Page 1
2364
Ecology, 84(9), 2003, pp. 2364–2377
? 2003 by the Ecological Society of America
ESTIMATING SPECIES RICHNESS: SENSITIVITY TO SAMPLE COVERAGE
AND INSENSITIVITY TO SPATIAL PATTERNS
ULRICH BROSE,1NEO D. MARTINEZ, AND RICHARD J. WILLIAMS
Romberg Tiburon Center, Department of Biology, San Francisco State University, 3152 Paradise Drive, Tiburon,
California 94920 USA
Abstract.
tionary and ecological theory from mass extinctions to island biogeography. Still, the factors
influencing the accuracy of estimators of species richness are poorly understood. We ex-
plored these factors by simulating landscapes that varied in species richness, relative abun-
dances, and the spatial distribution. We compared the extrapolations of nine nonparametric
estimators and two species accumulation curves under three sampling intensities. Com-
munity evenness of species’ abundances, sampling intensity, and the level of true species
richness significantly influenced bias, precision, and accuracy of the estimations. Perhaps
most surprisingly, the effects of gradient strength and spatial autocorrelation type were
generally insignificant. The nonparametric estimators were substantially less biased and
more precise than the species accumulation curves. Observed species richness was most
biased. Community evenness, sampling intensity, and the level of true species richness
influenced the performance of the nonparametric estimators indirectly via the fraction of
all species found in a sample or ‘‘sample coverage.’’ For each particular level of sample
coverage, a single estimator was most accurate. Choice of estimator is confounded by a
priori uncertainty about the sample coverage. Accordingly, researchers can extrapolate
species richness by various estimators and base the estimator choice on the mean estimated
sample coverage. Alternatively, the most reliable estimator with respect to community
evenness can be chosen. These predictions from our simulations are confirmed in two field
studies.
Key words:biodiversity; estimations; extrapolations; gradients; jacknife estimators; nonpara-
metric estimators; relative abundance distribution; sample coverage; spatial autocorrelation; species
accumulation curves.
The number of species in an area is critical to the development of evolu-
INTRODUCTION
Studies of biodiversity initiated the study of evolu-
tion and have continued to become more centrally im-
portant to ecology (Gaston 1996). However, the mea-
surement of biodiversity has been surrounded by a con-
siderable debate (Peet 1974, May 1975, Pielou 1975,
Magurran 1988, Colwell and Coddington 1994, Gaston
1996, Gotelli and Colwell 2001, Petchey and Gaston
2002). Historically, diversity has been measured by a
bewildering range of diversity indices that often consist
of two components: the number of species and the rel-
ative evenness of their abundances (Magurran 1988).
Both the compound character and the variety of dif-
ferent indices make comparisons across studies diffi-
cult. More recently, both components have generally
been measured separately, and much emphasis has been
put on species richness (Colwell and Coddington 1994,
Gaston 1996), which continues to be the most frequent
measure of biodiversity. Earlier, Whittaker (1972) pro-
posed ? diversity to be within-habitat species richness,
? diversity between-habitats species turnover, and ?
Manuscript received 11 September 2002; revised 22 Decem-
ber 2002; accepted 16 January 2003. Corresponding Editor: N. J.
Gotelli.
1E-mail: brose@sfsu.edu
diversity landscape species richness. Measurements of
? diversity or ‘‘complementarity’’ have subsequently
been treated separately from species richness (Magur-
ran 1988, Colwell and Coddington 1994, Colwell 1997,
Harte et al. 1999). In contrast to this, the separation of
? and ? diversity has been modified into a concept of
species richness on a continuous scale from local to
global (e.g., Ricklefs and Schluter 1993), and the meth-
odology of measuring local or global species richness
is not considered to be qualitatively different (Colwell
and Coddington 1994, Gaston 1996). This concept of
species richness is currently the most basic and often
used parameter, not only in biodiversity studies but also
in community and trophic ecology (Gaston 1996, Mar-
tinez et al. 1999, Williams and Martinez 2000). Despite
the centrality of species richness, many difficulties re-
garding its measurement are not well understood nor
agreed upon. In most empirical studies, the observed
number of species, Sobs, is used as a surrogate for the
true number of species, Strue. Still, Sobstypically ex-
cludes many rare species and seriously underestimates
Strue(Palmer 1990, Baltana ´s 1992, Colwell and Cod-
dington 1994, Martinez et al. 1999). This bias increases
with increasing Strueand decreases with mean species
detectability, which frequently invalidates hypothesis
testing (Palmer 1990, Boulinier et al. 1998). A large
Page 2
September 20032365
ESTIMATING SPECIES RICHNESS
number of extrapolation methods have been developed
to reduce this bias (reviews in Bunge and Fitzpatrick
1993, Sobero ´n and Llorente 1993, Colwell and Cod-
dington 1994, Colwell 1997), but their performance has
been compared in only few studies (Palmer 1990, 1991,
Baltana ´s 1992, Colwell and Coddington 1994, Walther
and Morand 1998, Longino et al. 2002). Most of these
methods can be classified as either species accumula-
tion curves or nonparametric estimators. Although gen-
erally less biased, the nonparametric estimators are sel-
dom used in biodiversity studies, where species ac-
cumulation curves are preferred (e.g., Chazdon et al.
1999).
Species accumulation curves extrapolate species
richness vs. sample size data to an asymptote of total
richness (Sobero ´n and Llorente 1993, Colwell and Cod-
dington 1994). The most often used species accumu-
lation curves are the exponential equation (Holdridge
et al. 1971) and the Michaelis-Menten model (Mi-
chaelis and Menten 1913). Their performance varies
with the underlying relative abundance distributions
(Keating and Quinn 1998), spatial aggregation of spe-
cies (Baltana ´s 1992) and habitat heterogeneity (Colwell
and Coddington 1994, Lande et al. 2000).
Nonparametric estimators are sampling theoretic ex-
trapolation methods that only require the number of
samples in which each species is found rather than any
parametric information about their abundance. They
were originally developed to estimate population sizes
with capture–recapture data in closed populations (Otis
et al. 1978, Burnham and Overton 1979). The technique
can be adapted to estimate species richness by substi-
tuting recaptures of species in different samples for
recaptures of individuals of the same species in time
between mark and recapture. The underlying models
assume species-specific detection probabilities that fol-
low an arbitrary distribution (e.g., model Mh, Otis et
al. 1978), which relaxes the implicit assumption of
equal detection probabilities made when using Sobs
(Boulinier et al. 1998). However, the distribution of
capture probabilities among individuals in a species’
population is likely to be more even than the distri-
bution of detection probabilities among species with
strong variations in their abundances. This could re-
strict legitimate use of the estimators to communities
with highly even relative abundance distributions. Fur-
thermore, the estimators that are based on model Mh
have a second assumption: that the species-specific de-
tection probabilities are constant for all sampling units,
which assumes spatially homogeneous species distri-
butions. This assumption is violated in almost every
type of community due to environmental gradients and
spatial autocorrelation (Legendre 1993).
In this study, we examine the sensitivity of different
estimators’ bias and accuracy to abundance distribu-
tions, strength of environmental gradients, spatial au-
tocorrelation, the level of Strueand sampling intensity
in simulated landscapes. The predictions of the sen-
sitivity analyses are tested in two field studies.
METHODS
Simulations
Rectangular communities were simulated using lat-
tices of 100 by 50 cells (x, y), each representing 1 m2.
Strueis the number of species present in these simulated
landscapes and Sobsis the number of species found in
samples taken from these landscapes. The landscapes
contained communities that systematically differed due
to Strue, relative abundance, spatial gradients, and spatial
autocorrelation. The combination of seven levels of Strue
(25, 50, 100, 150, 200, 250, and 500 species), three
relative abundance models, three gradient strengths,
and three types of spatial autocorrelation resulted in a
total of 189 different types of communities. Monte Car-
lo simulations with 100 replicates for each community
type were repeated for three sampling intensities, yield-
ing a total of 56700 simulated landscapes.
In each landscape, the abundance of each species,
which also equals the number of cells occupied, was
derived from relative abundance distributions possess-
ing an average of 200 individuals per species. We then
independently assigned each species i a probability of
occurrence pi(x, y) in every cell (x, y) of the lattice.
These pi(x, y) depend on the respective species-specific
probability functions of the gradients and the spatial
autocorrelation: pi(x, y) ? p
all individuals of a species are subjected to the same
spatial distribution of probabilities but the distribution
may be different for each species. For each species the
individuals were placed on the lattice according to pi(x,
y). First, the cells were ranked according to their pi.
Second, a number was sampled from a normal distri-
bution with a mean of zero, and a standard deviation
equal to the number of empty cells divided by two, and
the chosen number was converted to a rank equal to
the absolute value of the number. Third, an individual
was assigned to the cell with the chosen rank and that
cell was no longer available for occupation by an in-
dividual of the same species. The second and third steps
were repeated until the number of cells occupied
equalled the species’ abundance. This procedure was
repeated for all remaining species.
We estimated species richness under three intensities
of sampling from the 5000-cell communities: low (25
cells), intermediate (200 cells), and high intensity (500
cells). We applied 12 estimators of species richness to
each sample: nine nonparametric estimators, two spe-
cies accumulation curves, and Sobs. The performance of
these estimators was compared concerning bias, pre-
cision, and accuracy.
(x) ? p
(x, y) such that
ii
gradauto
Relative abundance distributions
We used three stochastic relative abundance models:
the broken-stick, random-fraction, and random assort-
Page 3
2366
ULRICH BROSE ET AL.
Ecology, Vol. 84, No. 9
FIG. 1.
with Sest? Strue. Regression details are given in Table 1. Key to abbreviations: Jsel? Jacksel; M-M ? Michaelis-Menten model;
Expo ? Exponential equation.
Mean estimated species richness, Sest, plotted against Strue. The solid line indicates an idealized unbiased estimator
ment models. Stochastic models generate much more
variable individual communities than deterministic
models do. We chose the stochastic models because
they represent more accurately natural communities
that are simultaneously influenced by various factors,
and they estimate the relative abundances of rare spe-
cies more precisely. This allows stochastic models to
encompass a broad range of community types (Nee et
al. 1991, Tokeshi 1993, Naeem and Hawkins 1994).
The broken-stick model represents a rarely observed,
extremely even community with low sequential dom-
inance and few rare species. The relative abundance of
Struespecies is determined by placing Strue?1 points
randomly on a line of unit length. Breaking the line at
those points yields Struesegments whose lengths rep-
resent the relative abundances of Struespecies (Mac-
Arthur 1957). Broken-stick abundance distributions
characterize some of the most even natural commu-
nities observed (Tokeshi 1993). The random-fraction
model yields comparatively uneven communities with
abundance distributions similar to Preston’s (1948) de-
terministic lognormal model (see Tokeshi 1993: Fig.
1). Randomly breaking a line of unit length into two
parts, and then repeating the process on randomly cho-
sen parts until Struesegments are produced, generates a
random-fraction community (Tokeshi 1993). The ran-
dom-assortment model yields more highly uneven com-
munities that are analogous to those described by geo-
metric series (Tokeshi 1993). The random-assortment
model assigns a random fraction of the line to the first
species, a random fraction of the remaining part of the
line to the second species, and so on until Struesegments
are produced.
For comparative purposes, we calculated community
evenness by the ratio of observed diversity to maxi-
mum diversity (Magurran 1988):
?
p ln p
i
ln S
?
i
evenness ?
where piis the proportion of individuals of species i
and S is the number of species.
Gradients
Gradients such as flooding regimes on shorelines can
be expressed as functions of the geographic coordinates
x and y (Legendre 1993). We simulated landscapes with
strong and weak gradients along the x coordinate as
well as landscapes with no gradient. Gradient strength
was expressed as the turnover distance dturnat which
on average all species were exchanged once (Jongmann
et al. 1995). A strong gradient was characterized by
one full turnover along the x coordinate of the lattice
(dturn? 100 cells), a weak gradient by a quarter turnover
along the lattice (dturn? 400 cells).
The individuals of a species i on a gradient are typ-
ically normally distributed with an optimum oi(the
mean) and an amplitude ai(standard deviation) (Jong-
mann et al. 1995). For every species i and every cell,
we simulated these probabilities as
22
i
p
(x) ? exp[?0.5(x?o) /a ].
grad
For every species i, we randomly derived the opti-
mum oifrom a uniform distribution with 0 ? oi? 100.
Given that one species turnover equals approximately
four standard deviations, the amplitude aiof every spe-
cies i was randomly sampled from a normal distribution
(mean ? dturn/4, standard deviation ? dturn/8).
ii
Spatial autocorrelation
Our simulations of spatial autocorrelation (Legendre
1993) vary the occurrence probabilities associated with
each cell based on variably located autocorrelation
sources. These probabilities were independently ap-
plied for every species i by locating between 1 and 10
autocorrelation sources on the lattice in the same way
as individuals, that is, either in a uniformly random
manner or according to p
igrad
source and in neighboring cells were established by
parameterizing a sill, range, and nugget. The sill is the
asymptotic occurrence probability at a certain range
from the nugget. The nugget indicates the maximum
. The probabilities at the
Page 4
September 20032367
ESTIMATING SPECIES RICHNESS
deviation from the sill at the autocorrelation source and
therefore determines the strength of autocorrelation.
Probabilities between the nugget and the sill are linear
functions of commonality dependent on the distance to
the autocorrelation source. We simulated communities
with biologically reasonable values (Mistral et al. 2000;
J. B. Wilson, personal communication) of long-range
autocorrelation (sill ? 0.27, nugget ? 0.72, and range
? 5 m) and short-range autocorrelation (sill ? 0.27,
nugget ? 0.62, and range ? 2 m). Species-specific
functions were derived to determine p
tocorrelation probabilities of occurrence for every cell
(x, y):
(x, y), the au-
iauto
p
(x, y)
iauto
0.033 ? 0.33d(x, y)
range ? 0.1
i
0
? c
0 ? d ? rangei
?
d ? range
i
where d(x, y) is the distance to the autocorrelation
source, rangeiis a species-specific range of autocor-
relation, and c is the autocorrelation strength that
equals 0.35 for short-range and 0.45 for long-range
autocorrelation. The values of rangeiwere randomly
sampled from normal distributions (short-range auto-
correlation: mean ? 2, lSD ? 1; long-range autocor-
relation: mean ? 5, lSD ? 2.5). In the model without
autocorrelation we defined for all species i: p
The spatial autocorrelation patterns were indepen-
dently created for every species i by (1) choosing 1–
10 autocorrelation sources from a uniform random dis-
tribution, (2) placing the sources on the landscape ac-
cording to the p
in the same way as individuals (see
igrad
spatial autocorrelation), and (3) calculating a new
landscape of probabilities of occurrence as pi(x, y) ?
p
(x) ? p
(x, y).
ii
grad auto
? 0.
iauto
Extrapolation of species richness
We compared estimations of two species accumu-
lation curves and nine nonparametric estimators (Ap-
pendix). This includes the two most frequently used
species accumulation curves, the exponential equation
(Holdridge et al. 1971) and the Michaelis-Menten mod-
el (Michaelis and Menten 1913). To minimize arbitrary
effects, cell sample order for accumulation curves was
randomized 100 times in each sample. The mean spe-
cies numbers S(n) were calculated for each value of n
cells sampled. Following Keating and Quinn (1998),
we used least squares nonlinear regressions with the
quasi-Newton algorithm to fit the Michaelis-Menten
model and the exponential equation to the mean S(n)
data. The often-discussed difference between number
of individuals and number of samples as measurements
of sampling intensity (Chazdon et al. 1999, Gotelli and
Colwell 2001) is minimal in our simulations, because
when averaged across all species the individuals are
evenly distributed across the lattice.
The nine nonparametric estimators used include the
first- to fifth-order jacknife estimators (Jack1–Jack5,
Burnham and Overton 1979), a selected and an inter-
polated jacknife estimator (Jacksel, Jackint, Burnham and
Overton 1979), the incidence-based estimator (Chao2,
Chao 1987), and the incidence-based coverage esti-
mator (ICE, Lee and Chao 1994). Since the original
formulations of Chao2 and ICE are undefined for zero
duplicates and when all infrequent species are uniques,
respectively, we used a corrected version of Chao2
(Colwell 1997; also see the Appendix) and replaced
ICE estimates by Chao2 estimates in the undefined cas-
es. To calculate Jacksel, the jacknife estimates are se-
quentially compared to the estimates of the next higher
order until the first estimate is reached that does not
significantly differ from the next (see Burnham and
Overton 1979 for the exact test statistics). Jackintis an
interpolated estimator between Jackseland the preced-
ing estimator.
Estimator evaluation
We based our estimator evaluation on bias, precision,
and accuracy. Positive and negative bias indicates over-
estimation and underestimation, respectively. Precision
is the closeness of repeated measurements of the same
quantity to each other (Sokal and Rohlf 1995). Accu-
racy is the closeness of a measured or computed value
to its true value (Sokal and Rohlf 1995). Overall bias
and precision were calculated for each estimator in all
simulations across all community types using linear
least-squares regressions of estimated species richness,
Sest, as a function of true species richness, Strue, forced
through an intercept of zero (method suggested by N.
Gotelli, personal communication). All data points fall
on a straight line with a slope of one for an unbiased
and perfectly precise estimator. We measured bias as
the difference between the observed slope of the re-
gression and the expected slope of one: Bias ? ob-
served slope ? 1. Precision was measured by the r2
values of the regressions.
We also calculated bias and inaccuracy for single
samples. In this case, bias was defined as the propor-
tional deviation of Sestfrom Strue:
S
? S
Strue
esttrue
bias ?
.
We measured inaccuracy using the square proportional
deviation of Sestfrom Strue:
2
S
? S
Strue
esttrue
inaccuracy ?
.
??
For groups of data sets, we calculated mean bias and
mean inaccuracy, and imprecision was measured using
the standard deviation of the bias. Bias, imprecision
and inaccuracy were measured as percentage of Strue.
RESULTS
Overall, no single estimator consistently performed
better than all others in each situation, but the non-
Page 5
2368
ULRICH BROSE ET AL.
Ecology, Vol. 84, No. 9
TABLE 1.
Seston Strueand correlation of the independent variables with the residuals.
General bias and precision of the estimators determined by linear regressions of
EstimatorBias Precision Rad†Grad‡Auto‡ Sampling†
Sobs
Chao2
Jack1
Jack2
Jack3
Jack4
Jack5
Jacksel
Jackint
ICE
M-M
Expo
?0.558
?0.372
?0.457
?0.404
?0.366
?0.336
?0.309
?0.363
?0.366
?0.466
?0.474
?0.588
0.679
0.762
0.733
0.765
0.784
0.788
0.764
0.785
0.784
0.735
0.664
0.677
?0.80
?0.64
?0.83
?0.82
?0.80
?0.74
?0.62
?0.72
?0.72
?0.75
?0.87
?0.85
0.00
?0.02
0.00
?0.01
?0.01
?0.02
?0.02
?0.02
?0.02
?0.01
0.01
0.01
?0.02
?0.02
?0.02
?0.02
?0.02
?0.02
?0.01
?0.02
?0.02
?0.02
0.00
?0.01
0.39
0.50
0.33
0.31
0.32
0.34
0.35
0.43
0.44
0.42
0.17
0.25
Notes: Bias ? slope ? 1; Precision ? r2. P ? 0.001 for all regressions. Standard errors of
all biases are 0.002. Spearman rank correlation coefficients are given in the last four columns.
Key to abbreviations: ICE ? incidence-based coverage estimator, M-M ? Michaelis-Menten
model, Expo ? exponential equation, Rad ? relative abundance distribution, Grad ? gradient
strength, Auto ? autocorrelation type.
† All values, P ? 0.001.
‡ All values not significant.
parametric estimators as a group did generally outper-
form the other estimators with respect to both bias and
precision. In contrast, several estimators consistently
performed poorly, most notably the most popular mea-
sure of biodiversity, Sobs. We first present the general
performance of the estimators followed by the effects
of abundance distribution and sample coverage on this
performance. Finally, we compare results from our sim-
ulated landscapes with results of similar analyses of
empirical data collected from field studies.
General performance of the estimators
We evaluated the general performance of the esti-
mators using linear regressions of Sestas a function of
Strue. All estimators were much more variable in the 25-
species simulations and comparatively less biased in
the500-species simulations. Furthermore, Chao2
showed an extraordinary high variance in these high-
richness simulations. Therefore, we avoided heteros-
cedasticity and nonlinearity in evaluations of bias and
precision by excluding results from landscapes with
Strue? 25 and Strue? 500. We also excluded the Strue?
50 samples from the analyses of Jack4 and Jack5 due
to high variances. All estimators were negatively bi-
ased, but only the exponential equation provided more
biased results than Sobs(Fig. 1). The Michaelis-Menten
model and ICE were the most biased of the remaining
estimators. The regression details are given in Table 1.
All estimators except the Michaelis-Menten model and
the exponential equation yielded more precise esti-
mates than Sobs. Due to their poor general performance,
the Michaelis-Menten model and the exponential equa-
tion were excluded from further analyses. Bias de-
creased with increasing order of the jacknife estima-
tors, resulting in Jack4 and Jack5 being least biased
while Jack4 and Jackselwere most precise (Table 1).
Surprisingly, Jackintwas more biased and imprecise
than Jacksel, and was also excluded from further anal-
yses. Chao2 performed intermediately with respect to
both bias and precision.
Regression residuals were significantly correlated
with relative abundance distributions and sampling in-
tensity, indicating that these two factors had a sub-
stantial additional impact on estimator performance
(Table 1). For most estimators, relative abundance dis-
tributions were twice as highly correlated with the re-
siduals as were sampling intensities. Surprisingly, gra-
dient strength and autocorrelation type were not cor-
related with the residuals in any of the situations an-
alyzed. Therefore, we further analyzed the impact of
the relative abundance distributions on estimator per-
formance in more detail.
Relative abundance distributions
The bias of all estimators was influenced by the rel-
ative abundance distributions as well as by sampling
intensity and Strue(Fig. 2). At each level of sampling
intensity and Strue, the differences in mean bias between
the three relative abundance distributions groups were
highly significant for all estimators (Kruskal-Wallis P
? 0.001). Sobswas reasonably unbiased only in very
even (broken-stick) communities under intermediate or
high sampling intensity, but Sobswas never the least
biased. In all uneven communities, all estimators were
substantially less biased than Sobs(Fig. 2). In general,
the estimators were up to 50-fold more biased in the
uneven (random-fraction) communities than in the even
(broken-stick) communities (Fig. 2). The estimates in
the highly uneven (random-assortment) communities
were generally extremely biased, with mean estimates
usually well below 50% of Strue. This increase in bias
coincided with a decrease in evenness from broken-
stick (mean ? 0.92, range 0.86–0.95) to random frac-
Page 6
September 20032369
ESTIMATING SPECIES RICHNESS
FIG. 2.
at different levels of Strueand sampling intensity. Results are depicted for broken-stick (solid circles), random fraction (open
circles), and random assortment (solid squares) communities. Jsel? Jacksel.
Influence of the relative abundance distributions on estimator bias (mean) and imprecision (standard deviations)
tion (mean ? 0.62, range 0.25–0.87) and random as-
sortment communities (mean ? 0.47, range 0.22–0.60).
Jack4 and Jack5 were generally the least biased es-
timators, but they were also the least precise, as in-
dicated by the large standard deviations of their mean
biases (Fig. 2). Chao2 and ICE were markedly impre-
cise in some data sets (e.g., Fig. 2, random assortment
communities with Strue? 250 under high sampling).
Jack1 and Jack2 were more biased, yet substantially
more precise, than Jack4 and Jack5. Jack3 and Jacksel
performed intermediately with respect to bias and pre-
cision. Jack1 and Jack2 performed best with respect to
accuracy in highly even communities, and Jack2,
Jack3, and Jack4 performed best with respect to ac-
curacy in the uneven and highly uneven communities.
However, estimator ranking also depended on sampling
intensity.
Sample coverage
Strue, sampling intensity (i.e., the number of samples
per community), and the community evenness medi-
ated by the relative abundance distributions influence
Sobs, and therefore the sample coverage, i.e., the pro-
portion of species present in the sample. Among all
simulations, the sample coverage was strongly corre-
lated with sampling intensity (Spearman rank corre-
lation, r ? 0.456) and community evenness (r ? 0.836),
but correlations with Strue(r ? 0.144) were surprisingly
weak (P ? 0.001 for all three Spearman rank corre-
lations). There were no significant correlations with
spatial autocorrelation type and gradient strength. Es-
timator bias was significantly correlated with all in-
dependent variables (Table 2). With respect to the cor-
relation coefficients, sample coverage was the most im-
Page 7
2370
ULRICH BROSE ET AL.
Ecology, Vol. 84, No. 9
TABLE 2.
planatory variables: Strue, sampling intensity, community evenness, and sample coverage.
Simple Spearman rank and partial correlations among estimator bias and the ex-
Estimator
Simple correlations
Strue
Sampling
intensityEvenness
Sample
coverage
Partial correlations
Strue
0.137
?0.065
?0.047
?0.032
?0.020
?0.011
0.019
0.048
Sampling
intensityEvenness
Chao2
Jack1
Jack2
Jack3
Jack4
Jack5
Jacksel
ICE
?0.033
?0.144
?0.139
?0.122
?0.090
?0.048
?0.094
?0.110
0.524
0.394
0.361
0.345
0.341
0.324
0.452
0.474
0.664
0.862
0.848
0.796
0.703
0.562
0.739
0.785
0.843
0.985
0.950
0.891
0.800
0.660
0.885
0.930
0.276
?0.243
?0.136
?0.047
0.015
0.045
0.147
0.162
?0.031
0.280
0.226
0.163
0.101
0.053
0.045
0.043
Notes: P ? 0.001 for all correlations; in partial correlations the effect of sample coverage
was removed prior to further correlation analyses.
portant independent variable and Struethe least impor-
tant. After the effect of sample coverage on estimator
bias had been removed in partial correlation analyses,
the other independent variables only had a minor direct
impact on the remaining variance of estimator bias de-
spite strong simple correlations (Table 2). Overall,
community evenness and sampling intensity controlled
sample coverage, which substantially influenced the
bias of the nonparametric estimators.
We analyzed the relationship between sample cov-
erage and estimator performance in more detail by plot-
ting estimator bias, imprecision, and inaccuracy in 1%
steps against sample coverage (Fig. 3). While bias and
inaccuracy decreased with increasing sample coverage,
imprecision was highest at intermediate levels of sam-
ple coverage. All estimators reduced the bias substan-
tially in comparison to Sobs, which, however, was most
precise (Fig. 3c–d). The least biased estimators, Jack3,
Jack4, and Jack5 (Fig. 3a–b), were markedly less pre-
cise than the other estimators (Fig. 3c–d), whereasmore
precise estimators like Jack1, Jack2, and ICE reduced
bias less than the other estimators did except Sobs(Fig.
3a–b). Chao2 was particularly imprecise under low
sample coverage, but was among the most precise es-
timators under high sample coverage (Fig. 3d). Chao2
reduced bias less under high sample coverage than the
jacknife estimators. At sample coverage ?95%, all
jacknife estimators had a positive bias. With respect to
accuracy, the estimators showed a more complex re-
sponse to sample coverage: the estimators’ curves in-
tersected, and for different levels of sample coverage,
different estimators were most accurate (Fig. 3e–f).
These intersections indicated that the most accurate
estimators were Jack5 up to a sample coverage of 26%,
Jack4 up to 38%, Jack3 up to 50%, Jack2 up to 74%,
Jack1 up to 96%, and Sobs? 96% (Fig. 3e–f, intersec-
tions are accurate ?1%).
Field studies
To test the predictions of our simulations, we ana-
lyzed data from two field studies on plant communities
that provided independent information about plant in-
ventories within plots and Strueof the entire habitats.
The first study was carried out on 12 temporary wet-
lands in East German farmland with a mean habitat
area of 1078 m2(Brose 2001). Systematic transect
searches (five times) were used to quantify Strue, and
nine 0.25-m2plots per study site were installed in a
stratified random design. The plant communities were
characterized by a mean Strueof 30 species (range: 19–
44 species) and a mean evenness of 0.92 (range: 0.91–
0.93), which is comparable to the 50-species broken-
stick communities. The mean sample coverage was
75%, which is between the mean values of the simu-
lated broken-stick communities sampled with low
(45%) and intermediate intensity (90%). The results
are very consistent with the simulations predicting that
(1) in broken-stick communities of 50 species, under
low to intermediate sampling intensity, the jacknife es-
timators are the least biased (Figs. 2 and 4a), (2) im-
precision increases with increasing order of the jacknife
estimators (Figs. 2, 3c–d, and 4a), and (3) under a
sample coverage of 75% Jack1 and Jack2 are the most
accurate estimators (Figs. 3e and 4a).
The second field study was carried out on five dry
grasslands on mountain slopes in the area of Pe ´cselyi,
?20 kilometers north of the Balaton in central Hun-
gary. The mean habitat area was ?1 km2. The vege-
tation was surveyed three times in 1999 and ranged
from subalpine communities to dry grasslands. On each
of the five grasslands, 40 1-m2plots stratified with re-
spect to altitude and soil depth were installed. Struewas
obtained by combining observed species lists with a
multiyear survey (Bauer 2000). Strue(mean: 133 spe-
cies, range: 83–177 species) was bracketed by our sim-
ulated Struevalues of 100 and 150 species. Their extreme
unevenness (mean ? 0.27; range: 0.26–0.27) is com-
parable to our random-fraction and random-assortment
communities. Sample coverage (mean 73%) is higher
than in any of the simulated uneven communities. Ran-
domly selecting 8 plots per grassland generated a less
sampled version of this data set. Species richness was
estimated in 20 replications and the mean estimates
were calculated for each estimator. Sample coverage
Page 8
September 20032371
ESTIMATING SPECIES RICHNESS
FIG. 3.
were determined for 1% ranges. Jsel? Jacksel.
Bias, imprecision, and inaccuracy of the nonparametric estimators as functions of sample coverage. All values
(mean 50%) was comparable to intermediate sampling
of our random-fraction simulations (53%, 50 species;
43%, 250 species). In accordance with our simulations,
the jacknife estimators were less biased than the other
estimators except Jack1, which was more biased than
Chao2 under intermediate sampling intensity with 8
samples (Figs. 1 and 4b–c). Most results are consistent
with our simulations predicting that (1) bias and pre-
cision decreased with increasing jacknife order (Figs.
1, 2, and 4b–c), (2) under a sample coverage of 50%
Jack3 is among the most accurate estimators (Figs. 2
and 4b), and (3) under 73% sample coverage Jack2
performs most accurately (Figs. 3e and 4c). However,
Jack4 and Jack5 under intermediate sampling and Jack3
under high sampling performed more accurately than
in our simulations.
DISCUSSION
Our general results have shown that most estimators
perform significantly better than Sobs, and that the non-
parametric estimators are substantially less biased and
more precise than the extrapolations of species accu-
mulation curves. This is consistent with previous stud-
ies (Palmer 1990, Baltana ´s 1992, Colwell and Cod-
dington 1994, and with some exceptions Walther and
Morand 1998 and Longino et al. 2002). In our study,
it was shown that these results hold for a large range
of communities with differences in evenness and spa-
tial patterns. In contrast to field studies, these param-
eters are not only exactly known in the simulated land-
scapes used in the present study, they were also sys-
tematically controlled as independent variables. Based
on this approach, it was shown that species richness
estimation depends on the relative abundance distri-
bution, Strue, and sampling intensity, but not on spatial
patterns.
Bias, precision and accuracy
Our results indicate that choosing estimators depends
on whether bias-reduction or precision is more impor-
tant. While Jack5 was the least biased and least precise
estimator, Sobswas most precise and most biased (see
Page 9
2372
ULRICH BROSE ET AL.
Ecology, Vol. 84, No. 9
FIG. 4.
temporary wetlands, and dry grasslands in (b) a reduced data set with eight samples per grassland, and (c) the full data set
with 40 samples per grassland. Jsel? Jacksel.
Bias (mean), imprecision (standard deviation of bias), and inaccuracy (mean ? 1 SD) of the estimators on (a)
Fig. 3). In our simulations as well as in the field studies,
both estimators would have yielded problematic re-
sults. Bias indicates systematic over- or underestima-
tion of the mean estimation when several independent
estimates are taken into account (Sokal and Rohlf
1995). Precision measures the similarity of independent
estimates, and therefore, it indicates the repeatability
of an estimate, whether or not it is biased. Biodiversity
studies typically prevent systematic evaluation of bias
and precision by restricting evaluation to one data set
rather than several independent estimates as in our
study. Accordingly, estimators used in such biodiver-
sity studies not only need to be as unbiased as possible,
they also need to be reliable in terms of precision in
order to reduce arbitrary variation while conducting
similar surveys or experiments. Accordingly, choosing
Page 10
September 20032373
ESTIMATING SPECIES RICHNESS
FIG. 5.
of three hypothetical estimators. In panels (b)–(d) these three
estimators were used for regression analyses representing a
hypothetical relationship y ? x: (b) Estimator1 (regression:
y ? x, r2? 0.68, P ? 0.01); (c) Estimator2 (y ? 0.6x, r2?
0.93, P ? 0.001); (d) Estimator3 (y ? 0.95 x, r2? 0.95, P
? 0.001).
(a) Bias (mean) and precision (standard deviation)
the least biased estimator may be appropriate for stud-
ies in which several independent data sets are available
to yield one mean estimate. This might be typical when
the concern is to estimate global biodiversity or the
species richness of an entire biome. In contrast to this,
field studies at smaller spatial scales are typically based
on single data sets, which makes precision critically
important. Our results indicate that there is a trade-off
between bias-reduction and precision of the estimators.
Following Sokal and Rohlf (1995), we feel that eval-
uations should be based on accuracy defined as a com-
bination of bias and precision. Similar to bias, accuracy
measures the deviation of the estimate from Strue, but
without the sign, which prevents underestimates from
averaging with overestimates to yield a very low mean
bias.
Fig. 5 illustrates this critical issue in term of three
hypothetical estimators: Estimator1 is unbiased, im-
precise and inaccurate (bias ? 0, imprecision ? 0.5,
inaccuracy ? 0.23), Estimator2 is biased, precise, and
inaccurate (bias ? ?0.4, imprecision ? 0.1, inaccuracy
? 0.17), Estimator3 is slightly more biased than Es-
timator1, slightly less precise than Estimator2, and
most accurate of the three estimators (bias ? ?0.05,
imprecision ? 0.15, inaccuracy ? 0.02). Although Es-
timator1 is unbiased on average, its individual esti-
mates are usually highly biased and over- or under-
estimate Struesubstantially (Fig. 5a). In contrast to this,
Estimator2 consistently underestimates Strue. Only the
accurate Estimator3 consistently provides unbiased es-
timates. To see the consequences of using biased, im-
precise, or accurate estimators, we then applied these
three estimators to a hypothetical study gathering data
that follows a relationship of y ? x (Fig. 5b–d). The
imprecise Estimator1 yielded a correct regression
slope, but a comparatively low coefficient of deter-
mination and a lower significance level than the other
estimators (Fig. 5b). The biased Estimator2 provided
a higher r2value, but markedly underestimated the re-
gression slope (Fig. 5c). Only the accurate Estimator3
yielded the approximately correct regression slope and
the highest r2value (Fig. 5d). In actual field studies
that include additional sources of variation, these ef-
fects are most likely even stronger. This example dem-
onstrates that basing estimator choice only on bias or
precision is potentially misleading, while accuracy
shows that the estimates are constantly close to the true
value. Our results indicate that accuracy benefits from
bias-reduction under low sample coverage when all es-
timators have a substantial negative bias, and precision
under high sample coverage when all estimators are
less biased.
Impact of the relative abundance distributions and
spatial patterns
Our results strongly corroborate the hypothesis that
uneven detection probabilities affect estimator perfor-
mance: the relative abundance distribution was the
most important factor influencing estimator bias. This
is consistent with Boulinier et al. (1998) who concluded
that the capture–recapture model Mh(Otis et al. 1978),
which assumes heterogeneous detection probabilities
of the species, is generally the most appropriate for the
estimation of species richness. Accordingly, the non-
parametric estimators based on model Mhare best suit-
ed to deal with heterogeneous detection probabilities
but, as our results indicate, these estimators are still
substantially biased. This bias is most likely caused by
the fact that the nonparametric estimators have origi-
nally been developed to estimate population size in
Page 11
2374
ULRICH BROSE ET AL.
Ecology, Vol. 84, No. 9
closed capture–recapture studies. However, detection
probabilities among different species are likely to be
more uneven than the detection probabilities of differ-
ent individuals of one species. Our results are also con-
sistent with Heltshe and Forrester (1983) and Baltana ´s
(1992), who found that the dispersion of the log-normal
distribution in communities affects the bias of the first-
order jacknife estimator. Our study has expanded this
finding to (1) all other estimators studied, (2) com-
munities with high species richness, (3) communities
of a broader evenness range, and (4) communities with
strong spatial gradients and autocorrelation. Therefore,
heterogeneous detection probabilities caused by un-
even relative abundance distributions pose a major
challenge for the extrapolation of species richness.
Even for the nonparametric estimators based on model
Mh, a more complex extrapolation procedure that cor-
rects for heterogeneous detection probabilities is re-
quired for comparing species richness and derived mea-
sures (e.g., connectance, Martinez et al. 1999) between
communities that vary in evenness. Until evenness-
corrected estimators are developed, our findings assist
the choice of appropriate estimators based on the even-
ness in the sample. We found that the most accurate
estimators are Jack1 and Jack2 in highly even broken-
stick and Jack2, Jack3, and Jack4 in the uneven random-
fraction or random-assortment communities. Jack2 ap-
pears to be the best overall estimator with respect to
accuracy.
To allow a detailed impact analysis of spatial pat-
terns, we distinguished spatial autocorrelation from
gradients. Gradients restrict the distribution of species
below the variation permissible in random and auto-
correlation communities. Therefore, the commonality
of two sampling plots on a gradient will be generally
lower than in plots lacking a gradient. Furthermore, we
differentiated the effects of the strengths of autocor-
relation and its range. However, our results suggest that
spatially heterogeneous species distributions mediated
by spatial autocorrelation and gradients do not affect
estimator bias. This is in contrast to previous studies
in which spatial autocorrelation strength had a small
but significant effect on estimator bias (Baltana ´s 1992,
Chazdon et al. 1998) or diversity indices (Fager 1972).
The estimator studies, however, varied autocorrelation
strength more than the present study, and testing the
effects of observed, randomized or moderately patchy
species distributions resulted mainly in nonsignificant
differences (Chazdon et al. 1998). Furthermore, Chaz-
don et al. (1998) have shown that abundance-based
estimators are more susceptible to be affected by patchy
distributions than the incidence-based estimators used
in the present study. Our results suggest that various
different spatial aggregation patterns in a realistic pa-
rameter space are not likely to have a substantial impact
on estimator performance. This is consistent with an
empirical study that concluded that capture–recapture
models assuming spatial heterogeneity are rarely ap-
propriate for the estimation of species richness (Bou-
linier et al. 1998). Accordingly, violating the assump-
tion of spatially homogeneous species distributions
does not seriously challenge the extrapolation of spe-
cies richness.
Sample coverage
In our study, no single estimator performed best
across the different levels of community evenness, Strue,
and sampling intensity, and these variables should be
addressed simultaneously while evaluating estimators.
Our results suggest that community evenness, Strue, and
sampling intensity indirectly influence the nonpara-
metric estimators through effects of sample coverage.
Surprisingly, there was only a minor influence of Strue,
which indicates that the nonparametric estimators are
well suited to estimate species richness in communities
with substantially different levels of Strue. Accordingly,
sample coverage mainly combines the effects of sam-
pling intensity and community evenness on the esti-
mators. With respect to sample coverage, studying a
more even community has the same effect as studying
an uneven community with higher sampling intensity.
Therefore, our results concerning sample coverage are
independent of the specific levels of community even-
ness and sampling intensity chosen, and represent the
full range of theoretically possible data sets. Knowing
the exact sample coverage in a case study would make
the decision for a specific estimator straightforward,
but, of course, unnecessary. However, when sample
coverage is approximately known, our results indicate
which estimators are most accurate (Fig. 6). Although
the bias of each estimator was consistent for each level
of sample coverage, bias-correction based on sample
coverage is not reliable, as it would add the error of
sample coverage estimation to the error of the species
richness estimates.
Sobsis usually not the best choice to measure species
richness, and in most data sets, ICE and Chao2 were
less accurate than the jacknife estimators. The low ac-
curacy of Chao2 was caused by a low precision under
low sample coverage and high bias under high sample
coverage. Surprisingly, Chao2, Jack3, Jack4, and Jack5
performed much more precisely in our general evalu-
ation (Table 1) than in the more detailed evaluations
(Figs. 2 and 3). This is caused by the fact that the data
sets with their most imprecise performances were ex-
cluded from the regression analyses in the general eval-
uation to avoid heteroscedasticity. Therefore, the gen-
eral precision values in Table 1 provide a relatively
flattering image of their precision, and the detailed re-
sults shown in Figs. 2 and 3 are more meaningful.
In empirical studies, there has been some disagree-
ment about the relative performance of the nonpara-
metric estimators: while Palmer (1990) concluded that
Jack1 is the least biased estimator, Colwell and Cod-
dington (1994) argued in favour of Chao2 and Jack2
and Walther and Morand (1998) preferred Chao2 and
Page 12
September 20032375
ESTIMATING SPECIES RICHNESS
FIG. 6.
coverage or the community evenness. Sobs? observed species richness. Thresholds are accurate within ?1%.
Two alternative ways to choose the most appropriate estimator based on either an estimated range of sample
Jack1. Our results indicate that this might stem from
the differences in their sample coverage. The field data
sets in the present study give an example that evalu-
ations under different sampling intensity, community
evenness, and consequently different sample coverage
yield different results concerning estimator ranking. In
contrast to this, in simulated landscapes, Strue, sampling
intensity, and community evenness are not only un-
questionably known, they are also controlled as inde-
pendent variables. This has led to the new estimator
performance model based on sample coverage in the
present study.
All jacknife estimators were positively biased above
95% sample coverage. This has previously been re-
ported in field studies by Heltshe and Forrester (1983),
Colwell and Coddington (1994), and Hellmann and
Fowler (1999), but with different magnitudes. In con-
trast to this, the jacknife estimators did not overesti-
mate in the studies of Palmer (1990, 1991) and Baltana ´s
(1992). Potential explanations include the following.
First, differences in the sample coverage between the
studies might be responsible. Second, particularly in
field studies, the determination of Strueis prone to un-
derestimation. Such lower Struemay account for some
of the jacknife estimator’s overestimations. In our sim-
ulations, however, the first- and second-order jacknife
estimators were less positively biased under high sam-
ple coverage than previously reported (e.g., Hellmann
and Fowler 1999).
Estimator choice
We will subsequently present a framework for the
choice of the most appropriate nonparametric estimator
depending on sample coverage (see Fig. 6). However,
unless the sample contains the entire community, one
can only guess about the sample coverage, which
makes the estimator choice more problematic. An es-
timator evaluation based on sampling intensity, i.e., the
percentage of total quadrats sampled (Hellmann and
Fowler 1999), only provides an incomplete framework,
since sampling intensity is only one of the factors in-
fluencing sample coverage. This framework remains
specific for the respective levels of community even-
ness, Strueand habitat area. As independent reliable in-
formation about Strueis scarcely obtainable in field stud-
ies, the most promising procedure to choose the ap-
propriate estimator is iterative (Fig. 6, path 1). First, a
range of Strueneeds to be calculated by various esti-
mators. Then, the most accurate estimator can be cho-
sen based on the resulting mean estimated sample cov-
erage. Applying this procedure to the field data sets
yields mean inaccuracies of 1.5% of Struefor the tem-
porary wetlands and 3.5% of Strueand 9.6% of Struefor
the dry grasslands in the full and reduced data set,
respectively. While these estimates are the most ac-
curate for the temporary wetlands, they are only slight-
ly outperformed by Jack2 and Jack3 on the dry grass-
lands. Compared to a decision based on community
evenness (Fig. 6, path 2), the decision path based on
estimated sample coverage (Fig. 6, path 1) is more
Page 13
2376
ULRICH BROSE ET AL.
Ecology, Vol. 84, No. 9
elaborate. For every level of sample coverage just one
estimator is most accurate. However, uncertainty of
sample coverage might force researchers to rely on path
2 if the estimated range of sample coverage is large.
In contrast to this, estimates of evenness are highly
accurate even when based on very few samples. For
example, under low sampling intensity evenness was
calculated with a mean absolute proportional deviation
of 9.8% in our simulated landscapes. Although the de-
cision based on evenness is more reliable, it does not
offer conclusive advice, as estimator choice still de-
pends on sampling intensity. In conclusion, we rec-
ommend that researchers examine both alternatives and
rely primarily on path 1. Although this procedure surely
does not satisfy the desire for a single generally ap-
plicable estimator, we do provide a method for choos-
ing estimators that should generally increase accuracy
beyond current methods and certainly beyond the wide-
spread measurement of species richness by Sobs. Beyond
choosing among available estimators, our in silico
methods could be used to evaluate and design new
estimators to be more highly accurate and also avoid
problems such as consistent negative bias at low sample
coverage.
CONCLUSIONS
Our analyses demonstrate that in most community
types the nonparametric estimators perform substan-
tially more accurately than Sobsor the species accu-
mulation curves. As specific spatial patterns and Strue
were of minor importance, our results appear spatially
scale invariant, making no differences between ? and
? diversity. Irrespective of scale, our results suggest
that species richness may be calculated accurately un-
der intermediate or high sampling intensity if the most
appropriate estimator is chosen. Community evenness
and sampling intensity, however, have a strong influ-
ence on estimator accuracy. With respect to these fac-
tors, there was not a single generally most accurate
estimator, and an optimal decision has to be based on
knowledge about the sample coverage that can only be
estimated from the samples. There are two possibilities
to deal with this dilemma (Fig. 6). First, species rich-
ness can be extrapolated by various estimators to base
the decision about the most appropriate estimator on
the estimated ranges for the sample coverage. Second,
the least sensitive estimator with respect to community
evenness can be chosen. Our study indicates a need for
nonnegatively biased estimators for low sample cov-
erage and a more complex extrapolation procedure that
corrects for the very heterogeneous detection proba-
bilities of species in natural communities. Until such
estimators are developed, we anticipate that our find-
ings will enhance the precision and accuracy of species
richness estimations, while assisting researchers to
choose the appropriate estimator for a particular com-
munity structure or sample coverage.
ACKNOWLEDGMENTS
We sincerely thank Jennifer Dunne, Robert K. Colwell,
Nicholas J. Gotelli, J. Bastow Wilson, Sa ´ndor Samu, Michael
Glemnitz, J. F. Quinn, and an anonymous reviewer for their
help and suggestions. U. Brose is supported by the German
Academy of Naturalists Leopoldina by funds of the German
Federal Ministry of Education and Science (BMBF-LPD
9901/8-44). N. D. Martinez and R. J. Williams are supported
by an NSF Biocomplexity Incubation Grant (DEB-0083929).
LITERATURE CITED
Baltana ´s, A. 1992. On the use of some methods for the es-
timation of species richness. Oikos 65:484–492.
Bauer, N. 2000. A Pe ´cselyi-medence e ´s peremhegyeinek bo-
tanikai vizsga ´lata. In J. Futo ´, editor. A Balaton-felvide ´k
terme ´szeti e ´rte ´kei I.:A Pe ´cselyi-medence e ´s ko ¨mye ´ke. Bal-
aton felvide ´ki nemz, Park igazgatosa ´ga, Veszprem, Hun-
gary.
Boulinier, T., J. D. Nichols, J. R. Sauer, J. E. Hines, and K.
H. Pollock. 1998. Estimating species richness: the impor-
tance of heterogeneity in species detectability. Ecology 79:
1018–1028.
Brose, U. 2001. Relative importance of isolation, area and
habitat heterogeneity for vascular plant species richness of
temporary wetlands in East-German farmland. Ecography
24:722–730.
Bunge, J., and M. Fitzpatrick. 1993. Estimating the number
of species: a review. Journal of the American Statistical
Association 88:364–373.
Burnham, K. P., and W. S. Overton. 1979. Robust estimation
of population size when capture probabilities vary among
animals. Ecology 60:927–936.
Chao, A. 1987. Estimating the population size for capture–
recapture data with unequal catchability. Biometrics 43:
783–791.
Chazdon, R. L., R. K. Colwell, and J. S. Denslow. 1999.
Tropical tree richness and resource-based niches. Science
285:1459a.
Chazdon, R. L., R. K. Colwell, J. S. Denslow, and M. R.
Guariguata. 1998. Statistical methods for estimating spe-
cies richness of woody regeneration in primary and sec-
ondary rain forests of northeast Costa Rica. Pages 285–
309 in F. Dallmeier and J. A. Comiskey, editors. Forest
biodiversity research, monitoring and modeling. UNESCO
and The Parthenon Publishing Group, London, UK.
Colwell, R. K. 1997. Estimates: Statistical estimation of spe-
cies richness and shared species from samples. Version 5.
User’s guide and application. published at: http://viceroy.
eeb.uconn.edu/estimates.
Colwell, R. K., and J. A. Coddington. 1994. Estimating ter-
restrial biodiversity through extrapolation. Pages 101–118
in D. L. Hawksworth, editor. Biodiversity: measurement
and estimation. The Royal Society, London, UK.
Fager, E. W. 1972. Diversity: a sampling study. American
Naturalist 106:293–310.
Gaston, K. J. 1996. Species richness: measure and measure-
ment. Pages 77–113 in K. J. Gaston, editor. Biodiversity:
a biology by numbers and difference. Blackwell Science,
Oxford, UK.
Gotelli, N. J., and R. K. Colwell. 2001. Quantifying biodi-
versity: procedures and pitfalls in the measurement and
comparison of species richness. Ecology Letters 4:379–
391.
Harte, J., S. McCarthy, K. Taylor, A. Kinzig, and L. Fischer.
1999. Estimating species–area relationships from plot to
landscape scale using species spatial-turnover data. Oikos
86:45–54.
Hellmann, J. J., and G. W. Fowler. 1999. Bias, precision and
accuracy of four measures of species richness. Ecological
Applications 9:824–834.
Page 14
September 20032377
ESTIMATING SPECIES RICHNESS
Heltshe, J. F., and N. E. Forrester. 1983. Estimating species
richness using the Jacknife Procedure. Biometrics39:1–11.
Holdridge, L. R., W. C. Grenke, W. H. Hatheway, T. Liang,
and J. A. Tosi. 1971. Forest environments in tropical life
zones. Pergamon Press, Oxford, UK.
Jongmann, R. H. G., C. J. F. Ter Braak, and O. F. R. Van
Tongeren. 1995. Data analysis in community and landscape
ecology. Cambridge University Press, Cambridge, UK.
Keating, K. A., and J. F. Quinn. 1998. Estimating species
richness: the Michaelis–Menten model revisited. Oikos 81:
411–416.
Lande, R., P. J. DeVries, and T. R. Walla. 2000. When species
accumulation curves intersect: implications for ranking di-
versity using small samples. Oikos 89:601–605.
Lee, S.-M., and A. Chao. 1994. Estimating population size
via sample coverage for closed capture–recapture models.
Biometrics 50:88–97.
Legendre, P. 1993. Spatial autocorrelation: trouble or new
paradigm?. Ecology 74:1659–1673.
Longino, J. T., J. Coddington, and R. K. Colwell. 2002. The
ant fauna of a tropical rain forest: estimating species rich-
ness three different ways. Ecology 83:689–702.
MacArthur, R. H. 1957. On the relative abundance of bird
species. Proceedings of the National Academy of Sciences
(USA) 43:293–295.
Magurran, A. E. 1988. Ecological diversity and its measure-
ment. Princeton University Press, Princeton, New Jersey,
USA.
Martinez, N. D., B. A. Hawkins, H. A. Dawah, and B. P.
Feifarek. 1999. Effects of sampling effort on character-
ization of food-web structure. Ecology 80:1044–1055.
May, R. M. 1975. Patterns of species abundance and diver-
sity. Pages 81–120 in M. L. Cody and J. M. Diamond,
editors. Ecology and evolution of communities. Harvard
University Press, Cambridge, Massachusetts, USA.
Michaelis, M., and M. L. Menten. 1913. Der kinetik der
invertinwirkung. Biochemische Zeitschrift 49:333–369.
Mistral, M., O. Buck, D. C. Meier-Behrmann, D. A. Burnett,
T. E. Barnfield, A. E. Scott, B. J. Anderson, and J. B.
Wilson. 2000. Direct measurement of spatial autocorrela-
tion at the community level in four plant communities.
Journal of Vegetation Science 11:911–916.
Naeem, S., and B. A. Hawkins. 1994. Minimal community
structure: how parasitoids divide resources. Ecology 75:
79–85.
Nee, S., P. H. Harvey, and R. M. May. 1991. Lifting the veil
on abundance patterns. Proceedings of the Royal Society
of London B, Biological Sciences 243:161–163.
Otis, D. L., K. P. Burnham, G. C. White, and D. R. Anderson.
1978. Statistical inference from capture data on closed an-
imal populations. Wildlife Monographs 62.
Palmer, M. W. 1990. The estimation of species richness by
extrapolation. Ecology 71:1195–1198.
Palmer, M. W. 1991. Estimating species richness: the second-
order Jacknife reconsidered. Ecology 72:1512–1513.
Peet, R. K. 1974. The measurement of species diversity. An-
nual Review of Ecology and Systematics 5:285–307.
Petchey, O. L., and K. J. Gaston. 2002. Functional diversity
(FD), species richness and community composition. Ecol-
ogy Letters 5:402–411.
Pielou, E. C. 1975. Ecological diversity. Wiley, New York,
New York, USA.
Preston, F. W. 1948. The commonness and rarity of species.
Ecology 29:254–283.
Ricklefs, R. E., and D. Schluter. 1993. Species diversity in
ecological communities. University of Chicago Press, Chi-
cago, Illinois, USA.
Sobero ´n, M. J., and B. J. Lorente. 1993. The use of species
accumulation functions for the prediction of species rich-
ness. Conservation Biology 7:480–488.
Sokal, R. R., and F. J. Rohlf. 1995. Biometry. Third edition.
W. H. Freeman, New York, New York, USA.
Tokeshi, M. 1993. Species abundance patterns and commu-
nity structure. Advances in Ecological Research 24:111–
186.
Walther, B. A., and S. Morand. 1998. Comparative perfor-
mance of species richness estimation methods.Parasitology
116:395–405.
Whittaker, R. H. 1972. Evolution and measurement of species
diversity. Taxon 21:213–251.
Williams, R. J., and N. D. Martinez. 2000. Simple rules yield
complex food web dynamics. Nature 404:180–183.
APPENDIX
A comparison of formulas for the estimators of species richness is available in ESA’s Electronic Data Archive: Ecological
Archives E084-057-A1.
Download full-text