We carried out a review of the literature available at the database ISI Web of Science (Thompson Scientific, 2009) using “fruit,” “crop,” “selection,” and “birds” as keywords to find published papers in any year (until 1 September 2015) or language. We then checked out the references of these studies to find earlier publications. With this list and the studies carried out by us and collaborators (Palacio et al., 2017a,b) we completed a list of 100 references (Tables S1, S2, Figure S1). We used the Pearson product-moment correlation coefficient as effect size measure to determine the strength and direction of bird-mediated selection on fruit crop size. Although directional selection is properly measured using the standardized directional selection differential i, which represents the covariance between relative fitness (mean = 1) and a standardized phenotypic trait (mean = 0, variance = 1; Brodie et al., 1995), the advantage of Pearson's correlation is that it is widespread as a statistical descriptor, and well-developed meta-analytical methods exist (Hedges and Olkin, 1985). Moreover, r is proportional to i:r=i/I, where I is the variance in relative fitness (i.e., opportunity for selection) and is a measure of the power to detect directional selection (Hersch and Phillips, 2004). We therefore selected studies that reported data on fruit crop size and any of three quantitative estimators deemed as surrogate for fitness measures (Brodie and Janzen, 1996; McGraw and Caswell, 1996; Kingsolver et al., 2001; Schupp et al., 2010): bird visitation rate, the number of fruit removed or the proportion of fruit removed. References without this information were excluded (Figure S1).

We defined a visit as the event in which a bird approached a plant and consumed at least one fruit. Several studies have widely used these three fitness components to assess bird-mediated selection on fruit display traits (e.g., Jordano, 1995; Martínez et al., 2007; Sobral et al., 2010; Palacio et al., 2014). We also assumed a positive covariation between short-term dispersal success (a fitness component) and long-term plant fitness (McGraw and Caswell, 1996; Muñoz et al., 2017). Values for Pearson's r were also obtained from one tail P-values when sample size was known (Rosenthal and DiMatteo, 2001), and Spearman's ρ were converted into Pearson's r following Rupinski and Dunlap (1996) as: r = 2sin(ρ × π/6). In several cases, we were able to compute r from raw data. In those studies that raw data were not reported, we estimated approximate values by processing published plots with DATA THIEF III version 1.6 (Tummers, 2010). The final database and the complete list of corresponding references are shown as Supplementary Material (Tables S1, S2).

Correlation coefficients were transformed to Zr using Fisher's algorithm (Hedges and Olkin, 1985). A single mean effect size was computed for species with more than one effect size by averaging Zr values (Rosenthal and DiMatteo, 2001) to account for dependence among species. Once we obtained and transformed effect sizes per species, we computed mean effect sizes for the three fitness components (bird visitation rate, number, and proportion of fruit removed). Within-species variances were computed as 1/(n−3), where n is sample size (Hedges and Olkin, 1985). For comparative purposes, and whenever possible, we computed standardized directional selection differentials. Overall effect sizes were considered statistically significant if 95% confidence intervals did not include zero.

We assessed publication bias both graphically, by funnel plots, and statistically by rank correlations between effect sizes and samples variances (Begg and Mazumdar, 1994). Spearman's rank correlations between effect size and sampling variance were non-significant for the proportion of fruit removed (ρ = −0.050, P = 0.779), but it was significant for the bird visitation rate (ρ = 0.450, P = 0.040) and the number of fruit removed (ρ = 0.511, P = 0.003), indicating the presence of publication bias (Figure S2). We therefore used the Copas selection model (Copas, 1999) to correct mean effect sizes. This model has two components: a model for the effect size, and a “selection” model that gives the probability that study i is published. A Pearson correlation between these two components estimated by maximum likelihood quantifies the magnitude of publication bias; the stronger the correlation, the higher the probability that more extreme effect sizes are published.

Phylogenetic relationships may lead to non-independence in effect sizes, due to shared phylogenetic history among taxa (Lajeunesse, 2009). We therefore constructed phylogenetic hypotheses using S.Phylomaker, based on the megaphylogeny of vascular plants PhytoPhylo (Qian and Jin, 2016; Figure S3). This phylogeny includes all extant families of gymnosperms and angiosperms, provides ages of all branches, and includes five times as many genera and 55 times more species than does the newest version of angiosperm supertrees derived from Phylomatic (i.e., tree version R20120829; Webb and Donoghue, 2005). We estimated phylogenetic signal in effect sizes with Pagel's λ according to a Brownian motion model of evolution where λ = 0 indicates no phylogenetic signal, and λ = 1 indicates high phylogenetic signal (Pagel, 1999). P-values were computed with 999 randomizations of species across the tips of the phylogeny. Pagel's λ values for correlations between fruit crop size and visitation rate, the number of fruit removed and the proportion of fruit removed were λ = 1.000 (P = 0.151), λ = 0.325 (P = 1.000), and λ = 0.406 (P = 1.000), respectively. This indicates that no effect size expressed phylogenetic inertia, and a phylogenetically controlled meta-analysis is thus unnecessary (Lajeunesse, 2009; Chamberlain et al., 2012).

To test for the heterogeneity of effect sizes we used Cochran's Q, which is computed as the weighted sum of squared differences between individual study species and the mean effect size (Hedges and Olkin, 1985). As a complement, we also computed the I² index to quantify the degree of heterogeneity in a meta-analysis, which measures the percentage of the total variability in effect sizes due to between-studies heterogeneity (Higgins and Thompson, 2002; Huedo-Medina et al., 2006). We then assessed the influence of factors on mean effect sizes to explain variation in phenotypic selection on fruit crop size. We fitted mixed-effects models using both categorical (fruit type: aril, drupe, or berry) and continuous variables (length of the fruiting season, mean fruit diameter, proportion of gulper, pulp consumer and seed predator species, and latitude) as fixed effects. Mixed- or random-effects meta-analysis assumes that the studies analyzed represent a random sample of effect sizes by including a component of between-study variation (study-specific random effect) into the uncertainty of the effect size parameters and their estimates (Hedges and Olkin, 1985). We selected these variables since most studies included information on these factors. The proportion (relative richness) of gulpers, pulp consumers, and seed predators were used instead of interaction frequency, as most studies did not report the latter. We considered “fleshy fruit” as any fleshy diaspore (dispersal unit of seeds plus fleshy structures covering the seeds; Herrera, 2002), so aril was categorized as a fruit type. Regarding bird foraging behavior, we categorized gulpers as species that swallow fruits whole, pulp consumers as those that mash or peck fruits, and seed predators as those that crush and typically destroy seeds (Foster, 1987; Levey, 1987). The length of the fruiting season and fruit diameter were log-transformed due to right-skewed distributions. The proportion of gulper and pulp consumer species were highly correlated (r < −0.90), so we only included the former. To assess the importance of explanatory variables, we adopted an information-theoretic approach based on Akaike's Information Criterion corrected for small sample sizes (AICc), which allows identifying best models given a certain set of plausible hypotheses (Johnson and Omland, 2004; Burnham et al., 2011). We fitted four non-nested main models per fitness component (bird visitation rate, number and proportion of fruit removed), which represented a trade-off between the classification of factors proposed here (plant traits, bird assemblage traits and environmental factors) and differences in sample sizes: a “plant trait” model 1 (fruit type + fruit diameter), a “plant trait” model 2 (length of the fruiting season), a “bird assemblage trait” model (proportion of gulper species + proportion of seed predator species), and an “environmental model” (latitude). For each of the four main models, all the possible nested models were ranked based on their AICc, and those models with the lowest AICc in a range such that ΔAICc <2 were identified as the best models. To complement this analysis and test the support of our results, we fitted a model for bird visitation rate, number and proportion of fruit removed including all variables.

All analyses were performed in R 3.4.2 (R Core Team, 2017), using the packages phytools (function phylosig; Revell, 2012), metafor (function rma; Vietchtbauer, 2010), metasens (function copas; Schwarzer et al., 2016), and glmulti (function glmulti; Calcagno, 2013).

Data of 50 plant species from 27 families (44 published and 5 unpublished studies) met the selection criteria to assess phenotypic selection on fruit crop size (see Figures S3, S4). Most families were represented by 1 or 2 species, but Rosaceae, Burseraceae, and Myristicaceae were represented by 7, 6, and 5 species, respectively. The median number of plants monitored per species within a publication was 17 (mean = 43.7, SD = 61.9, range = 5–278). Most species were represented by only one study; only data for Olea europaea (Oleaceae), Virola surinamensis (Myristicaceae), and Crataegus monogyna (Rosaceae) were available from 2, 3, and 4 studies, respectively. All continents (except for Asia and Antarctica) were represented in the meta-analysis, with most data from the Americas (31 studies) and Europe (12 studies; Figure S4). Significant positive effects (z-transformed values) of bird-mediated selection on fruit crop size were detected for bird visitation rate (mean bias-uncorrected effect size = 0.584, n = 21, 95% CI = 0.385–0.783; mean bias-corrected effect size = 0.400, n = 21, 95% CI = 0.188–0.613) and the number of fruit removed (mean bias-uncorrected effect size = 1.640, n = 32, 95% CI = 1.390–1.890; mean bias-corrected effect size = 1.300, n = 32, 95% CI = 1.181–1.419), but not for the proportion of fruit removed (mean bias-uncorrected effect size = 0.098, n = 34, 95% CI = −0.017 to 0.213; mean bias-corrected effect size = 0.097, n = 34, 95% CI = −0.011 to 0.211). Moreover, between-species heterogeneity was high (bird visitation rate I² = 80.10%, number of fruit removed I² = 94.59%; proportion of fruit removed I² = 71.94%) and significant for the three overall effect sizes (bird visitation rate Q = 73.573, df = 20; number of fruit removed Q = 497.432, df = 31; proportion of fruit removed Q = 98.103, df = 33; all P's < 0.0001). These results therefore justified our subsequent analyses of factors explaining variation in bird-mediated selection on fruit crop size.

Directional selection differentials were consistent with results using Pearson's correlation as effect size. Differentials showed positive directional selection for the three fitness components (bird visitation rate, number, and proportion of fruit removed). Selection strength on fruit crop size was stronger for the number of fruit removed (median = 1.030, range = 0.033–2.336, n = 31) and bird visitation rate (median = 0.473, range = −0.483–1.581, n = 23), and it was lowest for the proportion of fruit removed (median = 0.041, range = −0.261 to 0.586, n = 36). It should be noted that selection differentials measure both direct selection on a given phenotypic trait and indirect selection through correlations with other traits (Lande and Arnold, 1983), so these values should be interpreted with some caution.

Both plant (length of the fruiting season, mean fruit diameter) and bird (proportion of species belonging to the same functional group) traits, as well as environmental factors (latitude) explained variation in bird-mediated selection on fruit crop size (Tables 2, 3, Figure 2). For the three fitness measures higher selection strength was detected in plants with shorter fruiting seasons and higher proportion of gulper species (Table 3, Figure 2). Also, for the proportion of fruit removed, higher selection strength was detected in plants with larger fruits (Table 3, Figure 2). For the number and proportion of fruit removed, the proportion of seed predators explained variation in phenotypic selection on fruit crop size but showed opposite trends (Table 3, Figure 2). In particular, seed-dispersal systems with higher proportion of seed predators showed either higher or lower selection strength on fruit crop size for the proportion and number of fruit removed, respectively. Finally, selection strength increased toward the equator for bird visitation rate (Table 3, Figure 2). Models with all variables gave similar qualitative results despite their low sample size (Table S3).

The functional equivalence hypothesis proposes that different ecological filters (e.g., phenology, plant traits, environmental factors) remove species and leave those with similar trait values and functional roles, promoting similar selection regimes on plant traits (Zamora, 2000). Several studies have found an increase in selection strength resulting from an increase in the abundance of species with shared functional traits. For instance, Alcántara et al. (2007) found that ant assemblages with similar-sized species exerted similar selection pressures on elaiosome and seed size in 12 populations of Helleborus foetidus. In another study, Galetti et al. (2013) found that the functional extinction of large-gape seed dispersers correlated with an evolutionary reduction of seed size in 22 populations of Euterpe edulis in the Atlantic Forest from Brazil. In this case, defaunation acted as a filter on large species changing the evolutionary trajectory of seed size (Galetti et al., 2013). These studies suggest that changes in the abundance of functionally equivalent interacting partners stemming from ecological filters may be an important evolutionary force driving phenotypic trait variation, and therefore deserves further research.

For the three fitness components studied, two plant and bird traits (fruiting phenology and fruit-handling behavior) consistently played a role in bird-mediated selection strength on fruit crop size. As predicted by the functional equivalence hypothesis, factors that would filter bird functional traits promoted higher selection strength on fruit crop size. The timing of fruiting has been shown to play a role in regulating the abundance and diversity of frugivores in many communities (Fenner, 1998; Burns, 2004; Albrecht et al., 2015). A fruit display in a short time period would decrease assemblage diversity relative to extended fruiting seasons, acting as a temporal filter and increasing selection strength on fruit crop size (McKey, 1975; Howe and Estabrook, 1977; Howe, 1993; Fenner, 1998). Higher proportions of species from the same functional group also supported the functional equivalence hypothesis, under the idea that an increase in the proportion of species with shared functional traits promotes higher selection pressure on fruit traits. One limitation of our approach is that we considered the effects of legitimate seed dispersers and seed predators on plant fitness as equivalent. It is known that the effect of fruit removal by seed dispersers and predators are opposite in all aspects of plant recruitment, in which seed predators have a detrimental effect on plant fitness, contrary to seed dispersers (Levey, 1987; Schupp et al., 2010; but see Loayza and Knight, 2010; Tella et al., 2015). From our results, we can only assert that both gulpers and seed predators favor larger fruit crop sizes but, under the assumption that seed predators negatively affect fitness, higher proportion of seed predators entails a negative directional selection pattern on fruit crop size. This highlights the role of antagonistic interactions in shaping fruit display traits in multispecies assemblages (e.g., Alcántara and Rey, 2003; Gómez, 2004; Martínez et al., 2007; Siepielski and Benkman, 2007).

Another plant trait, i.e., mean fruit diameter, also explained variation in selection strength on fruit crop size. In particular, larger fruits promoted higher selection strength on fruit crop size for the proportion of fruit removed. Fruit size is considered a key trait at the fruit unit level in mutualistic interactions of frugivores and fleshy-fruited plants, as it constraints fruit-handling behavior and fruit consumption by seed dispersers (“the fruit size hypothesis”; Wheelwright, 1985; Levey, 1987). Under this hypothesis, it would be expected a decrease of frugivore richness with increased fruit size (Wheelwright, 1985; Jordano, 2000). Indeed, plant species of our database with the largest mean fruit sizes (Aglaia flavida: 150 mm, Pinus edulis: 45 mm) were consumed by one and two frugivores, respectively. In contrast, mean frugivore richness per plant species in those systems ranging from 1 to 10 mm in fruit diameter was 10.2 ± 6.4 bird species. Therefore, larger fruits would promote higher strength on fruit crop size, as a result of coupled selection pressures of a reduced bird assemblage (Herrera, 1985). A latitudinal effect on phenotypic selection on fruit crop size would also be expected, as both frugivore richness (Kissling et al., 2009) and fruiting seasons (Ting et al., 2008) increases toward the tropics which, interestingly, is explained by the same climatic variable (actual evapotranspiration). Nonetheless, we found a negative latitudinal trend of selection strength on fruit crop size, although of low magnitude. A possible explanation is that an unmeasured variable covaries with latitude that decreases selection strength toward the tropics. Recently, Dalsgaard et al. (2017) found opposed latitudinal patterns of network-derived measures of specialization and dietary specialization in bird-frugivore interactions. In particular, the proportion of obligate frugivores increased toward the tropics, but frugivorous birds divided the niche of fruiting plants more finely at high latitudes (Dalsgaard et al., 2017). Higher proportion of obligate frugivores within a seed-dispersal assemblage (potentially fulfilling similar functional roles; Loiselle et al., 2007) may drive higher selection strength on fruit crop size, an idea that deserves further research.

Finally, phylogenetic relationships played a minor role in bird-mediated selection strength on fruit crop size. Indeed, no phylogenetic signal was detected for any fitness component. This also supports the fruit crop size hypothesis, which states that birds promote positive selection on fruit crop size, regardless of plant identity.

Several conditions have been advocated to constrain reciprocal coadaptation among plants and seed dispersers, leading to diffuse coevolution (Wheelwright and Orians, 1982; Howe, 1984; Herrera, 1985; Tewksbury, 2002): (1) low selection intensity of dispersers on fruit traits, (2) inconsistent spatiotemporal bird-mediated selection, (3) antagonistic selection by fruit or seed predators, (4) correlations among fruit traits that prevent the evolution of mutual adaptations, and (5) differences in evolutionary rates between plants and dispersers. Two of these conditions, however, are not supported by our results. First, selection coefficients on fruit crop size fell within the range of other reported values of phenotypic selection in natural populations, yet for the number of fruit removed it would be considerably strong (most linear gradients range between −1 and +1, |median| = 0.16; Kingsolver et al., 2001). Therefore, fruit crop size represents a phenotypic trait under strong selection. Furthermore, it has shown to be highly heritable (h² = 83.0–99.0%; e.g., Manju and Sreelathakumary, 2002; de Moraes et al., 2005; Denton and Nwangburuka, 2011; Meena and Bahadur, 2014) indicating a strong response to selection. Second, we did not find strong negative correlations among fruit crop size and other fruit traits that may prevent reciprocal adaptations. Indeed, we found only positive correlations between fruit crop size and other fruit traits, such as mean fruit diameter (r_{Psychotria carthagenensis} = 0.11, n = 72; r_{Vassobia breviflora} = 0.10, n = 27; r_{Jodina rhombifolia} = 0.28, n = 17; r_{Passiflora caerulea} = 0.35, n = 8) and mean sugar concentration (r_{Psychotria carthagenensis} = 0.03, n = 72; r_{Vassobia breviflora} = 0.22, n = 27). Other studies in natural conditions have found no associations between fruit crop size and other fruit level traits (Michaels et al., 1988; Martínez et al., 2007; Palacio et al., 2014; but see Sallabanks, 1993). Although an increase in fruit production could compromise fruit quality (fruit size, pulp mass, nutrient content), this has been shown under experimental conditions (Barone et al., 1994; Berman and DeJong, 1996; Naor et al., 1999, 2008; Stopar et al., 2002; Roussos et al., 2011). Negative correlations between fruit crop size and other fruit traits may be obscured by the fact that fruit crop size is a complex trait constrained by several factors (Lee, 1988) including resource availability (Lescourret and Génard, 2003), plant age (Mechlia and Carroll, 1989; Marshall et al., 2010), flower display and pollination service (Herrera, 1991; Cunningham, 2000; Benavidez et al., 2013; Stournaras and Schaefer, 2017). The interaction between different factors affecting fruit crop size may reduce correlations with other fruit traits, potentially relaxing indirect selection on fruit crop size in nature (e.g., Jordano, 1995; Sobral et al., 2010; Palacio et al., 2014). By contrast, counteracting selection pressures exerted by seed predators may dilute the overall selection pressure imposed on fruit crop size, favoring the process of diffuse coevolution (Herrera, 1985). Our results support this idea, as an increase in the proportion of seed predators favored large fruit crops. Finally, seed dispersers may shape the evolution of seed size and number, as a result of a trade-off between fruit crop size, seed size, and seed number (Smith and Fretwell, 1974; Eriksson and Jakobsson, 1999; Sadras, 2007). In particular, fruit crop size is positively correlated with the total seed number of a plant, but seed number within a fruit is negatively correlated with mean seed size (Westoby et al., 1992; Parciak, 2002; Sadras, 2007). Therefore, selection on large fruit crop sizes should promote lower mean seed size. This suggests that seed dispersal may be more important than seed size for successful offspring recruitment (Parciak, 2002), although the existence of conflicting selection by other selection agents (e.g., seed predators) might explain the occurrence of an optimal seed size in some plant species without invoking a seed number-size trade-off (Gómez, 2004).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.