We conducted our study in the region of Nizanda, southern Mexico (16.66°N, 95.01°W; Figure 1). The region has a mean annual temperature of 27.7 °C and a mean annual precipitation of 902 mm year^–1, with a marked rainy season spanning from mid-May to mid-October (precipitation > 50 mm month^–1). Tropical dry forest (TDF) is the dominant vegetation type in the region and large undisturbed forest patches are still present in the landscape, alongside areas devoted to food production (mechanized and shifting agriculture, and cattle ranching) and extensive patches of secondary vegetation in different stages of succession. TDF in this region occurs on limestone and siliciclastic phyllite, two contrasting parental materials whose weathering results in different soil types. Limestone outcrops are associated with major faults in the terrain and can reach elevations of ca. 500 m above mean sea level (amsl). By contrast, the terrain on phyllite is hilly with much lower elevations (ca. 280 m amsl at most). These two substrates represent the main parental material of the most common soil types (Cambisols, Leptosols, and Regosols) on which the tropical dry forest develops (Miguel-Talonia, unpublished data). The physical characteristics of these soils such as color (an organic matter content indicator) and texture differ notably between the two substrates. Overall, limestone-derived soils are darker and have higher clay contents, and in some places of the limestone mountains the high clay content results in shrink-swell cracks during the dry season; consequently, water availability for plants is higher in limestone- than in phyllite-derived soils. Therefore, soil water content emerges as one of the major factors potentially driving differences for TDF development on these substrates (Covelo et al., 2017). In Nizanda, physiognomic differences between TDF developing on the two substrates are evident, and these seem to be linked to variations in structure and diversity between the two TDF subtypes.

In 2008 we established seven 500 m² permanent forest plots in each forest subtype (14 permanent plots in total; Figure 1). At present, we continue monitoring these plots but for this study we used data for the 2008–2020 period. The plots are subdivided in five 100-m² transects, each of which is further subdivided into four 25 m² quadrats. We used a stratified, nested sampling design for measuring trees in the plots. Stems with a diameter at breast height (DBH) ≥ 5 cm were measured in the whole 500 m² plot area, medium-sized stems with a DBH ≥ 2.5 cm were measured in a 375 m² plot subarea, while small stems with a DBH ≥ 1 cm were only measured in a 125 m² plot subarea. In each transect, 25 m² quadrats were randomly assigned a minimum DBH criterion (5, 2.5 or 1 cm). All trees meeting the inclusion criterion in the quadrat were tagged, identified to species (96.1 % of all trees), and their DBH and heights were measured. At each annual census, survival, DBH, and height were recorded, as well as new recruits meeting the inclusion criteria.

Based on the annual vegetation censuses of permanent plots, we estimated four forest structure variables [basal area, aboveground biomass (AGB), tree density, and canopy height], and three community diversity variables [species richness (⁰D), common species (¹D), and dominant species (²D)]. Aboveground biomass (AGB) was estimated by using a broadly used allometric equation (Chave et al., 2014) that integrates DBH, height, and species-specific wood density measurements. Canopy height was calculated as the average height of trees on the top canopy layer (details for canopy height estimation are provided in Supplementary Material 1). Community diversity values (⁰D, ¹D, and ²D) correspond to Hill numbers of order q = 0 (total species richness), q = 1 (typical species), and q = 2 (dominant species), which represent the effective number of species in the community when the weight given to rare species is gradually reduced (Chao et al., 2014). These diversity metrics were estimated with the vegan package (Oksanen, 2007) in R by using species abundances. All community variables were estimated on an annual basis per plot.

To assess substrate effects on TDF structure and diversity, we estimated the mean value for the seven response attributes per substrate on an annual basis. Next, 95 % confidence intervals around these means were constructed through bootstrapping with 1,000 iterations per estimation. Mean values and their confidence intervals were plotted against time to illustrate the temporal trajectories of structure and diversity attributes on both substrates.

To compare the dynamics in TDF attributes between substrates, we calculated a mean value per TDF attribute per substrate (including all plot/year combinations over the study period). Then, we standardized individual plot-year estimates by dividing these values by their substrate mean. This allowed us to express plot variability relative to their substrate mean, a procedure that controlled for larger variation caused by larger absolute values in forest attributes. Next, we estimated substrate mean values again with 95 % bootstrapped confidence intervals on an annual basis. Standardized TDF attributes were plotted against time to visualize the temporal trajectories of the TDFs growing on the two substrates. For all forest attributes, a value of 1 represented the substrate average value for the entire study period.

Next, we assessed quantitative differences in TDF dynamics between substrates by comparing plot deviations from their substrate mean using linear mixed-effects models (LMM). For each standardized TDF attribute, we computed the absolute deviations between individual plot values and their entire-period substrate mean on an annual basis (i.e., |1–standardized plot value|). We then fitted an LMM of individual plot-year deviations as a function of substrate, adjusting a random intercept for each permanent plot. By including a plot-dependent intercept, we controlled for the pseudo-replication arising from repeatedly measuring the same plots over time. To assess whether variation in TDF dynamics is determined by substrates, we extracted the substrate coefficient from the previously fitted LMMs and then estimated their 95 % bootstrapped confidence intervals with 1,000 iterations. Coefficient estimates and their confidence intervals were plotted separately for each forest attribute to show whether TDF dynamics was more intense either on limestone or on phyllite.

We also compared differences in demographic rates (net rate of change, recruitment, growth, mortality, and residence time) between substrates for basal area, AGB, tree density, and species richness (⁰D). These four variables are, to some extent, additive forest attributes which can be decomposed into their gain and loss demographic components. To this end, we analyzed the individual tree trajectories over time (2008–2020) and determined, on an annual basis, what demographic process a given individual was undergoing: either recruitment, survival, or death. Based on these demographic classification, we determined the contribution of an individual toward a specific state variable (positive, negative, or null), and then added those change rates per demographic process, year, and plot (Muñoz et al., 2021). We performed this procedure for all four state variables analyzed. Later, we computed residence times per state variable as in Galbraith et al. (2013) by using Equation 1.

where τ_Y is the residence time of a unit of the state variable Y, M¯Y is the average value for Y during a yearly period, and Y¯P is the productivity in Y during the yearly period resulting from the addition of recruitment and growth. We did not differentiate between biomass and density formulas as in Galbraith et al. (2013) to enable direct comparisons between the residence times of the four attributes analyzed. Annual demographic rates were averaged per plot and then per substrate to obtain substrate-level estimates. Bootstrapped confidence intervals (95 %, 1,000 iterations) were computed to compare demographic rates between substrates.

To assess whether sudden decreases and increases in TDF attributes were associated to El Niño and La Niña events (which are known to promote dry and rainy weather conditions, respectively), we ran non-parametric (Spearman) correlations between the Oceanic El Niño Index [ONI, referred to as NINO3.4 in van Oldenborgh et al. (2021)] and the net rates of change in all seven forest attributes. ONI values were retrieved from the National Oceanic and Atmospheric Administration historical El Niño/La Niña episodes (National Oceanic and Atmospheric Administration [NOAA], 2022). Net rates of change were retrieved from the analysis of demographic rates for basal area, AGB, tree density and species richness; for canopy height and diversity metrics ¹D and ²D, rates of change were simply computed as (Y_t+1–Y_t)/Y_t. Correlation coefficients and their respective P-values were reported per substrate. Although this analysis was not essential to test the core hypotheses of this work, we considered it would be appropriate to explain notorious tendencies on the dynamics of TDF attributes.

To visualize differences in species composition between TDF substrates, as well as their differences in dynamics, we used a non-parametric multidimensional scaling ordination (NMDS) of species composition and abundance data for all plots and years. Species were considered as variables and plot-year combinations were considered as observations. We used Bray-Curtis distance to estimate the dissimilarity matrices for the analysis, and projected the NMDS scores into a bidimensional scatterplot. Permanent plots from the same substrate were enclosed into an ellipse to show the ordination space occupied by each substrate based on its species composition. The NMDS ordination was performed using the vegan package in R (Oksanen, 2007).

To quantify differences in species composition dynamics between substrates, we computed three metrics of composition dynamics: annual mean distance, dispersion from centroid, and distance from initial to final point. Annual mean distance quantified the distance that any plot moved on average on the ordination space during an annual period. Dispersion from centroid quantified how broadly or tightly spread all datapoints belonging to each plot were on the ordination space. Distance from initial to final point, as its name implies, quantifies the distance between the 2008 and 2020 datapoints on the ordination space, and represents a proxy of circularity (i.e., floristic stability) in individual plot trajectories. To assess differences in these three metrics between substrates, we performed an LMM for annual mean distance and dispersion from centroid, and a standard linear regression for distance from initial to final point. In the three models, substrate was set as the fixed factor. For the LMMs, a plot-dependent random intercept was adjusted to account for the temporal autocorrelation between datapoints. Coefficients were extracted from the three models, and their 95 % bootstrapped confidence intervals were computed. The coefficients along with their confidence intervals were plotted to show differences in composition dynamics between substrates.

The datasets and the R code necessary to replicate all analyses, tables, and figures of this study were deposited in a Zenodo repository for reproduction and validation purposes (Muñoz et al., 2023).

The TDF on limestone had consistently higher values for almost all forest attributes than the TDF on phyllite except for canopy height. The largest differences between substrates were observed for diversity variables, with limestone TDF displaying higher diversity, while differences in structure attributes were less pronounced (Figures 2A–G).

The dynamic behavior of all TDF attributes was remarkably similar over time in terms of direction at both substrates, with limestone and phyllite TDF following nearly parallel trajectories (Figures 2A–G and Supplementary Figure 1). This weak differentiation between substrates was particularly evident for the standardized TDF attributes, which followed almost identical patterns over time on limestone and phyllite, both in terms of direction and magnitude of the variation on their dynamics (Supplementary Figure 1). Similarly, LMM modeling did not reveal a significant role of parental bedrock on the variability of TDF dynamics; the analysis of plot deviations in standardized TDF attributes showed that substrates did not yield large differences in dynamics for any of the attributes analyzed (Figure 2H).

Temporal trends in net change, recruitment, growth, and mortality were remarkably similar between substrates over time (Figures 3A–D). Also, the analysis of mean demographic rates did not reveal any differences between demographic rates for any of the four TDF attributes analyzed from this demographic perspective (basal area, AGB, tree density, and species richness; Figures 3E–H). Similarly, residence times, which were estimated from recruitment, growth, and attribute values, did not show differences for any of the four TDF attributes analyzed (Figures 3I–L).

The NMDS analysis based on species composition recorded over the study period revealed a clear segregation of the plots from the two forest types on the ordination space (Figure 4A). The overlap of the ordination spaces corresponding to the forests on limestone and phyllite was relatively small, pointing to large differences in the presence and abundance of species between substrates. Interestingly, however, the dynamics of floristic composition did not show such a between-substrate differentiation: all three NMDS-based metrics (annual mean distance, dispersion from centroid, and distance between initial and final points) were indistinguishable between substrates (Figure 4B). Therefore, despite clear differences in species composition between the two substrates, there is no evidence of differences in composition dynamics over time.

In the two forest subtypes, structure and diversity attributes increased around 2010–2011, remained high for several years, and then showed a sudden decrease around 2015–2016 (Figures 2A–G and Supplementary Figure 1). These generalized increments and reductions in vegetation development show some degree of association with both the rainy years produced by La Niña event of 2010–2012 and the drought years arising from El Niño event of 2015–2016 recorded in the region (Supplementary Figure 2).

Our results confirmed previous observations that structure and diversity attributes were generally higher in limestone than in phyllite old-growth TDF. The only exception to this pattern was canopy height, which showed higher values in phyllite (Figure 2), probably suggesting different tree architectures and allometric relationships across communities, given that trees in the forest on limestone are taller. The fact that structure and diversity differences were consistent over the entire study period (2008–2020) suggests that this is a permanent pattern rather than a peculiarity of exceptionally dry or wet years. More importantly, these results provide strong support to our first hypothesis, which states that limestone represents a more favorable environment for plant growth and thus for TDF development due to higher water content and nutrient availability, because lime materials enhance the availability and plant uptake of elements such as molybdenum, phosphorus, calcium, and magnesium (Brady, 1984).

Tropical dry forests are typically water limited, as they grow in areas with high temperatures, relatively low annual precipitation and long dry seasons (≥ 3 months, but often up to 6 months) (Sánchez-Azofeifa et al., 2005; Allen et al., 2017). Therefore, they should benefit from a substrate characterized by higher soil water content to maximize water uptake periods. Limestone seems to provide such an environment. Preliminary analyses showed that soils derived from limestone have higher clay and lower sand contents that soils derived from phyllite, with gravimetric humidity being 37 % higher in limestone soils (Sandoval-Granillo, 2020). Therefore, water limitation should be weaker on limestone than on phyllite, consequently yielding a greater potential for vegetation development (Terra et al., 2018). Our observations for four out of the five forest structure variables analyzed here (basal area, AGB, tree density and canopy height) support this idea.

The higher diversity at all levels analyzed (Hill numbers of order q = 0, 1, and 2) in limestone TDF can also arise from weaker limitations related to soil water content and fertility in this substrate. Limestone-derived soils have higher conductivity, Ca, and P concentrations, while phyllite-derived soils have a higher concentration of total N, nitrates, and ammonia (Sandoval-Granillo, 2020). Given the high proportion of legume species in TDFs and their frequent symbioses with nitrogen-fixing Rhizobium bacteria (Gei et al., 2018), nitrogen limitations are probably unimportant for these forests. Therefore, the higher diversity documented here for limestone might be, at least partially, related to higher soil water content but also to higher P availability, an element known to limit vegetation development on TDFs (Campo and Vázquez-Yanes, 2004) and other forest ecosystems (Hahm et al., 2014). Empirical observations from sites on the two substrates also suggest that limestone environments are rockier and highly heterogeneous, with many small and irregular patches of developed soils, shallow soils, and bare rock (Miguel-Talonia, unpublished data). Thus, limestone environments may provide a wider variety of microhabitats than phyllite, thus allowing more species to establish on this TDF community (Ribeiro et al., 2007; Díaz-Castellanos et al., 2022).

Unexpectedly, the dynamic behavior of TDF attributes was very similar between the two substrates, not only in terms of direction (increases and decreases in value) but also regarding its variability (magnitude of changes). In other words, changes in forest attributes over time were parallel in both forest types (Supplementary Figure 1), with all attributes showing non-significant differences in variation (Figure 2H). The finding of similar dynamics between limestone and phyllite TDF plots was also supported by the analysis of demographic rates. Rates of net change, recruitment, growth, mortality, and residence time were similar between substrates for the four forest attributes analyzed from a demographic perspective (basal area, AGB, density, and ⁰D). These results coincide with our third predicted scenario for TDF dynamics, which anticipated equally stable dynamics between substrates due to the soil-effects mechanism acting with an intensity similar to the composition-effects mechanism. This is the most intriguing finding of this study, as it suggests that, at least for our studied landscape, forest composition adapts to local environmental conditions (composition-effects mechanism) while ecosystem functioning remains spatially homogeneous even in patches with harsher environmental conditions (soil-effects mechanism).

The conclusion that soil- and composition-effects offset each other should be treated with care, as it relies on the assumption that species on phyllite TDF have rather conservative traits. However, the strong differences in structure and diversity, together with the soil differences observed by Sandoval-Granillo (2020), point to clearly different environmental conditions between substrates, as discussed earlier. As soils derived from phyllite seem drier and less fertile, one may reasonably assume that species in this substrate are better adapted to harsher environmental conditions (Díaz-Castellanos et al., 2022). For instance, communities in which conservative traits are prevalent in a Brazilian TDF showed higher productivity and biomass stocks than their counterparts with more acquisitive traits (Prado-Junior et al., 2016), which provides support to our hypothesis. Further research on the functional differences between the species occurring on limestone and phyllite TDF is needed to clarify if and how substrates influence forest functional profiles.

The result that in this region limestone TDF achieves an overall higher vegetation development without significant differences in dynamics is somewhat puzzling. Higher values in TDF attributes should arise with differing residence times, allowing limestone forests to keep a unit of a given attribute for a longer period in order to build higher attribute stocks. It should be noted that mortality was higher (more negative) for the variables analyzed on the demographic analysis; however, due to large within-substrate variation, results from the substrate-wise mortality comparison were not significant. Previous studies have shown that variation in attributes and rates increases with smaller samples (Galbraith et al., 2013). In this study, we used a surface area of 500 m² for the 14 permanent TDF plots. While this sampling area might be enough to obtain a reliable estimate for some TDF attributes, demographic rates may require larger samples (larger areas, longer study periods, and/or more sampling units) to be fully representative of the vegetation type they embody. Still, probably the differences in dynamics and demographic rates are small, and thus the noise-to-signal ratio does not allow to detect potential differences in dynamics between substrates. Therefore, for this site, it is essential to continue the monitoring for a longer period of time, as this would be the most feasible way of increasing the reliability of the estimations arising from this set of permanent plots.

In line with the results for TDF attributes, the NMDS demonstrated that species composition differed between substrates, with the respective floristic sets showing a relatively minor overlap on the ordination space (Figure 4A). Since the NMDS ordination was based on species abundances, year-to-year plot trajectories not only reflect differences in species presence/absence but also is species abundances between plots and substrates. Geology has been shown to affect species composition and abundance in other TDF sites (Durán et al., 2006), probably by differentially affecting soil fertility, soil water content and environmental physical heterogeneity. Soil fertility, which is largely dependent on parent material (Scholes, 1990; Augusto et al., 2017), has been shown to positively affect species dominance in TDFs (Peña-Claros et al., 2012) and to influence species and functional composition in tropical forests in general (Fayolle et al., 2012; Terra et al., 2018). Moreover, increasing soil water availability over time increased species richness and functional diversity in TDFs (Raymundo et al., 2019), and these results can be also extrapolated to a spatial dimension. Hence, lithological substrates, which determine soil spatial variability in TDFs (Pulla et al., 2016), can influence forest floristics through various mechanisms; thus, the finding of large differences in species composition and abundance is not surprising in these contrasting environments.

Contrary to the notorious compositional differences between the two forest subtypes, the NMDS ordination also revealed virtually identical dynamics regarding floristics in limestone and phyllite TDFs. The similarity in forest dynamics held regardless of the metric used to assess floristic dynamics (i.e., annual mean distance, dispersion from plot centroid, or the distance from initial and final points). In fact, the latter metric suggests a nearly circular pattern in the trajectories of most plots in the ordination space, irrespective of substrate. This observation implies a highly stable floristic composition, as is usually expected in old-growth forests under stable climatic regimes (Muñoz et al., 2021). The lack of differences in forest dynamics, even on these substrates associated with clearly dissimilar abiotic environments, is difficult to explain. Unfortunately, predicting species composition and its changes over time is extremely difficult, as the process of species assemblage in communities depends on too many biotic and abiotic factors and is of course species-specific (e.g., propagule and microhabitat availability, priority effects, nutrient limitations). Using a functional trait approach would likely be useful to understand floristic changes, as functional ecology links plant traits to plant responses to environmental stimuli. By doing so, changes in species composition and abundance could be framed in the context of the environments that these forests experience.