Volume 106, Issue 4 p. 1646-1659
Free Access

Pre-dispersal seed predation and pollen limitation constrain population growth across the geographic distribution of Astragalus utahensis

Kathryn C. Baer

Corresponding Author

Kathryn C. Baer

Department of Biological Sciences, University of Montana, Missoula, MT, USA

Pacific Northwest Research Station, USDA Forest Service, Anchorage, AK, USA


Kathryn C. Baer

Email: kcbaer2@gmail.com

Search for more papers by this author
John L. Maron

John L. Maron

Department of Biological Sciences, University of Montana, Missoula, MT, USA

Search for more papers by this author
First published: 18 January 2018
Citations: 22


  1. A central focus of ecology is to understand the conditions under which biotic interactions affect species’ abundance and distribution. Classic and recent studies have shown that biotic interactions can strongly impact local or regional patterns of species abundance, but two fundamental questions remain largely unaddressed for non-competitive biotic interactions. First, do the effects of these interactions on population performance change predictably with environmental context? Second, to what extent do population-scale effects contribute to limiting species’ geographic distributions?
  2. To address these questions, we experimentally assessed the extent to which pollen limitation and insect seed predators affected the fecundity and projected population growth rate (λ) of the native forb Astragalus utahensis. We studied populations at the centre and northern edge of the latitudinal range of A. utahensis that occur across a gradient in abiotic harshness characterized primarily by declining mean annual precipitation.
  3. Supplementing pollen and suppressing pre-dispersal seed predators increased seed production similarly within A. utahensis populations at the centre and northern edge of the range. Integral projection population models revealed that relaxing these checks on seed production tended to increase λ in most populations, regardless of their location within the range.
  4. Synthesis. Our results suggest that pollen limitation and insect herbivores limit population growth in A. utahensis similarly across the centre-to-north portion of its latitudinal distribution. However, because A. utahensis population growth barely reaches the level of replacement at the northern range edge, the reduction in λ resulting from these interactions may contribute to limiting expansion at the northern edge of A. utahensis’ latitudinal range.


Ecologists have long sought to understand the extent to which interactions affect the distribution and abundance of species. For plants, a lack of sufficient pollen receipt (Ashman et al., 2004; Knight et al., 2005) or herbivory (Kolb, Ehrlén, & Eriksson, 2007; Marquis, 1992; Strauss & Zangrel, 2002) commonly reduce reproductive output. However, tests of how these decrements in fecundity influence population growth (λ), particularly across varying environmental conditions, are rare (Katz, 2016; Maron, Baer, & Angert, 2014; Maron & Crone, 2006). The circumstances under which reduced seed set due to pollen limitation or herbivory translates to meaningful reductions in population growth remain relatively understudied. As a result, how these factors influence the distribution of plant species is unclear.

Many studies that document negative impacts of pollen limitation or herbivory on plants only examine effects upon a single vital rate (Crawley, 1989; Maron & Crone, 2006), which may not be reflective of population-level consequences (Ehrlén, 2003; Katz, 2016; Kolb, Leimu, & Ehrlén, 2007). For example, many studies have shown that insufficient pollen receipt can reduce seed set (i.e. pollen limitation; Ashman et al., 2004; Knight et al., 2005), but the conditions under which these reductions in seed production affect λ are not well understood. Pollen limitation can lead to decreased λ (Anderson, Kelly, Ladley, Molloy, & Terry, 2011; Bierzychudek, 1982; Law, Salick, & Knight, 2010; Parker, 1997; Price, Campbell, Waser, & Brody, 2008), or have no appreciable impacts on λ (Castro, Dostalek, van der Meer, Oostermeijer, & Munzbergova, 2015; Ehrlén & Eriksson, 1995; Feldman & Morris, 2011; García & Ehrlén, 2002; Knight, 2004). Similarly, a recent meta-analysis of the population-level outcomes of plant–herbivore interactions found that herbivory generally decreases λ, but often in ways that cannot be predicted based upon how herbivores influence individual demographic transitions (Katz, 2016). The lack of a correlation between the negative consequences of pollen limitation or herbivory for individual fecundity and its consequences for population abundance arises in part when reproduction is not a strong contributor to population growth rate. More studies are needed to establish the environmental conditions under which biotic interactions affect population-scale demographic performance (Maron et al., 2014).

This gap in knowledge is particularly acute when it comes to understanding whether non-competitive biotic interactions limit geographic ranges (Sexton, McIntyre, Angert, & Rice, 2009). These interactions may contribute to or maintain range limits in two non-mutually exclusive ways. First, they may suppress λ below the replacement rate beyond the edge of the distribution, effectively limiting the geographic area in which a species can maintain viable populations. Second, they may decrease population performance at the range boundary to a level where low numbers of propagules cause the species to become dispersal limited, while failing to have strong negative effects in populations in the interior of the range or at sites beyond the range edge. This could occur if the outcomes of biotic interactions are dependent upon geographically determined environmental context.

At small spatial scales, the impacts of pollen limitation and herbivory on λ can vary across gradients in the environment (e.g. pollen limitation: Feldman & Morris, 2011; Horvitz, Ehrlén, & Matlaga, 2010; Law et al., 2010; Parker, 1997; Price et al., 2008; herbivory: Kauffman & Maron, 2006; Kolb, Leimu, et al., 2007; Rose, Russell, & Louda, 2011; von Euler, Ågren, & Ehrlén, 2014). The per capita impacts of herbivory can also vary across broader elevational (Bruelheide & Scheidel, 1999; Miller, Louda, Rose, & Eckberg, 2009) or latitudinal gradients (Alexander, Price, Houser, Finch, & Tourtellot, 2007; Anstett, Naujokaitis-Lewis, & Johnson, 2014; Moreira, Abdala-Roberts, Parra-Tabla, & Mooney, 2015; Nunes, Cassin, & Kotanen, 2016; Pennings & Silliman, 2005). Yet, few studies have linked the individual-scale impact of pollinators or herbivores across sizable portions of the latitudinal range of a species to how these interactions influence population growth from the range centre to range edge. A notable exception are exemplary studies implicating pollen limitation as a contributor to declining population growth rate of the annual forb Clarkia xantiana across a gradient of increasing abiotic harshness characterized by declining precipitation from the centre to the eastern edge of its longitudinal distribution (Eckhart et al., 2011; Moeller, Geber, Eckhart, & Tiffin, 2012). In contrast to these studies, Louthan, Doak, and Angert (2015) suggested that biotic interactions should exert their strongest control on demographic performance and influence range limits where environmental conditions are relatively benign (the species interactions-abiotic stress hypothesis, SIASH). SIASH implies that biotic interactions are only likely to contribute to range limits if the abiotic environment is less harsh at the edge of the distribution than in its interior.

In this study, we determined the degree to which pollen supplementation and pre-dispersal seed predation affect plant fecundity in populations of the perennial forb Astragalus utahensis. We studied these interactions across a gradient of increasing abiotic harshness characterized primarily by declining precipitation from the centre to the northern edge of A. utahensis’ latitudinal range. We then used integral projection population models to explore how reductions in seed set resulting from these interactions affect A. utahensis’ population growth rate and how this varies from the range centre to its extreme northern edge. Our goal was to understand whether these interactions play a role in shaping the geographic distribution of A. utahensis.


Study system

Astragalus utahensis (Fabaceae) is a long-lived perennial forb. It produces racemes of 3–10 magenta flowers from late April to mid-June that develop into densely hairy seed pods (K. C. Baer, pers. obs.). Seeds disperse an average of ≤1 m from the parent plant via gravity dispersal. The plant is non-clonal; all reproduction is seed-based and seed production requires outcrossing via insect pollinators (Green, 1976). Its primary pollinators, regardless of range location, include large bee species in the genera Bombus, Anthophora and Eucera (K. C. Baer, pers. obs.). Astragalus utahensis is heavily attacked by a seed beetle (Acanthosceldies fraterculus) and a seed weevil (Tychius proxilus) that specialize on Astragalus (Green, 1976; G. Morse, pers. comm.).

The north–south geographical range of A. utahensis extends from southeastern Idaho, USA (roughly 43°N) to southern Utah, USA (roughly 38°N); the centre of the distribution is in northern Utah, USA (roughly 41°N; Figure 1). Populations included in this study were located at the centre of the distribution and the northern margin (Figure 1). Four central populations and four northern populations were included in this study. Northern study populations included the three northernmost populations identified from herbarium collection records and an additional population identified during a survey conducted along roadways at the northern range edge. All of the central populations and two of the northern populations (Wolverine and Reservoir; Figure 1) were included in the study in 2013 (Appendix S1). Two additional northern study populations (Nordic Center and Swan; Figure 1) were added in 2014 (Appendix S1). One of the four central populations (Ogden; Figure 1) was lost to a wildfire in fall 2014 and therefore excluded from analyses of the effects of experimental treatments on fecundity in 2015 and from all analyses of their effects on population growth rates (Appendix S1).

Details are in the caption following the image
Map showing locations of central and northern study populations. The distribution of Astragalus utahensis is shown in grey; the approximate boundaries of the distribution are represented by a grey dashed line

Study populations occurred across a gradient of decreasing mean annual temperature and precipitation from the centre of A. utahensis’ range to its northern boundary. Populations at the centre of the distribution had categorically higher predicted climate suitability scores (as calculated by a MaxEnt ecological niche model for this species; K. C. Baer, A. L. Angert & J. L. Maron, unpubl. data) than those at the northern range edge. Specifically, predicted climatic suitability was primarily a product of precipitation in the wettest quarter of the year (higher in central than northern populations), and to a lesser extent mean temperature in the warmest and driest quarters of the year, and isothermality (higher in central than northern populations), along with temperature in the wettest quarter of the year (higher in northern than central populations).

Pollen supplementation, pre-dispersal seed predator suppression and demographic monitoring

We applied experimental treatments to flowering plants in each population every 7–12 days throughout the flowering season (see Appendix S1 for specific treatment application years). During the first year in which each population was included in the study, we haphazardly marked 90–105 reproductive plants (defined by the presence of flowers or developing racemes) per population and randomly assigned them to one of three treatments: a pollen supplementation (hereafter, pollination) treatment, a pre-dispersal seed predator suppression (hereafter, insecticide) treatment, or a control treatment which was compared to both treatment groups. The northern Swan population contained few plants, so only 20–25 reproductive plants were marked and assigned to each treatment in that population. Any plants that died from one growing season to the next were replaced with another randomly chosen plant assigned to the same treatment at the beginning of the next growing season. The same treatment was administered to an individual reproductive plant throughout all years in which it was included in the study. We utilized a factorial design for the first 2 years of the study, but removed the “pollination + insecticide” treatment after determining that there were no interactive effects on any aspect of reproduction associated with the simultaneous application of both treatments (K. C. Baer & J. L. Maron, unpubl. data).

We provided supplemental pollen to focal plants by hand pollinating up to seven receptive flowers (determined by stigma protrusion beyond the cluster of anthers) per plant per visit, using a homogenized mixture of pollen from ≥10 unmarked plants within the same population collected on the same day as our visit. Pollen was applied using a camel hair paint brush. Although an ideal protocol is to pollinate all flowers on a plant to account for potential resource re-allocation (Knight, Steets, & Ashman, 2006), the number of flowers on a single A. utahensis plant (over 200 on some individuals on a given day) precluded doing so. We marked the calyxes of pollinated flowers using paint pens (Uchida Brand) so that fruits could later be identified and collected. Beginning in 2014, we also marked the calyxes of up to seven flowers blooming on the same day on control plants.

Plants assigned to the insecticide treatment were sprayed with an 8 ml/L solution of 0.0425% esfenvalerate (Asana XL, DuPont USA), a non-systemic surface pesticide that has been demonstrated to be effective against pre-dispersal weevil seed predators (Agrawal, Hastings, Johnson, Maron, & Salminen, 2012). Once dry, this insecticide should not deter or negatively affect pollinators. We found no significant differences in either growth or survival related to the application of the insecticide except a significant positive effect of insecticide application on survival in the northern Reservoir population; our findings echoed those of other studies which utilized this insecticide (Appendix S2; Agrawal et al., 2012; Root, 1996). Furthermore, in a greenhouse study there were no effects of esfenvalerate on A. utahensis growth (Appendix S2). Plants assigned to the control treatment experienced natural pollination and no herbivore exclusion, but were sprayed with an amount of water equivalent to that used in the insecticide treatment.

At the end of each growing season, we counted the total number of fruits produced by each marked reproductive plant. We collected all fruits produced by hand-pollinated flowers on pollen-supplemented plants and up to ten randomly selected fruits from each plant that received the insecticide or control treatments. If a plant produced fewer than ten fruits, all fruits were collected. After counting the number of seeds in each collected fruit, we estimated per capita seed production for each individual as the product of mean seeds per fruit and total fruits of that individual. We also recorded whether any of the collected fruits showed signs of pre-dispersal seed predation, including the presence of larvae, frass, or seeds which had been partially eaten or bored into. As pollen supplementation was performed on a subset of flowers per plant, we used a formula to extrapolate the effects of pollen supplementation on total per capita fruit production (Appendix S3).

In addition to collecting data on the fecundity of reproductive plants in each treatment group, we monitored the vital rates of plants in each population except the central Ogden population from 2014 to 2016 in order to forecast how pollen supplementation and insect seed predation affected projected population growth (Appendix S1). In each population, we marked 60–80 seedlings (newly germinated plants with cotyledons still attached) with individually numbered metal tags at the beginning of each growing season and followed their fate over a yearly time step. Seedling size was not measured, as it was similar across all study populations. Seedlings that survived to the following season were re-classified as juveniles (plants with no signs of cotyledons or developing racemes through the entire growing season) or reproductive plants and the area of their basal rosette was determined according to the formula for the area of an ellipse. We maintained 40–60 marked juvenile plants in each population during each year of the study, monitoring their rates of survival and changes in the area of their basal rosette across yearly transitions. Any juveniles that died or became reproductive were replaced with new marked juveniles. Starting in 2014, we also determined the basal rosette area of each reproductive plant assigned to an experimental treatment at the beginning of each growing season and continued to monitor the rates of survival, growth and fecundity among these individuals.

Seed addition experiments to determine emergence rates

We employed multi-year seed addition experiments to estimate emergence rates and determine whether populations of A. utahensis are seed-limited (sensu Andersen, 1989). If recruitment is limited by seed availability (seed limitation), increased seed rain should result in correlated increases in seedling emergence. Conversely, if the availability of safe sites for recruitment is more limiting, seedling emergence should asymptote at some level of natural seed production once all safe sites are occupied. At the end of the 2013 and 2014 growing seasons, we added seeds sourced from within each population to blocks containing four 0.5 × 0.5 m plots spaced 1 m apart and sown with seeds across a range of densities. The seed densities used were designed to span a biologically meaningful range of seed rain in an area equivalent to our plot sizes based upon estimates of total per capita seed production across populations. The number of seeds sown in the highest density subplots was increased in the second year of the experiment due to low rates of germination. The blocks were located ≥1 m from any existing plants in the population to minimize the chance of germination from an existing seed bank within seed addition plots. Blocks were haphazardly placed throughout each population; the distance between blocks varied according to population size but ranged from 5 to >100 m. In 2013, we randomly sowed plots within each block with 25, 50, 75 or 100 seeds, placing 10 replicate blocks in each of the central populations and the northern Nordic Center population and five replicate blocks in the northern Reservoir and Wolverine populations (Figure 1) due to scarcity of available seeds in these populations (Appendix S1). In 2014, we established 10 new replicate blocks of four plots randomly sown with 25, 50, 100 or 200 seeds in all study populations except the central Ogden population, which had burned that year. We monitored the rates of emergence in the spring of 2014–2016.

Population projection modelling

To evaluate the projected impact of pollen limitation and pre-dispersal seed predation on λ, we constructed integral projection models for each study population and experimental treatment over the 2014–2015 and 2015–2016 transitions (IPM; Easterling, Ellner, & Dixon, 2000; Ellner & Rees, 2006; IPMpack; Metcalf, McMahon, Salguero-Gómez, & Jongejans, 2013). Integral projection models are ideal for describing the population dynamics of species such as A. utahensis, whose life history cannot be clearly dived into discrete stage classes. Rather, IPMs allow for vital rates to be modelled using a continuous variable such as size. If the number of individuals of size x in a population at time t is described by nt (x), the number of individuals of size xʹ at time t + 1 can be described by an IPM according to the generic formula:
where minSize represents the minimum possible size of the individual, maxSize represents the maximum possible size and the kernel K is analogous to a more traditional stage-structured matrix, but contains ≥100*100 cells integrated according to the midpoint rule (Ellner & Rees, 2006; Merow et al., 2014). In a standard IPM, the kernel is comprised of functions describing the growth, survival, and fecundity of individuals according to their size:
Here, P represents a growth and survival matrix for individuals moving from size x to size x′ over one time step, and F represents a matrix describing the fecundity of these individuals over this step: the production of offspring of size x′ by parents of size x (Ellner & Rees, 2006; Merow et al., 2014). The P matrix in our models was comprised of logistic regressions with a logit link for binomial survival probabilities (s) and linear models for growth (g) that incorporated variance around the predicted relationship in the form of residual standard error. These models were based upon log-transformed basal rosette area and described by P (xʹ,x) = s(x) g(xʹ|x). The F matrix was comprised of models for the probability of flowering, per capita seed production, and offspring growth and survival. We used logistic regression models with a logit link for the binomial data predicting the probability of flowering and Quasi–Poisson regressions with a log link and an over-dispersion parameter to predict per capita seed production based upon log-transformed basal rosette area. Constants described the rates at which seeds germinated the following year or entered the seed bank. The F matrix was described by F (xʹ,x) = pflower(x) fseeds(x)pgerm. For pollen-supplemented plants, we estimated per capita seed production for whole-plant pollination as described above and in Appendix S3. To prevent biologically unrealistic estimates of per capita seed production, we capped the maximum value in each model at the maximum observed seed production (estimated seed production in the case of pollen supplementation) for any individual in the population, transition year and treatment pertaining to that model. We structured models according to the log-transformed basal rosette area of plants because size is generally a strong correlate of survival, growth and fecundity (Easterling et al., 2000; Merow et al., 2014). The minimum size included in the model was derived from the smallest plant ≥1 year old observed in any year of the study, and the maximum size was defined as 1.2 times the size of the largest plant observed in any year of the study in order to avoid unintentional eviction of the largest size classes from the model (Williams, Miller, & Ellner, 2012). We also included two discrete stages in the models to describe the dynamics of the seed bank and seedlings, whose size was not recorded. As we did not measure the rate at which seeds remained viable in the seed bank over a yearly time step (psbanksurv), we estimated this value based upon the mean value for this vital rate reported in a study on an Astragalus species with a similar life history to A. utahensis (Martin, 2010; Appendix S4). Seedling emergence rates from seeds produced the previous season (pgerm) were based upon mean germination after one winter in experimental seed addition plots. Emergence from the seed bank (psbank germ) was described by constants calculated from germination in seed addition plots after three winters (2 years in the seedbank), which was assumed to be representative of long-term emergence rates. In the Swan population, where germination after three winters from the seedbank was unknown, we estimated germination from the seedbank as mean germination in seed addition plots after two winters. The proportion of surviving seedlings (pseedling surv) was based upon yearly survival of marked seedlings; their size distribution was described by the mean (M) and standard deviation (SD) of the basal rosette area of surviving seedlings (frecruit size; Appendix S4). The structure of our integral projection models is described by the following equations:
where B (Equation 3.1) represents transitions into the discrete seed bank stage, S (Equation 3.2) represents the discrete seedling stage, and the K kernel (Equation 3.3) represents the growth and survival of all plants ≥1 year old. A life cycle diagram incorporating these transitions is presented in Appendix S4.

We parameterized our IPMs for each yearly transition and treatment in each population using demographic data collected during that transition, with the exception of constants describing recruitment from seeds produced the previous season and from the seedbank. These constants remained identical for each model within a given population. The same juvenile plants and seedlings were included in all datasets for a particular population and transition year regardless of the treatment applied to the reproductive adults in the dataset, as experimental treatments were applied solely to reproductive adults. Datasets for each combination of population, transition year and experimental treatment included demographic data pertaining to non-reproductive individuals in a population over a given transition as well as reproductive plants of a particular experimental treatment over the same transition.

We constructed separate IPMs for each transition and treatment in each of the study populations. In constructing these IPMs, we compared linear and quadratic models for each vital rate (survival, growth, the probability of flowering and per capita seed production). We ranked candidate models by their corrected Akaike index criterion (AICc) scores (Quasi-AICc, or QAICc, for models of per capita seed production to account for over-dispersion in seed counts). If ΔAICc or ΔQAICc > 2, we chose the model with the lower score unless it was a poor fit to small individuals or biologically unrealistic (Appendix S4). In cases where ΔAICc or ΔQAICc ≤ 2, we used model averaging (model.avg; MuMIn package; Barton, 2016) to construct models with parameters weighted by the relative AICc or QAICc scores of candidate models. Model terms, their coefficients from the model fit to the original dataset, and their AICc or QAICc scores are presented in Appendix S4. We used these models to build the P and F matrices, added these matrices together, discretized the continuous portions of each IPM into a 100*100 cell matrix, and added the discrete stages to this matrix in order to calculate projected population growth rates (λ; the dominant eigenvalue).

We estimated stochastic population growth rates based upon models for the 2014–2015 and 2015–2016 transitions for each treatment in each population by randomly applying one of these IPMs to a population vector 4,000 times (discarding the first 2,000 runs in order to avoid initial conditions influencing the outcome), and determining the value upon which population growth rate converged (stochGrowthRateSampleList, IPMpack, Metcalf et al., 2013, 2015). We used bootstrapping to estimate the variation around 500 estimates of λ for each treatment over each transition in each study population. To do this, we created 500 bootstrapped datasets of the same size as the original dataset via random selection of individuals with replacement from the original dataset (in accordance with methods outlined in Kalisz & McPeek, 1992). We used these bootstrapped datasets to construct 500 IPMs as described above, using the same model form as the original dataset and parameters calculated using the bootstrapped dataset. We calculated λ for each of the bootstrapped IPMs. To estimate the mean effect of the pollination treatment on population growth, we chose random values of λ from “pollination models” and subtracted randomly chosen values of λ from “control models” for the same transition, repeating this process without replacement 500 times to obtain 500 estimates of the change in lambda with supplemental pollination (ΔλP) for a particular population and transition. We did the same for comparisons between lambdas from “insecticide” and “control models,” obtaining 500 estimates of the change in lambda with insecticide treatment (ΔλI) in a population over a specific transition.

To estimate the mean effect of the experimental treatments on lambda over the course of the 2014–2015 and 2015–2016 transitions, we calculated the mean value of ΔλP and ΔλI for the combined set of 1,000 total estimates of each over the course of the two transitions (500 estimates for the 2014–2015 transition and 500 estimates from the 2015–2016 transition). We also estimated a 95% confidence interval around this mean, defined as the values lying between the 2.5th and 97.5th percentiles of the distribution. If the 95% confidence interval of ΔλP or ΔλI overlapped zero, we considered the difference among the control and experimental treatments to be non-significant (sensu Goldstein & Healy, 1995; Wolfe & Hanley, 2002). If the 95% confidence interval did not overlap zero, we considered the difference to be significant. We also calculated a 90% confidence interval defined by values lying between the 5th and 95th percentiles of the distribution, which we used to evaluate marginal significance. This methodology is similar to that used by Tenhumberg, Suwa, Tyre, Russell, and Louda (2015), and is preferable to the use of a t test as the standard error of the mean was not influenced by the number of replicate bootstrapped models. For populations in which a treatment had a significant effect on λ, we utilized life table response experiments to determine the vital rates responsible for this difference (Caswell, 2001; K. C. Baer & J. L. Maron, unpubl. data). We used simple linear regressions (lm and ANOVA, r base package) to evaluate whether ΔλP and ΔλI were correlated with the latitudinal position of study populations, and a Welch two-sample t test (t test, r base package) to test for differences in mean ΔλP or ΔλI in populations grouped by their range location (central and northern edge).

Statistical analyses

All statistical analyses were conducted using r Statistical Software version 3.3.1 (R Core Development Team, 2016). For each model, we tested for significance of all possible interactions between explanatory variables. If interactions were not significant, models were re-run without interaction terms.

To verify the efficacy of the insecticide treatment, we compared the mean proportions of plants that showed evidence of pre-dispersal seed predation among control and insecticide-treated plants in each population. To do this, we used a repeated measures linear mixed model (lmer, lme4 package, Bates, Maechler, Bolker, & Walker, 2015; lmerTest, lmerTest package, Kuznetsova, Brockhoff, & Christensen, 2016; ANOVA, r base package) with the log-transformed proportion of plants with signs of insect damage to fruits as the response variable, experimental treatment and year as fixed effects, and study population nested within year as a random effect to account for multiple measures across years in each population.

To evaluate whether pollen supplementation increased fruit production, we used a repeated measures linear mixed model to evaluate the effect of treatment (supplemental pollen vs. control), range location and year on the mean ratio of flowers that produced fruits, including population nested within year as a random factor to account for multiple measures across years in each population. We used identical models for determining the effect of pollen supplementation on population means for seed production per fruit for all populations and years and for determining the effects of insecticide on per capita fruit production and mean seeds per fruit. We determined whether the effects of the supplemental pollination and insecticide treatments on log-transformed mean per capita seed production varied among range locations using repeated measures linear mixed models with treatment, range location and year as fixed factors and population nested within year as a random factor to account for repeated measures within each population. We compared each experimental treatment to the control separately rather than using a single model including all treatments as there was no factorial pollination*insecticide treatment, although the results were similar when a single model was run with both experimental treatments included (K. C. Baer & J. L. Maron, unpubl. data).

Using the results of the seed addition experiment, we tested whether increasing seed input resulted in correlated gains in recruitment, which would be indicative of seed-limited recruitment. To do this, we used linear mixed models (lmer, lme4 package, Bates et al., 2015; lmerTest, lmerTest package, Kuznetsova et al., 2016; ANOVA, r base package) to compare cumulative germination at each site*sowing density combination after 3 years for seed additions performed in 2013, and after 2 years for seed additions performed in 2014. Models included range location, sowing density and their interaction as fixed factors and population as a random factor. We did not include seed additions in the central Ogden population in these analyses as the wildfire that passed through the site was expected to result in anomalous cumulative germination rates.


Effects of supplemental pollination and insecticide on seed production

Pollen supplementation had a positive effect on fruit production that was consistent across years and range locations (Figure 2a; mean increase in fruit production = 41% across all years and populations; F1,14 = 7.5, p < .05; p > .05 for all interaction terms). The mean number of seeds per fruit did not differ significantly among the control (4–13 mean seeds/fruit) and pollination (3–12 mean seeds/fruit) treatments, regardless of range location or year (F1,19.7 = 1.8, p = .2; p > .05 for all interaction terms). The estimated mean effect of supplemental pollination on per capita seed production across all years of the study ranged from an 8% decrease (in the northern Wolverine population) to a 291% increase (in the northern Nordic Center population). When all populations and years were included in the model, we estimated an average increase in per capita seed production of 78% with pollen supplementation (Figure 2b; F1,19.2 = 5.3, p < .04), a pattern that that did not vary according to range location (central or northern) or year (p > .05 for all interaction terms).

Details are in the caption following the image
(a) Mean (±SEM) change in the proportion of flowers that produced fruits in the control and hand (supplemental) pollination treatment groups over the 2014 and 2015 growing seasons in each study population (arranged by increasing latitudinal position from left to right). (b, c) Least squares mean (±SEM) number of seeds produced by control vs. (b) Hand pollination and (c) insecticide treatment plants at the centre of the distribution and the northern range edge. * denotes significance at the p < .05 level in generalized linear mixed models of log-transformed values. Plants included in the control group are identical for both tests

Seed beetles (A. fraterculus) and weevils (T. proxilus) attacked plants in populations at both the centre and the northern edge of the distribution. The insecticide treatment effectively decreased pre-dispersal seed predation (F1,29.6 = 9.6, p < .01) and effects of insecticide were consistent across the years of the study (year: F2,15.6 = 3.3, p = .06; treatment*year: F2,29.6 = 0.29, p = .75). Insecticide increased per capita seed production between 33% (in the central Ogden population) and 233% (in the central Providence and northern Nordic Center populations) when averaged across all years of the study, with an average increase of 115% across all populations and years (Figure 2c; F1,20 = 33.6, p < .0001). As with pollination, the effect of insecticide on per capita seed production was similar across all years and at the centre and northern edge of the range (p > .05 for all interaction terms). The increase in mean per capita seed production in the insecticide treatment was the result of both greater total fruit production (F1,20 = 23.9, p < .0001) and greater seed production per fruit (F1,24.8 = 4.3, p < .05) when all study populations were included in the model. Mean increases in per capita fruit production ranged from 13% (in the Ogden population) to 171% (in the Providence population). Changes in mean seeds per fruit when averaged across all years of the study ranged from a 13% decrease in the Providence population to a 35% increase in the Wolverine population. No interaction terms were significant in either model.

Plants in pollen supplementation and insecticide treatments did not differ significantly in their growth or survival as compared to control plants. However, in some populations and years growth decreased with higher seed production across all treatments applied at the whole-plant scale, indicating a cost of increased reproduction unrelated to experimental treatment (Appendix S2).

Seed addition effects on seedling recruitment

Cumulative seedling recruitment over multiple years showed a significant positive relationship with the number of seeds added to the subplot regardless of range location, indicating that populations are consistently seed-limited. However, the slope of this relationship was steeper within central than northern populations (Figure 3a; range location*sowing density: F1,16 = 7.9, p < .05). A similar pattern was seen with cumulative seedling recruitment after 2 years in seed addition plots added in 2014 (Figure 3b; range location*sowing density: F1,19 = 10.6, p < .01).

Details are in the caption following the image
Least squares mean (±SEM) cumulative recruits in seed addition plots in central and northern edge study populations over the course of (a) 3 years for seed addition plots established in 2013 and (b) 2 years for seed addition plots established in 2014

Impacts of experimental treatments on projections of stochastic population growth

Mean values of ΔλP over the two transitions were positive, albeit not significantly so, in all but one northern edge population (Figure 4a). In the populations showing a positive mean value of ΔλP, this impact was driven by higher projected seed production with supplemental pollination (K. C. Baer & J. L. Maron, unpubl. data). In the northern Nordic Center population, the non-significant negative mean value of ΔλP resulted from a decreased probability of flowering associated with pollination in the largest size classes within the model (K. C. Baer & J. L. Maron, unpubl. data). This likely represents a cost of increased reproduction in pollen-supplemented plants at Nordic Center, which showed the largest proportional increase in per capita seed production with pollination of any population. The impact of pollen supplementation on λ did not vary significantly across the latitude of study populations (Figure 4a; F1,5 = 3.0, p = .15). Furthermore, when populations within a given range location (i.e. range centre vs. range edge) were compared as groups, the effect of pollen supplementation on λ did not differ significantly among range locations (t = 1.7, df = 2.9, p = .19).

Details are in the caption following the image
Mean difference in projected population growth rate across latitude between (a) supplemental pollination treatment groups and control groups (ΔλP) and (b) insecticide treatment groups and control groups (ΔλI). Estimates are based on 1,000 bootstrapped values of asymptotic λ (500 estimates for each of the two yearly transitions; see Methods for details). Error bars represent the 95% confidence interval for the mean. The dashed line represents the point above which treatments increase λ, and below which treatments decrease λ [Colour figure can be viewed at wileyonlinelibrary.com]

Insecticide application was projected to significantly increase λ in one central and one northern study population and marginally increase λ in an additional northern population (Figure 4b; Table 1). In all of the remaining populations, mean values of ΔλI were projected to be positive, but not significantly greater than zero (Figure 4b). Regardless of statistical significance, the positive mean projected values for ΔλI in all study populations were the result of increased seed production (K. C. Baer & J. L. Maron, unpubl. data). The magnitude of mean ΔλI did not change across the latitudinal range of the study populations (Figure 4b; F1,5 = 0.2, p = .65); the same was true of comparisons among grouped central and northern sites (t = 0.6, df = 4.1, p = .59).

Table 1. Estimates of stochastic lambda (λs) for control and supplemental pollination or insecticide treatments across all populations. Treatments where λ is estimated to differ significantly from the control are indicated with an asterisk (*); marginal significance is indicated with a dagger (†). All other differences are non-significant
Site Latitude λs
Control Pollination Insecticide
Bountiful 40.903 1.442 1.541 1.745*
Uinta 41.091 1.511 2.067 1.705
Providence 41.690 1.155 1.307 1.345
Nordic Center 42.723 1.035 1.121 1.032
Swan 43.209 1.026 1.060 1.037
Wolverine 43.259 1.038 1.317 1.404*
Reservoir 43.285 1.027 1.114 1.16


Our results demonstrate that both pollen limitation and pre-dispersal seed predation were pervasive negative influences on per capita seed production. Amelioration of pollen limitation and pre-dispersal seed predation led to consistent increases in projected population growth at both the centre and the northern edge of the geographic distribution of A. utahensis, although these increases were generally not statistically significant. Below, we discuss these effects and their potential implications for the role of pre-dispersal seed predation and pollen limitation in constraining the northern extent of the A. utahensis’ latitudinal distribution in more detail.

Effects of pollen supplementation and insecticide application on seed production

Pollen limitation commonly limits plant fecundity (Ashman et al., 2004). Similarly, we found strong evidence for pollen limitation; A. utahensis seed output was enhanced by an average of 78% due to experimental pollen supplementation when averaged across all years in the eight populations we studied. Populations of A. utahensis at its northern range edge are generally less dense and smaller than those at the range centre (Baer, 2017; K. C. Baer & J. L. Maron, unpubl. data), characteristics which have been linked to higher pollen limitation in other species (Ågren, 1996; Ashman et al., 2004; Kunin, 1993; Wilcock & Jennings, 1999; e.g. Cheptou & Avendaño, 2006; Groom, 1998; Morgan, 1999). Northern populations also tend to have smaller floral displays (K. C. Baer, pers. obs.), another trait often associated with lower pollinator visitation (Ashman & Morgan, 2004; Bosch & Waser, 2001; Delph & Ashman, 2006; Harder & Johnson, 2009; Weber & Kolb, 2013). Nevertheless, we found no relationship between the magnitude of pollen limitation and A. utahensis density or population size (K. C. Baer & J. L. Maron, unpubl. data). Other factors associated with pollen limitation that we did not quantify, such as resource availability (Burd, 2008; Haig & Westoby, 1988) or relative abundance of heterospecific neighbours that could attract the generalist pollinators (Ghazoul, 2006; Groom, 1998; Nottebrock, Esler, & Schurr, 2013; Seifan, Hoch, Hanoteaux, & Tielboerger, 2014) that visit A. utahensis may be more predictive of variation in pollen limitation among populations than density or floral display.

Just as pollen limitation commonly constrains seed production, so too does pre-dispersal seed predation (Brody, 1991; Ehrlén, Käck, & Ågren, 2002; Kolb, Ehrlén, et al., 2007; Leimu, Syrjanen, Ehrlén, & Lehtila, 2002; Louda, 1982; Weppler & Stoecklin, 2006). Perhaps due to their relatively large seeds, Astragalus species often suffer substantial pre-dispersal seed loss to insects, ranging from roughly 5% to 78% (Boe & Johnson, 2008; Combs, Lambert, & Reichard, 2013; Combs, Reichard, Groom, Wilderman, & Camp, 2011; Green, 1976; Jordano, Haeger, & Rodriguez, 1990; Kaye, 1999; Lesica, 1995; Platt, Hill, & Clark, 1974). In keeping with a prior observational study of A. utahensis (Green, 1976), we found that pre-dispersal seed predation strongly impacted seed production in this species. Our results suggest that these impacts are pervasive across a large portion of its range.

We observed that plants often aborted flowers that received insufficient pollen or flowers and developing fruits that contained seed predator larvae. This contrasts in part with the findings of Green (1976), who demonstrated that the negative impacts of pre-dispersal seed predators on A. utahensis took place at the level of mean seed production per fruit. While we found that pre-dispersal seed predation does significantly decrease the mean number of seeds per pod, our experimental approach suggests that this difference is less dramatic than suggested by Green (1976). A possible explanation for this discrepancy is that Green (1976) did not apply an experimental treatment to reduce pre-dispersal seed predators in A. utahensis. Rather, he estimated per cent seed predation as the difference between the total production of viable and aborted seeds per pod and an estimated value for potential seed production per pod based upon locule number. This methodology attributed all decreases in seed production to pre-dispersal seed predation and may have overestimated its effects. In addition, Green (1976) did not estimate the effect of pre-dispersal seed predators on per capita fruit production, so we cannot evaluate the relative strength of this effect as compared to effects on mean seed production per fruit.

Effects of pollen supplementation and insecticide application on population growth

Although pollen limitation of seed production is widespread, its effects on λ remain understudied (Ashman et al., 2004), as does the question of how these effects may vary according to environmental context (Maron et al., 2014). We are aware of three studies which have tested effects of pollen limitation on λ under different abiotic contexts. While all studies found evidence for pollen limitation of population growth, they came to contrasting conclusions regarding the extent to which this outcome was dependent upon environmental context (Bierzychudek, 1982; Horvitz et al., 2010; Parker, 1997).

In contrast to the findings of Parker (1997) and Horvitz et al. (2010) but in keeping with those of Bierzychudek (1982), our results indicate that pollen limitation on average has negative, albeit quite variable, impact on λ in most study populations. The average projected increase in λ with pollen supplementation across all study populations was 0.15 ± 0.05, and the magnitude of this increase was not correlated with shifting abiotic conditions from the range centre to the northern edge (K. C. Baer & J. L. Maron, unpubl. data). The mean effect of pollen supplementation on per capita seed production over the course of the two transitions was not predictive of ΔλP in a population, indicating that studies of the impacts of biotic interactions on individual plants at range edges are unlikely to accurately predict their importance in limiting population growth at range edges. Our findings contrast with the predictions of SIASH (Louthan et al., 2015) that the effects of biotic interactions should be strongest where abiotic conditions are most benign. We did not observe higher pollen limitation in the centre of the distribution than at its northern edge, regardless of the fact that the central study populations were predicted to have categorically higher climatic suitability for the species than northern edge populations (K. C. Baer, J. L. Maron, & A. L. Angert, unpubl. data).

An important caveat to our results is that that our projection of ΔλP may be an overestimate. Since we supplemented pollen to only a subset of flowers per plant and mathematically extrapolated the effects of this treatment on seed production to the whole plant (Appendix S3), costs of reproduction associated with the projected change in per capita seed production were not necessarily reflected in the vital rates of pollen-supplemented plants. This is expected to be particularly true when the number of pollen-supplemented flowers represented a small proportion of the total flowers produced by an individual throughout the season. This may explain the relatively large projected increase in population growth with pollen supplementation in the Uinta population, where plants produced more flowers than plants in all other populations and as such a smaller proportion of flowers were pollinated. If all flowers had been pollinated, costs of higher reproduction may have arisen and been reflected in estimates of both per capita seed production and Δλ.

In contrast to the pollination treatment, the control and insecticide treatments were applied at the level of the whole plant and vital rates associated with these treatments incorporated associated costs of reproduction. In several study populations, the growth of both control and insecticide treatment plants with higher per capita seed production was often lower than that of plants with lower per capita seed production regardless of experimental treatment (Appendix S2). Despite this evidence of a cost of higher reproduction to yearly growth, we do not expect that this result qualitatively changes the patterns seen in our results for the pollen supplementation treatment or their interpretation, since yearly growth rate contributed relatively weakly to the determination of λ in all populations (K. C. Baer and J. L. Maron, unpubl. data). In addition, we observed no consistent costs of higher reproduction in plants treated with insecticide, which increased seed production by an amount similar to or greater than that estimated for pollen supplementation at the whole-plant level in all but one study population (Appendix S2). Therefore, we assume that costs of reproduction associated with pollen supplementation are relatively minimal, and our estimates of its effect on λ are reasonably accurate.

Evidence for significant impacts of pre-dispersal seed predation on plant population growth is more common than for pollen limitation. Across 26 studies, Katz (2016) found that amelioration of herbivory increased projected λ from 1.08 ± 0.36 to 1.28 ± 0.58, with a mean projected increase in λ due to amelioration of seed predation of 0.14 (95% CI: −0.10 to 0.35). Our results agree well with this estimate: insecticide application was projected to increase λ by an average of 0.19 ± 0.05 across all study populations. We are aware of no other studies of how herbivory influences plant population growth across a suite of populations from the centre to the edge of a species’ latitudinal geographic range. Miller et al. (2009) showed that the impact of herbivory on projected λ of Opuntia imbricata decreased with increasing elevation across a portion of its elevational range, a pattern in keeping with the predictions of SIASH that herbivore impacts should be greatest in areas of relatively low abiotic stress (Louthan et al., 2015). Similarly, other studies have found that per capita effects of herbivory often decline with increasing abiotic harshness (Louthan et al., 2015; but see Andrew & Hughes, 2005; Kim, 2014; Russell, Rose, & Louda, 2010). In contrast to the predictions of SIASH, we found that pre-dispersal seed predation affected λ similarly across the centre-north portion of its latitudinal distribution, regardless of climatic context. As with pollen supplementation, the exclusion of pre-dispersal seed predators was projected to lead to an increase in λ in all study populations, although unlike with pollen supplementation, this increase was statistically significant in two of the study populations and marginally significant in one.

Implication of biotic interactions for influencing A. utahensis’ northern range limit

As pre-dispersal seed predation and pollen limitation affected A. utahensis population growth similarly regardless of range location, changes in the outcome of these interactions are unlikely to be responsible for the observed decline in stochastic λ (λs) from the centre to the edge of its distribution (Table 1). Nevertheless, they may contribute substantially to the establishment of the northern range boundary. Decreases in λs due to pollen limitation or pre-dispersal seed predation are unlikely to affect the long-term viability of most central populations; values of stochastic lambda were well above replacement under ambient levels of pre-dispersal seed predation and pollination in two of the three central populations (Table 1). In northern edge populations, where projected values of stochastic lambda are much closer to 1 (Table 1), these interactions may lead to a lower probability of persistence during demographically unfavourable years. It should be emphasized that this conclusion assumes that the two transitions for which we present demographic data are representative of average conditions encountered in these populations over much longer time-scales. Ideally, an evaluation of the effects of pollen limitation or pre-dispersal seed predation on stochastic population growth and population persistence should be carried out over many transitions in order to accurately capture the degree of temporal stochasticity in the biotic and abiotic environments (Metcalf et al., 2015). Nonetheless, our results provide an important step forward in our understanding of the possible role for non-competitive biotic interactions in setting range limits of plants.

In addition to their potential effects on population persistence at the northern edge of the range, these biotic interactions may contribute to dispersal limitation of the northern extent of the range via their effects on the availability of seeds for colonizing suitable habitat beyond the northern edge of the current distribution (Baer, 2017). Future studies that examine the impacts of biotic interactions on projected population growth in transplant sites beyond the range edge would lend further insight into the role played by non-competitive biotic interactions in constraining species’ geographic distributions.


The authors thank Peggy Stolworthy for permission to perform work on private land, and the U.S. Forest Service for permission to perform work on public lands in Utah and Idaho. We also thank Zachary Baer, Christina Cain, Becky Fletcher, Allison Klocke, Mike Kimball, Eric Mohrmann and Jennifer Neville for their assistance with field work, Phillip Hahn, Nicole Hupp and Loralee Larios for assistance with statistics, and Jim Cane and Geoffrey Morse for assistance in identification of bee and beetle specimens. Assistance with integral projection model construction and troubleshooting was provided by Emilio Bruna, Johan Ehrlén, Roberto Salguero-Gómez and the Max Planck Institute for Demographic Research. Helpful comments from Ragan Callaway, Phillip Hahn, Kay Hajek, Ryan Hegstad, Rebecca Irwin, Winsor Lowe, Tom Martin and two anonymous reviewers improved the manuscript. This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under NSF grant no. DGE-1313190. J.L.M. was supported off of NSF grants DEB-0614406 and DEB-1553518.


    K.C.B. and J.L.M. developed the question and methodologies described in this paper; K.C.B. collected and analysed the data and led the writing of the manuscript; J.L.M. contributed critically to data analysis and drafts of the manuscript. Both authors gave final approval for publication.


    Data associated with this study are archived in the Dryad Digital Repository: https://doi.org/10.5061/dryad.g62pv (Baer & Maron, 2018).