Phenotypic Selection Varies with Pollination Intensity Across Populations of Sabatia Angularis

Research

Phenotypic selection varies with pollination intensity across populations of Sabatia angularis

Sarah L. Emel1, Steven J. Franks2 and Rachel B. Spigler1 1Department of Biology, Temple University, BioLife Building, 1900 N. 12th St, Philadelphia, PA 19122, USA; 2Department of Biological Sciences, Fordham University, Larkin Hall, 441 E. Fordham Road, Bronx, NY 20458, USA

Summary Author for correspondence: Pollinators are considered primary selective agents acting on plant traits, and thus variation Rachel B. Spigler in the strength of the plant–pollinator interaction might drive variation in the opportunity for Tel: +1 215 204 8855 selection and selection intensity across plant populations. Here, we examine whether these Email: [email protected] critical evolutionary parameters covary with pollination intensity across wild populations of Received: 9 December 2016 the biennial Sabatia angularis. Accepted: 5 April 2017 We quantified pollination intensity in each of nine S. angularis populations as mean stigmatic pollen load per population. For female fitness and three components, fruit number, fruit New Phytologist (2017) set (proportion of flowers setting fruit) and number of seeds per fruit, we evaluated whether doi: 10.1111/nph.14608 the opportunity for selection varied with pollination intensity. We used phenotypic selection analyses to test for interactions between pollination intensity and selection gradients for five Key words: fitness components, flowering floral traits, including flowering phenology. phenology, flowering synchrony, opportunity The opportunity for selection via fruit set and seeds per fruit declined significantly with for selection, petal area, phenotypic increasing pollen receipt, as expected. We demonstrated significant directional selection on selection, plant–pollinator interactions, multiple traits across populations. We also found that selection intensity for all traits depended Sabatia angularis. on pollination intensity. Consistent with general theory about the relationship between biotic interaction strength and the intensity of selection, our study suggests that variation in pollination intensity drives variation in selection across S. angularis populations.

populations (Benkman, 2013; Vanhoenacker et al., 2013). For Introduction mutualistic relationships such as plant–pollinator interactions, Understanding patterns of selection within and among natural decreased interaction strength is expected to result in reduced populations is a major goal in evolutionary biology. For plant mean absolute fitness of the population and, as a general conse- species that rely on pollinators for successful reproduction, quence, increased variance in relative fitness (Rundle & Vamosi, plant–pollinator interactions are predicted to play critical roles in 1996; Benkman, 2013; Vanhoenacker et al., 2013). It is this vari- shaping selection on traits related to pollen receipt and export ance in relative fitness that represents the opportunity for selec- (Harder & Barrett, 2006; Harder & Johnson, 2009). Indeed, tion (Crow, 1958; Arnold & Wade, 1984). Moreover, for traits experimental studies have demonstrated pollinator-mediated that influence fitness and are at the interface of the interaction, selection on a number of floral traits within populations (e.g. the strength of selection should also covary with biotic interaction Johnston, 1991; Caruso, 2000; Alexandersson & Johnson, 2002; strength (Benkman, 2013; Vanhoenacker et al., 2013), such that Sandring & Agren, 2009; Sletvold et al., 2010; Chapurlat et al., selection on floral traits should be greatest where plant–pollinator 2015; Lavi & Sapir, 2015). Given that pollinator and mate avail- interactions are weakest. Bartkowska & Johnston (2015) also ability are well known to be variable among populations, it fol- showed that this prediction arises from a simple arithmetic lows that selection on floral traits should covary with plant– relationship between selection intensity and one metric of plant– pollinator interaction strength among populations within species, pollinator interaction strength, pollen limitation (i.e. the relative yet few have evaluated this prediction (Vanhoenacker et al., increase in seed production from pollen-supplemented flowers or 2013; Sletvold & Agren, 2014; Bartkowska & Johnston, 2015). plants relative to controls; Larson & Barrett, 2000; Ashman et al., Specifically, both the opportunity for selection and selection 2004; Eckert et al., 2010). Others have considered the influence strength are predicted to increase with diminished plant–pollina- of plant–pollinator interactions on selection in terms of resource tor interactions. This expectation follows recent ecological theory availability, viewing pollen as a resource (Wilson, 1995; Rundle pointing more generally to spatial variation in biotic interaction & Vamosi, 1996), with the analogous prediction that, as pollen strength as key to explaining variation in selection across loads decline, selection on traits related to pollen import and

Ó 2017 The Authors New Phytologist (2017) 1 New Phytologist Ó 2017 New Phytologist Trust www.newphytologist.com New 2 Research Phytologist

export should increase (Rundle & Vamosi, 1996). This expecta- Materials and Methods tion should hold even if some pollen deposition is attributable to autonomous self-pollination. Therefore, here we use the more Study system general term ‘pollination intensity’. Importantly, these predictions may be met only where variation in pollination intensity is Sabatia angularis (L.) Pursh (Gentianaceae) is native to the east- relevant, that is, where seed production can be improved by ern USA and Canada. Widely distributed throughout its range, it increased pollination (Campbell & Halama, 1993; Campbell & is found in a diversity of habitats including marshes, dry prairies Bischoff, 2013; Vanhoenacker et al., 2013). Otherwise, beyond and glades, roadsides, and serpentine barrens. A biennial, its seeds some point, seed production and thus selection resulting from disperse in the fall, germinate in the spring and form rosettes that plant–pollinator interactions saturate with continuing increases overwinter until the following spring, when plants bolt. From in pollen receipt. July to August, the plants produce displays of showy pink, Previous studies provide insight into the traits most likely to protandrous flowers, which develop into many-seeded dry dehis- be selected for by pollinators and thus result in more reliable cent capsules. The self-compatible flowers are nectarless but do pollination (Knight et al., 2005). Pollinators are expected to have a scent and offer a pollen reward to a suite of generalist floral favor plants with traits that promote attraction, ranging from visitors including leaf-cutter bees (Megachilidae), sweat bees greater flower number, size, and floral display to color or (Halictidae), andrenid bees (Andrenidae), bumblebees (Apidae), greater overall plant size (Harder & Barrett, 2006; Harder & small carpenter bees (Anthophoridae), and hover flies (Syrphidae) Johnson, 2009; Campbell & Bischoff, 2013). Rarer are studies (Dudash, 1987; Spigler, 2007). of selection on floral phenological traits, but these traits may Prior work has demonstrated that pollen receipt varies widely also be under selection by pollinators (reviewed in Munguıa- among populations of S. angularis (Spigler & Chang, 2008) and Rosas et al., 2011). Of the few studies investigating selection that both fruit and seed production can be pollen limited on phenological traits, most have focused on date of first or (Dudash, 1993; Spigler & Chang, 2009). Much higher pollen last flower (e.g. Sandring & Agren, 2009; Sletvold & Agren, limitation occurs for seeds per fruit, given that a single fruit can 2014; Chapurlat et al., 2015; and see Munguıa-Rosas et al., produce anywhere between several and > 1000 seeds (Dudash, 2011). Other phenological traits such as flowering duration 1993; Spigler & Chang, 2008). Moreover, flower number per and synchrony may also be important. For example, pollinators plant varies markedly in this monocarpic species, with plants pro- might be more likely to visit plants that flower synchronously ducing anywhere from a single flower to upwards of 300. Thus, with the rest of the population (e.g. Augspurger, 1981; Rathke flower number not only probably influences pollinator attraction & Lacey, 1985), or alternatively, asynchronous flowering may via floral display, but also sets an upper limit on fruit number be favored where competition for pollinators is high (Elzinga and consequently female fitness. Together, fruit number and et al., 2007). Selection on these traits should be greatest where number of seeds per fruit are expected to lead to wide variation pollination intensity is lowest. in fitness among plants, particularly where pollen loads are low. Support for the hypotheses that the opportunity for selection Populations of S. angularis vary greatly in size (Spigler & Chang, and selection strength covary with pollination intensity thus far 2008), but importantly pollen limitation can occur in both small has been equivocal. The few studies evaluating the strength of and large populations (Dudash, 1993; Spigler & Chang, 2009; selection on a number of floral traits in relation to pollen limita- R. B. Spigler, unpublished) and pollen loads need not correlate tion across species provide general support (Ashman & Morgan, with population size or density (Spigler & Chang, 2008). Popu- 2004; Sletvold & Agren, 2014; Bartkowska & Johnston, 2015; lations range from mixed mating to highly outcrossing (mean = Trunschke et al., 2017). However, findings from studies of these multi-locus outcrossing rate (tm) 0.78 SD; Spigler et al., patterns across populations within species have been mixed 2010). (Rundle & Vamosi, 1996; Vanhoenacker et al., 2013; Sletvold & Agren, 2014; Bartkowska & Johnston, 2015). Study area In this study, we evaluated phenotypic selection on traits of the biennial Sabatia angularis, an insect-pollinated plant prone The current study was conducted across S. angularis popula- to pollen limitation (Dudash, 1993; Spigler & Chang, 2009), tions occurring in the rare serpentine barren grasslands of and evaluated the hypothesis that variation in pollination inten- southeastern Pennsylvania, USA. Populations throughout the sity, measured here as mean pollen receipt, predicts selection serpentine system vary widely in size, as is typical of the species across populations. First, we quantified pollination intensity and more generally (Spigler & Chang, 2008). Because we were the opportunity for selection within each of nine wild interested in capturing dynamics of populations spanning a S. angularis populations and asked whether they covary. Next, broad range of pollination conditions, we included all serpen- we performed phenotypic selection analyses, estimating selection tine populations located during 2014 (n = 9; Table 1). We strength (i.e. selection gradients; Lande & Arnold, 1983) for included in the study all individuals for which we could obtain five display and phenological traits and testing whether these both fitness and trait data; in populations of 70 individuals or gradients vary with pollination intensity across populations. We fewer, our sample size represented 67–87% of the population, conducted these analyses for total female fitness as well as its and in larger populations up to 115 randomly selected individ- fitness components. uals were included.

New Phytologist (2017) Ó 2017 The Authors www.newphytologist.com New Phytologist Ó 2017 New Phytologist Trust New Phytologist Research 3

Table 1 Summary information for nine study populations of Sabatia angularis, including geographic coordinates, population size, density as number of con- speciﬁcs in a 1-m radius (mean SD), and pollen receipt (mean SD)

Population Geographic coordinates Size Density Pollen receipt (grains)

BTM 39°43.9960N, 76°2.6500W 55 18.7 13.8 241.1 114.6 F9 39°43.8520N, 76°2.8260W 2155 14.2 16.5 338.9 123 FH 39°58.8030N, 75°36.0820W 502 1.5 1.5 203.1 220.8 HIS 39°44.2560N, 76°2.3170W 19 0.5 0.9 200 90.6 HM 39°59.9830N, 75°34.1160W 2094 6.5 6.6 172.5 100.3 MB 39°57.0500N, 75°39.7320W 33 0.2 0.4 293.6 154.2 PH 39°55.7830N, 75°34.1660W7014 11.5 283.7 230.6 SB2 40°0.5790N, 75°31.3240W 46 0.8 1.1 126.5 121.8 UB2 39°54.7000N, 75°42.8950W 325 2.3 2.5 257.2 134.4

study. There is no early-acting inbreeding depression in S. Estimating pollination intensity angularis; fruit set and number of seeds per fruit resulting from To characterize pollination intensity, we examined stigmatic pol- outcrossed and selfed pollinations are identical (Dudash, 1990; len loads. Previous work has shown that both mean fruit set (pro- Spigler et al., 2017). Finally, we note that neither population size portion of flowers setting fruit) and seeds per fruit increase with nor density is correlated with mean pollen receipt across our mean pollen receipt across S. angularis populations (Spigler & study populations (r = 0.18, P = 0.65, and r = 0.47, P = 0.20, Chang, 2008). In each study population, we selected a single respectively). flower at the bud stage on up to 20 randomly designated plants (mean 14.6 4.3 SD). Buds were selected once at least 20% of Measuring phenotypic traits and fitness flowers on a given plant had opened, and the stigmas were collected after flower wilting. In this way, pollen receipt on each At the outset of the study, we measured plant height and total stigma is not limited to a certain day but encompasses the entire number of buds (flowers) on all study plants. Each of these traits lifespan of that flower, ensuring full opportunities for pollen cap- reflects in part a plant’s resource status but also can influence ture. In addition, because plants were chosen randomly before apparency to pollinators (Peakall & Handel, 1993). To estimate flowering and varied in their flowering phenologies (see the petal area as a metric of flower size, we measured the length and Results section), the stigmas collected from these plants represent width of a haphazardly chosen petal on three flowers per plant. a random sample of pollen loads across the entire flowering sea- We standardized measurements of flowers within and across son within each population. This representation is confirmed by plants by only measuring flowers in the male phase in the pri- data on sampling dates; the mean per cent overlap between dates mary position of cymes on lateral branches and only after least of flower bud selection and total flowering period per population two flowers were open on each plant (i.e. flowers were measured was 73% ( 0.12 SD). In the laboratory, the bi-lobed stigmas relatively early in a plant’s flowering phenology but the first two were soaked in 8M NaOH for 1 h, and the autofluorescent pollen flowers open per plant were never used). Petal area was approxi- grains were visualized on an epifluorescence microscope at 94 mated using the formula for the area of an ellipse and petal length magnification. For each stigma, we enumerated the number of and width. We evaluated how well estimates from this calculation pollen grains on one of the two stigma lobes, selected at random, matched actual total petal area by collecting and photographing a following Spigler & Chang (2008). We then calculated mean sample of flowers (n = 24), determining the petal area of each pollen receipt for each population. We note that, because from digital photographs using IMAGEJ software (Schneider et al., S. angularis is capable of autonomous self-pollination in the 2012), and regressing this measure against estimated total area absence of pollinators (provided pollen has not been completely (estimated area of a single petal multiplied by 5 for the number removed before the female phase; R. B. Spigler, unpublished), of petals per flower). Estimated petal area strongly predicted total pollen per stigma could represent a mixture of pollinator- actual petal area (R2 = 0.80; P < 0.0001), and therefore we used mediated and autonomous pollen receipt. Thus, although the mean petal area per plant as measured in the field in all percentage of the contribution of autonomous self-pollination in subsequent analyses. In addition, we visited populations every 3– natural populations is low on average (R. B. Spigler, unpub- 4 d throughout the flowering season and upon each visit recorded lished), stigmatic pollen loads may not only reflect pollinator- the date of first flower, the number of open flowers for each mediated pollen deposition, and consequently plant–pollinator plant, and the date of last flower. From these data, we determined interactions sensu stricto. Instead, they may be better viewed in flowering duration and maximum and average daily floral display the context of pollen as a resource (e.g. Rundle & Vamosi, per individual. Daily floral display data were also used to calcu- 1996). We also note that in the case of variance in self-pollen late an index of flowering synchrony for each plant in relation to deposition, via either autonomous or pollinator-mediated selfing, the flowering schedule of its population. To accurately capture issues of pollen quality are not confounded with quantity in our the overlap of the entire flowering curve of an individual

Ó 2017 The Authors New Phytologist (2017) New Phytologist Ó 2017 New Phytologist Trust www.newphytologist.com New 4 Research Phytologist

compared with the population, we implemented the metric P < 0.001, respectively). Thus, we considered five traits in subse- developed by Elzinga et al. (2007), calculating flowering syn- quent phenotypic selection analyses: flower number, plant height, chrony for each individual as: petal area, flowering duration, and flowering synchrony. Pairwise "# ! trait correlations for each population are reported in Supporting Xk x Information Table S1. P t k ft Eqn 1 We estimated selection gradients through total female fitness, ¼ x t 1 t¼1 t fruit number, and number of seeds per fruit using a standard Lande & Arnold (1983) approach (Conner, 1996; Conner & (xt, the total number of open flowers on the focal individual on Hartl, 2004). Because population-specific gradients are often of census day t; ft, the proportion of the total number of flowers interest to other researchers, we evaluated selection gradients for open in the population on day t; k, the maximum number of our five focal traits for each population separately before explic- census days.) When the within-individual flowering peak matches itly testing whether selection intensity varied with pollen receipt the population peak, this metric approaches 1 (complete syn- across populations, our main hypothesis (described in the next chrony), whereas when the within-individual flowering peak paragraph). The population-specific selection gradient results are diverges from the population peak, the metric approaches 0 provided in Table S2. Here, we standardized each trait within (complete asynchrony). populations (e.g. Caruso, 2000; Caruso et al., 2003) to a mean of We estimated the following metrics of female fitness. We 0 and a standard deviation of 1 and relativized total female fitness counted total fruit number per plant, excluding fruits lost to pre- and its components by their within-population means as earlier dispersal seed predation, and estimated the mean number of seeds when calculating the opportunity for selection. To identify possi- per fruit. To estimate the mean number of seeds per fruit, we col- ble multicolinearity among predictor variables in these models, lected c. 30% of the total number of fruits produced per plant, we calculated variance inflation factors (VIFs) using the R pack- weighed them individually and converted fruit mass to seed num- age ‘CARS’. These results showed potential evidence of multicolin- ber using a linear multiplier following Spigler & Chang (2009). earity (VIFs between 4 and 7.2) among traits in three Total female fitness per plant was determined by multiplying populations. However, results from analyses removing the trait these fitness components. with the highest VIF in each of those populations rendered selec- We calculated the opportunity for selection in each population tion gradients that fitted within 95% confidence intervals of the as the variance in relative fitness (Crow, 1958; Arnold & Wade, original selection gradients. Thus, we retained all variables in the 1984) for total female fitness and its components, fruit produc- selection gradient analysis. tion as fruit number and fruit set (the proportion of flowers pro- To robustly test the hypothesis that selection gradients vary ducing fruits) and the number of seeds per fruit. By relativizing with pollen receipt, however, we used a combined phenotypic these fitness measures within populations (dividing each individ- selection analysis, in which all individual-level phenotypic traits ual value by the population mean), we ensured that the opportu- were standardized and female fitness, fruit number, and number nity for selection depended only on variance within populations, of seeds per fruit were relativized after pooling data from all pop- not differences in mean absolute fitness (Benkman, 2013). ulations (ter Horst et al., 2015). Phenotypic selection analyses for relative female fitness and each of its components were performed using linear mixed-effects models (using the MIXED procedure in Statistical analyses SAS 9.3 (SAS Institute, Cary, NC, USA)) with flower number, We used simple regression analyses to evaluate whether the plant height, petal area, flowering duration, flowering synchrony, opportunity for selection for each fitness metric varied as a func- and mean population pollen receipt as predictor variables. We tion of mean pollen receipt in R v.3.2.3 (R Development Core used VIFs to identify multicollinearity of variables in this analysis Team, 2015). Scatterplots of these data revealed a nonlinear rela- as described in the previous paragraph; all traits had low VIFs tionship between the opportunity for selection via number of and were thus all retained in the models. We further included all seeds per fruit and mean pollen receipt, as predicted by theory pairwise interactions between each trait and mean pollen receipt, more generally (Benkman, 2013; Vanhoenacker et al., 2013); which allowed us to explicitly test our hypothesis that the selec- therefore, we also evaluated a quadratic regression for opportu- tion gradient(s) varies with pollen receipt across populations. We nity for selection based on this fitness parameter. considered quadratic terms for individual-level traits but retained Before phenotypic selection analyses, we evaluated pairwise them only if significant (a ≤ 0.05). Recognizing that differences phenotypic correlations among all measured plant traits within in population size and/or mean density per population may be populations in R and used false discovery rate (FDR) to correct important in determining fitness, we initially considered models the significance levels for multiple comparisons within popula- including each of these variables, but in no case were they signifi- tions (Benjamini & Hochberg, 1995). We found that maximum cant and they were omitted from the final models. However, all floral display and average floral display were highly correlated main effects of individual-level phenotypic traits and their inter- with flower number in many populations (r = 0.76–0.98, actions with mean population pollen receipt were retained in the P < 0.0001 and r = 0.52–0.96, P < 0.001–0.003, respectively). analyses regardless of significance as they represent tests of our Similarly, petal width and length were both highly correlated main hypotheses. Population identity was included as a random with petal area (r = 0.80–0.97, P < 0.0001 and r = 0.77–0.95, term to account for differences in mean fitness across populations

New Phytologist (2017) Ó 2017 The Authors www.newphytologist.com New Phytologist Ó 2017 New Phytologist Trust New Phytologist Research 5 not related to variation in pollen receipt and because pollen Results receipt is a population-level variable. Initial evaluation of residu- als from the final models revealed heteroscedasticity. To account Opportunity for selection vs pollination intensity for heterogeneity of variances among populations, we further included population identity as a grouping factor in a ‘repeated’ We predicted that the opportunity for selection should be great- statement (Littell et al., 1996). Inclusion of this grouping factor est where pollination intensity is lowest. Supporting this predic- uses generalized least squares rather than ordinary least squares tion, we found that the opportunity for selection within for estimating parameters and is analogous to a weighted least populations as estimated for total female fitness, fruit number, squares regression, advocated for use in selection analyses when fruit set, and number of seeds per fruit tended to decline with variances among groups are unequal (Stanton & Thiede, 2005). increasing mean pollen receipt. Although we note that our power In addition, we recognized that some individuals had extremely to detect significance may generally have been limited, we found high fitness; to determine whether these or other individuals may that this relationship was significant for fruit set and number of be overly influential in our analyses, we examined influence diag- seeds per fruit (Fig. 1). For the opportunity for selection via num- nostics for each model in SAS using a combination of Cook’s D, ber of seeds per fruit, the relationship was nonlinear, as revealed covariance ratio, press statistic, and restricted likelihood distance by a strong quadratic relationship (Fig. 1d). The quadratic rela- scores to identify potentially influential data points. Only one tionship weakened (R2 = 0.71; P = 0.098) if the population with individual identified as ‘influential’ was truly a biological outlier the highest pollen receipt was omitted, but overall, a significant (very low fruit number compared with flower number because of (linear) relationship remained (R2 = 0.62; P = 0.020). We also broken branches); the remaining points were not. We present evaluated the relationships between mean pollen receipt and results from analyses removing the few (two to three) influential opportunity for selection via fruit set and opportunity for points for total female fitness and its components, as these results selection via number of seeds per fruit when the population with represent the selection gradients applicable to c. 99% of S. the lowest pollen receipt and highest variance in fitness, popula- angularis individuals sampled. However, we also indicate in the tion SB2, was removed. The significance of the relationship did results whether model terms are statistically significant when not remain for the opportunity for selection via number of seeds using the full data set, given that those few influential points are per fruit (R2 = 0.01; P = 0.84), but did remain for the opportu- nonetheless biologically real. nity for selection via fruit set (R2 = 0.58; P = 0.028). We recognize that there are a number of ways that one can approach evaluating differences in phenotypic selection across Strength of selection vs pollination intensity populations (or treatments). One choice is whether to relativize fitness and standardize traits within vs across (i.e. by pooling) pop- Selection via total female fitness Our phenotypic selection anal- ulations (discussed in ter Horst et al., 2015). While the hypothe- ysis revealed significant directional selection, on average, for sis of biotic interactions driving spatial variance in selection increased flower number, plant height, and petal area, and intensity explored here inherently implies comparisons of selec- marginally significant selection for increased flowering duration tion across populations and thus might best be evaluated by rela- (Table 2). Significant quadratic terms for plant height and petal tivizing/standardizing within populations, this approach also area further indicated that their effect on female fitness was non- presents some difficulties. One of these is that, if data are treated linear; at low trait values, increases in plant height and petal area in this way, then it may be difficult to tell if differences in selec- only modestly increased fitness, but the change in fitness acceler- tion gradients among populations are truly attributable to differ- ated as the trait value increased. Selection intensity on flower ences in the covariance between traits and fitness or attributable number as estimated by the selection gradient was three to six to differences in trait distributions among populations. In addi- times greater than that for any other trait. tion, specific to our own data set, evaluation of residual plots The slopes of the relationships between fitness and the traits revealed strong heteroscedasticity that was not resolved by flower number, plant height and flowering duration (i.e. the accounting for differences in variance among populations as it selection gradients) depended on mean population pollen receipt, was when data were relativized/standardized across populations. revealed by statistically significant interactions between these Strong heteroscedasticity coupled with uneven sample sizes (as a traits and pollen receipt in our analysis. The interaction between result of differences in population size in our case) can lead to flowering synchrony and pollen receipt fell just short of our incorrect standard errors and hypothesis tests, leading us to focus defined significance level (a = 0.05). Interaction plots (Fig. 2) on analyses based on data pooled across populations. However, reveal the nature of these interactions. Although caution must be the results of the analyses when evaluated in the two different used in extrapolating fitness surfaces beyond the range of the ways (within vs across populations) were similar (data not data, the surfaces based on our models demonstrate how the slope shown), with significant interactions between traits and pollen of the relationship between relative fitness and each trait varies receipt found in both analyses, revealing similar fitness surfaces. along the pollen receipt axis (Fig. 2). We found a significantly This similarity indicates that the choice of how one scales the positive interaction coefficient between flower number and pollen data does not influence our conclusions. We provide means and receipt (Table 2), such that, while total female fitness strongly variance for fitness estimated across all populations and within increased with flower number across the entire range of pollina- each population in Table S3. tion intensity examined, the slope of the relationship between

Ó 2017 The Authors New Phytologist (2017) New Phytologist Ó 2017 New Phytologist Trust www.newphytologist.com New 6 Research Phytologist

Fig. 1 Relationships between the opportunity for selection, estimated for (a) total female fitness, (b) fruit number, (c) fruit set, and (d) number of seeds per fruit, and mean pollen receipt across nine Sabatia angularis populations. Solid blue lines indicate significant best fit linear or quadratic regression, while dashed blue lines indicate a nonsignificant regression line.

Table 2 General linear mixed-effects model results including standardized selection gradients (b or c; estimate SE) and interactions with pollen receipt (estimate SE) for ﬁve traits via total female ﬁtness and its components across nine Sabatia angularis populations

Total female ﬁtness Fruit number Number of seeds per fruit

Den. Den. Trait Den. df Estimate P df Estimate P df Estimate P

† † Flower number 306 0.5672 0.1178 < 0.001 336 0.9957 0.0452 < 0.001 284 -0.1152 0.0997 0.249 † Plant height 144 0.1793 0.0819 0.030 143 0.0082 0.0354 0.818 244 0.2364 0.0923 0.011 † † Petal area 184 0.169 0.0831 0.043 113 0.0947 0.0312 0.003 232 0.0312 0.0883 0.724 † Flowering duration 199 0.1469 0.0757 0.054 227 -0.0172 0.0334 0.607 353 0.3257 0.084 < 0.001 † † Flowering synchrony 153 0.0966 0.0588 0.102 170 -0.0439 0.0295 0.139 186 0.1139 0.0698 0.104 Pollen receipt 4.55 0.0006 0.0005 0.267 3.8 0.0004 0.0003 0.265 8.18 -0.0004 0.0004 0.364 † Plant height2 161 0.0504 0.019 0.009 † Petal area2 313 0.0751 0.0221 < 0.001 † Flower number 9 pollen receipt 327 0.0019 0.0005 < 0.001 316 0.0005 0.0002 0.002 214 0.0004 0.0004 0.290 † Plant height 9 pollen receipt 200 -0.0007 0.0003 0.030 172 0 0.0001 0.773 293 -0.0008 0.0004 0.042 † † Petal area 9 pollen receipt 204 -0.0002 0.0003 0.459 139 -0.0003 0.0001 0.010 200 0.0005 0.0004 0.129 † Flowering duration 9 pollen 211 -0.0005 0.0002 0.044 243 0.0001 0.0001 0.555 246 -0.0006 0.0003 0.064 receipt † † Flowering synchrony 9 pollen 254 -0.0006 0.0003 0.052 201 0.0002 0.0001 0.123 376 -0.0012 0.0003 < 0.001 receipt Population 0.0033 0.0034 0.171 0.0019 0.0018 0.150 0.0019 0.0018 0.150

Model results represent data sets for each fitness metric with highly influential points removed (see the main text for details). P-values from these models † are shown with significant P-values at a = 0.05 indicated in bold, and significant P-values from models including all points denoted by . Quadratic terms for plant height and petal area were included where significant, and the estimates and SE were doubled as per Stinchcombe et al. (2008). Denominator (Den.) df are provided and all numerator df are 1. Population identity was included as a random effect; results presented represent Wald test results.

fitness and flower number steepened with increasing pollen with positive slopes between fitness and each trait at low pollen receipt (Fig. 2a). For plant height, however, the interaction coef- receipt that declined and became negative at high pollen receipt ficient was negative (Table 2); the slope of the fitness surface for levels. Thus, it was not simply a change in the magnitude of plant height decreased with increasing pollen receipt (Fig. 2b). selection but a change in direction along the pollination intensity For both flowering duration (Fig. 2c) and synchrony (Fig. 2d), gradient. negative interaction coefficients (Table 2) also indicated that selection gradients decreased with increasing pollen receipt. Inter- Selection via fitness components: fruit number and number of estingly, for these interactions, the fitness surfaces also twisted, seeds per fruit For the analysis of fruit number, there was a

New Phytologist (2017) Ó 2017 The Authors www.newphytologist.com New Phytologist Ó 2017 New Phytologist Trust New Phytologist Research 7

Fig. 2 Predicted fitness surfaces illustrating interactive effects between pollen receipt and standardized (a) flower number, (b) plant height, (c) flowering duration, and (d) flowering synchrony on relative total female fitness across nine Sabatia angularis populations. Interactions where P < 0.10 are presented (see Table 2). Surfaces were created using the R package ‘LATTICE’ from linear mixed-effects model predictions holding all other variables constant at their mean (mean = 0 for standardized traits). Color reflects z (fitness)-axis value. Note that the range of relative fitness values on each z- axis varies as it reflects the proportion of the variance in fitness explained by each trait after accounting for the contributions of the other parameters in the model. Inset scatterplots display the distribution of observed trait and pollen receipt values. strong and highly significant selection gradient for flower num- the nature of interactions between the effects of pollen receipt ber, which was at least an order of magnitude greater than selec- and both flowering duration (Fig. 3d) and synchrony (Fig. 3e) on tion through any other trait (Table 2). For the remaining traits, relative number of seeds per fruit matched the shapes of those for the selection gradient analysis for fruit number is akin to that for total female fitness; selection was greatest where pollen receipt proportion fruit set because we are accounting for variation in was low and declined with increasing pollen receipt, becoming total flower number. The only other significant selection gradient negative where pollen receipt was greatest. was for petal area (Table 2). We further found that selection gradients for both flower number and petal area varied depending Discussion on mean pollen receipt (Table 2). For flower number, this interaction was similar to, but weaker than, that seen for relative total The results presented here provide evidence that the opportunity female fitness (Table 2; Fig. 3a). The interaction between petal for selection and strength of phenotypic selection on floral and area and pollen receipt was negative (Table 2); the slope of the phenological traits vary in relation to pollination intensity in relationship between relative fruit number and petal area S. angularis. Few studies have evaluated these relationships within increased as pollen receipt declined (Fig. 3b; rotated view of sur- species (Rundle & Vamosi, 1996; Vanhoenacker et al., 2013; face provided in Fig. S1). We also detected a significant interac- Sletvold & Agren, 2014; Bartkowska & Johnston, 2015), and tion between flowering synchrony and pollen receipt, but only these few prior studies provided equivocal results. Our work when all data points were included in the analysis (Table 2). Here shows how these relationships operate for total female fitness as again the fitness surface mimicked that found for total female fit- well as its components and provides insight into how pollination ness, with positive selection gradients at low pollen receipt levels intensity shapes spatial variation in selection on floral traits. that declined and became negative at high pollen receipt levels. Moreover, the results support broader ecological theory about the There was also significant selection for greater plant height and role of biotic interaction strength in shaping fundamental evolu- longer flowering duration, on average, via number of seeds per tionary parameters (Benkman, 2013). fruit, and again we found that selection gradients for some traits varied with pollen receipt. Specifically, mean pollen receipt inter- Pollination intensity and the opportunity for selection acted significantly with plant height and flowering synchrony to determine number of seeds per fruit. This interaction was also Recent models predict that, when biotic interaction strength is marginally significant for flowering duration. In all cases, the low and acts as a driving evolutionary force, variation in relative interaction coefficients were negative (Table 2). For plant height, fitness and thus the opportunity for selection should increase we saw that the slope of the relationship between predicted rela- (Benkman, 2013; Vanhoenacker et al., 2013). Only a few other tive number of seeds per fruit and plant height was positive when studies besides ours have evaluated the influence of biotic interac- pollen receipt was low, but declined and became nearly flat when tion strength on the opportunity for selection within species pollen receipt was high (Fig. 3c). The fitness surfaces revealing (Benkman, 2013; Sletvold & Agren, 2014; Bartkowska &

Ó 2017 The Authors New Phytologist (2017) New Phytologist Ó 2017 New Phytologist Trust www.newphytologist.com New 8 Research Phytologist

Fig. 3 Predicted fitness surfaces for female fitness components, (a, b) fruit number and (c–e) number of seeds per fruit, illustrating interactions between pollen receipt and standardized (a) flower number, (b) petal area, (c) plant height, (d) flowering duration, and (e) flowering synchrony across nine Sabatia angularis populations. Interactions where P < 0.05 are presented (see Table 2). Surfaces were created using the R package ‘LATTICE’ from linear mixed-effects model predictions holding all other variables constant at their mean (mean = 0 for standardized traits). Color reflects z (fitness component)-axis value. Note that the range of relative fitness values on each z-axis varies as it reflects the proportion of the variance in fitness explained by each trait after accounting for the contributions of the other parameters in the model. Inset scatterplots display the distribution of observed trait and pollen receipt values.

selection in all three cases. However, in two previous studies on the influence of pollination intensity on the opportunity for selection within species, neither found a significant relationship (Sletvold & Agren, 2014; Bartkowska & Johnston, 2015). Our data are consistent with the hypothesis that pollination intensity is a significant driver of the opportunity for selection across S. angularis populations. We found that the opportunity for selection based on total female fitness and its components across our study populations tended to decline with increasing pollination intensity, significantly so for fruit set and number of seeds per fruit. It is interesting that this relationship for number of seeds per fruit was nonlinear, as is also predicted (Benkman, 2013; Vanhoenacker et al., 2013); that is, the opportunity for selection should decrease sharply with increased biotic interaction strength at low levels and then begin to decelerate at higher interaction strengths. Inclusion of additional populations, particularly at the extremes of the pollination intensity spectrum, is needed to more robustly evaluate the shape of this relationship. Nonethe- less, the work involved in conducting phenotypic selection analyses often limits the number of populations that can be included, and our sample size is comparable to or greater than that of similar studies (mean = 9.5; median = 5.5; n = 5; Benkman, 2013; Sletvold & Agren, 2014; Bartkowska & Johnston, 2015). Despite potentially limited statistical power, the lack of a relationship between pollen receipt and the opportunity for selection via total female fitness and fruit number is perhaps not surprising. Both are largely driven by variance in flower number, which itself is not influenced by biotic interactions, whereas fruit set and number of seeds per fruit are likely to be influenced directly by pollination intensity. This may in part explain why the two previous studies did not find significant relationships between the opportunity for selection and pollination intensity as they focused only on total female fitness. Of course, pollination intensity is probably not the only driver of the opportunity for selection. Other abiotic or biotic factors, including resource competition and predispersal seed predation, may be important in determining the selective environment. For example, population SB2 in this study had several times greater variance in fruit Johnston, 2015). Benkman (2013) presented three distinct biotic set than the rest of the populations (Fig. 1c). While it had the interactions, a Clark’s nutcracker (Nucifraga columbiana)– lowest mean pollen receipt, it also had the highest the rate of pre- whitebark pine (Pinus albicaulis) mutualism and two antago- dispersal seed predation rate by Lepidopteran larvae and high nisms, finding the expected relationship with the opportunity for levels of herbivory (data not shown), which could have inflated

New Phytologist (2017) Ó 2017 The Authors www.newphytologist.com New Phytologist Ó 2017 New Phytologist Trust New Phytologist Research 9 variation in relative fitness compared with the other study popu- studies on flowering duration; longer flowering duration under lations. Experimental work can best tease apart the contributions low pollination conditions can maximize opportunities for polli- of these biotic interactions to variation in fitness (Bartkowska & nator visitation and outcrossing, while in populations where pol- Johnston, 2012; Sletvold et al., 2015). lination conditions are sufficient to allow pollen receipt over a shorter period, longer duration may only serve to increase the chance of encountering antagonists (predispersal seed predators Pollination intensity and the strength of selection and florivores) or to extend fruit and seed production later into We found that selection via total female fitness varied with polli- the season when resources begin to deplete (reviewed in Elzinga nation intensity across populations of S. angularis. Specifically, et al., 2007). Resource depletion may be particularly important our results revealed significant interactions between mean pollen for fitness in S. angularis given declines in seed set, but not polli- receipt per population and flower number, plant height, and nator-mediated pollen receipt, across the season (R. B. Spigler, flowering duration, with weaker support for an interaction unpublished). A similar fitness surface for total fitness was between pollen receipt and flowering synchrony. We also found observed for the interaction between pollen receipt and flowering evidence of selection on petal area, but it was consistent across synchrony, indicating that selection favors flowering syn- populations regardless of pollination intensity. For a generalist- chronously with the population when pollen is scarce but asyn- pollinated plant such as S. angularis, one predicts that more chronously where pollination conditions are more favorable. This flowers, taller plant height, and larger flowers allow for greater pattern is also consistent with prior findings that flowering syn- pollinator attraction and thus higher fitness, as seen in many chronously with the rest of the population can promote pollina- previous phenotypic selection studies (e.g. Johnston, 1991; Con- tor attraction and outcrossing (e.g. Augspurger, 1981; Rathke & ner & Rush, 1997; Caruso, 2000; Maad, 2000; Thompson, Lacey, 1985; Marquis, 1988; Murawski & Hamrick, 1992) but 2001; Caruso et al., 2003). In fact, flower number, plant height also attraction of seed predators (e.g. Augspurger, 1981; Brody, and flower size have explicitly been shown to be under pollina- 1997; reviewed in Munguıa-Rosas et al., 2011) and strong tor-mediated selection through experimental manipulation (San- intraspecific competition for pollinators or even resources (e.g. dring & Agren, 2009; Sletvold et al., 2010; Sletvold & Agren, Gomez, 1993; Parra-Tabla & Vargas, 2007). Given evidence for 2014; Chapurlat et al., 2015; Lavi & Sapir, 2015). Thus, we reproductive Allee effects, competition for pollinators, and pre- predicted that selection intensity for these traits would be great- dispersal seed predation across S. angularis populations (Spigler est under low pollen receipt and become weaker as pollen receipt & Chang, 2008, 2009), we interpret our data to indicate that the increases. This prediction was met for selection on plant height negative effects of flowering synchronously begin to outweigh the via total female fitness. Selection on plant height was greatest in benefits where pollen receipt is sufficient. For these and all fitness populations with low pollen receipt, with the relationship surfaces presented here, however, we caution that interpretations between fitness and plant height weakening as pollen receipt may be complicated by the influence of correlational selection increased. Notably, this relationship was nonlinear. To the among traits (Schluter & Nychka, 1994), which has been exam- extent that plant height represents apparency to pollinators (e.g. ined more intensively in other work (Reynolds et al., 2010; Peakall & Handel, 1993), this suggests that fitness gains from Fenster et al., 2015). increased apparancy accelerate with increasing plant height. We By evaluating selection on fitness components, we were able to note that the disruptive selection at high pollen receipt seen in determine through which components selection acts and to detect the predicted surface is probably not biologically real but rather interactions between pollen receipt and selection intensity that attributable to the statistical inference of the model used to visu- may not have translated into patterns seen for total female fitness. alize the interaction (Fig. 2b). For flower number, there was Differences in the traits that influence each component were strong directional selection across all populations, and, although anticipated, because it takes very few pollen grains to set a fruit in the change was slight, selection intensity actually increased with S. angularis, while multiple pollinator visits are probably needed greater pollen receipt. While opposite to the general prediction, to maximize seed number per fruit (Spigler & Chang, 2008). this makes biological sense because, where pollen is not limiting, Similar to the pattern observed for total female fitness, selection plants with more flowers have greater opportunities to increase for flower number via fruit number increased with pollen receipt. their fitness, whereas, where pollen receipt is low, plants with However, we also identified an interaction between the effects of many flowers may not receive enough pollen to set all flowers pollen receipt and selection on petal area via fruit number that into fruits or to fertilize more seeds, placing a strong limit on was not evident for total female fitness. Positive selection for petal fitness. area was strong under poor pollination conditions and steadily Expectations for selection on flowering phenology may be decreased with increasing pollen receipt, suggesting relaxed selec- more complex than those for floral traits, and, in fact, the direction on petal size when pollen is abundant (Fig. 3b). We also tion of selection – not just strength – may be expected to change found that changes in selection on plant height, flowering dura- with pollination intensity (Elzinga et al., 2007; Munguıa-Rosas tion, and flowering synchrony via total female fitness with pollen et al., 2011). We found that, where pollen was scarce, selection receipt arise from variation in the number of seeds per fruit. In favored increased flowering duration in S. angularis, but where all these cases, the fitness surface illustrating the interaction pollen was plentiful, selection acted to decrease it (i.e. became between pollen receipt and plant height, flowering duration or negative). Such a fitness surface fits with expectations from other synchrony fruit mirrored those for total female fitness. A longer

Ó 2017 The Authors New Phytologist (2017) New Phytologist Ó 2017 New Phytologist Trust www.newphytologist.com New 10 Research Phytologist

flowering duration and greater synchrony, in particular, under discussion on phenotypic selection analyses. This manuscript was poor pollination conditions may allow individual flowers to accu- improved by comments from the Editor and anonymous review- mulate higher pollen loads and thus set more seeds per fruit. In ers. Funds for research were provided to R.B.S. by Temple fact, flowering duration and floral longevity are related in University. S. angularis (Spigler, 2017). Our findings join only a few others evaluating the relationship Author contributions between pollination intensity and selection intensity via total female fitness within species in natural systems (Vanhoenacker S.L.E. analyzed the data, interpreted results, and wrote the et al., 2013; Sletvold & Agren, 2014; Bartkowska & Johnston, manuscript; S.J.F. interpreted results and wrote the manuscript; 2015). A strength of our study is that we examined fitness com- and R.B.S. designed and performed the research, analyzed the ponents and explicitly tested for interactions between pollination data, interpreted results and wrote the manuscript. intensity and selection gradients using a comprehensive, robust analysis that accounts for variation within populations as well as References among them. However, we note that we examined variation in net selection. An important distinction is that two of the prior Alexandersson R, Johnson SD. 2002. Pollinator-mediated selection on flower- studies experimentally isolated pollinator-mediated selection by tube length in a hawkmoth-pollinated Gladiolus (Iridacae). Proceedings of the Royal Society B: Biological Sciences 269: 631–636. including pollen-supplemented plants and control plants Arnold SJ, Wade MJ. 1984. On the measurement of natural and sexual selection: (Sletvold & Agren, 2014; Bartkowska & Johnston, 2015). theory. Evolution 38: 709–719. Although net and pollinator-mediated selection can be highly Ashman T-L, Knight TM, Steets JA, Amarasekare P, Burd M, Campbell DR, et al. correlated (Sletvold & Agren, 2014; Chapurlat et al., 2015), Dudash MR, Johnston MO, Mazer SJ, Mitchell RJ 2004. Pollen limitation of plant reproduction: ecological and evolutionary causes and they may not show the same patterns. Sletvold & Agren (2014) consequences. Ecology 85: 2408–2421. found that net selection intensity was strongly related to pollen Ashman T-L, Morgan MT. 2004. Explaining phenotypic selection on plant limitation for one orchid species, consistent with our study and attractive characters: male function, gender balance, or ecological context? that of Vanhoenacker et al. (2013), but did not see this pattern Proceedings of the Royal Society B: Biological Sciences 271: 553–559. for pollinator-mediated selection. Bartkowska & Johnston Augspurger CK. 1981. Reproductive synchrony of a tropical shrub: experimental (2015) found positive, but nonsignificant, trends between studies on effects of pollinators and seed predators in Hybanthus prunifolius (Violaceae). Ecology 62: 775–788. strength of pollinator-mediated selection and pollen limitation Bartkowska MP, Johnston MO. 2012. Pollinators cause stronger selection than in Lobelia cardinalis. Importantly, all of these studies, including herbivores on floral traits in Lobelia cardinalis (Lobelieaceae). New Phytologist ours, only measured female fitness, but for a hermaphroditic 193: 1039–1048. plant male fitness should ideally also be examined, considering Bartkowska MP, Johnston MO. 2015. Pollen limitation and its influence on – that selection on traits via male and female fitness may differ natural selection through seed set. Journal of Evolutionary Biology 28: 2097 2105. (e.g. Campbell, 1989; Sahli & Conner, 2011; reviewed by Benjamini Y, Hochberg Y. 1995. Controlling the false discovery rate: a practical Delph & Ashman, 2006) and these differences may in fact relate and powerful approach to multiple testing. Journal of the Royal Statistical Society to pollination intensity (Ashman & Morgan, 2004). Clearly, B 57: 289–300. more studies are needed to evaluate the generality of these trends Benkman CW. 2013. Biotic interaction strength and the intensity of selection. – for both female and total fitness. Ecology Letters 16: 1054 1060. Brody AK. 1997. Effects of pollinators, herbivores, and seed predators on In conclusion, we identified novel support for pollination flowering phenology. Ecology 78: 1624–1631. intensity as a driver of spatial variation in the opportunity for Campbell DR. 1989. Measurements of selection in a hermaphroditic plant: selection as well as the intensity of phenotypic selection across variation in male and female pollination success. Evolution 43: 318–334. populations of a generalist-pollinated plant. We suspect that, Campbell DR, Bischoff M. 2013. Selection for a floral trait is not mediated by given variable pollination across years, these same patterns are pollen receipt even though seed set in the population is pollen-limited. Functional Ecology 27: 1117–1125. likely to occur when the pollination intensity varies across time Campbell DR, Halama KJ. 1993. Resource and pollen limitations to lifetime within populations, leading to fluctuating selection pressures seed production in a natural plant population. Ecology 74: 1043–1051. across years. We also emphasize the importance of evaluating Caruso CM. 2000. Competition for pollination influences selection on floral selection on less well-studied phenological traits in addition to traits of Ipomopsis aggregata. Evolution 54: 1546–1557. floral ones. Our results contribute to the growing body of knowl- Caruso CM, Peterson SB, Ridley CE. 2003. Natural selection on floral traits of Lobelia (Lobeliaceae): spatial and temporal variation. American Journal of edge about the ecological drivers of spatial variation in selection Botany 90: 1333–1340. and point to the need for further experimental study to isolate Chapurlat E, Agren J, Sletvold N. 2015. Spatial variation in pollinator-mediated and compare the relative importance of multiple abiotic and selection on phenology, floral display and spur length in the orchid biotic drivers shaping these relationships. Gymnadenia conopsea. New Phytologist 208: 1264–1275. Conner JK. 1996. Understanding natural selection: an approach integrating selection gradients, multiplicative fitness components, and path analysis. Acknowledgements Ethology Ecology & Evolution 8: 387–397. Conner JK, Hartl DL. 2004. A primer of ecological genetics. Sunderland, MA, We thank A. Collins, S. Olshevski, J. Wyatt, A. Ky, and espe- USA: Sinauer Associates Inc. cially M. Le Sage, for assistance in the field and lab and M. Sekor Conner JK, Rush S. 1997. Measurements of selection on floral trains in black – for assistance with R code. We also thank J. Connor for helpful mustard, Brassica nigra. Journal of Evolutionary Biology 10: 327 335.

New Phytologist (2017) Ó 2017 The Authors www.newphytologist.com New Phytologist Ó 2017 New Phytologist Trust New Phytologist Research 11

Crow JF. 1958. Some possibilities for measuring selection intensities in man. Reynolds RJ, Childers DK, Pajewski NM. 2010. The distribution and Human Biology 30:1–13. hypothesis testing of eigenvalues from the canonical analysis of the gamma Delph LF, Ashman T-L. 2006. Trait selection in flowering plants: how does matrix of quadratic and correlational selection gradients. Evolution 64: 1076– sexual selection contribute? Integrative and Comparative Biology 46: 465–472. 1085. Dudash MR. 1987. The reproductive biology of Sabatia angularis L. Rundle HD, Vamosi SM. 1996. Selection may be strongest when resources are (Gentianaceae). PhD thesis, University of Illinois, Chicago, IL, USA. scarce: a comment on Wilson. Evolutionary Ecology 10: 559–563. Dudash MR. 1990. Relative fitness of selfed and outcrossed progeny in a self- Sahli HF, Conner JK. 2011. Testing for conflicting and nonadditive selection: compatible, protandrous species, Sabatia angularis L. (Gentianaceae): a floral adaptation to multiple pollinators through male and female fitness. comparison in three environments. Evolution 44: 1129–1139. Evolution 65: 1457–1473. Dudash MR. 1993. Variation in pollen limitation among individuals of Sabatia Sandring S, Agren J. 2009. Pollinator-mediated selection on floral display and angularis (Gentianaceae). Ecology 74: 959–962. flowering time in the perennial herb Arabidopsis lyrata. Evolution 63: 1292– Eckert CG, Kalisz S, Geber MA, Sargent R, Elle E, Cheptou P-O, Goodwillie 1300. C, Johnston MO, Kelly JK, Moeller DA et al. 2010. Plant mating systems in a Schluter D, Nychka D. 1994. Exploring fitness surfaces. American Naturalist changing world. Trends in Ecology and Evolution 25:36–43. 143: 596–616. Elzinga JA, Atlan A, Biere A, Gigord L, Weis AE, Bernasconi G. 2007. Time Schneider CA, Rasband WS, Eliceiri KW. 2012. NIH Image to ImageJ: 25 years after time: flowering phenology and biotic interactions. Trends in Ecology and of image analysis. Nature Methods 9: 671–675. Evolution 22: 432–439. Sletvold N, Agren J. 2014. There is more to pollinator-mediated selection than Fenster CB, Reynolds RJ, Williams CW, Makowsky RM, Dudash MR. 2015. pollen limitation. Evolution 68: 1907–1918. Quantifying hummingbird preference for floral trait combinations: the role of Sletvold N, Grindeland JM, Agren J. 2010. Pollinator-mediated selection on selection on trait interactions in the evolution of pollination syndromes. floral display, spur length and flowering phenology in the deceptive orchid Evolution 69: 1113–1127. Dactylorhiza lapponica. New Phytologist 188: 385–392. Gomez JM. 1993. Phenotypic selection on flowering synchrony in a high Sletvold N, Moritz KK, Agren J. 2015. Additive effects of pollinators and mountain plant, Hormathophylla spinosa (Cruciferae). Journal of Ecology 81: herbivores result in both conflicting and reinforcing selection on floral traits. 605–613. Ecology 96: 214–221. Harder LD, Barrett SCH. 2006. Ecology and evolution of flowers. Oxford, UK: Spigler RB. 2007. The reproductive consequences of reduced population size in the Oxford University Publications. biennial Sabatia angularis (Gentianaceae). PhD thesis, University of Georgia, Harder LD, Johnson SD. 2009. Darwin’s beautiful contrivances: evolutionary Athens, GA, USA. and functional evidence for floral adaptation. New Phytologist 183: 530–545. Spigler RB. 2017. Plasticity of floral longevity and floral display in the ter Horst CP, Lau JA, Cooper IA, Keller KR, La Rosa RJ, Royer AM, Schultheis self-compatible biennial Sabatia angularis (Gentianaceae): untangling the EH, Suwa T, Conner JK. 2015. Quantifying nonadditive selection caused by role of multiple components of pollination. Annals of Botany 119:167– indirect ecological effects. Ecology 96: 2360–2369. 176. Johnston MO. 1991. Pollen limitation of female reproduction in Lobelia Spigler RB, Chang S-M. 2008. Effects of plant abundance on reproductive cardinalis and L. siphilitica. Ecology 72: 1500–1503. success in the biennial Sabatia angularis (Gentianaceae): spatial scale matters. Knight TM, Steets JA, Vamosi JC, Mazer SJ, Burd M, Campbell DR, Dudash Journal of Ecology 96: 323–333. MR, Johnston MO, Mitchell RJ, Ashman T-L. 2005. Pollen limitation of Spigler RB, Chang S-M. 2009. Pollen limitation and reproduction vary with plant reproduction: pattern and process. Annual Review of Ecology, Evolution, population size in experimental populations of Sabatia angularis and Systematics 36: 467–497. (Gentianaceae). Botany-Botanique 87: 330–338. Lande R, Arnold SJ. 1983. The measurement of selection on correlated Spigler RB, Hamrick JL, Chang S-M. 2010. Increased inbreeding but not characters. Evolution 37: 1210–1226. homozygosity in small populations of Sabatia angularis (Gentianaceae). Plant Larson BMH, Barrett SCH. 2000. A comparative analysis of pollen limitation in Systematics and Evolution 284: 131–140. flowering plants. Biological Journal of the Linnean Society 69: 503–520. Spigler RB, Theodorou K, Chang S-M. 2017. Inbreeding depression and Lavi T, Sapir Y. 2015. Are pollinators the agents of selection for the extreme large drift load in small populations at demographic disequilibrium. Evolution size and dark color in Oncocyclus irises? New Phytologist 205: 369–377. 71:81–94. Littell RC, Milliken GA, Stroup WW, Wolfinger RD. 1996. SAS system for Stanton ML, Thiede DA. 2005. Statistical convenience vs biological mixed models. Cary, NC, USA: SAS Institute Inc. insight: consequences of data transformation for the analysis of fitness Maad J. 2000. Phenotypic selection in hawkmoth-pollinated Platanthera bifolia: variation in heterogeneous environments. New Phytologist 166:319– targets and fitness surfaces. Evolution 54: 112–123. 338. Marquis J. 1988. Phenological variation in the neotropical understory shrub Stinchcombe JR, Agrawal AF, Hohenlohe PA, Arnold SJ, Blows MW. 2008. Piper arielanum: causes and consequences. Ecology 69: 1552–1565. Estimating nonlinear selection gradients using quadratic regression coefficients: Munguıa-Rosas MA, Ollerton J, Parra-Tabla V, De-Nova JA. 2011. Meta- double or nothing? Evolution 62: 2435–2440. analysis of phenotypic selection on flowering phenology suggests that early Thompson JD. 2001. How do visitation patterns vary among pollinators in flowering plants are favored. Ecology Letters 14: 511–521. relation to floral display and floral design in a generalist pollination system? Murawski DA, Hamrick JL. 1992. The mating system of Cavanillesia platanifolia Oecologia 126: 386–394. under extremes of flowering-tree density: a test of predictions. Biotropica 24: Trunschke J, Sletvold N, Agren J. 2017. Interaction intensity and pollinator- 99–101. mediated selection. New Phytologist 214: 1381–1389. Parra-Tabla V, Vargas MF. 2007. Flowering synchrony and floral display size Vanhoenacker D, Agren J, Ehrlen J. 2013. Non-linear relationship between affect pollination success in a deceit-pollinated tropical orchid. Acta Oecologica intensity of plant-animal interactions and selection strength. Ecology Letters 16: 131:26–35. 198–205. Peakall R, Handel SN. 1993. Pollinators discriminate among floral heights Wilson P. 1995. Variation in the intensity of pollination in Drosera tracyi. of a sexually deceptive orchid: implications for selection. Evolution 47: Evolutionary Ecology 9: 382–396. 1681–1687. R Development Core Team. 2015. R: a language and environment for statistical computing. R Vienna, Austria: Foundation for Statistical Computing. [WWW Supporting Information document] URL https://www.R-project.org/ [accessed 1 Feb 2016]. Rathke B, Lacey EP. 1985. Phenological patterns of terrestrial plants. Annual Additional Supporting Information may be found online in the Review of Ecology and Systematics 16: 179–214. Supporting Information tab for this article:

Ó 2017 The Authors New Phytologist (2017) New Phytologist Ó 2017 New Phytologist Trust www.newphytologist.com New 12 Research Phytologist

Fig. S1 Rotated predicted fruit number surface (corresponding Table S3 Summary fitness and trait values for nine populations to Fig. 3b) to illustrate interaction between standardized petal of Sabatia angularis area and pollen receipt across nine Sabatia angularis populations. Please note: Wiley Blackwell are not responsible for the content Table S1 Correlations (r) between plant traits for nine popula- or functionality of any Supporting Information supplied by the tions of Sabatia angularis authors. Any queries (other than missing material) should be directed to the New Phytologist Central Office. Table S2 Standardized selection gradients (b) for five traits via total female fitness and its fitness components for each of the nine study Sabatia angularis populations

New Phytologist is an electronic (online-only) journal owned by the New Phytologist Trust, a not-for-profit organization dedicated to the promotion of plant science, facilitating projects from symposia to free access for our Tansley reviews.

Regular papers, Letters, Research reviews, Rapid reports and both Modelling/Theory and Methods papers are encouraged. We are committed to rapid processing, from online submission through to publication ‘as ready’ via Early View – our average time to decision is <26 days. There are no page or colour charges and a PDF version will be provided for each article.

The journal is available online at Wiley Online Library. Visit www.newphytologist.com to search the articles and register for table of contents email alerts.

If you have any questions, do get in touch with Central Office ([email protected]) or, if it is more convenient, our USA Office ([email protected])

For submission instructions, subscription and all the latest information visit www.newphytologist.com

New Phytologist (2017) Ó 2017 The Authors www.newphytologist.com New Phytologist Ó 2017 New Phytologist Trust