TIME-OF-DETECTION METHOD FOR ESTIMATING ABUNDANCE

FROM POINT-COUNT SURVEYS

Author(s): Mathew W. Alldredge, Kenneth H. Pollock, Theodore R. Simons, Jaime A. Collazo,
and Susan A. Shriner
Source: The Auk, 124(2):653-664. 2007.
Published By: The American Ornithologists' Union
DOI: http://dx.doi.org/10.1642/0004-8038(2007)124[653:TMFEAF]2.0.CO;2
URL: http://www.bioone.org/doi/
full/10.1642/0004-8038%282007%29124%5B653%3ATMFEAF%5D2.0.CO%3B2

BioOne (www.bioone.org) is a nonprofit, online aggregation of core research in the biological,

ecological, and environmental sciences. BioOne provides a sustainable online platform for over 170
journals and books published by nonprofit societies, associations, museums, institutions, and presses.
Your use of this PDF, the BioOne Web site, and all posted and associated content indicates your
acceptance of BioOne’s Terms of Use, available at www.bioone.org/page/terms_of_use.
Usage of BioOne content is strictly limited to personal, educational, and non-commercial use.
Commercial inquiries or rights and permissions requests should be directed to the individual publisher as
copyright holder.

BioOne sees sustainable scholarly publishing as an inherently collaborative enterprise connecting authors, nonprofit publishers, academic
institutions, research libraries, and research funders in the common goal of maximizing access to critical research.
 The Auk 124(2):653–664, 2007
© The American Ornithologists’ Union, 2007.
Printed in USA.

TIME-OF-DETECTION METHOD FOR ESTIMATING

ABUNDANCE FROM POINT-COUNT SURVEYS
Mathew W. Alldredge,1,4 Kenneth H. Pollock,1 Theodore R. Simons,2
Jaime A. Collazo,2 and Susan A. Shriner3
1
Colorado Division of Wildlife, 317 West Prospect Road, Fort Collins, Colorado 80526, USA;
2
U.S. Geological Survey, North Carolina Cooperative Fish and Wildlife Research Unit, Department of Zoology,
Campus Box 7617, North Carolina State University, Raleigh, North Carolina 27695, USA; and
3
Department of Fishery and Wildlife Biology, Colorado State University, Fort Collins, Colorado 80523, USA

Abstract.—Point-count surveys are o en used to collect data on the abundance

and distribution of birds, generally as an index of relative abundance. Valid com-
parison of these indices assumes that the detection process is comparable over space
and time. These restrictive assumptions can be eliminated by estimating detection
probabilities directly. We generalize a recently proposed removal model for estimat-
ing detection probabilities using a time-of-detection approach, which can account
for more sources of variation in point-count data. This method is specifically
designed to account for variation in detection probabilities associated with singing
rates of birds. Our model accounts for both availability bias and detection bias by
modeling the combined probability that a bird sings during the count, and the prob-
ability that it is detected given that it sings. The model requires dividing the count
into several intervals and recording detections of individual birds in each interval.
We develop maximum-likelihood estimators for this approach and provide a full
suite of models based on capture–recapture models, including covariate models. We
present two examples of this method: one for four species of songbirds surveyed
in Great Smoky Mountains National Park using three unequal intervals, and one
for the Pearly-eyed Thrasher (Margarops fuscatus) surveyed in Puerto Rico using
four equal intervals. Models incorporating individual heterogeneity were selected
for all data sets using information-theoretic model-selection techniques. Detection
probabilities varied among count-time intervals, which suggests that birds may be
responding to observers. We recommend applying this method to surveys with four
or more equal intervals to reduce assumptions and to take full advantage of stan-
dard capture–recapture so ware. The time-of-detection approach provides a better
understanding of the detection process, especially when singing rates of individual
birds affect detection probabilities. Received 20 July 2004, accepted 14 June 2006.

Key words: abundance estimation, bird surveys, capture–recapture, detection

probability, point counts.

Estimación de la Abundancia en Puntos de Conteo Mediante el Método del Tiempo

de Detección

Resumen.—Es común usar censos en puntos de conteo para recolectar datos

de abundancia y distribución de aves, generalmente como índices de abundancia
relativa. Una comparación válida de estos índices supone que el proceso de detección
es comparable en el espacio y en el tiempo. Estas restricciones pueden eliminarse
mediante la estimación de las probabilidades de detección. Generalizamos un

4
E-mail: mat.alldredge@state.co.us

653
654 Alldredge et al. [Auk, Vol. 124

modelo de remoción propuesto recientemente para la estimación de probabilidades

de detección usando un enfoque basado en el tiempo de detección, el cual considera
diversas fuentes de variación en los datos obtenidos de puntos de conteo. Este método
fue diseñado específicamente para considerar la variación en las probabilidades de
detección asociadas con las tasas de canto de las aves. Nuestro modelo considera
tanto el sesgo de disponibilidad como el sesgo de detección mediante el modelado
de la probabilidad combinada de que un ave cante durante el conteo, y de que sea
detectada dado que canta. El modelo requiere dividir el conteo en varios intervalos
y registrar la detección de aves individuales en cada intervalo. Desarrollamos
estimadores de máxima verosimilitud para este enfoque y brindamos un paquete
completo de modelos basados en modelos de captura-recaptura, incluyendo modelos
de covariación. Presentamos dos ejemplos de este modelo: uno para cuatro especies
de aves canoras muestreadas en el Parque Nacional Great Smoky Mountains usando
tres intervalos desiguales y otro para la especie Margarops fuscatus, muestreada en
Puerto Rico usando cuatro intervalos iguales. Seleccionamos modelos que incorporan
la heterogeneidad individual para todos los datos usando técnicas de la teoría de
la información. Las probabilidades de detección variaron entre los intervalos de
tiempo, lo que sugiere que las aves podrían estar respondiendo a los observadores.
Recomendamos aplicar este método a los muestreos con cuatro o más intervalos
iguales para reducir las restricciones y para aprovechar al máximo los programas
de computación de captura-recaptura. El enfoque del tiempo de detección brinda
un mejor entendimiento del proceso de detección, especialmente cuando las tasas de
canto de las aves individuales afectan las probabilidades de detección.

Point-count surveys are routinely used to Here, we present a time-of-detection ap-

determine animal abundance, particularly for proach that uses a closed-population capture–
breeding birds (Thompson 2002). Nevertheless, recapture framework (Otis et al. 1978, Williams
statistically valid methods for collecting and et al. 2002), which provides a more general
analyzing such data are not common practice alternative to the removal method. The full
(Rosenstock et al. 2002). Many studies do not time-of-detection approach uses both first
correct for the detection process, relying instead detections and all subsequent detections of an
on raw count data as an index of abundance individual bird to estimate the probability of
(Ralph et al. 1995, Rosenstock et al. 2002). detection. Seber (1982) showed that capture–
Estimating detection probabilities is essential recapture models are more efficient (smaller
for obtaining reliable estimates of population variance) than removal models. Because the
abundance (Nichols et al. 2000, Farnsworth et time-of-detection method uses information
al. 2002, Thompson 2002). Farnsworth et al. on all detections of an individual bird, there
(2002) and Alldredge (2004) provided descrip- is greater ability to model potential sources
tions of the two primary components of the of variation in the data. However, capture–
detection process on avian point counts: avail- recapture models require that observers accu-
ability and detection given availability. Ignoring rately track individual birds throughout the
either of these components can lead to biased count. Errors in the capture history will bias
abundance estimates (Marsh and Sinclair 1989). the abundance estimates. We first outline the
When birds are detected by song or call, avail- field methods required to collect data suitable
ability is a function of singing rate and duration for this method and then describe the candi-
of the count. Farnsworth et al. (2002) advanced date models and the general form of the model.
a removal model to account for both of these We present example analyses using data from
components. The method is based on dividing a four equal-interval point counts, and from
count into multiple intervals and recording the three unequal-interval point counts, and dis-
interval in which an individual is first detected cuss potential uses and insights, particularly
and uses closed-population removal models to regarding the importance of singing rates in
estimate abundance. point-count surveys.
April 2007] Time-of-detection Method 655

Methods detection probabilities between the first detec-

tion of a bird and all subsequent detections may
Approach and candidate models.—Recording occur when observers key-in on previously
data on when individual birds are detected detected birds. It is also likely that detection
during a point-count survey is useful for con- probabilities differ between individual birds
structing a traditional detection history for of a given species. This may be from observ-
each bird (a one if it is detected in an interval, able differences between individuals, such as
a zero if it is not). Shriner (2001) used multi- differences in detection distance, or from unob-
colored pens to simultaneously map detections servable differences, such as topographic differ-
and record the time interval of detections. ences associated with the location of the bird,
Detections in intervals following initial detec- differences in singing rates, singing volume, or
tion were recorded by underlining previously other factors causing differences in detection
mapped birds. probabilities among birds. There are eight gen-
The full detection-history data for a time-of- eral capture–recapture models (Otis et al. 1978,
detection survey consist of the number of birds White et al. 1982, Pollock et al. 1990) that incor-
observed with each of the 2n – 1 (n = number porate these sources of variation.
of intervals) possible detection histories. For Model M0: Equal detection probability for all
example, if three periods are used, seven detec- individuals among all periods.
tion histories are possible (xw: x111, x110, x101, x011, Model Mt: Equal detection probability for all
x100, x010, and x001), compared with the three individuals but different detection prob-
possible detection histories for the removal abilities among periods.
method (x100, x010, and x001). Zeroes in the detec- Model Mb: Equal probability of first detec-
tion history represent either individuals that tion for all individuals among all periods
did not sing during a given interval (i.e., were and a unique probability of subsequent
not available) or individuals that sang but were detections that is equal for all individuals
not detected by the observer, which is why among all subsequent periods.
this method accounts for both availability and Model Mtb: Equal probability of first detec-
detection probability given availability. tion for all individuals but different among
The general assumptions for this method are periods and a unique probability of subse-
that (1) there is no change in the population of quent detection that is equal for all indi-
birds within the sample area during the point viduals but different among subsequent
count (population is closed); (2) observers periods.
accurately track individual birds throughout Model Mh: Unique probability of detection
the count (i.e., there is no double-counting); for each individual that remains constant
and (3) if counts are done with a limited radius, among all periods.
observers accurately assign birds to within the Model Mth: Unique probability of detection
radius used. Additional assumptions are made for each individual that differs among
with the various models that account for differ- periods.
ent sources of variation. Model Mbh: Unique probability of first cap-
The time-of-detection method provides mod- ture for each individual that remains
els that can account for sources of variation constant among periods and a unique
associated with the time interval (t), changes probability of subsequent detection that
in detection probability following first detec- remains constant among periods.
tion (b; behavior models of Otis et al. 1978), and Model Mtbh: Unique probability of first
individual differences in detection probability capture for each individual that differs
(h; heterogeneity models of Otis et al. 1978). among periods and a unique probability
Variation in detection probabilities associated of subsequent detection that differs among
with the interval arises from situations where periods.
the probability of detecting a bird changes Models Mb and Mbh are equivalent to the
during the count. For example, time variation removal models given by Farnsworth et al.
might arise in situations where observers affect (2002). Model Mtbh cannot be fit without several
the singing rates of birds over time (McShea assumptions to constrain the number of param-
and Rappole 1997). Similarly, differences in eters (Williams et al. 2002).
656 Alldredge et al. [Auk, Vol. 124

There are three possible estimators for mod- when counts consist of two or more equal
eling individual differences in detection proba- intervals. The “Huggins closed captures” data
bilities (heterogeneity models) when sources of type, which allows for individual covariates,
variation are not observable, but we restricted is used in MARK. Model selection is based
our analyses to finite mixture models. Finite on AIC. Akaike’s Information Criterion is an
mixture models may be biologically reasonable information-theoretic approach used to select
for this situation, because the mixtures can be the most parsimonious (best tradeoff between
explained by the proportion of the population squared bias and variance of parameter estima-
with a particular singing rate, which corre- tors) model that explains the variation in the
sponds to detection probabilities. Finite mixture data (Burnham and Anderson 2002). In many
models of individual differences are likelihood- situations, AICc should be used, which is a
based estimators (Norris and Pollock 1996, second-order AIC that corrects for small sample
Pledger 2000) that assume that a population can size (Burnham and Anderson 2002).
be subdivided into a finite number of groups Unequal-interval point-count analysis.—Standard
with distinct detection probabilities. The advan- capture–recapture programs and their suite of
tage of having maximum-likelihood-based esti- candidate models cannot accommodate point
mators is that (1) models can be analyzed with counts divided into unequal intervals, because
likelihood-ratio tests for evaluating sources of in this case interval detection probabilities
variation and (2) model selection techniques, must be modeled as a detection rate. Models
such as Akaike’s Information Criterion (AIC), that include variation in detection probabilities
can be used to select the most parsimonious among intervals could still be examined with a
model (Williams et al. 2002). standard model, but not models without time
Covariates can also be included in all the variation. For situations where detection rates
above models to account for individual differ- are constant, interval detection rates will still
ences in detection probabilities among birds. vary because of the differing lengths of each
These differences are caused by observable interval. Therefore, it is necessary to model
sources of variation (sources that can be mea- interval detection probabilities as a function
sured during the point count), such as detection of time and detection rate. For example, model
distance, time of day, time of year, weather con- M0 would indicate a constant detection rate, but
ditions, singing rate, etc. interval detection probabilities would differ
Detection probability is modeled as a function because of differences in interval lengths.
of individual covariates. An intercept (β0) and a One approach to modeling the detection
slope (β1) parameter are estimated to model the process assumes that singing rates follow a
detection process. If, for example, detection Poisson process and that the probability of
probability is modeled as a function of distance detecting all songs of a species is equal. Under
on the logit scale, then, a er transforming back these assumptions, the instantaneous detection
to the real parameter estimates, detection prob- rate (ϕi) or the Poisson detectability coefficient
ability plotted against distance will appear is used to model the probability that a bird
similar in shape to a half-normal. The original sings during a specified period. The probabil-
suite of models described can be viewed as ity of detecting an individual in time interval i
intercept-only models. The effect of the covari- of length ti using an instantaneous rates formu-
ate term on the slope parameter is modeled as lation (Seber 1982) is
either a constant effect (additive) over time or a
variable effect (interaction) over time. Thus, in p i = 1 − e − ϕiti (1)
addition to the original 7 models, an additional
14 models parameterized with either additive This formulation is consistent with that typi-
or interaction covariate effects are possible, for cally used for removal experiments where
a total of 21 conceptual models. ϕi corresponds to the “Poisson catchability
Equal-interval point-count analysis.—Standard coefficient” and ti corresponds to the effort on
capture–recapture so ware, such as CAPTURE the ith occasion (Otis et al. 1978, Seber 1982).
(White et al. 1982) or MARK (White and Farnsworth et al. (2002) used an alternative for-
Burnham 1999), can provide estimates of mulation, the discrete rate formulation, which
detection probabilities and population size assumes a constant per-minute detection rate
April 2007] Time-of-detection Method 657

(γi), so that the probability of detecting an indi- The full multinomial likelihood can be written
vidual in interval i of length ti was given by as

t
p i = γ ii (2) L( N , ϕi , νi ,| xw ) =
N − ∑ xw
π ww ⎛⎜ 1 − ∑ π w ⎞⎟ w
N!
∏
x
When point counts consist of equal intervals,
it is not necessary to use equations (1) or (2), ∏ xw ! ⎛⎜⎝ N − ∑
w
xw ⎞ ! w
⎟
⎠
⎝ w ⎠
w (5)
because interval detection probabilities are on
the same time scale and, thus, are estimated
directly. Where xw is the number of observations for each
Here, we present the likelihoods for the possible detection history, excluding the history
instantaneous rate formulation. For a full devel- for the individuals never detected and πw (equa-
opment of these models, including discrete rate tion 4) is the probability of observing each cap-
models, see Alldredge (2004). The models that ture history, excluding the probability of never
account for differences in subsequent detection seeing an individual.
probabilities (e.g., model Mb) model the change Because N cannot be directly observed, we
in the detection probability for subsequent condition on the total number of birds counted
detections (ν) as (xT) to make the problem more amenable to
numerical methods. This is necessary when
ci = 1 − e −(ϕi + ν)ti (3) using SURVIV (White 1983) but may not be
necessary when using other so ware packages.
where ci is the interval detection probability for To do this, the likelihood is decomposed into a
all subsequent detections. Excluding the models marginal distribution L1 and a conditional dis-
with individual differences in detection prob- tribution L2. Then, the detection probabilities
ability, model Mtb is the most general model can be estimated from the conditional distribu-
and all other models are constrained forms of tion, and abundance can be estimated from the
this general model. First we present this model, marginal distribution. The relationship between
and then we present the models with individual these likelihoods is L = L1 × L2. Marginal likeli-
differences. hood L1 can be written as
A count consisting of four intervals has 15
possible observable detection histories (xw: x1111, ⎛N⎞
x1110 , … , x0001). The set of possible detection L1 ( N | xT ) = ⎜ ⎟ ( πi )xT (1 − πi )N − xT (6)
probabilities for the observable capture histo- ⎝ xT ⎠
ries is denoted by πw. The expected values of the
counts for each detection history of the general where π• is the sum of the πw. The likelihood L2
model can be written as is conditional on the observed capture histories,
given by
E( x1111 ) = N(1 − e − ϕ1t1 )(1 − e −(ϕ2 + ν)t2 )
xw
⎛ xT ⎞ ⎛ πw ⎞
(1 − e −(ϕ3 + ν)t3 )(1 − e −(ϕ4 + ν))t4 ) = N π1111 L2 (ϕi , νi | xw ) = ⎜ ⎟∏⎜ ⎟ (7)
x
⎝ 1111 … x 0001 ⎠ w ⎝ πi ⎠
E( x1110 ) = N(1 − e − ϕ1t1 )(1 − e −(ϕ2 + ν)t2 )
(1 − e −(ϕ3 + ν)t3 )( e −(ϕ4 + ν)t4 ) = N π1110 From equation (7), the ϕi’s and ν’s can be esti-

(4) mated by maximizing the likelihood for the
observed data. The probability that an indi-
E( x1010 ) = N(1 − e − ϕ1t1 )( e −(ϕ2 + ν)t2 )
vidual is detected at least once during the count
(1 − e −(ϕ3 + ν)t3 )( e −(ϕ4 + ν)t4 ) = N π1010 (p̂T) can be calculated by


(8)
E( x0001 ) = N( e − ϕ1t1 )( e − ϕ2t2 )( e − ϕ3t3 )
(1 − e − ϕ4t4 ) = N π0001 The variance of p̂T is obtained for each estimator
by reparameterizing to estimate p̂T directly. This
can be done by solving one of the equations of
658 Alldredge et al. [Auk, Vol. 124

pT for the last interval parameter and using this the proportion of animals in each group times
in the likelihood. For example, if we solve equa- the group-specific capture and recapture prob-
tion (8) for ϕ4, we get abilities. For example, the expected value of the
count for the four-interval detection history for
1 ⎛ 1 − pT ⎞ individuals detected in all intervals is
ϕ4 = ln
t4 ⎜⎝ eϕ1t1 eϕ2t2 eϕ3t3 ⎟⎠ (9)
E( x1111 ) = N(λ(1 − e − ϕ11t1 )(1 − e −(ϕ21 + ν1 )t2 )
Replacing the final interval parameter with this (1 − e −(ϕ31 + ν1 )t3 )(1 − e −(ϕ41 + ν1 )t4 ) +
equation will allow for direct estimation of the (12)
probability that an individual is detected at least (1 − λ )(1 − e − ϕ12t2 )(1 − e −(ϕ22 + ν2 )t2 )
once and the associated variance. An alternative (1 − e −(ϕ32 + ν2 )t3 )(1 − e −(ϕ42 + ν2 )t4 )
approach is to use the delta method (Seber 1982)
and obtain estimators for the approximate vari- where λ is the proportion of animals in detec-
ances. Abundance is estimated from the likeli- tion group one, and the proportion of animals in
hood L1 as detection group two is 1 – λ as they must occur
in one of the groups, and the ϕij’s and νj’s are the
(10) probabilities of first detection and subsequent
detection in the ith interval for individuals in
The observed count (xT) is one realization of the jth group. There are similar expressions for
a random variable and, thus, has a variance the other xw. The conditional likelihood L(ϕij, νj ,
associated with it. Assuming that the observed λ| xw) is similar to equation (7):
count (xT) is from a binomial distribution and
that xT and p̂T are independent, an estimate of L2 (ϕij , ν j , λ | xw ) =
the variance of abundance is (Nichols et al. 2000,
x
Williams et al. 2002) ⎛ xT ⎞ ⎛ π w1 + π w 2 ⎞ w (13)
⎜x ⎟ ∏ ⎜ πi ⎟
⎝ 1111 ...x0001 ⎠ ∀w ⎝ ⎠
(11)
where π• is the probability of being detected
at least once during the entire count, πw1 is the
In general, four or more intervals are probability of being in the first group and hav-
required to parameterize heterogeneity mod- ing capture history w, and πw2 is the probability
els, unless very strong assumptions are made, of being in the other group and having capture
as in Farnsworth et al. (2002). We developed history w. Using this likelihood, the ϕij’s and νj’s
heterogeneity models based on ≥4 intervals. The can be estimated by maximizing the likelihood
assumptions required for the restricted hetero- for the observed data.
geneity models are given with the presentation The probability that an individual is detected
of the three interval examples. at least once during the count (p̂T) is calculated
We used a two-point finite mixture model using the estimated detection coefficients:
because a maximum-likelihood approach
allowed us to use information-theoretic model-
selection procedures. Pledger (2000) found that (14)
two-point mixtures generally provide the most
parsimonious models and best estimators. Our
approach is easily extended to more mixtures if for the equal interval formulation. The variance
appropriate. The model is presented using four for p̂T can again be calculated by reparameteriz-
sampling intervals. ing the model estimators to directly estimate p̂T.
Data from four interval point counts are Equations (10) and (11) are then used to estimate
summarized using the counts for the 15 observ- abundance and the corresponding variance.
able detection histories xw. Assuming that the Parameter estimation for unequal interval
population comprises only two groups (two data is done by maximizing the likelihood for
mixtures), the probabilities for each capture his- the observed data. This can be done in SURVIV
tory are calculated by summing the products of (White 1983). Model selection is based on AICc.
April 2007] Time-of-detection Method 659

Field Trials different colored pens for each interval. Point

counts were unlimited radius plots, and detec-
We present examples from two different field tion distance was recorded for all detections.
studies, one using point counts using four equal Example analyses are presented for Red-eyed
intervals and the other from point counts using Vireo (Vireo olivaceus), Black-throated Green
three unequal intervals. The first example, for Warbler (Dendroica virens), Ovenbird (Seiurus
the Pearly-eyed Thrasher (Margarops fuscatus), aurocapilla), and Hooded Warbler (Wilsonia
is based on point counts conducted in the karst citrina), all of which have relatively loud songs
belt of north-central Puerto Rico in 2003. Data and high singing rates. All models (except
were collected during the breeding season the covariate models) were evaluated using
(mid-February through May). Surveys were con- SURVIV. We truncated 10% of the data by omit-
ducted using three teams of two experienced and ting observations with the largest detection dis-
trained observers each. Forty-nine routes of 1–29 tances to omit outliers. Truncation would also
(average 9.7) point counts each were sampled, serve to define a detection radius for each spe-
for a total of 477 points. Point-count routes were cies if density estimates were required. Model
located away from human habitation along “low selection was based on AICc (Burnham and
use” trails. The first point was located 500 m Anderson 2002).
from the start of the trail and subsequent points We modeled heterogeneity using constrained
were located 200 m from the previous. forms of the two-point mixture models. These
Point counts were conducted from 0400 to models must be constrained so that all param-
0800 hours on days with suitable weather. Each eters are identifiable. Like Farnsworth et al.
count was conducted for 10 min. Counts were (2002), we constrained our models by fixing the
divided into four equal intervals, and the com- detection probabilities for one of the groups in
plete detection history was recorded for each the mixture to 1. This parameterization makes
individual by using different colored pens for the very strong assumption that all individu-
each interval. Point counts were fixed 100-m als in the fixed detection group are detected in
radius plots, and detection distance within this every interval of the survey.
plot was recorded for all detections. During
the previous year, distances at each plot were Results
flagged at 10-m intervals to give observers
visual references to help verify distance esti- Four-interval data set.—Sixteen of the 21 con-
mates during a count. ceptual models for the four-interval Pearly-eyed
The Pearly-eyed Thrasher data were ana- Thrasher data set gave reasonable results. The
lyzed using the full suite of models, including other five models had parameters that were
heterogeneity and detection-distance covariate not identifiable. The most parsimonious model
models. Both additive and interaction effects selected for this data set was model Mth (∆AICc
between detection distance and time were used. weight = 0.70), with the remaining support for
All analyses were done using MARK. Data model Mth (additive distance effect) (∆AICc
were not truncated, because a fixed radius plot weight = 0.30) (Table 1). For all models, the
was used. Model selection was based on AICc general form of the model and its two covari-
(Burnham and Anderson 2002). ate formulations were always ordered together
The other example is based on point-count according to AICc model selection. Heterogeneity
data collected in Great Smoky Mountains models were always “better” than models not
National Park in 1998. Analyses were restricted accounting for heterogeneity. Of the models that
to data from a single year and a single observer did not account for unobservable heterogeneity,
to avoid temporal and observer effects. Point models incorporating detection distance always
counts were conducted along pre-established had lower AICc values.
survey routes along trails (Shriner 2001). Counts Under the selected model, 29% of the popula-
were conducted in May and June between dawn tion had low detection probabilities, whereas the
and 1000 hours under acceptable environmental remaining portion had high detection probabili-
conditions. Counts were divided into 3-, 2-, and ties (Table 2). Low detection probabilities ranged
5-min intervals, and the complete detection his- from 0.09 to 0.56, and high detection prob-
tory was recorded for each individual by using abilities ranged from 0.70 to ∼1.00. Because the
660 Alldredge et al. [Auk, Vol. 124

Table 1. Values of ∆AICc for the four-interval Pearly-eyed Thrasher data set. A value of 0.0 indicates
the most parsimonious model that adequately fits the data. When ∆AICc weight > 0, ∆AICc
weight is given in parentheses. Distance is modeled as either an additive effect with time or an
interaction effect with time.

Model No distance Additive distance Interaction distance

M0 170.2 165.2 164.2
Mt 148.1 143.1 139.7
Mb 102.3 99.5 97.4
Mtb –a –a
Mh 72.2 70.3 53.4
Mbh 23.3 13.7 –a
Mth 0.0 (0.7) 1.7 (0.3) –a
a
Models not included because of unreasonable parameter estimates.

Table 2. Estimated detection probabilities pij , probability of group occurrence λj , and standard

errors (SE) for interval i and group j of the four-interval Pearly-eyed Thrasher data set based on
model Mth. Standard error for group two and interval two is not estimable.

Group 1 Group 2
Parameter Estimate (p̂ij) SE (p̂ij) Estimate (p̂ij) SE (p̂ij)
p 1j 0.21 0.036 0.78 0.030
p 2j 0.09 0.090 ∼1.0 –a
p 3j 0.47 0.050 0.84 0.023
p 4j 0.56 0.055 0.70 0.026
λj 0.29 0.037 0.71 0.037
a
Models not included because of unreasonable parameter estimates.

detection probability was essentially 1 for one for models incorporating both heterogeneity
of the heterogeneity groups during the second and behavior, which indicates that detection
interval, abundance for this group is equivalent probabilities for birds did not change a er first
to the count. Note that the standard error (SE) is detection.
not estimable for parameter estimates near the Interval-detection probabilities showed a
boundary. The total number of observations for consistent pattern; the shortest intervals had the
this data set was 520, and the estimated abun- smallest detection probabilities, and the longest
dance for the sampled area was 547 (SE = 8.6), intervals had the highest detection probabilities
which is 5% higher than the observed count. (Table 4). This was not true for the Red-eyed
Three-interval data set.—Heterogeneity mod- Vireo, which had a detection probability of 0.41
els were the most parsimonious for the three for the 3-min interval and 0.48 for the 2-min
unequal-interval data sets from Great Smoky interval. This may indicate an observer effect
Mountains National Park (Table 3). The ∆AICc that made this species less detectable during the
weights for all models without heterogeneity first interval. The overall probability of detect-
were always negligible, indicating no evidence ing an individual at least once during a 10-min
supporting these models. Model Mth was count was 0.92 (SE = 0.006) for the Ovenbird,
selected as the most parsimonious for three of 0.79 (SE = 0.021) for the Black-throated Green
the species, whereas Model Mh was selected for Warbler, 0.71 (SE = 0.031) for the Red-eyed
the Ovenbird. All species except the Ovenbird Vireo, and 0.65 (SE = 0.037) for the Hooded
showed evidence for time variation in instanta- Warbler. Comparing the observed counts to
neous rates of detection, indicating that detec- the estimated abundance in the sample area
tion probabilities do not remain constant for showed differences of 5% (Ovenbird) to 16%
the duration of counts. There was little support (Hooded Warbler).
April 2007] Time-of-detection Method 661

Table 3. Values of ∆AICc for the 11 time-of-detection models fit to the Great Smoky Mountains
National Park data sets. For the most parsimonious model for each data set, ∆AICc = 0.0. ∆AICc
weights (in parentheses) indicate the strength of the evidence for a given model compared with
the other models (the larger the number, the more evidence for that model).

Species M0 Mt Mb Mtb a Mh Mbh Mth

Red-eyed Vireo 75.3 55.5 75.0 – 7.7 9.7 0.0
(0.00) (0.00) (0.00) (0.02) (0.01) (0.97)
Black-throated Green Warbler 41.7 38.9 43.7 – 0.3 2.3 0.0
(0.00) (0.00) (0.00) (0.40) (0.14) (0.46)
Ovenbird 44.3 28.5 46.3 – 0.0 2.0 2.7
(0.00) (0.00) (0.00) (0.62) (0.22) (0.16)
Hooded Warbler 51.8 40.7 53.9 – 10.1 12.2 0.0
(0.00) (0.00) (0.00) (0.01) (0.00) (0.99)
a
Parameter estimates for model Mtb were not reasonable (SE > 1).

Table 4. Parameter estimates from the selected model for the three-interval Great Smoky Mountains
National Park point-count data sets. λ1 is the proportion of the population that is in group 1.
Detection probability (pij) is the probability of detecting an individual from group j in interval i.
Detection probabilities for group 2 (pt2) were fixed to 1. Standard errors (SE) are in parentheses.

Species Observed λ̂1 p̂ 11 p̂ 21 p̂ 31 p̂ t2 N̂

Red-eyed Vireo 397 0.66 0.41 0.48 0.55 1.0 457
(0.033) (0.039) (0.040) (0.040) (15.9)
Black-throated 377 0.71 0.52 0.40 0.63 1.0 409
Green Warbler (0.032) (0.040) (0.040) (0.039) (8.83)
Ovenbird 444 0.72 0.56 0.43 0.75 1.0 468
(0.036) (0.042) (0.044) (0.038) (7.03)
Hooded Warbler 274 0.72 0.45 0.35 0.46 1.0 318
(0.035) (0.045) (0.043) (0.046) (11.50)

Discussion simpler models should not be ignored. Clearly,

heterogeneity exists in these data, as evidenced
The time-of-detection method of modeling by the large differences in AIC among models;
point-count data collected from consecutive however, heterogeneity models should be inter-
intervals is a less restrictive approach for esti- preted with caution. Link (2003) demonstrated
mating detection probabilities than the removal that there are difficulties or inconsistencies in
method of Farnsworth et al. (2002). Recording distinguishing among alternative heterogeneity
the complete detection history of birds during models. We are confident that variation from
a point count provides a much larger suite individual differences in detection probability
of models that can incorporate more sources is a component of these data and must be mod-
of variation, such as time variation, than the eled, but we acknowledge that the estimates
removal model. With the exception of behavior from heterogeneity models may also be biased.
models that are equivalent to removal models Many factors cause unobservable individual
(Seber 1982), this approach is more efficient heterogeneity in detection probabilities. These
(smaller variance) than the removal approach. factors are o en related to variation in the
Heterogeneity models accounting for unob- behavior of individual birds (Burnham 1981,
servable differences in detection probabilities Johnson et al. 1986), such as age and breeding
among birds were selected as the best models status (mated, unmated, incubating, etc.). Other
for all data sets. The AIC model-selection tech- factors affecting detection probabilities include
niques tend to select the fuller models, such habitat differences associated with topography
as the heterogeneity models; therefore, the and vegetation, weather, time of day, time of
662 Alldredge et al. [Auk, Vol. 124

year, and the presence of predators or competi- Short-duration counts reduce violations of the
tors. Although stratification by habitat type and closure assumption. The probability that birds
standardization of the conditions under which move into or out of the sample area increases
point counts are conducted can eliminate some with the duration of the count. This does not
differences among points (Buckland et al. 1993, imply that all point counts should be arbitrarily
Nichols et al. 2000), unobservable heterogeneity short, because there is a tradeoff associated with
exists in all data (Burnham 1981). count duration and the need for four intervals
A particularly important source of hetero- to fit the full heterogeneity models. If intervals
geneity in auditory point counts is individual are too short, interval detection probabilities
variation in singing rate. Singing rates change will be small and the variance on abundance or
during the breeding season as birds form pairs, density estimates will be large. Careful consid-
begin incubation, and begin caring for nest- eration must be given to the appropriate length
lings (Wasserman 1977, Lein 1981). Therefore, of point counts for the species surveyed. This
asynchronous breeding inevitably results in may imply different survey protocols for differ-
temporal heterogeneity in detection probabili- ent species groups.
ties. Singing rates are also affected by habitat, Assumption (2): Individuals can be accurately
local abundance, and the proximity of observers tracked throughout the count (i.e., no double-
(McShea and Rappole 1997). counting). Double-counting results in abun-
Evidence for a time effect was found in four dance estimates that are too large. Problems
of the five data sets presented here. Therefore, it with double-counting are likely to increase
appears that the assumption of constant detec- as the length of the count increases because
tion rates under the removal model may not be of undetected movement of birds. Infrequent
valid. McShea and Rappole (1997) found that singing by individuals may also lead to double-
singing rates were affected by the presence of counting, because of limited ability of observers
an observer. Movement of an individual during to track these birds throughout the survey. Our
the count would also affect the detection process, method requires observers to track individuals
because as birds move away from observers they during the count, which should reduce double
are less likely to be detected. This assumption is counting. However, violations of this assump-
not necessary with our method, because there is tion related to bird movement are more likely as
sufficient information to model time variation count durations increase.
from the full detection history. Assumption (3): Distance estimates are accu-
The time-of-detection approach also reduces rate. Although observers are o en trained to
the number of assumptions required to esti- estimate detection distance or assign birds to
mate abundance. The removal method for three fixed-radius plots, the accuracy and precision
intervals had five assumptions (Farnsworth et al. of distance estimation on auditory point counts
2002), which could be relaxed to four if more than has not been rigorously assessed. We suspect
three intervals were used. The time-of-detection that, even with training, observers tend to over-
approach has only three assumptions for counts count individuals within fixed-radius plots,
conducted with more than three intervals. especially for louder species, and that the ability
Assumption (1): there is no change in the of observers to estimate detection distances for
population within the detection radius during singing birds may be poor.
the point count (closed population). Violations As with other point-count methods that
of the closure assumptions are more likely to allow for abundance estimation, this method
occur for longer-duration point counts and for requires collection of additional information
wide-ranging species. This method may not be during the point count. This extra demand on
applicable for wide-ranging species, such as observers may make the method impractical in
woodpeckers (Picidae) or crows (Corvus spp.), some situations. For example, if the number of
where movement during the count may be sig- birds and the number of species at a point are
nificant. Violations of the closure assumption large, it is unlikely that observers can accurately
are probably less of a problem for many small record the detection history of every bird. In
breeding songbirds that have relatively small, such cases, other methods, or limiting the col-
fixed territories during the breeding season lection of time-of-detection data to a subset of
(Farnsworth et al. 2002). species, might be more appropriate.
April 2007] Time-of-detection Method 663

Our findings suggest that singing rates have in Estimating Numbers of Terrestrial Birds
a strong influence on detection probabilities, (C. J. Ralph and J. M. Scott, Eds.). Studies in
but further work is needed in this area. The Avian Biology, no. 6.
time-of-detection approach we have described Burnham, K. P., and D. R. Anderson. 2002.
should prove useful in this effort. We recom- Model Selection and Multimodel Inference:
mend implementing this method with four or A Practical Information-theoretic Approach,
more equal intervals, because it permits the 2nd ed. Springer-Verlag, New York.
use of standard capture–recapture so ware Farnsworth, G. L., K. H. Pollock, J. D. Nichols,
and allows for the application of full two-point T. R. Simons, J. E. Hines, and J. R. Sauer.
mixture models. 2002. A removal model for estimating detec-
Farnsworth et al. (2002) recommended com- tion probabilities from point-count surveys.
bining the removal approach with the distance- Auk 119:414–425.
sampling approach to provide better estimates. Johnson, D. H., K. P. Burnham, and J. D.
We have done this by incorporating detection Nichols. 1986. The role of heterogeneity in
distance as a model covariate. Further develop- animal population dynamics. Proceedings
ment is needed to incorporate alternative forms of the International Biometrics Conference,
of the detection function. 13:1–15.
Finally, because point-count data usually con- Lein, M. R. 1981. Display behavior of Ovenbirds
sist of multiple species, we recommend using a (Seiurus aurocapillus) II. Song variation and
multiple-species modeling approach (Alldredge singing behavior. Wilson Bulletin 93:21–41.
2004, Alldredge et al. 2007). Modeling multiple Link, W. A. 2003. Nonidentifiability of popula-
species with similar detection probabilities may tion size from capture–recapture data with
provide more parsimonious models. This is heterogeneous detection probabilities.
especially useful for rare or hard-to-detect spe- Biometrics 59:1123–1130.
cies. However, it would also require additional Marsh, H., and D. F. Sinclair. 1989. Correcting
work to determine which factors are affecting the for visibility bias in strip transect aerial sur-
detection process to appropriately group species. veys of aquatic fauna. Journal of Wildlife
Management 53:1017–1024.
Acknowledgments McShea, W. J., and J. H. Ra ole. 1997. Variable
song rates in three species of passerines
This project was funded by the U.S. Geological and implications for estimating bird popu-
Survey, the National Park Service, and the Puerto lations. Journal of Field Ornithology 68:
Rico Department of Natural and Environmental 367–375.
Resources. We appreciate the suggestions by two Nichols, J. D., J. E. Hines, J. R. Sauer, F. W.
anonymous reviewers and D. Johnson. Fallon, J. E. Fallon, and P. J. Heglund.
2000. A double-observer approach for esti-
Literature Cited mating detection probability and abundance
from point counts. Auk 117:393–408.
Alldredge, M. W. 2004. Avian point-count Norris, J. L., III, and K. H. Pollock. 1996.
surveys: Estimating components of the Nonparametric MLE under two closed
detection process. Ph.D. dissertation, North capture–recapture models with heterogene-
Carolina State University, Raleigh. ity. Biometrics 52:639–649.
Alldredge, M. W., K. H. Pollock, T. S. Simons, Otis, D. L., K. P. Burnham, G. C. White, and
and S. A. Shriner. 2007. Multiple-species D. R. Anderson. 1978. Statistical inference
analysis of point count data: A more parsi- from capture data on closed animal popula-
monious modeling framework. Journal of tions. Wildlife Monographs, no. 62.
Applied Ecology 44:281–290. Pledger, S. 2000. Unified maximum likelihood
Buckland, S. T., D. R. Anderson, K. P. Burnham, estimates for closed capture–recapture mod-
and J. L. Laake. 1993. Distance Sampling: els using mixtures. Biometrics 56:434–442.
Estimating Abundance of Biological Pollock, K. H., J. D. Nichols, C. Brownie, and
Populations. Chapman and Hall, London. J. E. Hines. 1990. Statistical inference for
Burnham, K. P. 1981. Summarizing remarks: capture–recapture experiments. Wildlife
Environmental influences. Pages 324–325 Monographs, no. 107.
664 Alldredge et al. [Auk, Vol. 124

Ralph, J. C., J. R. Sauer, and S. Droege, Eds. White, G. C. 1983. Numerical estimation of
1995. Monitoring bird populations by point survival rates from band-recovery and
counts. U.S. Department of Agriculture, biotelemetry data. Journal of Wildlife
Forest Service General Technical Report Management 47:716–728.
PSW-GTR-149. White, G. C., D. R. Anderson, K. P. Burnham,
Rosenstock, S. S., D. R. Anderson, K. M. Giesen, and D. L. Otis. 1982. Capture–recapture
T. Leukering, and M. F. Carter. 2002. removal methods for sampling closed popu-
Landbird counting techniques: Current prac- lations. Los Alamos National Laboratory
tices and an alternative. Auk 119:46–53. Publication LA-8787-NERP, Los Alamos,
Seber, G. A. F. 1982. The Estimation of Animal New Mexico.
Abundance and Related Parameters, 2nd ed. White, G. C., and K. P. Burnham. 1999. Program
Edward Arnold, London. MARK: Survival estimation from popula-
Shriner, S. A. 2001. Distribution of breeding tions of marked animals. Bird Study 46
birds in Great Smoky Mountains National (Supplement):S120–S139.
Park. Ph.D. dissertation, North Carolina Williams, B. K., J. D. Nichols, and M. J.
State University, Raleigh. Conroy. 2002. Analysis and Management
Thompson, W. L. 2002. Towards reliable bird of Animal Populations. Academic Press, San
surveys: Accounting for individuals present Diego, California.
but not detected. Auk 119:18–25.
Wasserman, F. E. 1977. Mate attraction func-
tion of song in the White-throated Sparrow.
Condor 79:125–127. Associate Editor: D. H. Johnson