We searched the literature (peer-reviewed papers but also reports and unpublished theses) for field studies investigating the floristic composition in both ancient and post-agricultural deciduous forests in landscapes across Europe. Studies were only included if they met the following criteria (adapted from Verheyen et al. 2003c): (i) the inventoried forests were situated in lowland or lower mountainous (< 600 m a.s.l.) Europe; (ii) frequency data of the forest plants were available for all (or a representative subsample of) ancient and post-agricultural forest patches in the landscape; (iii) the land-use history and sufficient information on the other predictor variables (see below) of both forest types or the region was given and (iv) canopy closure had already occurred in post-agricultural forests (thus, the analyses were restricted to closed-forest ecosystems). Eighteen geographically non-overlapping studies from nine countries (from central Italy in the south to central Sweden in the north and from the UK in the west to Poland in the east; see Fig. 1) complied with these four criteria and yielded 2231 species × study combinations (693 species) from 1098 ancient and 1370 post-agricultural data collection units (patches or plots; Table 1 and S1 in Supporting Information). However, to avoid a biased estimation of the parameters in the statistical models for species with low frequency, we only focused on those species that were cited in more than five studies. This reduced the species × study combinations to 1332 (155 species).
To explore the effects of life-history traits related to plant colonization on the RR of post-agricultural forest, we used the emergent groups (EG) of plant species (sensuLavorel et al. 1997) identified by Verheyen et al. (2003c) according to 13 reproductive, vegetative and phenological traits: seed mass, seed size, seed shape, seed production, dispersal type, seed bank, germination requirements, age of first reproduction, growth form, life cycle, vegetative spread, maximum height and flowering phenology. Based on these traits, each species was classified into one of four (herbs) or three (graminoids) EG as delineated by Verheyen et al. (2003c) by combining Gower’s Similarity Coefficients, non-metric multidimensional scaling and clustering with Ward’s method (see Verheyen et al. 2003c for more information). For the herbs the four groups were (i) short-lived herbs, (ii) tall perennials with heavy seeds, (iii) tall perennials with light seeds and (iv) small perennials with heavy seeds; for the graminoids the three groups were (i) large, summer-flowering graminoids, (ii) small, summer-flowering vegetatively spreading graminoids, and (iii) early flowering graminoids (the characteristic life-history traits per EG are shown in Table S2 in Supporting Information and the EG per species in Table S3 in Supporting Information). Only those species for which > 50% of the life-history traits were available were used in the analyses. The 18 studies then yielded a final set of 90 species (71 herbs and 19 graminoids) and 812 species × study combinations (Tables 1 and S3 in Supporting Information). The EG were calculated across all species and individual species can thus display a deviating value for some life-history traits. The species nomenclature follows Wisskirchen & Haeupler (1998).
To conduct the interregional analyses, four groups of predictor variables were gathered for each study. These were related to habitat availability, time available for colonization, climate and soil and light characteristics. Similar to Vellend (2003), we focus on habitat loss and not habitat fragmentation sensu stricto (cf. Fahrig 1997, 2003) as information to calculate the absolute patch sizes or patch isolation distances was unavailable for most studies. Therefore, to account for habitat availability and time available for colonization, data on the total proportion of forest in the landscape (ancient plus post-agricultural; TF) and of ancient and post-agricultural forest separately (AF and PAF, respectively) as well as the mean and maximum post-agricultural forest age were collected instead. It should be kept in mind, however, that the amount of forest cover is not necessarily correlated with spatial isolation. Forest cover data were provided by the authors, compiled from the original publications, taken from Vellend (2003) or, in the case of Zacharias (1994) and Jakubowska-Gabara & Mitka (2007), obtained by digitizing maps provided in the original publication using Image J (Rasband, W.S., US National Institutes of Health, Bethesda, MD, USA, http://rsb.info.nih.gov/ij/). The total forest (TF) and AF cover in the landscape encompasses a broad range: 5.4–50.7% and 0.9–27%, respectively. The TF cover was significantly correlated with the AF cover (r = 0.489, P = 0.039) and the post-agricultural forest cover (r = 0.859, P < 0.001) across our studies (n = 18). The mean and maximum age of the post-agricultural forests (further referred to as the mean and maximum colonization time, respectively) was mostly available in the original publications or provided by the authors. We calculated the mean colonization time as the average of the minimum and maximum colonization time within the landscape when the mean was not explicitly mentioned. It should be noted that forest age sensu stricto is not the main focus here. Due to data availability in the different regions, an AF in one region can be younger than a post-agricultural forest in another region (Table S1 in Supporting Information). Although an old post-agricultural forest is probably more likely to be already colonized, we consider forest continuity here (i.e. whether a forest has always been forest or cleared for agriculture at some point). Hence, we take both the mean and maximum colonization time of the post-agricultural forests in a region into account. The mean and maximum colonization time of the post-agricultural forests varied between 24–135 and 54–240 years (excl. Peterken & Game 1984). The British study by Peterken & Game (1984) forms a notable exception because much older detailed maps are available in Britain (Goldberg et al. 2007; mean and maximum colonization time of the post-agricultural forests up to 201 and 370 years in Peterken & Game 1984). Exclusion of this study, however, yielded similar statistical results (data not shown).
Next, climate data were obtained from the NewLocClim 1.10 software (FAO 2005) using nearest-neighbour interpolation with ten weather stations. We gathered latitude, longitude and altitude data of the centre of each study region and used these values to deduce mean annual temperature (MAT; 1961–1990), mean annual precipitation (MAP) and potential evapotranspiration (PET). Also the effects of latitude and longitude themselves were tested for. Subsequently, an aridity index was calculated according to FAO (2005) as the ratio of MAP and PET. The MAT and aridity index varied between 6.6–15.0 °C and 0.74–2.58, respectively.
Finally, as detailed soil data or light measurements were not available for most studies so that we could rigorously test for their effects, we used the mean frequency-weighted Ellenberg indicator values (Ellenberg et al. 1992) for nutrients (mNj), reaction (mRj), moisture (mFj) and light (mLj) of the AF as rough proxies for the local soil nutrient availability, soil acidity, soil moisture and light availability, respectively. Ellenberg values are known to be very good correlates of in situ measured environmental characteristics in AF but not in post-agricultural forests (Dzwonko 2001a). Therefore we only used the Ellenberg values of the AF as proxy for these environmental variables to allow for a general interregional ranking in soil and light characteristics. For mNj, for example, this was calculated as:
- (eqn 1)
with Freqspi,j the frequency of species i in the AF in study j and Ni the Ellenberg N value for species i. The calculated mN values ranged between 2.5 (nutrient-poor soils) and 5.8 (nutrient-rich), the mR between 2.7 (acid soils) and 4.9 (more neutral), the mF between 2.6 (dry soils) and 6.0 (moister) and the mL between 3.1 (forests with low light availability in the understorey) and 4.9 (higher light availability) (Table S1 in Supporting Information). Hereafter, the mN, mR, mF and mL values are referred to as soil nutrient availability, soil acidity, soil moisture and light availability, respectively.
The RR calculated for each species × study combination as the risk ratio with binary data (2 × 2 tables) in standard meta-analytical procedures (Borenstein et al. 2009):
- (eqn 2)
with RRij being the RR for species i in study j and PAFij and AFij the percentages of data collection units (patches or plots; Table 1) occupied by species i in post-agricultural and AF in study j, respectively. Eqn 2 includes a correction (+ 0.01) in both the numerator and denominator to account for zero-percentages in both forest types. Zero values of RR thus correspond to equal percentages of the species in ancient and post-agricultural forest, whereas positive and negative values correspond to a lower and higher affinity to ancient than to post-agricultural forest, respectively. A species with an RR = −1, for instance, showed a percentage in AF that was approximately 2.7 times the percentage in post-agricultural forest.
Subsequently, the effect of the life-history traits of each species (EG) on RR was tested with mixed models in R 2.11.0, using the lmer function of the lme4 library (R Development Core Team 2010). According to Zuur et al. (2009), we first selected the optimal random-effects structure based on a likelihood ratio test between models with a similar fixed component (no predictor variables included), but a different random component. The optimal model included both study and species as non-nested random effects. Modelling the hierarchical nature of the data using two non-nested random effect terms in a mixed model then leads to partial pooling across the different levels (Qian et al. 2010), and hence, this takes the possible autocorrelated characteristics of (i) species from the same study region and (ii) similar species in different regions into account. Next, we compared the null model (only including the two non-nested random effects) with a model that included the EG (species level) (χ2-test statistic with likelihood ratio test; Zuur et al. 2009); these analyses were conducted for all species together and afterwards also separately for herbs and graminoids.
Next, to quantify the interregional variation in the RR for each species, we calculated the coefficient of interregional variation (CIVRR) of species i as , with SDi the standard deviation and i the mean RR for species i (hence, one CIVRR value per species). A correction factor of 1 was added to the denominator to prevent CIVRR values from skyrocketing for species with an RR close to zero. We then performed a one-way anova using a general linear model (GLM) with Bonferroni post hoc test (using spss 15.0) to investigate whether there were differences in the CIVRR between the different EG; again, herbs and graminoids were analysed separately.
To explore the effects of all the environmental predictor variables on the RR, namely (i) the proportion of total, ancient and post-agricultural forest in the landscape, (ii) the mean and maximum colonization time, (iii) latitude, (iv) longitude, (v) temperature, (vi) aridity index, (vii) soil nutrient availability, (viii) soil acidity, (ix) soil moisture and (x) light availability (all study-level), we again applied mixed modelling in R following a similar approach as above. The model again included study and species as non-nested random effects and the null model (only including random effects) was compared with a model that included one of the predictor variables (on a one-by-one basis mainly to avoid multicollinearity problems) to test the significance of that particular variable (χ2-test statistic with likelihood ratio test; Zuur et al. 2009). We also estimated the percentage of variation explained by adding the predictor variables to the null model through calculations of the ratio of the difference in residuals between the null model and the final model over the residuals of the null model (Hox 2002). Finally, we tested for additive models and interactions among all significant predictor variables with the life-histories of the species (EG) by comparing (i) a model that included the EG plus that particular predictor variable with a model that included only the EG as main effect (i.e. test of the additive effects) and (ii) a full factorial model with a model that included the EG plus that particular predictor variable (i.e. test of the interaction term). Again, the χ2-test statistic with likelihood ratio test was used for this purpose.