Accepted Manuscript

Title: Sampling completeness in seed dispersal networks:

When enough is enough

Author: José M. Costa Luı́s P. da Silva Jaime A. Ramos

Ruben H. Heleno

PII: S1439-1791(15)00125-5
DOI: http://dx.doi.org/doi:10.1016/j.baae.2015.09.008
Reference: BAAE 50922

To appear in:

Received date: 23-12-2014

Revised date: 18-8-2015
Accepted date: 16-9-2015

Please cite this article as: Costa, J. M., da Silva, L. P., Ramos, J. A., and Heleno, R.
H.,Sampling completeness in seed dispersal networks: When enough is enough, Basic
and Applied Ecology (2015), http://dx.doi.org/10.1016/j.baae.2015.09.008

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
 1 RESEARCH ARTICLE

2 Title: Sampling completeness in seed dispersal networks: When enough is enough

3 José M. Costa1,2*, Luís P. da Silva1,2, Jaime A. Ramos2 and Ruben H. Heleno1

t
5 1 Centre for Functional Ecology, Department of Life Sciences, University of Coimbra,

ip
6 Coimbra, Portugal.

cr
7 2 MARE – Marine and Environmental Sciences Centre, Department of Life Sciences,

8 University of Coimbra, Coimbra, Portugal.

us
9

an
10

11 * Corresponding author. Tel.: +351 239 855 210

M
12 E-mail address: jmgncosta@gmail.com

13
d
p te
ce
Ac

1
Page 1 of 32
13

14 Abstract

15 Ecological networks are an increasingly popular tool to explore community

16 assembly rules and frame practical conservation issues. However, most described

t
17 networks vary largely in sampling effort, hampering the distinction of true

ip
18 biological patterns from artefacts caused by poor sampling. Identifying entire seeds

cr
19 in the droppings of mist-netted birds is generally considered a preferred sampling

us
20 method for building unbiased, quantitative seed dispersal networks. We retrieved

21 seeds from the droppings of 936 mist-netted birds captured during five days in seven

an
22 sites in Portugal and estimated sampling completeness as the diversity of seed

23 species, disperser species, and links detected with respect to those predicted by
M
24 Chao 2 estimator. In one of those sites, sampling effort was extended to 25 days to

25 evaluate the sensitivity of ten network structure descriptors to increasing sampling

d

26 effort. After five sampling days we detected 93% of the seed species, 97% of the
te

27 disperser species, and 79% of the links predicted by Chao 2, however sampling for

28 25 days resulted in the detection of more seeds, dispersers, and links than those
p

29 estimated at day 5. Most network descriptors only began to stabilize around day 8,
ce

30 except for Connectance and Weighted Connectance that stabilized earlier. Similarly,

31 only after 8 days most networks descriptors significantly departed from the
Ac

32 confidence interval estimated by null models exclusively constrained by species

33 abundances, thus reflecting independent ecological patterns. Nestedness was the

34 only exception, as it never departed from the null models. We suggest that Chao 2

35 may slightly underestimate the real diversity and that in our case at least eight

36 sampling days were needed to build a sound seed dispersal network as 67% of the

37 seeds, 88% of the dispersers, and 71% of the links were detected. Our results have

2
Page 2 of 32
38 important implications for the interpretation of seed dispersal networks because

39 under-sampled networks may produce biased descriptors that do not suitably

40 characterize the focal communities.

41

t
ip
cr
us
an
M
d
p te
ce
Ac

3
Page 3 of 32
41 Zusammenfassung
42 Ökologische Netzwerke werden immer beliebter als eine
43 Methode, um die 'assembly rules' für Gemeinschaften zu
44 untersuchen und praktische Naturschutzregeln zu
45 entwerfen. Indessen variieren die meisten Netzwerke
46 erheblich hinsichtlich des Probenahmeaufwandes, was die

t
47 Unterscheidung zwischen tatsächlichen biologischen

ip
48 Mustern und Artefakten, die sich aus unzureichender
49 Probenahme ergeben, erschwert. Die Bestimmung von

cr
50 Samen aus dem Kot von Vögeln aus Japannetzfängen gilt
51 als eine bevorzugte Sammelmethode, um unverfälschte,

us
52 quantitative Samenausbreitungsnetzwerke zu konstruieren.
53 Wir separierten Samen aus dem Kot von 936 gefangenen
54 Vögeln (sieben Fangplätze in Portugal; Fangzeitraum: i.d.R.

an
55 fünf Tage) und bestimmten die Vollständigkeit der
56 Erfassung als die Diversität der festgestellten Samenarten,
57 Vogelarten und Verbindungen in Relation zum nach Chao 2
M
58 berechneten Schätzwertes. An einem der Fangplätze wurde
59 die Probenahme auf 25 Tage ausgedehnt und die Reaktion
60 von zehn Netzwerkdeskriptoren auf die zunehmende
d

61 Stichprobengröße ermittelt. Nach fünf Probetagen erhielten

te

62 wir 93% der Samenarten, 97% der Vogelarten und 79% der
63 vorhergesagten (Chao 2) Interaktionen. Der Fangzeitraum
p

64 von 25 Tagen erbrachte indessen noch mehr Samenarten,

65 Vogelarten und Interaktionen. Die meisten
ce

66 Netzwerkdeskriptoren begannen sich nach etwa acht Tagen

67 zu stabilisieren. Nur Konnektanz und gewichtete
68 Konnektanz stabilisierten sich früher. In gleicher Weise
Ac

69 wichen die meisten Netzwerkdeskriptoren erst nach acht

70 Tagen signifikant von den Vertrauensbereichen ab, die für
71 Nullmodelle mit limitierten Artenabundanzen berechnet
72 wurden, was unabhängige ökologische Muster anzeigt. Die
73 'nestedness' war die einzige Ausnahme und wich in keinem
74 Fall von den Nullmodellen ab. Wir meinen, dass Chao 2 die
75 tatsächliche Diversität leicht unterschätzen könnte und dass
76 in unserem Fall mindestens acht Fangtage benötigt wurden,
77 um ein vernünftiges Ausbreitungsnetzwerk zu konstruieren,
78 da 67% der Samenarten, 88% der Vogelarten und 71% der
4
Page 4 of 32
79 Interaktionen (bezogen auf den 25-tägigen Fangzeitraum)
80 entdeckt wurden. Unsere Ergebnisse haben wichtige
81 Konsequenzen für die Interpretation von
82 Samenausbreitungsnetzwerken, weil unzureichend
83 beprobte Netzwerke verfälschte Deskriptoren hervorbringen
84 können, die zu unbrauchbaren Beschreibungen der

t
85 untersuchten Gemeinschaften führen.

ip
86
87

cr
88

us
89 Keywords: Asymptotic estimators, Cumulative sampling, Endozoochory, Food-webs,

90 Mist-nets, Sampling effort

91

an
M
d
p te
ce
Ac

5
Page 5 of 32
 91

92 Introduction

93 The study of community-wide species interaction networks experienced a

94 tremendous growth in recent years and proved a powerful tool to explore many

t
95 processes in ecology, being particularly valuable in disentangling the relationships

ip
96 between the structure and function of nature’s “entangled bank” [Darwin 1859]

cr
97 (Heleno, Garcia, Jordano, Traveset, Gómez et al. 2014; Olesen, Dupont, Hagen,

98 Rasmussen & Trøjelsgaard 2012). The use of ecological networks deepened our

us
99 understanding of important conservation issues such as environmental degradation (e.g.

an
100 Traveset, Heleno, Chamorro, Vargas, McMullen et al. 2013; Tylianakis, Tscharntke &

101 Lewis 2007) and ecological restoration (e.g. Heleno, Lacerda, Ramos & Memmott
M
102 2010).

103 Seed dispersal is one of the research areas where ecological networks attracted
d

104 greatest interest (e.g. Carlo & Yang 2011; Donatti, Guimarães, Galetti, Pizo, Marquitti
te

105 et al. 2011; Heleno, Olesen, Nogales, Vargas & Traveset 2013; Mello, Marquitti,

106 Guimarães, Kalko, Jordano et al. 2011). The dispersal of seeds away from the mother
p

107 plant is a key process, allowing plants to avoid competition, find suitable conditions for
ce

108 germination, and expand their distribution range (Traveset, Heleno & Nogales 2014).

109 Due to their high mobility, frugivorous birds are the main seed dispersers in most
Ac

110 ecosystems (Herrera 1984; Traveset et al. 2014). Avian seed dispersal data can be

111 collected by: (1) recording fruit consumption by birds on focal plants (Olesen,

112 Bascompte, Dupont, Elberling, Rasmussen et al. 2011), (2) identifying entire seeds in

113 the droppings of mist-netted birds (Heleno et al. 2013), and (3) identifying seeds in

114 droppings collected in the field and identifying the disperser species with molecular

115 techniques (González-Varo, Arroyo & Jordano 2014).

6
Page 6 of 32
116 Most network structure descriptors are affected by sampling effort to some degree

117 (Blüthgen 2010; Blüthgen, Fründ, Vázquez & Menzel 2008; Rivera-Hutinel,

118 Bustamante, Marín & Medel 2012), with qualitative indices being more sensitive to

119 sample size than quantitative analogues (Banasek-Richter, Cattin & Bersier 2004).

t
120 Specifically, poor sampling underestimates the real diversity of links, truncating

ip
121 estimated trophic breath and leading to a biased network structure (Blüthgen et al.

cr
122 2008). On the other hand, as implementing such a holistic approach is inherently highly

123 labour-intensive, it is important to know when further effort will not significantly

us
124 increase the accuracy of the community description, thus avoiding unnecessary work

125 load (Hegland, Dunne, Nielsen & Memmott 2010). This effort is of utmost importance

126
an
to allow meaningful cross-study comparisons (Heleno et al. 2014).
M
127 Richness estimators based on species and link accumulation curves are a powerful

128 way to evaluate sampling completeness (Chacoff, Vázquez, Lomáscolo, Stevani,

d

129 Dorado et al. 2012; Olito & Fox 2015; Rivera-Hutinel et al. 2012), where the number of
te

130 missing species is estimated as those needed to reach a theoretical asymptote (Chacoff

131 et al. 2012). Although some statistical methods have been suggested to ease the problem
p

132 (e.g. Bartomeus 2013), there are no satisfactory solutions for poor sampling, and the
ce

133 sampling effort needed to build comprehensive seed dispersal networks from which

134 theoretical and applied conclusions can be drawn has not yet been evaluated. In this
Ac

135 study we aim to evaluate (1) the effect of sampling effort on the completeness of seed

136 dispersal networks based on the analysis of droppings from mist-netted birds, and (2)

137 the sampling effort needed to build high quality seed dispersal networks so that reliable

138 network descriptors can be calculated.

139

7
Page 7 of 32
140 Materials and methods

141 Study sites and data collection

142 On five consecutive days in the first half of September 2012, birds were captured in

143 seven sites throughout Portugal ranging from agro-forestry systems to secondary native

t
ip
144 forest (Fig. 1) (see Costa, Ramos, da Silva, Timoteo, Araújo et al. 2014). In each

145 site/day, 80-100 m of mist-nets (according to the site specific conditions) were operated

cr
146 during five hours after dawn. Nets were visited every 30 minutes and captured birds

us
147 were individually placed in ringing bags to produce droppings (Heleno et al. 2013).

148 Intact seeds were later extracted from the droppings and identified under a dissecting

an
149 microscope by comparison with a reference collection. Interaction frequency was

150 quantified as the number of droppings from each bird species containing at least one
M
151 intact seed of any of the plant species. We built quantitative seed dispersal networks for

152 each site by considering the cumulative samples collected up to day i. We used the
d

153 number of days as a measure of sampling effort because there were no significant
te

154 differences in the number of droppings collected per day (average ± SD droppings

155 collected per day = 124.8 ± 22.7, see Appendix A: Table 1). The availability of fleshy
p

156 fruits was independently assessed at each site by counting all ripe standing fruits along
ce

157 three 2x25m transects parallel to the mist-nets. In the site with highest fruit diversity
Ac

158 (n=10; Fig. 1, Site A - Larçã), sampling continued for another consecutive 20 days.

159

160

161 Network descriptors

162 The consistency of network structure descriptors to increasing sampling effort was

163 evaluated with the data collected during five days in all sites (Table 1) and with the data

8
Page 8 of 32
164 collected during 25 consecutive days at the site with the highest fruit diversity. We

165 explored the effect of increasing sampling effort in six qualitative and four quantitative

166 network descriptors. Qualitative descriptors included: Connectance, the proportion of

167 realised links from all possible links in the network (Jordano 1987); Nestedness

t
168 (NODF) (Almeida-Neto, Guimarães, Guimarães Jr, Loyola & Ulrich 2008), which

ip
169 reflects the degree of organization of interactions around a core of generalist species;

cr
170 Disperser richness; Seed richness; Link richness; Qualitative Modularity, reporting the

171 existence of clusters of tightly interacting species (Olesen, Bascompte, Dupont &

us
172 Jordano 2007). Qualitative Modularity was estimated with the algorithm QuaBiMo

173 (Dormann & Strauss 2014) using both binary (qualitative) and weighted (quantitative)

174
an
matrices. Although this algorithm has been specifically developed for weighted
M
175 interaction matrices, we applied the same algorithm to binary versions of the original

176 matrices in order to evaluate the effect of the input data (binary/weighted). The
d

177 quantitative descriptors included: Weighted Connectance, the ratio between linkage
te

178 density (mean number of links per species) and the number of species in the network

179 (Bersier, Banasek-Richter & Cattin 2002); Weighted Nestedness (WNODF), as NODF
p

180 but takes into account interaction frequency (Almeida-Neto & Ulrich 2011); Network
ce

181 specialisation index (H2’), measuring the degree of the partner’s selectivity as the

182 departure from a theoretical non-discrimination of interactions (Blüthgen, Menzel &

Ac

183 Blüthgen 2006); Quantitative Modularity, as Qualitative Modularity but based on the

184 original, weighted interaction matrices. The significance of each network descriptor was

185 assessed by comparison with the 95% confidence interval of a set of 1000 matrix

186 randomizations (100 for Modularity) with marginal totals equal to the observed matrix.

187 This allows us to distinguish if a certain parameter is driven by biological/ecological

188 characteristics of the interacting species or results simply from random interaction

9
Page 9 of 32
189 patterns driven mostly by species abundances. Piecewise regressions were used to

190 estimate the sampling effort needed to achieve asymptotic values (i.e. slope of partial

191 regression line not significantly different from zero).

192 Network descriptors were calculated with packages bipartite 2.05 (Dormann,

t
193 Gruber & Fründ 2008) and vegan 2.2 (Oksanen, Blanchet, Kindt, Legendre, Minchin et

ip
194 al. 2015), while piecewise regressions were performed with package segmented 0.5

cr
195 (Muggeo 2008) in R 3.0.2 (R Development Core Team 2014).

196

us
197 Sampling completeness

an
198 The minimum expected asymptotic richness of seed species, disperser species and

199 links was estimated with the Chao 2 , a non-parametric estimator based on the
M
200 proportion of uniques (species captured on a single day) relative to duplicates (species

201 captured on two days) (Chao 1984; Colwell & Coddington 1994). The expected
d

202 richness was computed using the program EstimateS 9.1.0 (Colwell 2013). The Chao 2
te

203 estimator was chosen for being more robust to reduced sample size (Colwell et al.

204 1994), performing considerably better than other estimators (Walther & Moore 2005).
p

205 Sampling completeness was considered as the percentage of observed richness

ce

206 relative to the estimated asymptotic richness. By extrapolation of the estimated

Ac

207 rarefaction curves (Colwell, Chao, Gotelli, Lin, Mao et al. 2012), we estimated the

208 additional number of sampling days needed to detect 80%, 90%, and 100% of the total

209 estimated richness.

210 To evaluate whether sampling completeness is affected by fruit diversity across

211 sites, a generalized linear model with quasibinomial error distribution (due to data

212 underdispersion) was used.

213

10
Page 10 of 32
214

215 Results

216

217 Network descriptors

t
ip
218 During the five sampling days, different network descriptors showed different

219 trends regardless of whether they were based on qualitative or quantitative links (Fig.

cr
220 2). Both Connectance and Weighted Connectance of all networks stabilized at day 3.

us
221 NODF, Weighted NODF and H2’ showed a high site-specific variability (Fig. 2).

222 All but two (Seed Species and Link Richness) descriptors stabilised before the end

an
223 of the study at the site sampled for 25 days (see Appendix A: Table 2) but different

224 metrics showed distinct patterns in respect to increasing sampling effort. The number of
M
225 detected seed species and links was still increasing as revealed by the positive slopes of

226 the partial regressions (seeds: 0.19 < β 95% CL < 0.42; links: 0.62 < β 95% CL < 0.80) (Fig.
d

227 2, see Appendix A: Table 2). However, disperser’s diversity stabilised at day 8 with no
te

228 further detected species (-0.02 < β 95% CL < 0.02, Fig. 2, see Appendix A: Table 2). Both
p

229 Connectance and Weighted Connectance remained relatively stable throughout the
ce

230 study period (Connectance = 40% - 50% (-0.02 < β 95% CL < 0.0005, Fig. 3A); Weighted

231 Connectance = 21% - 27% (-0.001 < β 95% CL < 0.0002, Fig. 3B)). NODF stabilized after
Ac

232 day 8, (day 8-21: -0.01 < β 95% CL < 1.17; day 21-25: -5.09 < β 95% CL < 2.01; Fig. 3C, see

233 Appendix A: Table 2), while Weighted NODF tended to stabilize at day 8 but formally

234 reached its asymptotic value only on day 16 (-0.51 < β 95% CL < 0.58, Fig. 3D). The

235 Qualitative and Quantitative Modularity as well as Network Specialization H2’

236 stabilized around day 8 (Qualitative Modularity: -0.004 < β 95% CL < 0.00002;

237 Quantitative Modularity: -0.002 < β 95% CL < 0.0008; H2’: -0.004 < β 95% CL < 0.001)

238 (Fig. 3E, F, and H; see Appendix A: Table 2).

11
Page 11 of 32
239

240 Sampling completeness

241 After five sampling days, we detected, on average, 93% (range: 71% - 100%) of the

242 estimated seed species, 97% (92% - 100%) of the estimated disperser species, and 79%

t
ip
243 (57% - 100%) of the estimated links on the seven study sites (Table 1). Fruit diversity at

244 a given site did not significantly affect sampling completeness (GLM: β = - 0.06, t = -

cr
245 0.464, p = 0.662; see Appendix A: Fig. 1). After the first five sampling days in Larçã,

us
246 we detected only 58% of the seed species, 75% of the disperser species, and 45% of the

247 links detected within 25 days. The detection rate improved to 67% of the seeds, 88% of

an
248 the dispersers, and 71% of the links by day 8. After 25 sampling days, we estimated to

249 have recorded all seed species (n=12), all disperser species (n=8), and 73% (38 out of
M
250 52) of the estimated number of links at site A (Table 1). The observed richness in this

251 site after 25 days was higher than that estimated using Chao 2 at day 5 for all levels
d

252 considered (Seeds, Dispersers, and Links; Table 1). While the final number of detected
te

253 seed species and links was within the confidence interval estimated after five sampling

254 days, the same was not true for disperser species (Table 1). A total of 50 days were
p

255 considered necessary to detect 90% of the links and 72 days would be needed to detect
ce

256 all links (Table 2). An animation of accumulation of species and links at the most
Ac

257 intensively sampled site (Larçã) is available as supporting information (see Appendix

258 B).

259

260 Discussion

261 In this study we show that while some network descriptors can be accurately

262 estimated with lower sampling effort (number of Disperser Species, Connectance, and

263 Weighted Connectance), most network metrics only started to stabilize after eight mist-

12
Page 12 of 32
264 netting days (number of Links and Seed Species, NODF, Weighted NODF, Modularity

265 (both Quantitative and Qualitative), and H2’).

266

267 Network properties

t
268 While five sampling days suitably captured the biological diversity across the seven

ip
269 study sites, it was not sufficient to deliver a completely sound representation of the

cr
270 community structure measured by all but two network descriptors (Connectance and

271 Weighted Connectance). Contrary to our expectation we could not detect a trend for

us
272 quantitative indices performing better to lower sampling efforts. In fact, Weighted

273 NODF took slightly longer to stabilize than its qualitative counterpart.

274
an
Both Connectance and Weighted Connectance remained relatively stable
M
275 throughout the study period and therefore were accurately predicted even with

276 low/moderate sampling effort. However, only after day 7/8 the network is significantly
d

277 less connected than we would expect according to a random association among species
te

278 (Fig. 3). Still, given the small difference between observed and predicted values, such

279 significance may be explained by the intrinsic mathematical behaviour of the used null
p

280 model, which tends to generate matrices more connected than that observed (Dormann,
ce

281 Fründ, Blüthgen & Gruber 2009). Additionally, these descriptors showed similar trends

282 in all of our study sites, suggesting that such robustness to sample size is not network-
Ac

283 specific but also that these descriptors are of limited informative value to compare seed

284 dispersal networks. Previous studies on pollination networks found a tendency for an

285 initial increase in Connectance towards an asymptote that might be reached at relatively

286 low (Nielsen & Bascompte 2007) or higher (Rivera-Hutinel et al. 2012) levels of

287 sampling completeness. Our results suggest that, despite both Connectance and

288 Weighted Connectance of seed dispersal networks may be stable even with a low

13
Page 13 of 32
289 sampling effort, they are of limited use if we want to compare at least similar seed

290 dispersal networks. Moreover, despite its wide use in ecological network studies and

291 apparent robustness to sample size, Connectance seems to have no relation between its

292 value and the network conservation status (Heleno, Devoto & Pocock 2012). Further

t
293 work is required to confirm the extent of this generalization.

ip
294 The two measures of Nestedness (NODF and Weighted NODF) were more

cr
295 sensitive to sampling effort than Connectance. Asymptotic values of NODF were

296 achieved earlier (day 8) than for Weighted NODF (day 16), but overall our results

us
297 coincide with those of Rivera-Hutinel et al. (2012) who found NODF to be relatively

298 stable if at least 30% of the network had been sampled. However, the observed network

299
an
was not significantly nested (Fig.3) even after 25 sampling days. This pattern was not
M
300 the result of sampling limitations, as both NODF and Weighted NODF values were

301 relatively stable, with as low as half of the full sampling effort and thus not to be
d

302 expected to change with a higher sample size, but likely driven by species abundances.
te

303 Taken altogether our results question the ecological relevance of Nestedness indices

304 (see also James, Pitchford & Plank 2012). Modularity has been increasingly used in
p

305 network studies as it is strongly related to Nestedness (Fortuna, Stouffer, Olesen,

ce

306 Jordano, Mouillot et al. 2010).We found both Qualitative and Quantitative Modularity

307 to stabilize after 7 and 8 sampling days, respectively, and to encompass real ecological
Ac

308 information as it cannot be explained by null models based solely on species

309 abundances. However, it is important to note that wile Quantitative Modularity revealed

310 a network more modular than expected by chance, the analyses based on binary

311 matrices revealed the opposite pattern, reinforcing the value of weighted interaction

312 networks in the capacity of this algorithm to detect modules (Dormann & Strauss 2014).

313 Network Specialization H2’ stabilized around day 6 to 8 and was higher (i.e. more

14
Page 14 of 32
314 specialized) than expected under random species associations. These results corroborate

315 the stability of this metric to moderate sampling effort (Blüthgen et al. 2006;

316 Schleuning, Böhning-Gaese, Dehling & Burns 2014). These three descriptors (NODF,

317 Weighted NODF, and H2’) showed high site-specific trends, and while they seem to

t
318 convergence to a narrower range during the five sampling days (Fig. 2), this sampling

ip
319 effort is insufficient to provide a clear pattern and highlight the meaningfulness of these

cr
320 descriptors to compare similar seed dispersal networks. The breakpoints identified by

321 the piecewise regression (i.e. those that maximise the variability explained by the

us
322 regression lines) should also be regarded as conservative given that quasi-asymptotic

323 values are often reached a few days before the estimated breakpoint (Fig. 3).

324
an
From Fig. 3, and considering the sampling effort and data quality trade-off, a
M
325 minimum of eight sampling days are required to adequately sample our seed dispersal

326 network (detection of 67% for seed species, 88% for disperser species, and 71% for
d

327 links). After this day, all descriptors either reach an asymptote or have a highly reduced
te

328 slope and show a consistent pattern of significance with respect to null models. Our

329 results are restricted to seven sites in Portugal and to a particular sampling method
p

330 (identification of seeds on the droppings of mist-netted birds), so further studies testing
ce

331 different sampling strategies and different interaction types, particularly in hyper-

332 diverse ecosystems, are important to test the generality of these results.
Ac

333 An assessment of each descriptor’s quality is critical to derive meaningful network

334 descriptors that reveal true ecological attributes of biological communities and not

335 mathematical artefacts resulting from poor sampling or exclusively from species

336 abundances. Since rare species and links require a higher sampling effort to be recorded,

337 the typically low evenness of species abundances alone can drive some observed

338 properties of the structure of some networks (Vázquez & Aizen 2004). Network

15
Page 15 of 32
339 descriptors can be more or less sensitive to these sampling artefacts (Blüthgen 2010;

340 Blüthgen et al. 2008; Rivera-Hutinel et al. 2012), thus its values may either reflect

341 sample effects or real nature processes. The use of appropriate null models is a way to

342 disentangle if the processes structuring the observed networks are mostly mathematical

t
343 or biological (e.g. Araujo, de Almeida, Cardoso & Corso 2010; Vázquez et al. 2004;

ip
344 Vázquez, Chacoff & Cagnolo 2009; Vázquez, Melián, Williams, Blüthgen, Krasnov et

cr
345 al. 2007). These analyses surpass the scope of this study but are crucial if one wants

346 descriptors that mirror true biological patterns.

us
347 The efficiency of mist nets to capture birds depends on several factors, such as

348 habitat structure, weather or bird behaviour and size (Estades, Escobar, Tomasevic,

349
an
Vukasovic & Páez 2006; Pagen, Thompson & Burhans 2002). Nevertheless, mist-
M
350 netting is likely the best method to construct quantitative seed dispersal networks as it is

351 largely free of observer bias and allows great taxonomic resolution (species level with
d

352 few exceptions), allows the detection of inconspicuous dispersers and interactions,
te

353 makes the study of individual food choice possible through individual marking (e.g.

354 rings), and allows the evaluation of the effect of seed ingestion on its viability (Costa et
p

355 al. 2014; Heleno, Ross, Everard, Memmott & Ramos 2011; Traveset et al. 2014;
ce

356 Traveset, Robertson & Rodríguez-Pérez 2007).

357
Ac

358 Sampling completeness

359 Five sampling-days detected consistently > 71% of the dispersed seed species and >

360 92% of the disperser species across the seven sites. We estimate that during the same

361 period the majority of the links (79%) were also detected, but the sampling

362 completeness was lower (down to 57%) on the most diverse site (Larçã).

16
Page 16 of 32
363 As expected, the thorough detection of links required higher sampling intensity

364 than the detection of the species. This is partly a mathematical consequence, once that

365 new links require new species but species do not necessarily report new links, and partly

366 because some links are likely very rare even if they involve common species. A similar

t
367 effect was already described for pollination networks (Chacoff et al. 2012). However,

ip
368 due to the generally lower size of seed dispersal networks, with lower diversity of

cr
369 dispersers, i.e. animals/plants ratio lower in seed dispersal than in pollination networks,

370 one may expect that seed dispersal studies require less effort to attain a similar sampling

us
371 completeness (Blüthgen, Menzel, Hovestadt, Fiala & Blüthgen 2007; Guimarães,

372 Machado, Aguiar, Jordano, Bascompte et al. 2007).

373
an
In Larçã, 25 sampling days resulted in the detection of a higher richness of seeds,
M
374 dispersers, and links than that estimated by Chao 2 with only five sampling days.

375 However, since this estimator computes the minimum expected richness (Chao 1984;
d

376 Colwell et al. 1994) and the seeds and link richness detected after 25 days lay within the
te

377 confidence interval of that estimated after five days, we consider that this estimator

378 performed relatively well for seeds and links. For dispersers, however, the richness
p

379 detected after 25 days lay above the confidence interval estimated after five sampling
ce

380 days. This, at least apparent, underestimation of Chao 2 might reflect a poor

381 performance of the estimator for species richness due to low sample size (five capture
Ac

382 occasions) (Walther et al. 2005). This apparent underestimation can be largely

383 explained by an increase in the availability of ripe fruits (advancing fruit phenology)

384 and new potential dispersers (bird migration) entering an open community.

385 We estimated that all species and 73% of the links between plants and avian

386 dispersers were detected in Larçã with 25 sampling days. Interestingly, some species

387 pairs that are present in this site, but apparently not interacting, are known to interact

17
Page 17 of 32
388 elsewhere: e.g. Sylvia atricapilla with Daphne gnidium, S. melanocephala with Smilax

389 aspera (Olesen et al. 2011). We believe that the independent study of species and link

390 distributions (e.g. interaction distribution modelling) holds a large potential to

391 understand community assembly rules in ecology.

t
392

ip
393 Concluding remarks

cr
394 Our results suggest that while some descriptors can be accurately estimated with

395 lower sampling effort, at least eight sampling days were needed to accurately describe

us
396 the structure of our seed dispersal network based on the droppings of mist-netted birds.

397 More studies on the effect of sampling intensity on network descriptors are needed to

398
an
allow a generalization of the conclusions on the effort required to get a realistic
M
399 overview of the seed dispersal networks’ structure and to critically assess sampling

400 limitations in previous studies. The reproducibility of the results is a central tenet of
d

401 experimental science, and there is no reason why this should not be applied to
te

402 ecological networks studies, for which the identification of what is a sound sampling

403 effort is paramount.

p

404
ce

405 Appendix A. Supplementary data

406 Supplementary data associated with this article can be found, in the online version, at
Ac

407 XXXXX.

408

409 Acknowledgements

410 We are grateful to the ringers that contributed to the dataset (Sérgio Timóteo, Pedro M.

411 Araújo, António Rosa, Paulo Tenreiro, and Marcial Felgueiras) and to the Coimbra

412 Botanical Garden, particularly to Arménio Matos for his help with seed identification.

18
Page 18 of 32
413 We also thank to two anonymous reviewers and the editor Klaus Hövemeyer whose

414 comments greatly improved the manuscript. JMC, LPS, and RH were supported by

415 Fundação para Ciência e Tecnologia (SFRH/BD/96292/2013, SFRH/BD/77746/2011,

416 and IF/00441/2013, respectively) and RH by the Marie Curie action (FP7-PEOPLE-

t
417 2012-CIG- 321794).

ip
418

cr
419

420 References

us
421 Almeida-Neto, M., Guimarães, P., Guimarães Jr, P.R., Loyola, R.D., & Ulrich, W.

an
422 (2008). A consistent metric for nestedness analysis in ecological systems: reconciling

423 concept and measurement. Oikos, 117, 1227-1239.

M
424 Almeida-Neto, M., & Ulrich, W. (2011). A straightforward computational approach for

425 measuring nestedness using quantitative matrices. Environmental Modelling &

d

426 Software, 26, 173-178.

te

427 Araujo, A.I.L., de Almeida, A.M., Cardoso, M.Z., & Corso, G. (2010). Abundance and

428 nestedness in interaction networks. Ecological Complexity, 7, 494-499.

p

429 Banasek-Richter, C., Cattin, M.-F., & Bersier, L.-F. (2004). Sampling effects and the
ce

430 robustness of quantitative and qualitative food-web descriptors. Journal of Theoretical

431 Biology, 226, 23-32.

Ac

432 Bartomeus, I. (2013). Understanding linkage rules in plant-pollinator networks by using

433 hierarchical models that incorporate pollinator detectability and plant traits. PLoS ONE,

434 8, e69200.

435 Bersier, L.-F., Banasek-Richter, C., & Cattin, M.-F. (2002). Quantitative descriptors of

436 food-web matrices. Ecology, 83, 2394-2407.

19
Page 19 of 32
437 Blüthgen, N. (2010). Why network analysis is often disconnected from community

438 ecology: A critique and an ecologist's guide. Basic and Applied Ecology, 11, 185-195.

439 Blüthgen, N., Fründ, J., Vázquez, D.P., & Menzel, F. (2008). What do interaction

440 network metrics tell us about specialization and biological traits? Ecology, 89, 3387-

t
441 3399.

ip
442 Blüthgen, N., Menzel, F., & Blüthgen, N. (2006). Measuring specialization in species

cr
443 interaction networks. BCM Ecology, 6, 9.

444 Blüthgen, N., Menzel, F., Hovestadt, T., Fiala, B., & Blüthgen, N. (2007).

us
445 Specialization, constraints, and conflicting interests in mutualistic networks. Current

446 Biology, 17, 341-346.

447
an
Carlo, T.A., & Yang, S. (2011). Network models of frugivory and seed dispersal:
M
448 Challenges and opportunities. Acta Oecologica, 37, 619-624.

449 Chacoff, N.P., Vázquez, D.P., Lomáscolo, S.B., Stevani, E.L., Dorado, J., & Padrón, B.
d

450 (2012). Evaluating sampling completeness in a desert plant–pollinator network. Journal

te

451 of Animal Ecology, 81, 190-200.

452 Chao, A. (1984). Nonparametric estimation of the number of classes in a population.

p

453 Scand. J. Statist., 11, 265-270.

ce

454 Colwell, R.K. (2013). EstimateS: Statistical estimation of species richness and shared

455 species from samples. Version 9. User's Guide and application published at:
Ac

456 http://purl.oclc.org/estimates.

457 Colwell, R.K., Chao, A., Gotelli, N.J., Lin, S.-Y., Mao, C.X., Chazdon, R.L., &

458 Longino, J.T. (2012). Models and estimators linking individual-based and sample-based

459 rarefaction, extrapolation and comparison of assemblages. Journal of Plant Ecology, 5,

460 3-21.

20
Page 20 of 32
461 Colwell, R.K., & Coddington, J.A. (1994). Estimating terrestrial biodiversity through

462 extrapolation. Phil. Trans. R. Soc. Lond. B, 345, 101-118.

463 Costa, J.M., Ramos, J.A., da Silva, L.P., Timoteo, S., Araújo, P.M., Felgueiras, M.S.,

464 Rosa, A., Matos, C., Encarnação, P., Tenreiro, P.Q., & Heleno, R.H. (2014).

t
465 Endozoochory largely outweighs epizoochory in migrating passerines. Journal of Avian

ip
466 Biology, 45, 59-64.

cr
467 Darwin, C. (1859). On the origin of species by means of natural selection, or the

468 preservation of favoured races in the struggle for life. London: John Murrey.

us
469 Donatti, C.I., Guimarães, P.R., Galetti, M., Pizo, M.A., Marquitti, F.M.D., & Dirzo, R.

470 (2011). Analysis of a hyper-diverse seed dispersal network: modularity and underlying

471 mechanisms. Ecology Letters, 14, 773–781.

an
M
472 Dormann, C.F., Fründ, J., Blüthgen, N., & Gruber, B. (2009). Indices, Graphs and Null

473 Models: Analyzing Bipartite Ecological Networks. The Open Ecology Journal, 2, 7-24.
d

474 Dormann, C.F., Gruber, B., & Fründ, J. (2008). Introducing the bipartite Package:
te

475 Analysing Ecological Networks. R News.

476 Dormann, C.F., & Strauss, R. (2014). A method for detecting modules in quantitative
p

477 bipartite networks. Methods in Ecology and Evolution, 5, 90-98.

ce

478 Estades, C.F., Escobar, M.A.H., Tomasevic, J.A., Vukasovic, M.A., & Páez, M. (2006).

479 Mist-nets versus point counts in the estimation of forest bird abundances in south-
Ac

480 central Chile. Ornitologia Neotropical, 17, 203-212.

481 Fortuna, M.A., Stouffer, D.B., Olesen, J.M., Jordano, P., Mouillot, D., Krasnov, B.R.,

482 Poulin, R., & Bascompte, J. (2010). Nestedness versus modularity in ecological

483 networks: two sides of the same coin? Journal of Animal Ecology, 79, 811-817.

21
Page 21 of 32
484 González-Varo, J.P., Arroyo, J.M., & Jordano, P. (2014). Who dispersed the seeds? The

485 use of DNA barcoding in frugivory and seed dispersal studies. Methods in Ecology &

486 Evolution, 5, 806-814.

487 Guimarães, P.R., Machado, G., Aguiar, M.A.M.d., Jordano, P., Bascompte, J., Pinheiro,

t
488 A., & Reis, S.F.d. (2007). Build-up mechanisms determining the topology of

ip
489 mutualistic networks. Journal of Theoretical Biology, 249, 181-189.

cr
490 Hegland, S.J., Dunne, J., Nielsen, A., & Memmott, J. (2010). How to monitor

491 ecological communities cost-efficiently: The example of plant–pollinator networks.

us
492 Biological Conservation, 143, 2092-2101.

493 Heleno, R., Devoto, M., & Pocock, M. (2012). Connectance of species interaction

494
an
networks and conservation value: Is it any good to be well connected? Ecological
M
495 Indicators, 14, 7-10.

496 Heleno, R., Garcia, C., Jordano, P., Traveset, A., Gómez, J.M., Blüthgen, N., Memmott,
d

497 J., Moora, M., Cerdeira, J., Rodríguez-Echeverría, S., Freitas, H., & Olesen, J.M.
te

498 (2014). Ecological networks: delving into the architecture of biodiversity. Biology

499 Letters, 10.

p

500 Heleno, R., Olesen, J.M., Nogales, M., Vargas, P., & Traveset, A. (2013). Seed-
ce

501 dispersal networks in the Galápagos and the consequences of plant invasions.

502 Proceedings of the Royal Society B-Biological Sciences, 280, 20122112.

Ac

503 Heleno, R.H., Lacerda, I., Ramos, J.A., & Memmott, J. (2010). Evaluation of

504 restoration effectiveness: community response to the removal of alien plants. Ecological

505 Applications, 20, 1191-1203.

506 Heleno, R.H., Ross, G., Everard, A., Memmott, J., & Ramos, J.A. (2011). The role of

507 avian ‘seed predators’ as seed dispersers. Ibis, 153, 199-203.

22
Page 22 of 32
508 Herrera, C.M. (1984). A study of avian frugivores, bird-dispersed plants, and their

509 interaction in mediterranean scrublands. Ecological Monographs, 54, 1-23.

510 James, A., Pitchford, J.W., & Plank, M.J. (2012). Disentangling nestedness from

511 models of ecological complexity. Nature, 487, 227-230.

t
512 Jordano, P. (1987). Patterns of mutualistic interactions in pollination and seed dispersal:

ip
513 connectance, dependence asymmetries, and coevolution. The American Naturalist, 129,

cr
514 657-677.

515 Mello, M., Marquitti, F., Guimarães, P.R., Kalko, E., Jordano, P., & Aguiar, M.d.

us
516 (2011). The missing part of seed dispersal networks: structure and robustness of bat-

517 fruit interactions. PloS one, 6, e17395.

518
an
Muggeo, V.M.R. (2008). segmented: an R package to fit regression models with
M
519 broken-line relationships. R News, 8, 20-25.

520 Oksanen, J., Blanchet, F.G., Kindt, R., Legendre, P., Minchin, P.R., O'Hara, R.B.,
d

521 Simpson, G.L., Solymos, P., Stevens, M.H.H., & Wagner, H. (2015). vegan:
te

522 Community Ecology Package. R package version 2.2-1.

523 Olesen, J.M., Bascompte, J., Dupont, Y.L., Elberling, H., Rasmussen, C., & Jordano, P.
p

524 (2011). Missing and forbidden links in mutualistic networks. Proc. R. Soc. B, 278, 725-
ce

525 732.

526 Olesen, J.M., Bascompte, J., Dupont, Y.L., & Jordano, P. (2007). The modularity of
Ac

527 pollination networks. PNAS, 104, 19891-19896.

528 Olesen, J.M., Dupont, Y.L., Hagen, M., Rasmussen, C., & Trøjelsgaard, K. (2012).

529 Structure and dynamics of pollination networks: the past, present, and future. In: S.

530 Patiny (Ed.), Evolution of Plant-Pollinator Relationships (pp. 374-391). Belgium:

531 Cambridge University Press.

23
Page 23 of 32
532 Olito, C., & Fox, J.W. (2015). Species traits and abundances predict metrics of plant–

533 pollinator network structure, but not pairwise interactions. Oikos, 124, 428-436.

534 Pagen, R.W., Thompson, I., Frank R., & Burhans, D.E. (2002). A comparison of point-

535 count and mist-net detections of songbirds by habitat and time-of-season. Journal of

t
536 Field Ornithology, 73, 53-59.

ip
537 R Development Core Team. (2014). R: A language and environment for statistical

cr
538 computing. Vienna, Austria: R Foundation for Statistical Computing.

539 Rivera-Hutinel, A., Bustamante, R.O., Marín, V.H., & Medel, R. (2012). Effects of

us
540 sampling completeness on the structure of plant–pollinator networks. Ecology, 93,

541 1593-1603.

542
an
Schleuning, M., Böhning-Gaese, K., Dehling, D.M., & Burns, K.C. (2014). At a loss for
M
543 birds: insularity increases asymmetry in seed-dispersal networks. Global Ecology and

544 Biogeography, 23, 385-394.

d

545 Traveset, A., Heleno, R., Chamorro, S., Vargas, P., McMullen, C.K., Castro-Urgal, R.,
te

546 Nogales, M., Herrera, H.W., & Olesen, J.M. (2013). Invaders of pollination networks in

547 the Galápagos Islands: emergence of novel communities. Proceedings of the Royal
p

548 Society B: Biological Sciences, 280.

ce

549 Traveset, A., Heleno, R.H., & Nogales, M. (2014). The ecology of seed dispersal. In:

550 R.S. Gallagher (Ed.), Seeds: The ecology of regeneration in plant communities (pp. 62-
Ac

551 93). Oxfordshire, UK: CABI.

552 Traveset, A., Robertson, A.W., & Rodríguez-Pérez, J. (2007). A review on the role of

553 endozoochory on seed germination. In: A.J. Dennis, E.W. Schupp, R.A. Green, & D.A.

554 Westcott (Eds.), Seed dispersal: theory and its application in a changing world. (pp. 78-

555 103). Wallingford, UK: CABI Publishing.

24
Page 24 of 32
556 Tylianakis, J.M., Tscharntke, T., & Lewis, O.T. (2007). Habitat modification alters the

557 structure of tropical host-parasitoid food webs. Nature, 445, 202-205.

558 Vázquez, D.P., & Aizen, M.A. (2004). Asymmetric specialization: a pervasice feature

559 of plant-pollinator interactions. Ecology, 85, 1251-1257.

t
560 Vázquez, D.P., Chacoff, N.P., & Cagnolo, L. (2009). Evaluating multiple determinants

ip
561 of the structure of plant–animal mutualistic networks. Ecology, 90, 2039-2046.

cr
562 Vázquez, D.P., Melián, C.J., Williams, N.M., Blüthgen, N., Krasnov, B.R., & Poulin, R.

563 (2007). Species abundance and asymmetric interaction strength in ecological networks.

us
564 Oikos, 116, 1120-1127.

565 Walther, B.A., & Moore, J.L. (2005). The concepts of bias, precision and accuracy, and

566
an
their use in testing the performance of species richness estimators, with a literature
M
567 review of estimator performance. Ecography, 28, 815-829.

568
d

569
te

570
p
ce
Ac

25
Page 25 of 32
570

571

572

573 Table 1. Percentage of disperser species, seed species, and links detected during five

t
574 consecutive sampling days across the seven sites. Estimated values of richness were

ip
575 computed using the Chao 2 estimator. Estimates marked with * were computed based

cr
576 on the classic formula of Chao 2 because the coefficient of variation for incidence

577 distribution > 0.5. O – observed richness, E (95% C.I.) – estimated richness (95%

us
578 confidence intervals), % - sampling completeness (=O/E*100).

an
Site # days Seeds Dispersers Links
O E (95% C.I.) % O E (95% C.I.) % O E (95% C.I.) %
A - Larçã 25 12 12 (12-16) 100 8 8* (8-21) 100 38 52 (42-93) 73
A - Larçã 5 7 7 (7-12) 100 6 6 (6-7) 100 17 30 (20-68) 57
B - Atenor 5 5 7* (5-24) 71 8 8 (8-15) 100 13 18 (14-40) 72
C - Nozelos 5 7 8 (7-19) 88 11 12 (11-21) 92 18 23 (19-44) 79
M
D - Santa Maria de Aguiar dam 5 7 7 (7-18) 100 11 12 (11-24) 92 25 29 (26-45) 86
E - Achete 5 10 11 (10-22) 91 11 11 (11-15) 100 33 41 (35-62) 81
F - Freixo do Meio 5 2 2* (2-3) 100 5 5 (5-6) 100 6 6 (6-14) 100
G - Bensafrim 5 4 4 (4-5) 100 6 6 (6-14) 100 10 13 (10-28) 77
579
d
p te
ce
Ac

26
Page 26 of 32
579

580 Table 2. Number of additional sampling days needed to detect 80%, 90%, and 100% of

581 the diversity estimated by the Chao 2 estimator.

582

t
Target sampling completeness

ip
Network
Sampling period 80% 90% 100%
descriptor
5 days (mean of 7 sites) 0 0 2
Seed species

cr
25 days (1 site) 0 0 0
5 days (mean of 7 sites) 0 0 1
Disperser species
25 days (1 site) 0 0 0
5 days (mean of 7 sites) 1 2 5

us
Link diversity
25 days (1 site) 8 25 47
583

an
M
d
p te
ce
Ac

27
Page 27 of 32
583

584 Figure captions

585

586 Fig. 1. Location of the study sites. A) Larçã, Souselas; B) Atenor, Miranda do Douro;

t
587 C) Nozelos, Torre de Moncorvo; D) Santa Maria de Aguiar dam, Figueira de Castelo

ip
588 Rodrigo; E) Casais da Estrada, Achete; F) Herdade de Freixo do Meio, Foros de Vale

cr
589 Figueira; G) Bensafrim, Lagos. Site A was sampled for 25 consecutive days; all other

590 sites were sampled for five days.

us
591

592 Fig. 2. Effect of sampling effort (measured as the number of days) on the Connectance

593

594 an
(A), Weighted Connectance (B), Nestedness (NODF (C) and Weighted NODF (D)),

Modularity (Qualitative (E) and Quantitative (F)), Specialization (H2’ (H)), richness of
M
595 seeds (I), dispersers (J), and links (G) detected in seven seed-dispersal networks

sampled simultaneously during five consecutive days throughout Portugal. Missing

d
596
te

597 values correspond to days/sites where the network was too small to allow the

598 calculation of those network descriptors. For the definition of network descriptors see
p

599 the Materials and methods section.

ce

600

601 Fig. 3. Effect of the sampling effort (measured as the number of days) on the
Ac

602 Connectance (A), Weighted Connectance (B), Nestedness (NODF (C) and Weighted

603 NODF (D)), Modularity (Qualitative and Quantitative), Specialization (H2’ (H)),

604 richness of seeds (I), dispersers (J), and links (G) detected on the seed-dispersal network

605 at Larçã (Fig. 1). The dashed lines indicate the 95% confidence interval of 1000 runs

606 (100 for Modularity) of a null model based exclusively on species abundances. The

607 fitted lines represent the theoretical values estimated by piecewise linear regressions,

28
Page 28 of 32
608 except for Connectance, which was fitted with linear regression, with the grey area

609 indicating the 95% confidence interval. The dashed vertical line at day 8 indicates the

610 minimum adequate sampling effort suggested for this site, while the full line at day 5

611 indicates the duration of the sampling period on the other sites. For the definition of

t
612 network descriptors see the Materials and methods.

ip
613

cr
614

615

us
616

617

618
an
M
619

620
d

621
te

622

623
p

624
ce

625

626
Ac

627

628

629

630

631

632

29
Page 29 of 32
633

634

635

t
ip
cr
us
an
M
d
p te
ce

636
Ac

637 Fig. 1

638
639

30
Page 30 of 32
639

t
ip
cr
us
an
M
d
p te
ce
Ac

640

641 Fig. 2

642

643

31
Page 31 of 32
 t
ip
cr
us
an
M
d
p te
ce

644
Ac

645 Fig. 3

646

647

32
Page 32 of 32