OIKOS 107: 610
/618, 2004

Modeling vertical beta-diversity in tropical butterfly communities

Thomas R. Walla, Steinar Engen, Philip J. DeVries and Russell Lande

Walla, T. R., Engen, S, De Vries, P. J. and Lande, R. 2004. Modeling vertical beta-
diversity in tropical butterfly communities.
/ Oikos 107: 610
/618.

We present a novel analytical method for assessing spatial and temporal structure in
community samples that is useful for comparing large data-sets that include species
abundance data. The model assumes that species numbers in two samples are drawn
from a bi-variate Poisson log-normal species abundance distribution and parameters
from the fitted distribution are estimated to assess community structure. We assessed
three tropical butterfly data-sets for spatial structure in the vertical dimension, and
tested for changes in structure as a result of temporal variance, disturbance regimes,
and geographic location. Our results indicate that the vertical dimension is a major
structural component in tropical forest butterfly communities that varies little through
time and is not measurably affected by small-scale disturbances. However, there is
evidence that the degree of vertical structure may vary among geographic regions.
These results are discussed in terms of the mechanisms maintaining vertical structure,
and the implications of changes in forest architecture on butterfly communities.

T. R. Walla, Dept. of Biology, Mesa State College, 1100 North Ave, Grand Junction, CO
81501, USA.
/ S. Engen, Dept. of Mathematical Sciences, Norwegian Univ. of Science
and Technology, NO-7491, Norway.
/ P. J. DeVries, Dept. Biological Sciences, Univ. of
New Orleans, New Orleans, LA 70148, USA ([email protected]).
/ R. Lande, Dept. of
Biology, Univ. of California at San Diego, La Jolla, CA 92093, USA.

The distribution of species in time and space is vital for abundance is critical for understanding community
describing ecological communities and developing hy- dynamics.
potheses to explain the evolution and maintenance of Species respond to the environment in different ways,
species diversity (Gleason 1926, Elton 1966, MacArthur and spatial scaling or habitat specialization may vary
1972, Whittaker 1975, Wiens 1984, Tilman 1994, Brown among species in communities. As such, changes in
1995). Environmental heterogeneity and habitat com- structure among locations within and between commu-
plexity are important determinants of community com- nities are generally continuous graded shifts in species
position, and species responses to environmental composition as opposed to sharply defined boundaries
variation provide insight into ecological and evolution- (Pielou 1975, Wilson and Mohler 1983, Williams et al.
ary processes such as speciation, adaptation, dispersal, 1999). Measuring spatial variation is challenging because
and colonization (Hubbell 2001). Recent investigations most communities are comprised of differing numbers of
have also emphasized the influence of environmental related species that vary in abundance, and include many
heterogeneity on community composition as a critical rare species (Preston 1948, Colwell and Coddington
factor affecting the measurement of species diversity, 1994, Novotny and Basset 2000), and evaluation of
species area relationships, and the relationship between spatial structure requires estimates of the number of
local and regional diversity patterns (Srivastava 1999, individuals at different locations. Hence, rare species are
Loreau 2000, Gering and Crist 2002, Wagner and Wildi often excluded from parametric analyses, and species by
2002). Therefore, spatial variation in relative species species assessments. Non-parametric tests for homo-

Accepted 13 May 2004

Copyright # OIKOS 2004
ISSN 0030-1299

610 OIKOS 107:3 (2004)

geneity of sample composition (chi-square) may include change under varying degrees of natural and anthro-
all species and ascertain the significance of structure, but pogenic disturbance. Third, we assess the generality of
they do not indicate the degree of structure present. vertical structure among geographic regions by compar-
Beta-diversity (MacArthur 1965, Whittaker 1975), which ing the vertical structure of rainforest butterfly commu-
is related to measures of community similarity (Lande nities from three different sites, two in eastern Ecuador,
1996), describes the difference in species diversity and one in Costa Rica. Finally, we discuss the results of
observed between samples. When measured as a function our analyses in the context of ecological and evolu-
of species richness and relative species abundance, beta- tionary patterns of community structure, and conserva-
diversity can be used to measure community structure in tion.
space or time. However, no single measure of beta-
diversity is universally accepted and few methods sup-
port statistical evaluation (Magurran 1988, Vellend
Methods
2001, Gering et al. 2003, Lande et al. 2003).
The distribution of species at particular heights within Study communities
forest communities, referred to here as vertical structure,
is a classic phenomena in ecology (Allee 1926, Bates Adult butterflies in the family Nymphalidae that are
1944, MacArthur 1958, 1965, Elton 1973, Erwin 1983, attracted to and feed on the juices of rotting fruits
Stork 1988, Basset et al. 1992, Wolda 1992, Richards constitute a feeding guild known as fruit-feeding nym-
phalids (DeVries and Walla 2001). Data-sets on fruit-
1996). Vertical structure among samples is manifested as
feeding nymphalids from three rainforest community
high beta-diversity in the vertical spatial dimension, and
studies were used in our analyses: two from the upper
accounts for a significant component of species diversity
Amazon of Ecuador, La Selva Lodge (DeVries and
in neotropical fruit-feeding nymphalid butterfly commu-
Walla 2001) and Jatun Sacha (DeVries et al. 1997), and
nities (DeVries et al. 1997, 1999, DeVries and Walla
one from the Atlantic lowlands in Costa Rica, Finca La
2001). Vertical structure has also been shown among
Selva (DeVries 1988). All three sites are considered
mimetic groups of neotropical butterflies that are not
evergreen wet (Costa Rica) or superwet (Ecuador)
attracted to fruit-baited traps (Papageorgis 1975, Becca-
rainforests (Richards 1996), and have characteristic wet
loni 1997). Thus accounting for vertical structure is an
and dry seasons with 3
/5 m of rain per year.
essential underlying concern for studies seeking to
Similar sampling methods were applied in all three
measure diversity in natural communities. Nevertheless,
communities, where traps were placed in replicated pairs,
due to limitations in analytical techniques and avail-
with one in the canopy (15
/25 m above ground) and one
ability of appropriate data-sets, vertical structure among
in the understory (1 m above ground), and baited with
butterfly species (and other insects) has not been fermented bananas (DeVries 1988, DeVries and Walla
investigated in terms of how spatial or temporal variance 2001).
may influence them, or under what conditions vertical At La Selva Lodge the traps were sampled monthly
structure becomes inconsequential. for five years from August 1994
/July 1999 in natural
Here we present a method for modeling structure in forest habitat. The Jatun Sacha study was designed to
species rich communities that is analogous to measure- sample equally from four different habitat types: pri-
ment of beta-diversity. We model structure between a mary, secondary, hi-graded forest, and forest edge
pair of community samples as a parametric distribution adjacent to pasture. At Jatun Sacha traps were sampled
among species with the expected frequency of species in monthly for one year from August 1992
/August 1993.
the canopy and understory as the variables of interest, At Finca La Selva five canopy and five understory traps
and we estimate the model parameters by fitting a bi- were sampled daily for approximately 2 months from
variate Poisson log-normal species abundance distribu- October 1979
/December 1979.
tion to two community samples. Unlike most techniques,
our novel method of evaluating beta-diversity between
community samples makes full use of the information in
samples without excluding species on the basis of sample
The model
size, and it allows sample pairs to be evaluated statisti- We model the structure observed in samples from two
cally for differences in the degree of vertical structure points in space or time, and apply the model to canopy
among communities. and understory samples. The vertical structure of each
After deriving the model we apply it to several large species is reflected in the proportion of individuals in
neotropical butterfly data-sets to understand how ver- that species observed in the canopy. For convenience we
tical beta-diversity varies in space and time. We first use the expected frequency of a species in the canopy p
assess the long-term stability of vertical structure in a as a proportion of the pooled sample to measure vertical
community sampled from intact rainforest. We then test structure and model the distribution of p among species
the null hypothesis that vertical structure does not to describe vertical structure at the community level.

OIKOS 107:3 (2004) 611

 A derivation of the model used in our analyses is
presented in the appendix. To summarize, we assume
samples from each forest height are drawn from a
Poisson log-normal species abundance distribution.
The abundance pair for each species (canopy and
understory) may then be considered drawn from the
bi-variate Poisson log-normal species abundance distri-
bution where every species has a probability of occurring
in the canopy (p) versus the understory (1/p). Max-
imum likelihood was then used to estimate the para-
meters of the fitted bi-variate distribution, allowing
some species to be absent from either the understory
and/or canopy samples.
The number of individuals observed in the canopy for Fig. 1. Plot of the hypothetical p distributions where correla-
a given species may be considered as the number of tion r is varied from maximum to minimum possible values.
successes in a bi-nomial sampling experiment with The four other parameters in the model are held constant.
When r is high the p distribution is a bell shaped curve. When r
parameters N and p, where N is the number of trials is low, the mass of the distribution is concentrated in the tails.
(individuals sampled) and p is the probability of a Vertical gray lines represent selected a level (a/0.2).
success (capture in the canopy). The observed frequency
of each species in the canopy (number of individuals in for each of five parameters as well as the distribution
the canopy/number of individuals sampled from that of p.
species) provides an estimate of p for each species. It is not required that sampling effort be equal in the
However, the distribution of p among species, denoted two samples compared, since this will not affect the form
g(p), is obtained from the bi-variate distribution of of the species abundance distributions. However it is far
abundances among species as shown in the appendix. more informative if both samples are large enough to
Here g(p) is a function of the five parameters (means, accurately model species abundance distributions; small
variances, and correlation) from the fitted joint canopy sample sizes result in large uncertainty.
and understory species abundance distribution. A simplified representation of differences in the form
The model distribution g(p) among species is a of the p distribution between sample pairs from different
probability density function expressing the probability communities is performed by partitioning the distribu-
that a species chosen at random from the community tion into sections corresponding to three categories of
(independent of species abundance) will exhibit a given vertical stratification: understory species, unstratified
frequency in the canopy. If the model shows the mass of species found at both heights, and canopy species. The
the distribution to be concentrated near 0.5, there is a cutoffs for the categories (referred to here as a) can be
high probability that a randomly selected species will be determined based on the biology of the study organisms
evenly distributed between canopy and understory; in or the interest of the researcher. Here we consider species
other words, species specialized to one height class are with 20% or less of their abundance in canopy samples
rare. However, if the mass of the distribution is are vertically stratified understory species. Thus an a
concentrated in the tails, it forms a U shape, and there level of 0.2 was chosen and applied to both tails of the
is a high probability that a randomly selected species is distribution. Expected p values between 0.2 and 0.8 in
found mostly at one height. Examples of potential forms the middle section of the distribution correspond to
of the p distribution are presented in Fig. 1 where all unstratified species found frequently in both canopy and
parameters are held constant with the exception of the understory samples. Values less than 0.2 correspond to
correlation r between samples. All five parameters of the understory species and values greater then 0.8 corre-
bi-variate species abundance distribution affect the spond to canopy species. A species chosen at random
distribution of p, thus the example is a simplification. from the community has a probability of being in each of
Confidence limits for the model are generated through the three stratification categories that are equivalent to
parametric bootstrapping; that is, repeatedly simulating the areas under each of the three a-defined sections of
new sets of data from the estimated model and the distribution (Fig. 1). The results were not sensitive to
recalculating the parameter estimates from each simu- a, since values of a from 0.1 to 0.3 yielded similar results.
lated data-set or bootstrap sample to approximate the To compare the form of the p distribution among
sampling distribution of the parameter estimates. For communities graphically, the areas under each of the
each bootstrap sample the five parameters are estimated three curve sections are considered as vectors in an
by maximum likelihood and the p distribution is equilateral triangle with axes corresponding to vertices
modeled; the distribution of parameter estimates among labeled Canopy, Understory, and Both located as a
the bootstrap samples provides measures of uncertainty single point in the triangle. If areas under each section of

612 OIKOS 107:3 (2004)

the p distribution are equivalent, the point will plot in levels of anthropogenic disturbance ranging from un-
the center of the triangle. If a randomly chosen species is disturbed forest to pasture edge.
more likely to be a ‘‘canopy species’’ then it will be closer To evaluate the consistency of vertical structure
to the Canopy vertex. among geographic regions where forest structure is
Conversely when the mass of the distribution is similar, we compared the vertical structure at La Selva
concentrated at 0.5, the point will be closer to the Lodge (Ecuador), Jatun Sacha (Ecuador), and Finca La
Both vertex, and when the mass of the distribution is Selva (Costa Rica).
concentrated near p /0, the point will be closer to the
Understory vertex (Fig. 2). The confidence region for the
bootstrap replicates is plotted on the triangle to indicate
the degree of uncertainty.
Results
Sample size and number of species sampled from each
community are reported in Table 1. The high species
richness from the Ecuadorian samples was characteristic
Comparisons of most Amazonian forest sites, and the Costa Rica site
was comparatively species poor due to its geographical
To test the hypothesis that vertical structure was location, shorter sampling period, and small sample size.
temporally stable, we modeled five years of monitoring Fitting the La Selva Lodge samples to the model
data from La Selva Lodge. The degree of stratification resulted in a U-shaped g(p) distribution that indicated a
modeled for each year was then compared to the five strongly stratified species community (Fig. 3). We found
year combined sample using parametric bootstrapping no evidence to reject the null hypothesis that the
to estimate confidence intervals for the parameter distribution of vertical structure among species was
estimates. consistent across all five years. The vertical structure
Effects of natural tree fall disturbances on vertical was similar among years despite considerable annual
structure were tested by sub-dividing the La Selva Lodge variance in overall abundance of butterflies. However,
trap samples into those that were adjacent to large tree the variance of the species abundance distributions in the
fall gaps for more than one year during the study period canopy and the understory increased through time, and
and those that were never adjacent to tree fall gaps. the correlation r decreased after the first year (Table 2).
These two groups were then assessed for differences in These observations reflect subtle changes in the distribu-
vertical structure. tion of abundance among species at both forest heights
To test for effects of human-induced disturbance on that did not significantly affect the vertical structure of
butterfly vertical structure, we compared samples from the community (Fig. 4).
the four habitats at Jatun Sacha representing different We found no evidence that natural disturbance
affected vertical distribution at La Selva Lodge. Traps
from within closed canopy forest and traps in tree fall
gaps showed no significant differences in vertical struc-
ture (Fig. 4, Table 2).
The four habitat types at Jatun Sacha showed no
significant differences in vertical structure (Fig. 4). Only
the edge habitat had distinctive parameter estimates,
showing a comparatively larger variance of the unders-
tory fitted species abundance distribution, and a greater
correlation, r (Table 2).
Comparison among geographic regions suggested that
vertical structure may vary regionally. Although Jatun
Sacha and La Selva Lodge had similar vertical structure,

Table 1. Summary of data-sets used for analysis of vertical

structure.

LSL JS FLS
Ecuador Ecuador Costa Rica
Fig. 2. Triangular plotting surface showing the vertical struc-
ture present in the hypothetical communities from Fig. 1. Canopy sample size 5840 1173 129
Canopy species richness 87 86 31
Higher correlations between canopy and understory abundance Understory sample size 6021 5517 53
distributions will move the community closer to the Both vertex. Understory species richness 96 105 22
Strongly stratified communities will plot near the center of the Total sample size 11 861 6690 182
triangle. Greater numbers of Canopy or Understory species will Total species richness 128 130 46
shift the community closer to one of the lower vertices.

OIKOS 107:3 (2004) 613

 Table 2. Parameter estimates for all communities and community subsets. LSL /La Selva Ecuador, JS/Jatun Sacha Ecuador, mc /mean of canopy, mu /mean understory, s2c /variance

/0.29
/0.10
/0.13
/0.01

/0.18
/0.11
/0.13
/0.20
/0.35
/1.00
0.00
0.10
0.08

0.08
/

0.39
0.51
0.44
0.49
0.52
0.54
0.56
0.40
0.40
0.67
0.50
0.47
0.41
0.32
canopy, s2u /variance understory, r /correlation between canopy and understory / and / signs denote 95% confidence intervals based on 500 bootstrap replicates.

/

/0.22
0.08
0.25
0.25
0.25
0.31
0.35
0.39
0.23
0.16
0.44
0.24
0.19
0.10
r

3.14
3.65
3.90
5.17
5.27
6.21
6.24
4.36
4.73
5.22
3.54
3.01
3.30
0.52
/

10.22
12.20
12.25
13.06
13.01
11.65
13.85
10.16
9.25
9.81

9.55

8.21
9.47
4.68
Fig. 3. Plot of the probability density function of the p

/
distribution for La Selva Lodge total community. The concen-
tration of mass at the tails of the distribution shows consider-
able vertical structure among species. The y-axis is truncated at

5.64
6.63
7.14
6.63
8.90
9.33
9.95
8.55
7.13
9.67
6.84
5.51
6.44
2.30
s2u
30.

the La Selva CR site showed a significantly different

0.25
4.73
5.00
6.13
6.10
6.91
6.42
6.10
3.10
2.02
2.53
2.68
2.95
0.97
/
pattern (Fig. 5, Table 2). In Costa Rica the distribution
of p showed more mass under the canopy tail of the
curve compared to the understory. In other words, a

9.75

7.07
6.77
8.11
8.38
9.07
5.08
11.84
11.17
12.92
12.77
13.70
13.01
12.45
/
species chosen at random in Costa Rica had a higher
probability of being a canopy species relative to the
Ecuadorian sites. Thus Costa Rican fauna had a

10.44
10.78
10.92
11.14
7.37
8.55
9.38
8.55

4.84
4.20
4.94
5.25
5.96
2.64
s2c

significantly greater proportion of canopy species than

the other two sites.
/1.84
/1.45
/2.34
/2.04
/2.10
/0.78
/1.57
/2.04
/0.06
/2.19
/2.19
/1.84
/2.48
/2.57
/

Discussion
Natural patterns of spatial structure in communities
/0.12

/0.05

/0.23
/0.39
0.17
0.55
0.12
0.08
0.99
0.36
1.37
0.45
0.21
have been described in terms of alpha and beta-diversity, 0.22
/

where alpha diversity describes the average diversity

within homogenous sites and beta-diversity describes
diversity between sites (MacArthur 1965, Whittaker
/0.74
/0.39
/1.39
/0.39
/1.42

/0.82
/1.31
/0.86
/0.91
/0.74
/1.40
/1.26
0.13

0.71
mu

1975, Wilson and Mohler 1983, Lande 1996). Ecologists

have not previously developed methods of measuring
spatial changes in relative abundance of species that
simultaneously account for unequal sample sizes, sam-
/2.53
/2.11
/3.13
/2.15
/2.18
/1.22
/1.95
/2.77
/0.91
/2.61
/2.87
/2.80
/3.02
/1.84
/

pling error, and the range of abundances among species

including large proportions of singletons (reviewed
Magurran 1988). The primary contribution of this
work is the presentation of a novel method for analyzing
/0.27
/0.02
/1.05
/0.15
/0.18

/0.06
/0.87
/0.52
/0.79
/0.66
/0.73
0.38

0.33

0.29
/

differences in the relative abundance of species among

samples. We used a community-level analysis and
derived a model that maximally utilizes the information
in the data. The model provides a description of the
/1.28
/1.08
/2.47
/1.08
/2.06
/0.45
/1.28
/2.23
/0.22
/1.43
/1.72
/1.66
/1.90
/0.51
mc

structure between two samples that may be interpreted

as a measure of beta-diversity, independent of sample
size, and may be generalized to evaluate temporal and
spatial community structure. The model also can be used
to generate measures of statistical uncertainty in the
Costa Rica
JS primary
LSL forest

JS higrade
JS second
LSL total
LSL gaps

results. Given these properties, we expect it may prove

LSL Y1
LSL Y2
LSL Y3
LSL Y4
LSL Y5

JS total
JS edge
Sample

useful for quantifying the magnitude of differences in

species abundances among areas separated by ecological

614 OIKOS 107:3 (2004)

 pattern confirms that vertical structure is an important
component of tropical forest butterfly communities
(DeVries and Walla 2001).
Despite dynamic changes in community abundance
among years presumably caused by disturbance to the
forest by large storms at the La Selva Lodge (DeVries
and Walla 2001), we found that the overall vertical
structure of the butterfly community remained unaf-
fected. Our model of g(p) distribution for each year was
similar in form to the total community and showed a
large proportion of the distribution concentrated at the
tails (Fig. 3). We also found that natural small-scale
disturbance due to tree falls did not disrupt vertical
structure within the intact forests. No significant differ-
ence in structure was observed between samples taken
adjacent to tree fall gaps and those within the forest (Fig.
4, Table 2). This contrasts with observations suggesting
Fig. 4. Vertical structure of communities in different distur-
bance regimes. Dashed ellipse includes 95% of the 500 bootstrap that canopy species may show a propensity to fly at
estimates for the total La Selva Lodge community. None of the ground level in forest gaps (DeVries 1988). Our results
communities plotted exhibit significantly different vertical suggest small-scale disturbances are not sufficiently
structures compared to other Ecuadorian communities.
frequent to significantly alter the overall community
vertical structure.
or environmental gradients. This approach has been At Jatun Sacha where four habitats represented
generalized by Engen et al. (2002a), and Lande et al. varying degrees of anthropogenic disturbance we found
(2003). that vertical structure was equivalent in all habitats.
This study provides the first comparisons of vertical Even at the forest edge adjacent to the cattle pasture
structure among tropical communities. We found sur- there was no evidence of a change in the distribution of p
prisingly robust vertical structure among fruit-feeding among species. The raw data showed that some common
nymphalid butterfly communities; neither small-scale canopy species were more abundant in the understory at
disturbance nor temporal fluctuations in abundance the forest edge, but concurrently some rare species were
influenced vertical structure. The consistency of this more frequent in the canopy than in less disturbed
habitats. The mean and variance of the p distribution for
the edge sample is similar to the other three habitats,
thus generating a non-significant difference in vertical
structure (Table 2, Fig. 4).
Current hypotheses explaining vertical structure
in tropical butterflies cite mechanisms associated
with potential differences in predator communities
(Papageorgis 1975), light levels (DeVries 1988), and
host plant availability at different heights in the forest
(Beccaloni 1997). Our results suggest that vertical
stratification is not likely to be maintained solely by
differential light levels, which are known to be associated
with canopy cover (Richards 1996). If this were the case
we would have expected to see a significant number of
canopy butterflies captured at the forest floor near gaps
and along the forest edge (Table 2, Fig. 4). We cannot
dismiss the role of predators and host plant availability
since either factor could potentially be maintained in the
face of changing habitat structure. It is also possible that
Fig. 5. Community comparison of vertical structure among the
three locations. Dashed ellipse includes 95% of the 500 boot- under both natural and anthropogenic disturbance,
strap estimates for the total Finca La Selva community. The individuals descend to the forest floor in gaps but they
results indicate significant differences between the Finca La are simply not attracted to baits in this alternate habitat.
Selva site and the Ecuadorian sites. Finca La Selva has more
species in the canopy compared to the understory and slightly In any case, we present strong evidence that fruit-feeding
less vertical stratification. behavior shows a strong relationship with height in the

OIKOS 107:3 (2004) 615

forest, regardless of the forest maturity or uniformity of 1996, Engen et al. 2002a, b, Diserud and Engen 2002,
canopy structure. Lande et al. 2003). This study indicates that while
Community disturbance or pollution may increase the vertical patterns in species distributions are robust to
variance of the species abundance distribution by changes in forest architecture, the underlying compo-
releasing some species to high abundance while others nents of the canopy and understory communities may
become or remain relatively rare (Patrick et al. 1954, vary among regions. Future comparative work using
Gray and Mirza 1979, Ugland and Gray 1982). This similar community-level analyses may reveal further
pattern was noted in the Jatun Sacha butterfly commu- patterns among ecoregions that provide a broader
nity (DeVries et al. 1999) where there was a significant understanding of community structure useful for testing
increase in the variance of the understory species ecological theory and its application to conservation.
abundance distribution at the forest edge. However, we
note that in the same study there was no measurable Acknowledgements
/ This work was possible due to the field
effect at the other levels of disturbance (i.e. higrade and assistance of C. Dingle, C. Dunn, C. Funk, H. Greeney, R.
Guerra, R. Hill, G. Hualinga, Jose Hualinga and M. Lysinger.
secondary habitats). We suggest that changes in com- Thanks to Eric Schwartz, creator of La Selva Lodge, for
munity structure that are reflected by an increased supporting the biologists involved in this project. We thank N.
variance of the species abundance distribution may Martin and C. M. Penz for comments on drafts of this
manuscript. We acknowledge the Population Biology
only be evident in cases where disturbance is particularly Foundation, National Science Foundation (DEB-9806779,
severe. Although this investigation does not address DEB-0096241, DEB-0313653), the Norwegian Research
the impacts of broad scale habitat disturbance as Council, and the National Geographic Society for financial
support. The Univ. of California San Diego, Univ. of
induced by clear-cutting and forest conversion to Trondheim, Univ. of Leiden, and Milwaukee Public Museum
agriculture, it shows that intermediate disturbance is provided logistical support. We dedicate this paper to the
unlikely to affect the degree of vertical structure in diversity of work produced by S. Gimelfarb and S. Lacy.
butterfly communities.
It is well known that geographic location has a large
influence on the richness and structure of butterfly
communities. Analysis of a limited data-set from Costa References
Rica showed that butterflies exhibit a vertical distribu- Allee, W. C. 1926. Measurement of environmental factors in the
tion of species significantly different from those in the tropical rainforest of Panama.
/ Ecology 7: 273
/302.
Ecuadorian samples (Fig. 5). Like the other sites, most Basset, Y., Aberlenc, H-P. and Delvare, G. 1992. Abundance
species were specialized to one height, but the prob- and stratification of foliage arthropods in a lowland rain
forest of Cameroon.
/ Ecol. Entomol. 17: 310
/318.
ability that a species selected at random from the Costa Bates, M. 1944. Observations on the distribution of mosquitoes
Rica sample is a canopy species was much higher. This in a tropical forest.
/ Ecology 25: 159
/170.
suggests that the canopy fauna in Costa Rica may be Beccaloni, G. W. 1997. Vertical stratification of ithomiine
butterfly (Nymphalidae: Ithomiinae) mimicry complexes:
more important in determining overall community the relationship between adult flight height and larval host-
composition and diversity than in Ecuadorian faunas. plant height.
/ Biol. J. Linn. Soc. 62: 313
/341.
It may be that during the short period of the Costa Brown, J. H. 1995. Macroecology.
/ Univ. of Chicago Press.
Rican study, many of the understory species were not Bulmer, M. G. 1974. On fitting the Poisson log-normal
distribution to species abundance data.
/ Biometrika 30:
present. Nonetheless, effects of seasonal bias are miti- 651
/660.
gated somewhat by the theoretical underpinnings of the Colwell, R. K. and Coddington, J. A. 1994. Estimating
species abundance distributions that permits accounting terrestrial biodiversity through extrapolation.
/ Philos.
Trans. R. Soc. (B) 345: 101
/118.
for the species that are not observed in samples DeVries, P. J. 1988. Stratification of fruit-feeding nymphalid
(Appendix 1). The canopy bias of species preference in butterflies in a Costa Rican rainforest.
/ J. Res. Lepidopt.
Costa Rica is probably not due to limited sample sizes 26: 98
/108.
DeVries, P. J. and Walla, T. R. 2001. Species diversity and
since this only affects the degree of certainty in estimates, community structure in neotropical fruit-feeding butterflies.
but not the expectations. Neither is it due to intraspecific
/ Biol. J. Linn. Soc. 74: 1
/15.
variation among regions in height preference; inspection DeVries, P. J., Murray, D. and Lande, R. 1997. Species diversity
of the raw data shows that all species shared among sites in vertical, horizontal, and temporal dimensions of a fruit-
feeding nymphalid butterfly community in an Ecuadorian
exhibited similar patterns of vertical structure. rainforest.
/ Biol. J. Linn. Soc. 62: 342
/364.
DeVries, P. J., Walla, T. R. and Greeney, H. F. 1999. Species
diversity in temporal and spatial dimensions of fruit-feeding
butterflies from two Ecuadorian rainforests.
/ Biol. J. Linn.
Soc. 68: 333
/353.
Conclusions Diserud, O. H. and Engen, S. 2000. A general and dynamic
species abundance model embracing the log-normal and
The comparative analyses performed here were possible gamma models.
/ Am. Nat. 155: 497
/511.
through an integration of long-term studies of species Elton, C. S. 1966. Animal ecology.
/ October House, New
York.
diversity and advances in statistical analyses of commu- Elton, C. S. 1973. The structure of invertebrate populations
nity structure (Engen 1978, 2001, Engen and Lande inside Neotropical rain forest.
/ J. Anim. Ecol. 42: 55
/104.

616 OIKOS 107:3 (2004)

Engen, S. 1978. Stochastic abundance models.
/ Chapman and Srivastava, D. S. 1999. Using local-regional richness plots to test
Hall. for species saturation: pitfalls and potentials.
/ J. Anim.
Engen, S. 2001. A dynamic and spatial model with migration Ecol. 68: 1
/6.
generating the log-Gaussian field of population densities. Stork, N. E. 1988. Insect diversity: facts, fiction and specula-

/ Math. Biosci. 173: 85
/102. tion.
/ Biol. J. Linn. Soc. 35: 321
/337.
Engen, S., Lande, R., Walla, T. R. et al. 2002a. Analyzing Tilman, D. 1994. Competition and biodiversity in spatially
spatial structure of communities by the two-dimensional structured habitats.
/ Ecology 75: 2
/16.
Poisson log-normal species abundance model.
/ Am Nat. Ugland, K. I. and Gray, J. S. 1982. Lognormal distributions and
160: 60
/73. the concept of community equilibrium.
/ Oikos 39:
Engen, S. and Lande, R. 1996. Population dynamic models 171
/178.
generating the log-normal species abundance distribution. Vellend, M. 2001. Do commonly used indices of beta-diversity

/ Math. Biosci. 132: 169
/183. measure species turnover?
/ J. Veg. Sci. 12: 545
/552.
Engen, S., Lande, R. and Sæther, B.- E. 2002b. Migration and Wagner, H. H. and Wildi, O. 2002. Realistic simulation of the
spatio-temporal variation in population dynamics in a effects of abundance distribution and spatial heterogeneity
on non-parametric estimators of species richness.
heterogeneous environment.
/ Ecology 83: 570
/579.

/ Ecoscience 9: 241
/250.
Erwin, T. L. 1983. Beetles and other insects of tropical forest
Whittaker, R. H. 1975. Communities and ecosystems.
/ Mac-
canopies at Manaus, Brazil, sampled by insecticidal fogging. Millan.

/ In: Sutton, S. L., Whitmore, T. C. and Chadwick, A. C. Wiens, J.A. 1984. On understanding a non-equilibrium world:
(eds), Tropical rain forest: ecology and management. Black- myth and reality in community patterns and processes, In:
well Scientific Publications, pp. 59
/76. Strong, D. (ed.), Ecological communities: conceptual issues
Fisher, R. A., Corbet, A. S. and Williams, C. B. 1943. The and the evidence. Princeton Univ. Press. pp. 439
/457.
relation between the number of species and the number of Williams, P. H., de Klerk, H. M. and Crowe, T. M. 1999.
individuals in a random sample from an animal population. Interpreting biogeographical boundaries among Afrotropi-

/ J. Anim. Ecol. 12: 42
/58. cal birds: spatial patterns in richness gradients and species
Gering, J. C. and Crist, T. O. 2002. The alpha-beta-regional replacement.
/ J. Biogeog. 26: 459
/474.
relationship: providing new insights into local-regional Wilson, M. V. and Mohler, C. L. 1983. Measuring composi-
patterns of species richness and scale dependence of tional change along gradients.
/ Vegetatio 54: 129
/141.
diversity components.
/ Ecol. Lett. 5: 433
/444. Wolda, H. 1992. Trends in abundance of tropical forest insects.
Gering, J. C., Crist, T. O. and Veech, J. A. 2003. Additive
/ Oecologia 89: 47
/52.
partitioning of species diversity across multiple spatial
scales: implications for regional conservation of biodiversity.

/ Conserv. Biol. 17: 488
/499.
Gleason, H. A. 1926. The individualistic concept of the plant
association.
/ Bull. Torrey Bot. Club 53: 7
/26. Appendix
Gray, J. S. and Mirza, F. B. 1979. A possible method for
detecting pollution-induced disturbance on marine benthic If ln l is a normally distributed variable with mean m and
communities.
/ Mar. Pollution Bull. 10: 142
/146. variance s2 and the distribution of X conditioned on l is
Grundy, R. M. 1951. The expected frequencies in a sample of an
animal population in which the abundances are log-nor- the Poisson distribution with parameter l, then the
mally distributed.
/ Biometrika 38: 427
/434. unconditional distribution of X is the Poisson log-
Hubbell, S. P. 2001. The unified neutral theory of biodiversity normal distribution
and biogeography.
/ Princeton Univ. Press.
Lande, R. 1996. Statistics and partitioning of species diversity, P(Xx; m; s2 q(x; m; s2 )
and similarity among multiple communities.
/ Oikos 76: 
exp(zsx  mx  e(zsm) )
5
/13.
Lande, R., Engen, S. and Sæther, B.-E. 2003. Stochastic
population dynamics in ecology and conservation.
/ Oxford

g  x!
f(z)dz

Univ. Press. where f(z) is the standard normal distribution.

Loreau, M. 2000. Are communities saturated? On the relation-
ship between a, b and g diversity.
/ Ecol. Lett. 3: 73
/76. Consider a community of s species with abundances
MacArthur, R. H. 1958. Population ecology of some warblers of l1, l2, . . . ls. In the log-normal species abundance model
northeastern coniferous forests.
/ Ecology 39: 599
/619. the abundances are modeled as independent observa-
MacArthur, R. H. 1965. Patterns of species diversity.
/ Biol. tions from the log-normal distribution (Preston 1948,
Rev. 40: 510
/533.
MacArthur, R. H. 1972. Geographical ecology.
/ Harper and Grundy 1951, Bulmer 1974, Engen and Lande 1996),
Row. which is equivalent to the ln l being normally distrib-
Magurran, A. E. 1988. Ecological diversity and its measure- uted, say N(m, s2). A dynamic process leading to this
ment.
/ Chapman and Hall. model was given by Engen and Lande (1996). For a
Novotny, V. and Basset, Y. 2000. Rare species in communities of
tropical insect herbivores: pondering the mystery of single- given abundance l, the corresponding number of in-
tons.
/ Oikos 89: 564
/572. dividuals observed is commonly assumed to be Poisson
Papageorgis, C. A. 1975. Mimicry in neotropical butterflies. distributed with parameter lv, where v is a parameter

/ Am. Sci. 63: 522
/532. characterizing the sampling effort. Since ln(lv) is
Patrick, R., Hohn, M. and Wallace, J. 1954. A new method for
determining the pattern of the diatom flora, Noctulae normally distributed with mean m/lnv and variance
Natura Acad. Nat. Sci.
/ Philadelphia 259: 1
/12. s2 it follows from the above definition that the observed
Pielou, E. C. 1975. Ecological diversity.
/ Wiley. number of individuals for each species are independent
Preston, F. W. 1948. The commonness and rarity of species.
variables with the Poisson log-normal distribution (x;

/ Ecology 29: 254
/283.
Richards, P. W. 1996. The Tropical Rain Forest, 2nd edn. m/lnv, s2). Notice that the first parameter in this

/ Cambridge Univ. Press. distribution depends on the sample size, while the second

OIKOS 107:3 (2004) 617

parameter, which is the variance of the log of the normal distribution, the straightforward generalization
abundances, takes the same value regardless the sam- of the one-dimensional case leads to the two-dimen-
pling effort. sional Poisson log-normal distribution with parameters
Since s is usually unknown, we only consider the (mA/lnvA, mB/lnvB, s2A, s2B, r). Again, we would have
observed number of individuals for the observed species. to consider the truncated form, including only species
The distribution of the number of individuals then that are observed at least at one of the sites.
follows the zero truncated Poisson log-normal distribu- In order to estimate the parameters in the two-
tion dimensional case, the likelihood function for the zero-
truncated bi-variate distribution has to be computed by
q(x; mlnv; s2 )=(1q(0; mlnv; s2 ))
performing the two-dimensional integrations numeri-
defined for x /1, 2, . . . The maximum likelihood cally, and finally the likelihood function must be
estimation of the parameters of this distribution based maximized numerically with respect to the unknown
on S observed species was derived by Bulmer (1974). The parameters. Uncertainties are most conveniently found
idea of using the truncated distribution was first put by performing parametric bootstrapping, as described in
forward by R.A. Fisher in the classical paper on log- the main text.
series species abundance distribution (Fisher et al. 1943). For a given species with given abundances in two
If the abundance structures at two different locations, samples, say (l, n), the number of individuals observed
say A and B, are to be analyzed, one may define a two- at site A, say X, given the total number of individuals
dimensional abundance model giving a joint description observed from this species, say N /X/Y, is bi-nomially
of the two sites. Suppose there are s species in the joint distributed with parameters N and p/l/(l/n), which is
community, and let (li, ni), for i/1, 2, . . .s, be the equivalent to logit (p)/lnl/lnn. Hence, the distribu-
abundances of the species at A and B, respectively. The tion of logit (p) among species is the normal distribution
natural generalization of the log-normal species abun- with mean m/mA mB and variance s2 /s2A/s2B /
dance model is then to make the assumption that for 2rsAsB. From this we find the distribution of p among
each pair of log of abundances at A and B, (lnli, lnni), species to be
are generated by a bi-variate normal distribution. The     2 
corresponding pair of abundances then has the bi-variate 1 1 p
g(p) pffiffiffiffiffiffi exp  ln m
log-normal distribution. Assuming that the sampling 2psp(1  p) 2s2 1p
intensities at A and B are vA and vB, respectively, the
observed number of individuals at A and B for this set of Notice that if the sampling intensities at A and B are
species is a sample of independent observations from the the same, that is, vA /vB, then the parameter m and s2
corresponding bi-variate Poisson log-normal distribu- are estimable by observations from the bi-variate Poisson
tion. log-normal distribution. Otherwise, vA/vB must be
Writing (mA, mB, s2A, s2B, r) for the expectations, known in order to estimate the parameter m in the
variances and the correlation of the underlying bi-variate distribution of p.

618 OIKOS 107:3 (2004)