An understanding of the climate conditions governing spatial variation in the reproductive performance of plants can provide important information about the factors characterizing plant community structure and influencing fitness in natural plant populations, especially in the context of climate warming [3,16,17,24]. As pointed out by Moles et al. [23], if we do not know how climate affects present‑day patterns in ecological traits and processes, then it will be very difficult indeed to predict the potential effects of climate change on the earth’s biota.

In the present study, a clear variation was found in the reproductive performance of D. glomerata populations along the elevational gradient. Resource investment in reproduction (i.e., the product of mean seed number and mean seed mass per individual) decreased by 53% from warmer sites (where annual mean temperature is ca. 12.9 °C, and elevation ranged from 300 to 500 m a.s.l.) to colder sites (where annual mean temperature is ca. 6.6 °C, and elevation ranged from 1,400 to 1,600 m a.s.l.). Similarly, several authors [15,16,17,24] documented a lower reproductive performance in cold climate‑adapted populations (e.g., higher latitude/elevation) than in warm climate-adapted populations (e.g., lower latitude/elevation). As previously indicated by Dainese [17], climate change along the elevation gradient determines a morphological differentiation of reproductive traits in D. glomerata, with low-elevation individuals having larger inflorescences and larger seed masses than higher elevation individuals, meaning a higher reproductive output in warmer/longer growing seasons than in colder/shorter growing seasons. These results suggest that longer growing seasons and/or warmer climates in lower elevation sites promote higher photosynthetic rates, generating greater plant allocation toward reproductive performance than colder climates in higher elevations [17,24]. This means that cold climate-adapted populations are functioning on lower levels of photosynthesis compared to warm climate-adapted populations, determining a lower reproductive output [24].

Although the results of this study clearly indicate that reproductive performance of D. glomerata changes along the elevational/climatic gradient, recent studies have shown that this response is to some extent species-dependent. For instance, De Frenne et al. [3,15] reported a contrasting effect of temperature on reproductive performance in forest herbs due to their different phenology, distribution and life form. Hence, there is a need for studies that compare species responses to temperature, since results cannot just be transferred from model species [3].

Considering the influence of climate variables on the reproductive performance of D. glomerata, this study shows a greater importance of environmental energy than water availability, confirming the primary importance of temperature in mountain systems [4]. The negative relationship between water availability during the growing season and reproductive performance might be due to the co-variation between elevation/temperature and precipitation.

Bioclimatic variables that provide information about climatic conditions during the harshest times of the year for plant performance (e.g., minimum temperature of the coldest quarter or precipitation of the driest month), are relatively weakly correlated with resource investment in reproduction. Moles et al. [23] found a primary influence of variables that capture information about the quality of the growing season compared to variables that capture information about difficult times, when growth is low or entirely ceased. Therefore, it would be advisable to take into account a range of climatic variables, especially those that consider the growing season, when studying potential plant responses to climate change.

However, in interpreting elevational gradient studies, one must consider that besides climate change, the variation of elevation can also modify local conditions (i.e., plant competitor, seed predator density, herbivorous species) [4,5]. The cascading effects of change in local conditions could indirectly modify the reproductive performance of D. glomerata. Thus, the observed variation of the resource investment in reproduction of D. glomerata may be attributable to both direct and indirect effects. Moreover, in a context of climate change, an increment in reproductive performance of competitive species, such as D. glomerata, could affect plant population dynamics and the community structure of neighboring plant communities, according to the total competition hypothesis [17,25,26]. As hypothesized by Dainese [17], the higher reproductive performance of competitive species may allow them to outcompete smaller and less competitive species.

To generate a climatic gradient, an elevation gradient was used as a ‘space for time substitution’. The dataset from Dainese [17] was used to test the relationship between the reproductive performance of D. glomerata and a range of climatic variables. In a previous study, Dainese [17] has used long-term mean climatic variables and considered only limited annual trends (i.e., annual mean temperature and annual precipitation) to test the influence of climate change on the reproductive performance of D. glomerata. In this study, a climate model was constructed using local meteorological stations and considering monthly climate data for the study year and the two previous years (2006–2008). Moreover, several bioclimatic variables were generated to take into account the quality of the growing season, as well as annual trends and limiting environmental factors. In Dainese [17], 60 grasslands were sampled along an elevational gradient (from 319 to 1,662 m a.s.l.) on the southern border of the European Alps (Province of Trento—NE Italy; N45°55'–46°28'; E11°10'–11°53') (see Supplementary Figure S2). The grasslands were distributed across three different districts: (i) Valsugana, (ii) Primiero, and (iii) Fiemme and Fassa Valley. The mean annual temperature at the minimum altitude considered (319 m a.s.l.) was 12.9, and 5.4 °C at the maximum altitude (1,662 m a.s.l.). The mean annual rainfall varied from 700 to 1,500 mm year⁻¹.

At each site in 2008, three 1 m × 1 m plots were randomly located, and all fertile stems of D. glomerata were collected and counted. Seeds were extracted and weighed for each plot (n = 180). The total seed mass per individual was determined by dividing the seed production per plot by the total number of fertile stems per plot. One hundred seeds from each plot were then randomly sampled and weighed and the mean seed mass calculated. Finally, the total number of seeds per individual was calculated as the total seed mass per individual divided by the mean seed mass.

Temperature and precipitation data were extracted from 31 local meteorological stations (with a minimum mean distance between meteorological stations and sampling sites of 4.9 km), covering an elevation range between 182 and 2,063 m a.s.l. (see Supplementary Figure S2). Averaged monthly precipitation and temperature for the period 2006–2008 were used to generate 19 bioclimatic variables, which represent annual trends, seasonality and extreme or limiting environmental factors. All bioclimatic variables were grouped into three categories: (i) environmental energy, including annual mean temperature, mean diurnal range, maximum temperature of the warmest month, minimum temperature of the coldest month, mean temperature of the wettest quarter, mean temperature of the driest quarter, mean temperature of the warmest quarter and mean temperature of the coldest quarter; (ii) water availability, including annual precipitation, precipitation of the wettest month, precipitation of the driest month, precipitation of the wettest quarter, precipitation of the driest quarter, precipitation of the warmest quarter and precipitation of the coldest quarter; and (iii) climatic seasonality, including temperature annual range, isothermality, temperature seasonality and precipitation seasonality.

Since the environmental energy variables were highly correlated with elevation (average r = −0.92; range: −0.97 ÷ −0.70), except for mean diurnal range, ordinary kriging with external drift was used as the interpolation method [27]. First, the fitted environmental energy variables were estimated from a simple regression with elevation. Then, the residuals were interpolated using ordinary kriging and, subsequently, summed with the fitted temperature from the regression with elevation. The same procedure was also used for some water availability variables due to their strong correlation with elevation, such as annual precipitation (r = 0.65), precipitation of the wettest month (r = 0.63) and precipitation of the wettest quarter (r = 0.78). Instead, the other variables, which were not correlated with elevation, were interpolated using ordinary kriging. The geostatistical interpolations were computed using the Geostatistical Analyst extension for ArcGIS 9.3 (ESRI).

Seed number and seed mass were used to calculate a compound measure of reproductive success: ‘resource investment in reproduction’ (RIR), as the product of mean seed number and mean seed mass per individual [3,17].

Due to the hierarchical nature of the data, linear mixed models (LMMs) were constructed for all regression analyses to account for dependencies within hierarchical groups through the introduction of random effects, therefore maximizing the statistical power. Models included bioclimatic variables as fixed effects and district and site-within-district as random effects. The models were validated by the analysis of residuals, to assess homogeneity and to verify normality [28]. Response variables were log-transformed to satisfy assumptions of normality. Furthermore, to avoid problems of heterogeneity, a variance model with different variances for each level of a stratification variable (in this case district) was applied when necessary [28]. Models were estimated using the ‘nlme’ package [29] in R, version 2.14.1 [30] with the restricted maximum-likelihood (REML) estimation method. The strength of the association between bioclimatic variables and RIR was calculated from the variance components of the mixed-effects models using a level-1 pseudo-R² statistics [31].

A combined model was developed using the following methods: (i) the best individual predictor for RIR of a bioclimatic group was selected in the model; (ii) to avoid the multi-co-linearity between the predictors of the same environmental category, only one variable from each category was used in the model; and (iii) all the possible combinations of predictors following the above criteria were examined using a multi-model inference within an information theoretic framework [32]. The information-theoretic approach compared the fit of all possible candidate models obtained by the combination of the predictors using second-order Akaike’s information criterion (AICc). The AICc is a measure of relative model fit, proportional to the likelihood of the model and the number of parameters used to generate it. The best fitting model is the one with the lowest AICc. In a set of n models, each model i can be ranked using its difference in AICc score with the best-fitting model (ΔAICci = AICc_i–AICc minimum). The difference in AICc values indicates the relative support for the different models. A model is usually considered plausible if its ΔAICc is below 2 [32]. For each model i, an Akaike’s weight (w_i) was calculated, which is the probability that model i would be selected as the best fitting model if the data were collected again under identical circumstances [32]. Akaike’s weight should be interpreted as a measure of model selection uncertainty. The multi-model inference analyses were performed using the ‘MuMIn’ package [33] implemented in R.

Finally, hierarchical partitioning (HP) [34] was used to evaluate the relative importance of predictors, given that the use of Akaike’s model weights has been recently criticized in cases of co‑linearity between the variables included in model selection [35]. HP addresses the presence of co‑linearity by determining the independent contribution of each explanatory variable to the response variable and separates it from the joint contribution, resulting from the correlation with other variables [34]. HP was conducted using the ‘hier.part’ package [36] for R. A normal error distribution and R² as a measure of goodness-of-fit were used.