G3), and conditionally on the other predictors being included in the model, G3 was sensitive to soil water content, whereas G1 and G2 were not. The two groups dominated by deciduous trees (G4, G7) showed opposite responses to the length of the dry season and to soil water content. As for the fastest-growing group G9, it included a substantial proportion of evergreen species, but these, for example M. cecropioides (J.-L. Doucet, pers. comm.), could actually be leaf-exchangers, that is, species that hold their leaves for about 1 year, replacing them just before or after they drop the previous year's foliage (Reich 1995).

Therefore, our results are consistent with the idea that leaf habit should be treated as a continuous variable that is, leaf life span (Williams, Bunyavejchewin & Baker 2008) to explain drought tolerance. Further studies on tree species drought tolerance should focus on leaf longevity rather than on leaf habit.

Sensitivity to Drought or Sensitivity to Light? The Ambiguous Effect of Dry Seasons

Two species groups (G4, G9) had unexpected growth responses to the length of the dry season. These exceptions may follow from the greater light availability during the dry season at M'Baïki, due to little cloud cover (Graham et al. 2003). For those species like pioneers (group G9) that are highly sensitive to light availability, a longer dry season (for the same amount of rainfall during the dry season) may mean more light, and thus better growth.

One species group (G7) had an unexpected growth response to the average soil water content. This counter-intuitive response could be explained by a reduction in light availability due to increased rainfall, combined with an adaptation to seasonal drought. The fast-growing species in group G7 are mostly deciduous and large, leaf fall being a drought avoidance strategy during the dry season (Worbes 1999; Schöngart et al. 2002) that offsets the increased drought sensitivity due to their size (Condit, Hubbell & Foster 1995; Van Nieuwstadt & Sheil 2005; Nepstad et al. 2007; Phillips et al. 2010). Because they have high growth rates (and so high light requirement) during the favourable growth period, these species may be sensitive to reductions in light availability associated with cloud cover.


The combination of an unmatched data set on adult tree growth, unique in terms of the number of observations, length of the time series and location (African semi-deciduous moist forests), with the use of a powerful, original modelling method, have given us new insights in the potential effects of drought on tropical tree communities. Overall, we found that tree growth decreased when drought increased and that shade-tolerant species, characterized by a low potential growth rate and thus a conservative strategy of resource use, were the least sensitive to drought. Consequently, a shift in forest composition towards slow-growing shade-tolerant evergreen species might occur if drought intensity increased. This prediction is in apparent contradiction with the shift towards deciduous NPLD species found in Ghanaian forests following long-term low-intensity droughts (Fauset et al. 2012). This contradiction might arise from the ambiguous effect of dry seasons, when reductions in rainfall are associated with greater light availability. This underlines that accurate predictions for the future of tropical forests under climate change should take into account the variation in light availability associated with drought.


We wish to thank the ARF Project (Appui à la Recherche forestière), its six funding partners (AFD – Agence Française de Développement, CIRAD – Centre de Coopération Internationale en Recherche Agronomique pour le Développement, ICRA – Institut Centrafricain de Recherche Agronomique, MEFCPE – Ministère Centrafricain des Eaux et Forêts, Chasse, Pêche, chargé de l'Environnement, SCAC/MAE – Service de Coopération et d'Actions Culturelles, and SCAD – Société Centrafricaine de Déroulage), Laurent Cerbonney, Emilien Dubiez, the field workers at the M'Baïki station who participated in data collection and data capture, and Alfred Ougou who provided rainfall data from the Boukoko station. We are grateful to Jean-Louis Doucet for his expertise on Central African moist forest tree species' characteristics and ecology. The water balance model was developed in the context of the CoForChange project, funded by the ERA-Net BiodivERsA, with the French National Research Agency (ANR, 2008-29489-62704-33) as part of the 2008 BiodivERsA call for research proposals.