See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/229796934

Statistical Estimation and Model Selection of Species‐Accumulation

Functions

Article  in  Conservation Biology · March 2005

DOI: 10.1111/j.1523-1739.2005.00453.x

CITATIONS READS
66 368

2 authors:

Eloisa Díaz-Francés Jorge Soberón

Centro de Investigación en Matemáticas (CIMAT) University of Kansas
18 PUBLICATIONS   163 CITATIONS    182 PUBLICATIONS   20,674 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Collaborative Forest Management in central Uganda View project

Ecology of Mexican Butterflies View project

All content following this page was uploaded by Jorge Soberón on 22 January 2019.

The user has requested enhancement of the downloaded file.

Statistical Estimation and Model Selection of
Species-Accumulation Functions
ELOÍSA DÍAZ-FRANCÉS∗ AND JORGE SOBERÓN†‡
∗
Department of Probability and Statistics, Center of Research in Mathematics (CIMAT) A.P. 402, Guanajuato, Gto. 36000, Mexico
†Department of Ecology of Biodiversity, Institute of Ecology, National University of Mexico, and National Commission on
Biodiversity (CONABIO) Insurgentes-Periférico 4903, Tlalpan, 14010, Mexico, D.F.

Abstract: Soberón and Llorente (1993) proposed pure-birth stochastic processes as theoretical models for
species-accumulation curves, and these processes have frequently been used to describe the progress of biological
inventories. We describe, in algorithmic form, an alternative statistical analysis based on a likelihood approach
(Dı́az-Francés & Gorostiza 2002) that provides mathematical rigor to the ideas in Soberón and Llorente
(1993) and improves the estimation of the models by incorporating the facts that the variance of the error
is not constant and that the observations are correlated. Additionally, we used the likelihood ratios between
candidate models as an objective procedure for model selection, allowing comparison between the goodness
of fit of various models. The software for these statistical methods can now be downloaded off the Internet. We
used two examples of butterfly data sets to illustrate the use of the methods and the software.

Key Words: model comparison, pure-birth process

Estimación Estadı́stica y Selección de Modelos de Funciones de Acumulación de Especies
Resumen: Soberón y Llorente (1993) propusieron procesos estocásticos de nacimientos puros como mode-
los teóricos para curvas de acumulación de especies, y estos procesos han sido usados frecuentemente para
describir el progreso de inventarios biológicos. Describimos, en forma algorı́tmica, un análisis estadı́stico alter-
nativo basado en un método probabilı́stico (Dı́az-Frances & Gorostiza 2002) que aporta rigor matemático a
las ideas en Soberón y Llorente (1993) y mejora la estimación de los modelos mediante la incorporación del he-
cho que la varianza del error no es constante y que las observaciones están correlacionadas. Adicionalmente,
utilizamos las proporciones probabilı́sticas entre modelos candidatos como un procedimiento objetivo para
la selección de modelos, que permite la comparación de la calidad del ajuste de varios modelos. El software
para estos métodos estadı́sticos puede ser descargado de la Internet. Utilizamos dos ejemplos con conjuntos de
datos de mariposas para ilustrar el uso de los métodos y el software.

Palabras Clave: comparación de modelos, proceso de nacimientos puros

Introduction the number of remaining uncollected species can be ob-

The problem of mathematically describing the process
of discovery of new species in an inventory has be-
come pressing, given the rate of disappearance of habitats
worldwide. Assessment of the degree of completeness of
the exploration process and extrapolations to estimate
tained from such mathematical descriptions. There are
many ways of attacking this problem (reviewed in Colwell
& Coddington 1994). A popular method is based on the
idea of fitting equations such as the Michaelis-Menten or
the von Bertalanffy models to effort-species data sets (ac-
cumulated units of collecting effort versus accumulated

‡Address correspondence to J. Soberón, email jsoberon@xolo.conabio.gob.mx

Paper submitted October 3, 2003; revised manuscript accepted July 1, 2004.

569

Conservation Biology, Pages 569–573

Volume 19, No. 2, April 2005
570 Model Selection of Species-Accumulation Functions
Table 1. Species-accumulation functions considered in Soberón and
Llorente (1993).∗
Model S(t; a,b) Asymptote
Exponential (a/b) (1 − e −bt ) a/b
Clench at/(1+bt) a/b
Logarithmic (1 − e−b )−1 log (1 + (1 − e −b ) at) — under consideration can be compared for a given data set
∗ The effort is represented by t, and a and b are the parameters to be
fitted.
number of observed species; see Clench 1979; Lamas et
al. 1991).
Soberón and Llorente (1993, henceforth Soberón and
Llorente) provide a theoretical justification for using such
parametric models by applying the ideas of the pure-birth
stochastic process theory. They assume that the accumu-
lated number of different species are the states of a pure-
birth process, where Y(t) represents the state (accumu-
lated number of species) after t units of effort, and that
the corresponding species-accumulation function can be
modeled as the mean function S(t; θ) of this process,
where θ is the vector of parameters and t is usually given
in time or effort units. They allow biological interpreta-
tion of three specific families of birth rates (also called col-
lecting functions), assuming that the probability of find-
ing a new species depends on the size of the list and the
time already spent in the field. From these birth rates, in
each case, they derived the mean functions of the process
Y(t) (Table 1) for θ = (a, b). Soberón and Llorente fitted
a selected S(t; θ) to the observed Y(t) by using nonlin-
ear regression procedures that assume uncorrelated er-
rors with constant variance. The authors acknowledge
that this procedure is far from ideal and ask for statistical
improvements. They also note the difficulty of the pro-
cedure and the strong need for development of objective
procedures for model selection.
Dı́az-Francés and Gorostiza (2002) (henceforth Dı́az-
Francés and Gorostiza) comment on some mathematical
inconsistencies they found in Soberón and Llorente. They
propose a statistical analysis that incorporates corrections
to these errors, in addition to presenting an objective pro-
cedure for model comparison and selection for a given
data set. Dı́az-Francés and Gorostiza show that the func-
tions S(t; θ) obtained by Soberón and Llorente are gener-
ally not means of pure-birth processes but that they can
be well approximated by the mean function of a suit-
able nonhomogeneous pure-birth process B(t), which
is close to the observed process Y(t). Dı́az-Francés and
Gorostiza suggest the use of the species-accumulation
function S(t; θ) proposed by Soberón and Llorente and
an upper bound U(t; θ) for the (unknown) variance of
B(t) to make statistical inferences about θ. They propose
the use of likelihood-based statistical methods to create a
nonlinear regression model with mean S(t; θ) and normal
errors, possibly correlated and such that the correlation
Conservation Biology
Volume 19, No. 2, April 2005
Dı́az-Francés & Soberón
between errors decreases with time. Statistical inference
about the model parameters is greatly improved by using
this upper bound for the variance and by considering an
additional parameter ρ that accounts for possible corre-
lation between the observations. A likelihood function is
then obtained for each model. As a result, relevant models
through their corresponding likelihood ratio and a useful,
objective, and quantitative tool for model comparison and
selection is provided.
Our purpose here is to give, for the first time in the bi-
ological literature, an outline of the alternative statistical
method proposed by Dı́az-Francés and Gorostiza and to
introduce its implementation, now available in freeware.
The work of Soberón and Llorente has been referenced
frequently in faunistic and floristic studies, and at least
two of the three families of species-accumulation func-
tions presented there are used widely. Therefore, it is im-
portant to provide better estimation methods for these
models. These estimation methods can now be fitted to
data sets through user-friendly software available free on
the Internet at http://cimat.mx/info.php?m=1&ind=5.
This software uses the likelihood ratios between candi-
date models among the pure-birth models proposed by
Soberón and Llorente and Dı́az-Francés and Gorostiza to
select the best model for a given data set. Further com-
parisons with other, different parametric stochastic mod-
els can be made through corresponding likelihood ratios.
A by-product of this freeware is a likelihood plot of the
total number of different species (TNS) for the models
that reach an asymptote, which gives information on how
plausible different values of TNS are for the given data.
We used this software to fit models from Table 1 to two
butterfly data sets.
Methods
A species-accumulation function is a nondecreasing curve
that represents the expected accumulated number of dif-
ferent species encountered within a certain geographi-
cal area as a function of a measure of the effort (usually
time or person-hour units) to collect them. The observed
effort-species data set is a collection of pairs {t i ,Y(t i )}, i =
1, . . . n, where t i is the successive effort units and Y(t i ) is
the accumulated number of different species seen up to t i ;
consequently, the observed data are always nondecreas-
ing. The exact time of appearance of a new species is not
recorded. The available information is only that Y(t i ) −
Y(t i−1 ) new species appeared within the time interval t i −
t i−1 and that frequently the increments Y(t i ) − Y(t i−1 ) are
>1.
The first two models proposed by Soberón and Llorente
(exponential and Clench, Table 1) describe families of
species-accumulation functions that are bounded (i.e.,
they are useful to describe situations in which TNS, which
Dı́az-Francés & Soberón
is directly related to the asymptote of the curve, will be
registered eventually). In other words, the area is small,
the taxa are well known, or both, or the collectors ac-
cumulate experience, which increases the plausibility of
detecting new species as more time is spent in the field.
In contrast, the logarithmic model is unbounded and is
useful to describe situations in which the area is large,
the taxa are poorly known, or both.
To use the freeware that implements the method pro-
posed by Dı́az-Francés and Gorostiza, it is necessary to
select a function S(t; θ) from Table 1, where θ = (a, b), or
to propose a valid function S(t; θ) with biological mean-
ing. The data can be input from a spreadsheet or a text file.
To obtain estimates for the parameter θ, the log-likelihood
function for θ under the proposed model by Dı́az-Francés
and Gorostiza is maximized. The proposed estimates for
θ are the values of the parameters that make the observed
data set most probable. These are called maximum likeli-
hood estimates because they maximize the log-likelihood
function. An upper bound for the variance of the asso-
ciated pure-birth process is also used in this estimation
process. The variance of the process is not constant,
and it usually cannot be calculated easily. Incorporating
the proposed upper bound for the variance, however,
greatly improves the inferences about the model parame-
ters. Also, as a by-product, the parameter ρ that describes
the strength of the correlation between adjacent observa-
tions is also estimated. In the approach presented here,
any positive correlation between pairs of observations (if
present) is inversely related to the distance in time or
effort units between them.
The estimated TNS has to be relevant to the observed
species-accumulation data; therefore, it must be an inte-
ger equal to or larger than the total number of observed
species Y(t n ). The estimation methods must be condi-
tioned on this fact.
If the selected S(t; θ) is bounded—exponential or
Clench—the model can be reparameterized in terms of
TNS and a likelihood plot can be obtained that gives in-
formation about the plausibility of different values of TNS
for the observed data. Also, likelihood intervals can be ob-
tained by drawing a horizontal line at a certain height in
the plot of the likelihood function; thus, any value within
a given likelihood interval is more likely to be the true
TNS than any other value outside it. Usually likelihood
intervals of height 0.01 to 0.15 are used for interpre-
tation purposes. The lower endpoint of the 0.01 inter-
val is often considered the smallest likely value for TNS,
whereas the upper endpoint of this interval is consid-
ered the largest plausible value for TNS in the light of the
observed data. Under certain conditions, likelihood inter-
vals of 0.15 achieve an approximate 95% confidence level.
The maximum likelihood estimate of TNS is included in
all the likelihood intervals.
Different species-accumulation functions S G (t; θ G ) and
S H (t; θ H ) for the same data set, {t i ,Y(t i )} (i = 1, . . . n),
Model Selection of Species-Accumulation Functions 571
can be compared by calculating their corresponding like-
lihood ratio,
L G (θ G )/L H (θ H ) = x,
where the likelihood functions are evaluated at their cor-
responding maximum likelihood estimates (MLE), θ G and
θ H , for models G and H, respectively. For example, if x =
3, the observed data set is three times more probable un-
der model G than under model H, so model G is preferred
over H for this specific data set. The value of x indicates
how many times more (or less) plausible model G is for
the observed data than model H; that is, the likelihood
ratio provides a continuous plausibility rating scale to as-
sess different models (Fisher 1973). Any departure from
1—the value for which both models fit equally well—is
evidence in favor of one of the two models. The issue is
then how strong the evidence has to be to prefer a given
model. This may vary depending on the consequences
of selecting a model. The recommendation is that if one
model is to have priority over others, the evidence should
be strong, as it is in the case for La Calera butterflies in the
examples that follow. In contrast, for those cases where
the evidence is not strong enough to clearly favor a given
model, additional data must be collected to acquire more
information for model selection. If at least one of the mod-
els considered describes the data set adequately from sta-
tistical and biological points of view, the selection of a
model can be based on this quantitative approach. This
process must always be complemented with diagnostic
checks to assess the goodness of fit of the “best” model
to the data, in order to verify that the assumptions of the
stochastic model are reasonable and that they hold at least
approximately for the data. Additionally, external biologi-
cal reasoning should provide support to the model under
consideration in light of the observed data, in the sense
that the model should give reasonable answers to the
questions being asked and its features and assumptions
should describe the biological characteristics of the data
reasonably well. The important point is that the model
should be adequate for the purpose at hand.
Examples
La Calera and Atoyac Butterfly Examples
La Calera, a tropical semideciduous rain forest in Jalisco,
Mexico, is near the biosphere reserve of Manantlán. As a
part of a butterfly inventory of Jalisco, collections were
made in La Calera. After 53 days of work, 240 differ-
ent species had been observed (Vargas et al. 1996). The
observed species-accumulation data and the estimated
species-accumulation curves S(t; θ) under the three dif-
ferent models given in Table 1 were fitted, using the free-
ware program, to the La Calera data (Fig. 1 & Table 2). The
best-fitting model was the exponential model, for which
Conservation Biology
Volume 19, No. 2, April 2005
572 Model Selection of Species-Accumulation Functions
Figure 1. Fit of three models to the La Calera
butterflies data set (open circles, accumulated species
number; continuous line, the fit to the exponential
model; broken line, fit to logarithmic model; dotted
line, fit to Clench model).
the observed data set was 767 times more probable than
under the next best-fitting model (Clench model). Also,
the exponential model was more than 50 million times as
probable as the logarithmic model for this data set. This is
a case where there is overwhelming evidence that favors
the exponential model over the other two models. The
software also allows displaying a likelihood plot of the
total number of species under the exponential model. In
this case, although the estimate of the TNS was 247, any
value within the 1% likelihood interval (240,257) is a plau-
sible value given the data set. Values outside this interval
can be overlooked because they are highly implausible
for the observed sample.
Vargas et al. (1994) reported results of 3 years of sam-
pling butterflies over a large transect at Atoyac, from 300
to 2500 m above sea level (asl). This range covers semide-
ciduous rain forest to pine forest. After 152 time-effort
units (person-days), 342 different species were observed.
The three models were fitted to this dataset (Fig. 2 & Ta-
ble 3). The best-fitting model was the exponential, which
was only 2.02 times more probable than the second-best-
fitting model, Clench, and 2.34 times more probable than
the logarithmic model. Although the data set slightly fa-
vors the exponential model, this is a case in which the
evidence in favor is clearly not strong. Nevertheless the
three models predicted contrasting long-term behaviors.
Table 2. Estimated parameters for models given in Table 1 for La Calera butterflies.∗
Model a b
Exponential 15.092 0.061
Clench 14.990 0.042
Logarithmic 10.986 0.005
∗ Abbreviations: TNS, total number of species; LR, likelihood ratio. The a and b are fitted parameters, and ρ is the correlation of errors.
Conservation Biology
Volume 19, No. 2, April 2005
Dı́az-Francés & Soberón
Figure 2. Fit of three models to the Atoyac butterflies
data set (open circles, accumulated species number;
continuous line, fit to the exponential model; broken
line, fit to logarithmic model; dotted line, fit to Clench
model).
The TNS under the exponential model was 361 species in
contrast to the TNS predicted under the Clench model of
509, and the logarithmic was unbounded. To be able to
discriminate sharply between the models, additional data
would need to be collected.
Discussion
Besides the models presented in Table 1, other equations
for S(t; θ) could be used for a species-accumulation func-
tion. The only requirements for the pure-birth model of
Dı́az-Francés and Gorostiza are that they must be posi-
tive, nondecreasing, and differentiable such that S(0) = 0
and that the derivative S’(t; θ) tends monotonically to 0
as t tends to infinity. The variance of the corresponding
approximating pure-birth process, as described in Dı́az-
Francés and Gorostiza, is generally not constant, and its
behavior should be taken into account to improve the
estimation procedure.
Ideally, effort units should be based on accumulated
number of individuals because the carriers of taxonomic
information are individuals (Gotelli & Colwell 2001). The
pure-birth models we propose assume that the effort is
represented as a continuous quantity, with the unit se-
lected such that the probability of observing one new
ρ TNS LR 1/LR
0.653 247 1 1
0.601 357 0.0013 767
0.688 — 1.97e–08 50,709,864
Dı́az-Francés & Soberón
Table 3. Estimated parameters for models given in Table 1 for Atoyac
butterflies.∗
Model a b ρ TNS LR 1/LR
Exponential 6.159 0.017 0.713 361 1 1
Clench 6.543 0.013 0.637 509 0.4939 2.02
Logarithmic 7.273 0.006 0.573 — 0.4268 2.34
∗ Abbreviations: TNS, total number of species; LR, likelihood ratio.
The a and b are fitted parameters, and ρ is the correlation of errors.
species within a small effort interval is much larger than
that of not seeing a new species or that of observing more
than one new species in that interval. Also, the conditions
throughout the observation period should be constant or
at least should not change significantly, and the same cap-
turing procedures should be used consistently.
Considering observed individuals as effort units
changes the setting to discrete time models. It is possi-
ble to consider observed individuals as effort units in a
continuous time model similar to a pure-birth process by
assuming that the individuals appear (or are collected)
approximately uniformly in time. This is consistent with
describing the appearance of individuals (independently
of their species) as a Poisson process in continuous time,
where the parameter of the process is the average number
of individuals observed in a unit of time, and replacing the
pure-birth process with a process that increases by one
whenever a new species is observed in the underlying
Poisson process.
This setting may not be reasonable when individuals
appear in a nonhomogeneous fashion (e.g., when more
are seen at different seasons). If special care is taken to
ensure the “uniformity” of appearance of the individuals
in time, the models in Dı́az-Francés and Gorostiza could
represent such data sets reasonably well. Nevertheless, a
modified model based on observed individuals as effort
units should be analyzed in more detail in future work.
The likelihood-ratio procedure per se is not enough to
find a good model for the data. At least one of the mod-
els entertained must adequately fit the data set. There is
no need, however, to test all reasonable models against
known standards such as complete or nearly complete in-
ventories for a wide variety of taxa and localities (Colwell
& Coddington 1994). The only requirement is to have a
reasonable probabilistic model in terms of a well-defined
parameter, so that model can be compared with other
models with the same characteristics. The selected best
model, among those considered, should always be subject
to standard diagnostic tests—as in any nonlinear regres-
sion procedure—to assess the fit of this model, in its own
right, to the observed data set (see Bates & Watts 1988:
section 3.7; Seber & Wild 1989: sections 4.6 & 5.5).These
diagnostic procedures check that data do not flagrantly
contradict any of the model assumptions.
Our results describe an appropriate application of pure-
birth-process theory to improve the estimation of species-
View publication stats
Model Selection of Species-Accumulation Functions 573
accumulation functions. The estimated values of parame-
ters with the method of Dı́az-Francés and Gorostiza are nu-
merically different from those presented in Soberón and
Llorente because they are based on methods with differ-
ent assumptions. The statistical analysis proposed by Dı́az-
Francés and Gorostiza gives mathematical rigor to the
ideas presented in Soberón and Llorente, and the freeware
provides a useful and practical method for estimating,
comparing, and selecting different species-accumulation
functions for a given data set.
Acknowledgments
This work was partially supported by the Consejo Na-
cional de Ciencia y Tecnologı́a (CONACYT) of Mexico
grants 32156-E and 37130-E. We thank L. Gorostiza and
D.A. Sprott for reading a preliminary version of the pa-
per and for useful comments. We also thank J. Golubov
for several comments that helped clarify the text and J.
Llorente, who kindly provided us with many butterfly
data sets. We thank J. Ramón Domı́nguez and the Centro
de Investigación en Matemáticas (CIMAT) software de-
velopment team for producing the species-accumulation
freeware. Finally, we thank R.K. Colwell and the referees
for constructive comments that helped to improve this
work.
Literature Cited
Bates, D. M., and D. G.Watts. 1988. Nonlinear regression analysis and its
applications. John Wiley & Sons, New York.
Clench, H. 1979. How to make regional lists of butterflies. Some
thoughts. Journal of the Lepidopterist’s Society 33:216–231.
Colwell, R. K., and J. A. Coddington. 1994. Estimating terrestrial bio-
diversity through extrapolation. Philosophical Transactions of the
Royal Society of London, B. 345:101–118.
Dı́az-Francés, E., and L. G. Gorostiza. 2002. Inference and model compar-
ison for species accumulation functions using approximating pure-
birth processes. Journal of Agricultural, Biological, and Environmen-
tal Statistics 7:29–43.
Fisher, R. A. 1973. Statistical methods and scientific inference. Hafner
Publishing, New York.
Gotelli, N. J., and R. K. Colwell. 2001. Quantifying biodiversity: proce-
dures and pitfalls in the measurement and comparison of species
richness. Ecology Letters 4:379–391.
Lamas, G. R., K. Robbins, and D. J. Harvey. 1991. A preliminary survey
of the butterfly fauna of Pakitza, Parque Nacional del Manú, Perú,
with an estimate of its species richness. Publicaciones del Museo de
Historia Natural, Universidad de San Marcos, Perú 40:1–19.
Seber, G. A. F., and C. J. Wild. 1989. Nonlinear regression. John Wiley &
Sons, New York.
Soberón, J., and J. Llorente. 1993. The use of species accumulation func-
tions for the prediction of species richness. Conservation Biology
7:480–488.
Vargas, I., J. Llorente, and A. Luis. 1994. Listado lepidopterofaunı́stico de
la Sierra de Atoyac de Alvarez en el estado de Guerrero: notas acerca
de su distribución local y estacional (Rhopalocera: Papilionoidea).
Folia Entomologica Mexicana 86:41–178 (in Spanish).
Vargas, I., A. Luis, and J. Llorente. 1996. Butterflies of the state of Jalisco,
Mexico. Journal of the Lepidopterists Society 50:97–138.
Conservation Biology
Volume 19, No. 2, April 2005