Pollinators exert natural selection on flower size and floral display in Penstemon digitalis


Author for correspondence:
Amy L. Parachnowitsch
Tel: +1 607 339 3153
Email: alp43@cornell.edu


  • A major gap in our understanding of floral evolution, especially micro-evolutionary processes, is the role of pollinators in generating patterns of natural selection on floral traits. Here we explicitly tested the role of pollinators in selecting floral traits in a herbaceous perennial, Penstemon digitalis.
  • We manipulated the effect of pollinators on fitness through hand pollinations and compared phenotypic selection in open- and hand-pollinated plants.
  • Despite the lack of pollen limitation in our population, pollinators mediated selection on floral size and floral display. Hand pollinations removed directional selection for larger flowers and stabilizing selection on flower number, suggesting that pollinators were the agents of selection on both of these traits.
  • We reviewed studies that measured natural selection on floral traits by biotic agents and generally found stronger signatures of selection imposed by pollinators than by herbivores and co-flowering plant species.


Pollinators are often thought to be driving floral evolution (Fenster et al., 2004). Indeed, pollinator specialization seems to have driven rapid evolution in some systems (e.g. Kay et al., 2005) and is the main hypothesis put forth to explain the diversity of flowering plants (e.g. Fenster et al., 2004). Moreover, key innovations that allow greater pollinator specialization seem to lead to diversification, as is the case for nectar spurs in Aquilegia (Hodges, 1997). Pollination syndromes (a collection of floral traits associated with attracting a particular group of pollinators) can also explain floral-trait variation (Wilson et al., 2004), suggesting that pollinator specialization has been an important force in floral evolution. Although pollinators seem to be significant drivers of floral evolution on a macro-evolutionary scale, and natural selection on floral traits is common (although not consistent) on a micro-evolutionary scale (Harder & Johnson, 2009), a major gap in our knowledge is whether pollinators are the agents of natural selection within populations (Ashman & Morgan, 2004).

Because flower size and display size are likely to be attractive to pollinators, these floral phenotypes are often assumed to be the result of selection by pollinators (Barrett & Harder, 1996; Ashman & Morgan, 2004). Furthermore, a recent review of phenotypic selection on floral traits shows that selection for larger flowers is common, and selection for flower production even more so (Harder & Johnson, 2009). However, it is also clear that floral signals used for pollinator attraction can be perceived and used by other organisms (Raguso, 2009), making them particularly vulnerable to conflicting selection (Strauss & Irwin, 2004). For example, antagonists such as herbivores and predispersal seed predators, as well as abiotic factors, can drive natural selection on floral traits (Strauss & Whittall, 2006). To conclusively determine the agents of selection, the selective environment must be manipulated (Wade & Kalisz, 1990; Conner & Hartl, 2004). However, only six studies have employed this approach to test the role of pollinators in natural selection.

We explicitly set out to test whether pollinators were acting as agents of selection on floral traits of Penstemon digitalis. Pollinators are thought to have played an important role in the diversification of the genus Penstemon (Wilson et al., 2004) and following the traditional pollination framework this could suggest that pollinators are acting as selective agents in particular Penstemon species. Therefore, we compared natural selection in open- and hand-pollinated P. digitalis to assess whether pollinators were exerting selection on floral traits. To gauge the generality of our findings, we evaluated the role of pollinators as selective agents for multiple species by reviewing studies that specifically manipulated pollination and measured natural selection. We then compared pollinators with two other potential agents of selection (herbivores and co-flowering species).

Materials and Methods

Study system

Penstemon digitalis Nutt. ex Sims (Plantaginaceae) is a native wild flower found in the meadows and prairies of North America. The flowers are protandrous and, while bagged to prevent pollinators, do not set fruit, although geitonogamy within a plant is possible (A. L. Parachnowitsch, unpublished). This plant can also be pollen-limited (Zorn-Arnold & Howe, 2007). Flowers are visited by small- to large-bodied bees throughout the geographical range (Clinebell & Bernhardt, 1998; Mitchell & Ankeny, 2001; Dieringer & Cabrera, 2002). The flowers are mainly white with variable purple striping, and sticky trichomes cover the flowers and flowering stems (but not the leaves). An unidentified micro-lepidopteran is a predispersal seed predator in both Ohio (Mitchell & Ankeny, 2001) and New York (A. L. Parachnowitsch, pers. obs.).

Field experiment

In June 2008, we selected a total of 300 P. digitalis plants in an old-field population in Tompkins County, New York (42°26.428′N 76°25.743′W). To limit bias, we used transects to choose individuals for inclusion in the study. Plants were paired (5 m apart) along four parallel transects in two spatial blocks of 150 plants (c. 25 m apart) and assigned to either an open-pollinated or a hand-pollinated treatment. The spatial blocks were chosen to encompass variation within the population. Our population is situated on a gentle slope ending in a valley, and the two blocks were qualitatively different from each other. The lower block floods more frequently and is generally more open, but the P. digitalis density is higher (A. L. Parachnowitsch, pers. obs.).

Open-pollinated plants were left unmanipulated. Hand-pollinated plants were supplemented with pollen every 2–3 d throughout flowering (eight times for c. 3 wk). Field-collected pollen was applied to the stigmas with wooden toothpicks. We collected pollen from plants at least 5 m away on the day of the pollinations to ensure outcrossing. Generally, two to four donors were used per plant per day, depending on the number of open flowers. Pollinated flowers were also marked with correction fluid at the base to ensure that each flower was supplemented at least once. Hand-pollination allowed us to assess whether plants were pollen limited (Ashman et al., 2004) and to determine whether pollinators were the agents of selection on floral traits (Conner & Hartl, 2004). We expected that selection on floral traits would be stronger in open-pollinated plants than in hand-pollinated plants if pollinators were the agents of selection because hand-pollinations remove the benefit of being attractive to pollinators through two mechanisms. Hand-pollinated plants are no longer pollen limited and, in addition to receiving excess pollen, hand-pollinations provide ‘haphazard’ contributions of pollen that may differ from that of pollinator-deposited pollen.

Phenotypic measurements

We estimated seven phenotypic traits that we hypothesized could be under natural selection in this species: flower size, flower colour, total number of flowers, flower density, aborted flowers, plant height and plant biomass. These traits were not chosen to represent an exhaustive list of traits potentially under selection by pollinators in this species but rather were the hypothesized pollinator-selected traits that we were able to measure in addition to manipulating pollination. For example, floral scents could also be cues under selection by pollinators (Raguso, 2009) and we are studying selection on scents in a separate study. We did not measure flowering phenology, which could also experience pollinator-mediated selection (Sandring & Ågren, 2009), because we wanted to limit our impacts on pollinator behaviour by reducing our visitation to the plants (e.g. Cahill et al., 2001; Hik et al., 2003).

We measured flower size by recording six dimensions of the flower for three haphazardly chosen flowers per plant (Fig. 1). Whenever possible, the three flowers were measured over multiple days (generally 2 d per plant) to capture within-plant variation. Moreover, to reduce error, the same researcher performed all floral measurements. To estimate the visual display of the petals, we measured the width and length of the centre lower lip where the petals are not fused. Pollinating bees enter the flower tube, brushing the sexual organs with their backs (Dieringer & Cabrera, 2002); therefore, we measured the width and length of the bell of the tube where the pollinator body fits into the flower, as well as the full length of the tube. Penstemon flowers have a constricted floral tube around the ovary, which could limit access to the nectaries, as well as to the ovaries (potentially important for predispersal seed predators), and therefore we measured the width of the constriction. We then took a plant average for the measured flowers and reduced the six measurements into a single size variable by calculating the geometric mean (as in Williams & Conner, 2001; Parachnowitsch & Caruso, 2008). The geometric mean was strongly correlated with the first principal component of the six flower measurements (= 0.85, < 0.0001) and the patterns of selection were similar, regardless of whether we used the geometric mean or principal component (data not shown). However, we present the geometric mean for its ease of interpretation (i.e. selection on the mean is selection on overall size).

Figure 1.

 Floral morphology and colour variation in Penstemon digitalis. The two flowers represent extremes of the floral colour phenotypes.

P. digitalis flowers vary in the presence, quantity and intensity of purple striping on the corolla (Fig. 1). The purple colour appears black under ultraviolet (UV) light and may act as a nectar guide for pollinators (Silberglied, 1979) and therefore could be under selection (e.g. Irwin & Strauss, 2005). We counted the number of purple lines on flowers measured for size and scored the intensity of the colour on a four-point scale (0 = no lines, 1 = light purple, 2 = medium purple, 3 = dark purple). To give a single numeric value to colour, we multiplied the number of lines by the intensity. Again, these values were averaged per plant to give a plant estimate of colour.

We also measured display size by counting the total number of flowers per plant based on end-of-the-season estimates (easily assessed from the senesced flowering stalk). Display size is commonly assumed to be attractive to pollinators, and the total number of flowers was positively correlated with daily display size for a close population of P. digitalis in 2006 (= 0.637, < 0.0001, = 61). The floral architecture of P. digitalis varies from dense inflorescences to a more open plan (A. L. Parachnowitsch, pers. obs.) and we estimated this trait as the number of flowers/length of the inflorescence. Predispersal seed herbivores may select for higher abortion rates (Thompson & Cunningham, 2002) and because P. digitalis has both high abortion rates and predispersal seed herbivores, we measured the total number of aborted flowers. Plant height can be a target of selection (Cariveau et al., 2004; Parachnowitsch & Caruso, 2008) and we measured the final plant height at the end of the growing season. Some pollinator-attractive traits, such as flower size and display size, could be correlated with general plant vigour, and therefore selection on these traits may be a result of correlations with vigour rather than selection by pollinators (Andersson, 1996). We estimated selection on vigour by measuring aboveground biomass (dry mass in g).

After senescence, when all fruits had matured, we collected plants for estimates of fitness and predispersal seed damage. To estimate the number of damaged fruits, we assessed five randomly chosen fruits per plant and then multiplied the proportion of damaged fruits by the total number of fruits to give the number of damaged fruits per plant. Predispersal seed damage caused by predators was scored as either present or absent based on evidence found within the fruit (larvae, pupae and/or frass). Female fitness was estimated by first measuring the fruit diameter (mm) of all the fruits. Fruit diameter was an accurate predictor of seed number when plant identity was treated as a random factor for both undamaged (R2 = 0.75, F1,526 = 710, <0.0001) and damaged (R= 0.72, F1,116 = 26, < 0.0001) fruits in a nearby population counted in 2006. Predispersal seed herbivores generally consumed half the seeds (undamaged fruits: 111 ± 2 seeds, = 527; damaged fruits: 56 ± 5 seeds, = 117). Therefore, we calculated female fitness as the total diameter of undamaged fruits plus one-half of the diameter of damaged fruits. Although P. digitalis is perennial, local populations in Tompkins Co., NY, are generally semelparous (A. L. Parachnowitsch, pers. obs.), suggesting that our measures were lifetime fitness for all but the most robust plants. We divided fitness by the treatment mean to give relative fitness values for each plant.

Statistical analyses

Plant loss and missing data led to a total sample size of 281 plants (= 141 open-pollinated; = 140 hand-pollinated). We used Pearson’s correlations to test for relationships among our traits. All analyses were conducted using SAS version 9.2 (SAS Institute Inc., Cary, NC, USA).

Pollinators could affect absolute fitness in our plants in two ways that we measured. First, they could alter the number of fruits that successfully matured (fruit set = successful fruits/total number of flowers). Second, they could alter the number of seeds per fruit (fruit size). We used these two measures to determine whether there was pollinator limitation in P. digitalis. We compared fruit set and mean fruit size between open- and hand-pollinated plants using an ANOVA model, with spatial block and plant pair (within block) as random effects. Both fruit set and mean fruit size met the assumptions of ANOVA and therefore were not transformed. Additional ANOVA models of the same form were tested for block differences in the seven phenotypic traits.

We measured directional (β) and nonlinear or quadratic (γ) selection gradients using multivariate regression models of standardized traits (mean of 0, variance of 1) on relative fitness, as in Lande & Arnold (1983). The multivariate models control for correlations among the traits included in the analysis and therefore measure direct selection, rather than total selection, on each trait. Thus, selection gradients allow identification of the targets of selection among the measured traits (Conner & Hartl, 2004). Standardized traits, relative fitness and selection coefficients were calculated within each pollination treatment. Directional selection models included the seven phenotypic traits. Nonlinear selection measures potentially stabilizing or disruptive selection and was estimated using regression models that included the linear and quadratic terms for the seven traits. The nonlinear selection gradients reported are a doubling of the regression coefficients (Stinchcombe et al., 2008). The patterns of selection were qualitatively the same regardless of whether or not we included block as a random factor in the model, so, for simplicity, we present selection estimates from the models without block.

To determine whether pollinators were selecting on our seven phenotypic traits, we used analysis of covariance (ANCOVA) to compare the selection between treatments (Sokal & Rohlf, 1995). If pollinators were the agents of selection then we would expect selection to be weaker in the hand-pollinated treatment (Fig. 2a). The multivariate model included the linear and quadratic terms for all of the traits, as well as a categorical term for the treatment and all the interaction terms with pollination treatment. Again, block did not affect the pattern of selection, so the models without block are presented.

Figure 2.

 Comparisons of directional selection (s or β) on floral traits in experiments that manipulated an agent of selection: pollinators (b), herbivores (c) or co-flowering species (d). The general pattern of the plots follows that shown in (a): +, selection is stronger when the agent is present; −, selection is stronger when the agent is absent; reversals, = change in the direction of selection. Trait classifications: flower display inline image, flower morphology (size) inline image, nectar inline image, flower type inline image, phenology inline image, petal colour inline image.

Literature survey

To assess the general effect of pollinator-driven phenotypic selection across many taxa and studies, we reviewed selection studies that specifically tested for pollinators as selective agents by manipulating pollination through open-pollinated and hand-pollinated plants using the method described by Lande & Arnold (1983). This method allows comparisons to be made among studies and has been used in a number of broad surveys of natural selection (Kingsolver et al., 2001; Geber & Griffen, 2003; Harder & Johnson, 2009; Siepielski et al., 2009). We first searched papers included in a recent review of phenotypic selection on flowers (Harder & Johnson, 2009), as well as an expanded data set of Geber & Griffen (2003) that was collected to examine variation in selection (M. A. Geber & A. L. Parachnowitsch, unpublished). An additional literature search with Web of Science used the key words ‘selection’ and ‘pollen limitation’ or ‘hand pollination’.

To compare selection by pollinators with selection by other agents on floral traits, we searched for papers which manipulated an agent of selection in an analogous way to hand pollinations (agent present/agent absent) and measured some aspect of selection on floral traits. Here we first searched through the expanded database for appropriate studies (Geber & Griffen, 2003). Additional papers were found by searching for papers citing Lande & Arnold (1983) because we assumed that papers which measured selection using their methodology would cite the paper. We then examined papers with titles that suggested they would meet our criteria. We do not assert that this is an exhaustive search of all published examples; however, we attempted to be as thorough as possible. Although there are examples that manipulate abiotic factors (such as nutrients or shade) and measure selection on floral traits, we excluded these from our comparisons for two reasons. Manipulations of abiotic factors examine a gradient of a selective agent (not the complete absence vs presence), and predictions about the strength of selection in each treatment are difficult to generalize.

We then categorized studies by the three agents with reported selection estimates on floral traits: pollinators, herbivores or co-flowering species. Whenever possible, we used selection gradients (β), which controlled for correlations among traits and estimated direct selection. We recorded selection estimates in the presence and absence of the agent of selection and categorized traits as: display (generally the number of flowers), flower morphology (estimates of flower size and distances within a flower), phenology, flower type (male or female or protandry), colour, and nectar (nectar production). For three studies, selection was estimated in multiple populations, years and/or treatments, and we averaged across the replicates to avoid biasing our data set to any particular study. To increase our power to detect trends in the pollinator manipulations, we included our results from P. digitalis. See the Supporting Information presented in Table S1 for further details.

We estimated the general trend of the relationship of selection between treatments, with and without agents, using regression. To test whether the regression slope was significantly different from 1 (the equivalency line where selection was equal in the presence and absence of the agent), we used the ‘test’ option in PROC REG (Littell et al., 2002). If the slope was significantly greater than 1, selection was stronger in the presence than in the absence of the agent. The opposite was true if the slope was significantly less than 1.


General biology

The seven phenotypic traits were variable within the population. Our mean (range) flower size estimate was 7.78 (6.48–9.44) mm. On our numeric scale, the mean (range) flower colour was 21 (0–48). Interestingly, flower colour was the only trait that did not correlate with any of the others (> 0.48). The mean flower production was 26 flowers per plant; however, this trait was quite variable (range: 7–73). Plants generally aborted nine flowers, although some aborted none while others aborted as many as 23. Those plants with higher flower numbers generally had higher abortion values (= 0.661, < 0.0001). The mean (range) plant height was 76 (38–119) cm, with 13 cm of that being the inflorescence (range: 3.5–29). Mean (range) biomass was 1.89 (0.40–5.29) g, and generally all the size measurements (flower size, flower number, two lengths and biomass) were positively correlated with each other.

We also found phenotypic variation between the two blocks. Plants produced smaller (7.71 vs 7.86 mm) and fewer (24 vs 28) flowers in the lower part of the field. The lower block was also generally significantly smaller (for inflorescence length, height and biomass). Conversely, aborted flowers and flower colour did not vary across the field.

Predispersal seed damage by predators was extremely low in this population; only five plants showed evidence of such damage, each with only one of five fruits attacked. Although all of these plants were in the hand-pollinated treatment, and most (four of five) were in the upper block, there were too few damaged plants to establish conclusively whether our hand-pollinations increased attack or whether damage was spatially variable.

Pollen limitation

We found no pollen limitation in our population of P. digitalis. Fruit set did not differ between open-pollinated and hand-pollinated plants (F1,280 = 1.22, = 0.27); however, spatial variation was found in fruit set between our two blocks (F1,280 = 26.11, < 0.0001). Likewise, the mean fruit size per plant did not differ between the pollination treatments (F1,280 = 0.46, = 0.50), but differed at the block level (F1,280 = 15.62, < 0.0001). Fruit set (0.70 vs 0.60) and mean fruit size (4.77 vs 4.57 mm) was higher in the lower block.

Natural section

Natural selection differed between the open-pollinated and hand-pollinated plants (Table 1). In open-pollinated plants, we found significant directional selection on four of seven phenotypic traits. Plants with larger, more and fewer aborted flowers, as well as larger size, had higher relative fitness. We found no detectable directional selection on flower colour, inflorescence length or plant height. There was also significant stabilizing selection on flower number, and disruptive selection on floral density in the open-pollinated plants, but no significant quadratic selection was detected on the other five traits. However, the sample size in our treatments may have limited our power to measure nonlinear selection. Conversely, we found significant phenotypic selection on only three of seven traits in the hand-pollinated plants. There was directional selection for more flowers, fewer aborted flowers and larger plants, but no significant quadratic selection in the hand-pollinated population. Furthermore, natural selection was significantly stronger in the open-pollinated plants for flower size and flower number, as well as stabilizing selection on flower number, suggesting that pollinators were agents of selection on these traits. Selection on flower size was marginally different between our two treatments (< 0.06), but when we used inflorescence length rather than our composite flower-density trait, selection on flower size was the same within treatments and was significantly different between treatments (< 0.05). We found that selection for fewer aborted flowers and greater biomass did not differ between pollination treatments, suggesting that pollinators were not the agents of selection for these two traits.

Table 1.   Comparisons of natural selection gradients (directional, β; and quadratic, γ) between open-pollinated (= 141) and hand-pollinated (n = 140) Penstemon digitalis
Phenotypic traitβFPγFP
  1. Flower size is the geometric mean of six flower measurements; flower colour was calculated by multiplying the number of purple lines with the intensity (see the text for details). Floral density is equal to the number of flowers/inflorescence length.

  2. Phenotypic selection (± 1 SE) is followed by analysis of covariance (ANCOVA) statistical testing of whether the selection estimates differed between treatments. Bold indicates significant selection within a pollination treatment.

Flower size0.030 ± 0.0100.0007 ± 0.0083.580.060.008 ± 0.005−0.012 ± 0.0071.170.28
Flower colour−0.003 ± 0.009−0.005 ± 0.0070.080.780.004 ± 0.007−0.006 ± 0.0070.000.98
Number of flowers0.745 ± 0.0180.601 ± 0.01443.85< 0.0001−0.068 ± 0.009−0.006 ± 0.0069.050.03
Floral density−0.004 ± 0.0100.009 ± 0.0090.720.400.026 ± 0.0090.012 ± 0.0060.680.41
Aborted flowers−0.319 ± 0.012−0.306 ± 0.0100.930.34−0.008 ± 0.007−0.006 ± 0.0060.040.84
Plant height−0.014 ± 0.012−0.0002 ± 0.0111.540.22−0.0006 ± 0.0060.014 ± 0.0070.580.45
Biomass0.077 ± 0.0170.060 ± 0.0120.000.990.024 ± 0.009−0.012 ± 0.0072.520.11

Survey of experimental evidence for biotic agent-mediated natural selection

In addition to the current study, we found 15 studies that manipulated an agent of selection and measured selection on floral traits for 12 species spanning 11 plant families (Table 2). Six additional studies explicitly tested, by manipulating the pollination environment, whether pollinators were agents of selection (Andersson, 1996; Galen, 1996; Totland et al., 1998; Fishman & Willis, 2008; Parachnowitsch & Caruso, 2008; Sandring & Ågren, 2009), five studies manipulated herbivory (Juenger & Bergelson, 1998, 2000; Gómez, 2003; Juenger et al., 2005; Wise & Cummins, 2007) and a further four studies manipulated the presence of co-flowering species (Caruso, 2000, 2001; Moeller & Geber, 2005; Smith & Rausher, 2008), although one study examined facilitation by congeners rather than competition for pollinators (Moeller & Geber, 2005). Measures of flower morphology were by far the most common, followed by display and phenology. Very few studies examined flower type, colour or nectar traits.

Table 2.   Summary of a literature survey of experimental manipulations of agents of natural selection on floral traits
Number in surveyComplete data setManipulation
PollinatorsHerbivoresCo-flowering species
  1. Pollinator manipulation refers to experiments with open-pollinated vs hand-pollinated treatments (and include this study); herbivores were manipulated either by simulating herbivory or by excluding herbivores; and co-flowering species were experimentally present or absent. Selection gradients (which control for correlations among traits) are presented when available.

Plant families11733
Selection estimates5323921
 Floral morphology2912314
 Flower type3012

Natural selection was stronger when pollinators were present rather than absent (Fig. 2b) and the slope of the regression line was significantly different from 1 (F1,22 = 5.78, = 0.026). Conversely, selection in the absence of co-flowering species was stronger than in their presence, and this trend was not affected by the nature of the study (competition vs facilitation) (Fig. 2d; F1,20 =5.85, = 0.026). Finally, natural selection on flower traits was equivalent in the presence or absence of herbivory (Fig. 2c; F1,9 = 0.33, = 0.58). Among the differences observed in selection, most biotic alterations of selection resulted in changes in the strength (but not in the direction) of selection; reversals in the direction of selection were most prevalent when comparing the presence and absence of co-flowering species (seven comparisons; Fig. 2d).


We found pollinator-mediated selection in P. digitalis on two traits thought to be important for pollinator attraction: flower size and display size (Table 1). Selection for larger flowers was present in open-pollinated plants, but not in hand-pollinated plants. Natural selection on flower morphology selected for larger values (Fig. 2; Harder & Johnson, 2009) and fits with the general hypothesis that pollinators select for larger flowers because they are more conspicuous and/or may be associated with larger rewards (e.g. Blarer et al., 2002). Natural selection for larger displays is also common (Fig. 2; Harder & Johnson, 2009) and we found stronger directional selection on flower number in open-pollinated P. digitalis (Table 1). Because flower number sets the upper limit for fruit number, in many systems such as P. digitalis, there is a direct, positive relationship between flower and fruit number. Thus, positive selection would be expected on flower number independently of pollinator-mediated selection. However, we found stronger directional selection in the open-pollinated plants, suggesting that the difference in strength of selection was the result of an added benefit of larger displays attracting pollinators. We also found stabilizing selection on flower number in the open-pollinated plants, but not in the hand-pollinated plants, suggesting a cost of having too many flowers. Large displays can have increased geitonogamy, which could be costly through reduced fitness of selfed seeds (Harder & Barrett, 1995). Because hand-pollinations probably reduced the number of selfed seeds, we would expect the cost of large displays to be likewise reduced, suggesting that the difference in stabilizing selection between our treatments was a result of stabilizing selection by pollinators.

We found natural selection by pollinators despite the lack of pollen limitation. Although stronger selection via female fitness can be correlated with a greater degree of pollen limitation (Ashman & Morgan, 2004), selection by pollinators is not always associated with pollen limitation. Galen (1996) found selection by pollinators without pollen limitation, and two studies found pollen limitation, but no selection by pollinators (Andersson, 1996; Totland et al., 1998). Therefore, there may be an increased likelihood of finding selection by pollinators in pollen-limited populations because selection is probably stronger. However, this cannot be assumed to be true for all populations.

Field estimates of phenotypic selection can be biased as a result of environmental covariance between traits and fitness (Rausher, 1992). We attempted to control for this bias by physically pairing our treatments and using a blocked design. There were phenotypic differences between our two blocks which suggest that pollination, competition and/or resources may have differed in these parts of the population. However, when we included block as a random factor in our selection models, the significant patterns remained, suggesting that an environmental gradient was not responsible for the overall selection we found.

If biotic agents frequently exert natural selection on plants, then we would expect stronger selection when the agents are interacting with the plant, compared to when they are experimentally removed. For pollinators specifically, we expect selection on floral traits to be stronger when pollinators are selecting plants than when experimenters hand-pollinate them. Indeed, we found that natural selection on floral traits was stronger in the presence than in the absence of pollinators (Fig. 2b). Furthermore, we do not necessarily expect herbivores to be strong selective agents on flowers, although they could influence floral evolution in a number of ways (Strauss & Irwin, 2004). We found that selection on floral traits was equivalent regardless of whether herbivores were present or absent (Fig. 2c). However, there were few selection coefficients with which to test this hypothesis, so these findings should be interpreted with caution. Co-flowering species could either have competitive or facilitative effects on the focus species, which could lead to divergent or convergent evolution of floral traits (Caruso, 2001). Therefore, it is difficult to predict across systems whether selection should be stronger with or without a co-flowering species. When co-flowering species were manipulated, selection was stronger in the absence of the co-flowering plant (Fig. 2d). However, there were also many more reversals when co-flowering species were removed (i.e. positive selection became negative, or vice versa), suggesting that community context could alter selection on floral traits (Caruso, 2000). Surprisingly, we found significant trends for manipulations of both pollinator and co-flowering species, despite our small sample size and the fact that selection is frequently weak in natural populations (Kingsolver et al., 2001; Knapczyk & Conner, 2007). Indeed, the majority of the selection coefficients included were close to zero and not significant. However, our ability to detect trends suggests that these patterns could be strong in nature.

Our survey data could be biased by three major limitations. First, although our complete data set included plants spanning multiple functional pollinator groups, pollinator manipulations have been limited to hymenopteran and dipteran pollinators. Pollinator shifts have been proposed as a major driver of speciation (e.g. Fenster et al., 2004) and some shifts seem more common than others. For example, bee-to-bird pollination is more common than the reverse (Thomson & Wilson, 2008). However, it is uncertain whether one functional group would exert stronger selection on floral traits than another. Thus, the addition of plants pollinated by other functional groups, such as birds, beetles, moths, etc., may alter the pattern we detected. Second, although our data set spans almost as many families and species as studies, it only includes herbaceous plants. Studying selection in herbaceous plants is a bias common to the plant literature; however, it is important to note that it may affect our ability to generalize the results to all flowering plants (e.g. Geber & Griffen, 2003). Because natural selection by pollinators is of general interest in floral evolution, measuring selection by pollinators in nonherbaceous plants and/or nonbee/fly-pollinated plants will provide further insights into their role as agents of natural selection. Lastly, our survey was necessarily limited to studies that measured the selection coefficients described by Lande & Arnold (1983) and directly manipulated an agent of selection. This allowed direct comparisons to be made across studies but could also introduce bias.

We were unable to directly compare the impacts of pollinators with selection by other floral visitors, such as florivores (McCall & Irwin, 2006) or predispersal seed herbivores (these data have simply not been collected, but see Wise & Cummins, 2007). In particular, predispersal seed predators are expected to conflict with pollinator selection on floral traits because they should visit fertilized flowers (or those that will be fertilized). Moreover, they can exert selection on floral traits in multiple systems (Pilson, 2000; Cariveau et al., 2004; Rey et al., 2006; Parachnowitsch & Caruso, 2008), making a direct comparison between seed predator- and pollinator-mediated selection a relevant question. For our population of P. digitalis, we found no evidence that predispersal seed predators influenced selection on flower size or number, or affected selection on aborted fruits. Only five of the 281 plants in this experiment showed evidence of any fruit damage, and our results were robust to excluding these plants (data not shown). Predispersal seed herbivores can be much more frequent in P. digitalis (Mitchell & Ankeny, 2001; A. L. Parachnowitsch, unpublished) and it is possible that they do exert natural selection in other populations and/or years.

An important limitation of our studies and those reviewed here is that they generally only measure female fitness in hermaphroditic plants in one season. Natural selection can vary over time (Siepielski et al., 2009). Therefore, to fully understand the evolutionary trajectory for a population, selection should be measured in multiple years. Moreover, if most of the natural selection on flowers by pollinators or other agents occurs via male function, then these types of experiments may fail to detect their impact (see Conner (1997) for a theoretical discussion and counter example). Although few natural-selection studies examine both male and female fitness estimates, the general pattern from the literature does not support the male-function hypothesis. That is, selection on attractive floral traits was not always stronger via male fitness, but rather selection through male and female fitness was context-dependent (Ashman & Morgan, 2004). Therefore, hand-pollinations can provide valuable information about selection by pollinators but are still only half of the story. Manipulations of the agents of selection via both male and female fitness are necessary to give a more complete picture and reveal functional differences.


Flowers offer a conundrum to evolutionary ecologists. On the one hand it seems fairly obvious that flowers are adapted to their pollinators (reviewed in Harder & Johnson, 2009) and that shifts in pollination systems have led to diversification in plants (e.g. Kay et al., 2005). Moreover, there are a limited number of examples of pollinator-mediated natural selection, such as those found in P. digitalis. However, the majority of selection estimates on floral traits from manipulations of pollination do not support selection by pollinators (Fig. 2b), and studies that have used path analyses often find that pollinators weakly effect fitness (Conner et al., 1996; Gómez, 2000; Rey et al., 2006; Ashman & Penet, 2007). Harder & Johnson (2009) argue that natural-selection studies may be measuring the effects of pollinators during periods of relative stasis and that we should expect selection by pollinators on novel traits or in new environments. In other words, the lion’s share of pollinator-mediated natural selection shaping flowers may have occurred during and shortly after speciation. While this view needs further testing, we are left to explain what the agents of selection are in cases when there is significant selection on floral traits, but not by pollinators (e.g. Andersson, 1996; Totland et al., 1998; Fishman & Willis, 2008; Parachnowitsch & Caruso, 2008; Sandring & Ågren, 2009). If pollinators truly exert most of their selection on flowers during cladogenesis, then an interesting question follows. During a period of stasis, how far from these adaptive peaks can nonpollinator agents push floral traits within a particular population or species?


We thank A. Freitag for field assistance, C. Elston, J. Porter and W. Tomascelli for data collection, A. Agrawal, M. Geber and K. Zamudio for discussions and feedback, and three anonymous reviewers for helpful comments on the manuscript. Our study was funded by the Botanical Society of America (Graduate Student Research Award), the Cornell Department of Ecology and Evolutionary Biology (Student Research Fund) and the National Science Foundation (NSF-DEB 0717139).