continuous-time spatially explicit capture-recapture models, with an application to a jaguar...
TRANSCRIPT
Acc
epte
d A
rtic
le
Date of submission: 11th April 2014
Running title: Continuous-time SECR
Number of words: approx 7000
Number of tables: 5
Number of gures: 3
Number of referen es: 43
Continuous-time spatially expli it
apture-re apture models, with an appli ation
to a jaguar amera-trap survey
David Bor hers
1
Greg Distiller
2 *
Rebe a Foster
3,4
Bart Harmsen
3,4
Lorenzo Milazzo
5
1. S hool of Mathemati s and Statisti s, Centre for Resear h into E ologi al
and Environmental Modelling, The Observatory, Bu hanan Gardens, Uni-
versity of St Andrews, Fife KY16 9LZ, UK
2. Statisti s in E ology, Environment and Conservation (SEEC), Department
of Statisti al S ien es, University of Cape Town, Private Bag X3, Rondebos h
7701, South Afri a
3. Panthera, 8 West 40th Street, 18th Floor, New York, NY 10018, USA
4. Environmental Resear h Institute (ERI), University of Belize, PO Box 340,
Belmopan, Belize, Central Ameri a
5. University of Cambridge, Cambridge Infe tious Diseases (CID), Madingley
Road, Cambridge CB3 0ES UK
*Corresponding author. Greg Distiller
This article has been accepted for publication and undergone full peer review but has not been
through the copyediting, typesetting, pagination and proofreading process, which may lead to
differences between this version and the Version of Record. Please cite this article as doi:
10.1111/2041-210X.12196
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le Summary
1. Many apture-re apture surveys of wildlife populations operate
in ontinuous time but dete tions are typi ally aggregated into
o asions for analysis, even when exa t dete tion times are avail-
able. This dis ards information and introdu es subje tivity, in
the form of de isions about o asion denition.
2. We develop a spatio-temporal Poisson pro ess model for spatially
expli it apture-re apture (SECR) surveys that operate ontinu-
ously and re ord exa t dete tion times. We show that, ex ept in
some spe ial ases (in luding the ase in whi h dete tion prob-
ability does not hange within o asion), temporally aggregated
data do not provide su ient statisti s for density and related
parameters, and that when dete tion probability is onstant over
time our ontinuous-time (CT) model is equivalent to an exist-
ing model based on dete tion frequen ies. We use the model
to estimate jaguar density from a amera-trap survey and on-
du t a simulation study to investigate the properties of a CT
estimator and dis rete-o asion estimators with various levels of
temporal aggregation. This in ludes investigation of the ee t on
the estimators of spatio-temporal orrelation indu ed by animal
movement.
3. The CT estimator is found to be unbiased and more pre ise than
dis rete-o asion estimators based on binary apture data (rather
than dete tion frequen ies) when there is no spatio-temporal or-
relation. It is also found to be only slightly biased when there
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le is orrelation indu ed by animal movement, and to be more ro-
bust to inadequate dete tor spa ing, while dis rete-o asion es-
timators with binary data an be sensitive to o asion length,
parti ularly in the presen e of inadequate dete tor spa ing.
4. Our model in ludes as a spe ial ase a dis rete-o asion estim-
ator based on dete tion frequen ies, and at the same time lays a
foundation for the development of more sophisti ated CT mod-
els and estimators. It allows modelling within-o asion hanges
in dete tability, readily a ommodates variation in dete tor ef-
fort, removes subje tivity asso iated with user-dened o asions,
and fully utilises CT data. We identify a need for developing
CT methods that in orporate spatio-temporal dependen e in de-
te tions and see potential for CT models being ombined with
telemetry-based animal movement models to provide a ri her in-
feren e framework.
Key-words: density estimation, data aggregation, su ien y, animal
movement, statisti al methods
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le Introdu tion
Monitoring elusive spe ies like forest-dwelling arnivores that o ur at low
densities and range over wide areas is di ult, as is monitoring spe ies that
are di ult to trap or dete t visually, but many su h spe ies are of high
onservation on ern. The estimation of density and dete tability are om-
mon in population monitoring and are ru ial parameters for elusive spe ies
(Thompson, 2004; Stuart et al., 2004).
Te hnologi al advan es in non-invasive sampling methods su h as s at
or hair olle tion for geneti analysis, amera-trapping, and a ousti dete -
tion have led to new ways of apturing individuals, that are parti ularly
useful for spe ies that are di ult to survey using more onventional meth-
ods. Camera-trap arrays in parti ular have be ome an important tool for
assessment and management of spe ies that are hard to dete t by other
means. They have been used with apture-re apture (CR) methods for vari-
ous spe ies that are individually identiable from photographs, in luding
felids, anids, ursids, hyaenids, pro ynoids, tapirids and dasypodids (see ref-
eren es in Foster & Harmsen, 2012), and are in reasingly used with spatially
expli it apture-re apture (SECR) methods. SECR in orporates spatial in-
formation on apture lo ations and allows inferen es to be drawn about the
underlying pro ess governing the lo ations of individuals, as well as providing
estimates of density without the need for separate estimates of the ee tive
sampling area (Eord, 2004; Bor hers & Eord, 2008; Eord, Bor hers &
Byrom, 2009a; Royle et al., 2009).
When a CR study ( onventional or SECR) involves physi ally apturing
3
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le individuals, the survey pro ess generates well-dened sampling o asions.
But when there is no physi al apture and release, as in the ase of amera-
trap studies or a ousti CR surveys that re ord times of dete tion, there is no
well-dened sampling o asion. In su h ases, aggregating data into apture
o asions introdu es subje tivity (the o asion length hosen by the analyst)
and dis ards some information (the exa t times of apture). Continuous-time
(CT) models avoid these problems. The earliest CT model is probably that
of Craig (1953). It assumed equal at hability for all individuals at all times,
but has sin e been extended to allow individual heterogeneity (Boy e et al.,
2001). Most CT estimators use martingale estimating fun tions (Be ker,
1984; Yip, 1989; Be ker & Heyde, 1990; Yip, Huggins & Lin, 1995; Yip
et al., 1999; Lin & Yip, 1999; Yip & Wang, 2002; Hwang & Chao, 2002),
although Nayak (1988), Be ker & Heyde (1990) and Hwang, Chao & Yip
(2002) developed maximum likelihood estimators (MLEs) and Chao & Lee
(1993) developed a CT estimator based on sample overage. Of the CT mod-
els in the literature, only the models of Nayak (1988), Yip (1989) and Hwang,
Chao & Yip (2002) estimate dete tability and abundan e simultaneously, in
ommon with the estimator we develop below.
There is no onsensus on the utility of CT estimators. Some authors
on luded that the use of a dis rete-o asion model to analyze a CT data-
set will bias estimates (Yip & Wang, 2002; Barbour, Pon iano & Lorenzen,
2013); while others found no benet over dis rete-o asion estimators (e.g.
Chao & Lee, 1993; Wilson & Anderson, 1995). We show in Appendix C,
that Chao & Lee's (1993) on lusion that apture times are uninformative
about abundan e does not hold in general. We develop CT SECR likelihoods
4
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le and MLEs whi h, unlike any existing CT models, are parameterised in terms
of population density rather than abundan e, and ru ially, in orporate a
model for (unobserved) individual a tivity entres. An a tivity entre is the
entre of gravity of an individual's lo ations over the duration of the sur-
vey. It may have biologi al interpretation, e.g. as a home range entre, but
need not. Our model is obtained by formulating the SECR survey pro ess
as a nonhomogeneous spatio-temporal Poisson pro ess, in whi h events are
dete tions at points in spa e and time (see Cook & Lawless, 2007, for a om-
prehensive overview of re urrent event pro esses, of whi h Poisson pro esses
are a spe ial ase). We note that the intera tion of re urrent event pro esses
and CR methods was identied by Chao & Huggins (2005, p85) as a fruitful
area for future resear h, whi h we explore in this paper.
We fo us on the development of SECR methods for data from amera-trap
surveys, but our models are also appropriate for any CR survey with on-
tinuous sampling in whi h apture times are re orded. This in ludes a ousti
SECR surveys su h as those des ribed in Eord, Dawson & Bor hers (2009b),
Dawson & Eord (2009) or Bor hers et al. (2013), although re apture iden-
ti ation may be di ult in su h surveys. CT SECR models also provide
the basis for varying-eort SECR methods su h as those of Eord, Bor hers
& Mowat (2013) and we show how dis rete-o asion models arise from a CT
formulation.
We assume independen e between aptures, onditional on a tivity entre
lo ation. This is almost ertainly violated to some degree in amera-trap
studies be ause individuals are only dete ted when they move within range
of a amera, and an individual's lo ations are temporally orrelated. We
5
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le investigate sensitivity to violation of this assumption.
We develop a general CT SECR likelihood and then evaluate models un-
der the assumptions of onstant dete tion probability over time and onstant
intensity of the spatial point pro ess governing the number and lo ations of
a tivity entres. In this ase the CT likelihood is equivalent to a single-
o asion Poisson ount likelihood spanning the whole survey, in whi h the
data are the apture frequen ies of ea h individual at ea h dete tor (see Ap-
pendix C). Simulations are used to ompare the CT estimator for this ase
with dis rete-o asion estimators based on binary data within o asions and
dete tors (i.e. in whi h ea h individual's apture frequen y at ea h dete tor
within an o asion is redu ed to a binary indi ator of dete tion or not). We
ompare various levels of temporal aggregation both when the assumption of
independen e holds and when it is violated by simulating orrelated animal
movement.
Methods
Study site and amera trapping of jaguars
Jaguars (Panthera on a) are near threatened and their global population is
de lining (Caso et al., 2008), and population monitoring is di ult be ause
they o ur at low densities, range widely and are elusive, often inhabiting
thi k habitat. Over the past de ade, the hallenge of dete ting jaguars for
population estimation has been fa ilitated by amera traps, following the
work of Silver et al. (2004); however, reliable and robust density estimates of
jaguars are rarely obtained (see Foster & Harmsen, 2012; Tobler & Powell,
6
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le 2013).
We estimate male jaguar density in the Co ks omb Basin Wildlife San -
tuary in Belize from amera-trap data. The san tuary en ompasses 490 km
2
of se ondary tropi al moist broadleaf forest at various stages of regenera-
tion following anthropogeni and natural disturban e (for more details see
Harmsen et al., 2010b). To the west, the san tuary forms a ontiguous forest
blo k with the prote ted forests of the Maya Mountain Massif (≈ 5,000 km
2
of forest). To the east the san tuary is partially buered by unprote ted
forest beyond whi h is a mosai of pine savannah, shrub land and broadleaf
forest, inter-dispersed with villages and farms. Jaguars are found through-
out this lands ape (Foster, Harmsen & Don aster, 2010). There are 65 km
of trails, all within the eastern part of the san tuary (see Figure 1). Male
jaguars routinely walk the trail system; trail use overlaps extensively. Al-
though they frequently leave the trails to move through the forest, they are
rarely dete ted o-trail by amera traps (Harmsen et al., 2010a). Nineteen
paired amera stations (Panthera am v3) were deployed along the trail net-
work within the eastern basin and maintained for 90 days (April to July
2011). Neighbouring stations had an average spa ing of 2.0 km (1.1 to 3.1
km). The ameras used have an enfor ed 8 se ond delay between onse utive
photos and digital photographi data were downloaded every two weeks.
Continuous-time models for proximity dete tors
We refer to both apture and dete tion as dete tion, and to all dete tors,
in luding traps, as dete tors.
We model the pro ess generating dete tions as a nonhomogeneous Poisson
7
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le pro ess (NHPP) in spa e and time, in whi h the events are dete tions. This
model is governed by a hazard of dete tion, whi h an vary with lo ation and
time and is equal to the expe ted number of dete tions per unit time at the
lo ation and time. The model allows this hazard to hange with lo ation (to
a ommodate lower apture probability for individuals with a tivity entres
farther from dete tors, for example), time (to a ommodate dierent animal
a tivity levels at dierent times of day, for example) and other explanatory
variables, but we assume initially that the hazard for any individual at any
time is independent of the individual's apture history up to that time. We
develop a model and estimator under this assumption and then investigate
its performan e when the hazard of dete tion does depend on where and
when individuals were previously dete ted (as a onsequen e of orrelated
animal movement). Parameters and standard errors were estimated using a
Newton-type algorithm with fun tion nlm in R (R Core Team, 2013).
Continuous-time proximity dete tor likelihood
We onsider a survey in whi h we have an array of K ontinuously-sampling
proximity dete tors. The key dieren e between the likelihood presented here
and a dis rete-o asion SECR likelihood is that there are no longer o asions,
but instead a ontinuous survey running for a period (0, T ). Instead of a
apture history of length S (if there were S o asions) for the ith individual at
dete tor k, we have ωik aptures of the individual at dete tor k at times tik =
(tik1, . . . , tikωik). For brevity, we let t = tik (i = 1, . . . , n; k = 1, . . . , K)
denote the set of all dete tion times.
To simplify presentation of the likelihood, we start by onsidering a single
8
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le individual (the ith individual), with a tivity entre at xi, and a single de-
te tor (the kth dete tor). By onsidering tik to be a realisation of an NHPP
in time, we obtain an expression for the probability density fun tion (pdf)
of tik. This is analogous to the probability of obtaining a apture history at
dete tor k for individual i in a dis rete-o asion formulation, but it is worth
noting that both the length of tik (i.e. ωik) and the times that it ontains
are random variables, whereas for a dis rete-o asion survey the length of a
apture history is xed by the number of o asions used.
Using the pdf of tik, we obtain an expression for the probability of dete t-
ing individual i at all. Then, following Bor hers & Eord (2008), we assume
that the (unobserved) a tivity entres of dete ted individuals are realisations
of a ltered spatial NHPP, with lter at a lo ation x being the dete tion
probability at that lo ation. This gives rise to a joint probability model for
the number of individuals dete ted (n) and their dete tion times at dete tors
(t). When onsidered as a fun tion of the unknown model parameters, this
is the CT SECR likelihood fun tion.
The pdf of tik
In a CT formulation, the dis rete-o asion expression for the probability
of dete ting individual i at dete tor k on an o asion is repla ed by a de-
te tion hazard fun tion that spe ies the expe ted dete tion rate per unit
time at dete tor k of an individual, at time t, given the lo ation of its a tiv-
ity entre (xi for individual i). This hazard fun tion hk(t,xi; θ) depends on
the distan e from dete tor k to the individual's a tivity entre (xi), and we
make this expli it when we spe ify models for hk(t,xi; θ), but for notational
brevity we do not make the dependen e on distan e expli it otherwise. The
9
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le hazard fun tion has an unknown ve tor of dete tion fun tion parameters θ.
The survivor fun tion for individual i at dete tor k over the whole survey
(the probability of individual i not being dete ted on the dete tor by time
T ) is Sk(T,xi; θ) = exp(
−∫ T0 hk(u,xi; θ) du
)
, and hen e 1 − Sk(T,xi; θ) is
the probability of dete tion in (0, T ).
In a similar vein, the link between a CT dete tion hazard and a dis rete-
o asion dete tion probability, pks(xi; θ) for dete tor k on o asion s, of
length Ts = ts−ts−1 (running from ts−1 to ts) is: pks(xi; θ) = 1−exp(
−∫ tsts−1
hks(t,xi; θ))
.
Under the assumption of a onstant hazard, solving this for hks(t,xi; θ) leads
to a hazard with the following form:
hks(xi; θ) = − log 1− pks(xi; θ) /Ts (1)
In our analyses, we use hazard fun tions in whi h pks(xi; θ) for a spe ied
o asion length has a half-normal form: pks(xi; θ) = g0,Tsexp−dk(xi)2/(2σ2)
,
where dk(xi) is the distan e from xi to dete tor k and g0,Tsand σ are paramet-
ers to be estimated. The parameter g0,Tsis the probability that an individual
with a tivity entre lo ated at a dete tor is dete ted in a time period of length
Ts, while σ is the s ale parameter of the half-normal dete tion fun tion, and
determines the range over whi h an individual is dete table.
Assuming that, onditional on the a tivity entre lo ation, the times of
dete tions are independent, we an model the number of dete tions, ωik, and
the dete tion times tik as a nonhomogeneous Poisson pro ess with pdf as
follows (see Cook & Lawless, 2007, Theorem 2.1 on page 30, and Equation
10
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le (2.13) on page 32):
fk(tik|xi; θ) = Sk(T,xi; θ)ωik∏
r=1
hk(tikr,xi; θ) (2)
(We omit ωik from the LHS for brevity, be ause given tik, ωik is known.)
For those familiar with the denition of hazard fun tions in terms of the pdf
and survivor fun tion, we note that hk(tikr,xi; θ) = fk(tikr|xi; θ)/Sk((tikr −
tik(r−1)),xi; θ).
Equation (2) is a generalisation of the expression obtained by Hwang,
Chao & Yip (2002, p43) for their model Mt (given by their expression for
Li), and of the expression of Chao & Huggins (2005, p80) for the ase in
whi h hk(tikr,xi; θ) is onstant. Both sets of authors use λ for the hazard,
whereas we use h( ). In addition to lo ating these models in the Poisson
pro ess literature, our generalisation involves the in lusion of x as a latent
variable (see below), and extension to multiple dete tors.
Multiple dete tors and individuals
An individual's overall dete tion frequen y a ross the K dete tors is de-
noted ωi· =∑K
k=1 ωik and the event ωi· > 0 indi ates dete tion by some
dete tor. The ombined dete tion hazard over all dete tors at time t is
h·(t,xi; θ) =∑K
k=1 hk(t,xi; θ), and the overall probability of dete tion in
(0, T ) over all dete tors is p·(xi; θ) = p(ωi·>0|xi; θ) = 1−S·(T,xi; θ), where
S·(T,xi; θ) = exp(
−∫ T0 h·(t,xi; θ)
)
is the overall survivor fun tion.
Following Bor hers & Eord (2008), we assume that a tivity entres o ur
independently in the plane a ording to an NHPP with intensity D(x;φ) at
x, where x is any point in the survey region and φ is a ve tor of unknown
11
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le parameters that govern the intensity and hen e the distribution of a tivity
entres. D(x;φ) is the expe ted a tivity entre density at x. A tivity entres
are assumed not to move during the survey.
The likelihood for φ and θ is the joint distribution of the number of
animals aptured n, and the density of the out ome ωik events, at times
tik1 < . . . < tikωikr, for all i and k. This an be written in terms of the
marginal distribution of n and the onditional distribution of t given n.
L(θ,φ | n, t) = P (n | φ, θ)ft(t|n, θ,φ) (3)
If individuals are dete ted independently, n is a Poisson random variable
with parameter λ(θ,φ) =∫
A D(x;φ)p·(x; θ) dx, where p·(x; θ) is the prob-
ability of an individual with a tivity entre at x being dete ted at all and
the integral is over all points x in the area A in whi h a tivity entres o ur.
We obtain an expression for ft(t|n, θ,φ) in Appendix A and show that,
given the dete tion times of all dete ted individuals, t, the likelihood for φ
and θ is
L(φ, θ|n, t) =exp−λ(φ,θ)
n!
n∏
i=1
∫
AD(xi;φ)S·(T,xi; θ)
K∏
k=1
ωik∏
r=1
hk(tikr,xi; θ) dx
(4)
The integral is over all possible a tivity entre lo ations that ould have led
to a dete tion on the survey. (See se tion 4.2 of Bor hers & Eord, 2008, for
a brief dis ussion of this issue.)
It is sometimes useful to write this likelihood in terms of the marginal
distribution of ωik and the onditional distribution of tik given ωik. The
12
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le apture frequen y over the survey for individual i at dete tor k has a Pois-
son distribution with parameter Hk(xi; θ) =∫ T0 hk(t,xi; θ) dt (see Cook &
Lawless, 2007, Equation (2.15) on page 32):
Pk(ωik|xi; θ) =Hk(xi; θ)
ωik exp−Hk(xi;θ)
ωik!=
Hk(xi; θ)ωikSk(T,xi; θ)
ωik!(5)
It follows from this and Equation (2) that the onditional pdf of dete tion
times tik, given ωik, for the ith animal is
ft|ω,k(tik|ωik,xi; θ) = ωik!ωik∏
r=1
hk(tikr|xi; θ)
Hk(xi; θ)(6)
and hen e that
L(φ, θ|n, t) =e−λ(θ,φ)
n!
n∏
i=1
∫
AD(xi;φ)
K∏
k=1
Pk(ωik|xi; θ)ωik∏
r=1
ft|ω,k(tik|ωik,xi; θ) dx
(7)
Constant density, onstant hazard likelihood
One simple spe ial ase arises when density is uniform i.e. the spatial Poisson
pro ess governing a tivity entre lo ations is homogeneous, and the dete -
tion hazards do not hange with time. In this ase D(x;φ) is a onstant
(D), λ(θ,φ) = Da(θ) (where a(θ) =∫
A p·(x; θ) dx is the ee tive survey
area), hk(t,xi; θ) = hk(xi; θ) and so Hk(xi; θ) = Thk(xi; θ). As a result,
Equation (6) be omes ft|ω,k(tik|ωik,xi; θ) = ωik!∏ωik
r=1 T−1
and the likelihood
Equation (7) simplies to
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le L(D, θ|n, t) =
(
Dne−Da(θ)
n!
n∏
i=1
∫
A
K∏
k=1
Pk(ωik|xi; θ) dx
)(
n∏
i=1
K∏
k=1
ωik!ωik∏
r=1
1
T
)
(8)
The se ond term in large bra kets is the pdf of the dete tion times t, given
the dete tion frequen ies ωik (i = 1, . . . , n; k = 1, . . . , K). Sin e this term
does not ontain any model parameters, and the term in the rst large bra k-
ets ontains the parameters and the dete tion frequen ies ωik but not the
dete tion times, it follows (from the Fisher-Neyman Fa torization Theorem
for su ien y) that in this ase the dete tion frequen ies are su ient stat-
isti s for the parameters θ and D, i.e., that no information about D is lost
by dis arding the dete tion times. Chao & Lee (1993) obtained this result
for the ase with a single dete tor, no x, and a model parameterised in terms
of abundan e N rather than density D but the result does not hold in
general, as we dis uss below.
Relationship to Poisson ount models
One an see from Equation (8) that under the above assumptions, the CT
model gives rise to a Poisson ount model, sin e Pk(ωik) is a Poisson prob-
ability mass fun tion (pmf) with rate parameter Tshks(xi; θ) and the term
in the se ond large bra ket an be negle ted be ause it is onstant with re-
spe t to the parameters. This model was proposed in Appendix A of Eord,
Dawson & Bor hers (2009b), and a similar model was proposed by Royle
et al. (2009). The CT model provides an interpretation of the Poisson rate
parameter of these models as the umulative dete tion hazard, Tshks(xi; θ).
14
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le Both Eord, Dawson & Bor hers (2009b) and Royle et al. (2009) proposed
modelling Tshks(xi; θ) using the same fun tional forms as are used for the
probability of dete ting an individual within an o asion, but allowing an
inter ept greater than 1. This may be a reasonable strategy but model-
lers should be aware that pks(xi; θ) and hks(xi; θ) ne essarily have dierent
shapes be ause pks(xi; θ) = 1 − exp −Tshks(xi; θ). So for example, as-
suming a half-normal shape for the Poisson rate parameter (as do Royle
et al., 2009): Tshks(xi; θ) = λ0 exp −dk(xi)2/σ2 generates a shape for the
dete tion fun tion pks(xi; θ) that is not half-normal.
La k of su ien y of dete tion frequen ies
In Appendix C we show that the dis rete-o asion Bernoulli, binomial and
Poisson models are redu ed-data versions of the CT model and that the
dete tion histories or dete tion frequen ies Ω = ωiks (i = 1, . . . , n; k =
1, . . . , K; s = 1, . . . , S, where S is the number of o asions in the survey) are
not in general su ient statisti s for θ and φ, i.e. that in general dete tion
time is informative. A notable ex eption is when the dete tion hazard is
onstant within o asions. In this ase, dis arding dete tion times does not
dis ard information about population density. Redu ing ounts to a binary
event does however dis ard information and the onsequen es of this redu -
tion are explored in the rst simulation below.
Models with variable effort
CT models readily a ommodate dete tors that are operational for dierent
periods of time, by setting the dete tion hazard at a dete tor to zero while
15
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le the dete tor is out of operation (i.e. hk(t,x; θ) = 0 if dete tor k was not
operating at time t). This is the basis of the varying eort model for binary
dete tors and ounts in Eord, Bor hers & Mowat (2013). It is more di ult
to a ommodate dete tors when their time of failing is unknown, and we do
not address that problem here.
Simulation testing amera-trap estimators
If animals are dete ted as a onsequen e of their moving lose to a dete tor,
the expe ted dete tion rate of the animal at any given dete tor depends on
when and where the animal was last dete ted, be ause animals' lo ations
are temporally orrelated. But if times between dete tions are su iently
long that the pdf of an animal's lo ation at time t is independent of where it
was last dete ted, then we an model the pro ess generating dete tion times
using the model derived above.
In addition, to investigate the properties of the CT estimator for amera-
trap surveys when dete tions are temporally orrelated, we ondu ted sim-
ulations using a generating pro ess in whi h probability of dete tion at any
dete tor depends on when and where an animal was last dete ted. Be ause
our fo us is on the appli ation of CT estimators to jaguar amera-trap sur-
vey data, we model dependen e by onstru ting individual movement models
that are plausible for jaguar movement.
For all simulation s enarios, we also investigated the properties of a binary
dis rete-o asion proximity dete tor SECR model (in whi h ωiks is binary)
using a half-normal dete tion fun tion and a range of aggregation levels: (i)
360 o asions of length Ts = 6 hours (ii) 90 o asions of length Ts = 24 hours
16
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le (iii) 30 o asions of length Ts = 72 hours, and (iv) 18 o asions of length
Ts = 120 hours. In all ases we assumed onstant intensity, D, in the survey
region.
The density estimates obtained from tting a variety of models to the jag-
uar survey data (see below) were used to inform the simulation s enarios and
led to the use of D = 4 individuals per 100 km
2. The movement model (be-
low) is not parameterised in terms of the parameters g0 and σ that are used
for the independen e simulation model. In order to fa ilitate omparison of
results a ross the two kinds of simulations, the values used for these para-
meters in the independen e simulations were based on estimates obtained by
tting the CT model to data from the movement simulations. All simula-
tions involved twenty dete tors arranged on a 4x5 grid. Three dierent grid
spa ings were used: 1,500 m (62.5% of σ), 3,000 m (125% of σ), and 4,500
m (187.5% of σ).
Independen e simulations
Fun tion sim.popn from the R pa kage se r (Eord, 2013) was used to sim-
ulate populations of N individuals with a tivity entres x1, . . . ,xN , where N
is sto hasti and varied from simulation to simulation (see Fewster & Bu k-
land, 2004, for details of the simulation method). A time- onstant hazard
hk(xi; θ) (k = 1, . . . , K) was assumed for the duration of the survey (0, T )
and the frequen y with whi h individual i was dete ted by dete tor k (ωik)
was obtained as a draw from a Poisson distribution with rate parameter
T × hk(xi; θ). Dete tion times were generated by independent draws from a
uniform distribution on (0, T ).
17
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le Separate dete tion hazards were used to simulate data for the dierent
aggregation levels, so as to have dis rete-o asion dete tion fun tions with
half-normal forms at ea h level (to fa ilitate dis rete-o asion estimation with
the se r pa kage). The orresponding CT model estimator was also applied
to ea h simulated dataset. For an o asion of length Ts, hks(xi; θTs) is equal
to Equation (1), where pks(xi; θTs) = g0,Ts
exp−dk(xi)2/(2σ2), dk(xi) is the
distan e from xi to dete tor k, θ6 = (g0,6 = 0.01, σ = 2, 400), θ24 = (g0,24 =
0.05, σ = 2, 400), θ72 = (g0,72 = 0.15, σ = 2, 400) and θ120 = (g0,120 =
0.25, σ = 2, 400).
Movement model simulations
A tivity entres were simulated as above. Animals were moved in time steps
of one hour a ording to the movement model des ribed below. Whenever
a path rossed a 250 × 250 m grid ell ontaining a dete tor, dete tion o -
urred with probability PGC (where subs ript GC stands for Grid Cell as
these probabilities operate at the level of grid ells). The lower PGC , the
less temporally lustered are dete tions and the loser the model is to the
independen e model. This is be ause a orrelated time series be omes less
orrelated when randomly thinned.
The movement model: a mixture of random walks There is onsid-
erable literature on the modeling of animal movement using various types
of random walks (Morales et al., 2004; Codling et al., 2008; Smouse et al.,
2010; Langro k et al., 2012). Movement an onveniently be simulated by
generating a step length and a turning angle from appropriate distributions
18
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le at ea h time step (Morales et al., 2004). Dierent turning angle distributions
produ e dierent types of random walks. For example, for a biased random
walk (BRW) the turning angle for the next step is drawn from a distribu-
tion with mean oriented towards some fo al point whereas for a orrelated
random walk (CRW) the mean is the urrent dire tion of movement.
In order to reprodu e both the tenden y of individuals to persist in the
dire tion in whi h they are urrently moving and the tenden y not to stray
too far from their a tivity entre, we onstru ted a movement model in whi h
individuals transition sto hasti ally between two states. In the rst state
(state s = 0) they follow a CRW, and in the se ond state (state s = 1) a
BRW with bias towards an a tivity entre. The probability of being in a
parti ular state depends on the distan e the individual is from its a tivity
entre. Spe i ally, the state s of an individual at distan e d from its a tivity
entre has the following Bernoulli distribution s|d ∼ γ(d)s(1−γ(d))1−s, where
γ(d) is the probability of being in state 1, whi h is modelled as a logisti
fun tion of distan e: γ(d) = (1+expα+βd)−1. For example, the parameter
values used in the simulation (Table 1) result in a probability of being in state
0 (CRW) of 0.95 at a distan e of zero, 0.5 at a distan e of 2,500 m, and 0.05
at a distan e of 5,000 m from the a tivity entre.
The size of ea h movement step (l) is a draw from a gamma distribution
with mean step length a and standard deviation b: l ∼ Gamma(a2
b2, b
2
a). The
turning angle (φ), given that the individual is in state s, is drawn from a
Von Mises distribution φ|s ∼ von Mises(µ(s), κ(s)) where µ(s) is the mean
and κ(s) the on entration parameter. Values for κ(s) lose to zero gener-
ate something lose to a uniform distribution on the ir le, whereas larger
19
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le values will pla e in reasing mass around the mean value. Table 1 shows the
movement model parameter values used in the movement simulations and
Figure 2 shows a realisation of the model over T = 2, 160 hours (90 days).
Be ause the obje tive is to generate orrelated dete tion times rather than to
make inferen es about the underlying behaviour, a large standard deviation
for step length was used to ree t movement of a range of behavioural states
(resting, hunting, travelling et ).
Results
Belize Jaguar survey
Females have very dierent home range sizes to males, so that modeling
both sexes would require separate parameters for ea h sex. Be ause very few
female jaguars were dete ted, this analysis only uses male dete tions. A total
of 207 dete tions of 17 identiable males were made, of whi h 6 males (35%)
were dete ted at least twi e in a day at the same dete tor. Not all dete tors
operated for the 90 days of the survey therefore we set the dete tion hazard
for the dete tor to zero for all of the failed days in the ase of the CT model,
and removed the dete tor from o asions on whi h it was out of a tion in the
ase of the dis rete-o asion model.
In addition to models with no ovariates, models with a behavioural re-
sponse in the form of separate g0 and/or σ parameters before and after rst
dete tion were onsidered. This led to models with a behavioural response
for g0 but not σ being sele ted on the basis of AICc (see Table 2). Estimates
from a dis rete-o asion model with Ts = 24 hours are shown in Table 3,
20
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le together with estimates from the omparable CT model. Varian es were
estimated using the inverse of the observed Fisher information matrix and
normality was assumed to estimate onden e intervals.
Simulation Results
Independen e simulation results The estimated biases of estimators as
a per entage of true parameter values (%Bias) are shown together with their
standard errors for the three dierent dete tor spa ings in Table 4. The per-
entage bias in density for the dis rete-o asion models was parti ularly high
for losely spa ed traps, more so when many short o asions were used. The
number of individuals dete ted per survey in reased with dete tor spa ing
be ause greater spa ing leads to a larger area being sampled.
With 1,500 m dete tor spa ing 12% of individuals were dete ted at least
twi e a day at the same dete tor, on average. This redu ed to 9% and 6% with
spa ing of 3,000 m and 4,500 m, respe tively. These values are substantially
lower than the 35% observed in the survey data.
Movement model simulation results Table 5 summarises the results
from the movement simulations. The pattern of bias is similar to that from
the independent simulations though the bias is worse when PGC is high ( om-
pare results from Table 4 for 125%σ with the orresponding spa ing in Table 5
when PGC = 0.9).
Redu ing PGC in the movement simulation ee tively thins the dete tion
21
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le time pro ess and onsequently redu es temporal orrelation. For spa ing of
3,000 m, when PGC = 0.9, on average 85% of individuals were dete ted at
least twi e in a day at the same dete tor and this proportion drops to 52%
when PGC = 0.3. Results were similar for the larger dete tor spa ing.
Dis ussion
Belize jaguar survey results
The per entage of individuals that are dete ted at least twi e in a day at the
same amera gives a rough indi ation of spatio-temporal lustering. The dif-
feren es between the real data (35%) and the simulated movement data (52%
and 85%) suggest that the Belize jaguar data are loser to satisfying the inde-
penden e assumptions of the ontinuous-time (CT) model than are either of
the movement simulation s enarios. Although the assumption of independ-
ent a tivity entres is not realisti for territorial spe ies, no SECR model
yet exists that a ommodates the repulsive ee t that a tivity entres of
territorial spe ies exert on one another's a tivity entres, and modelling this
is hallenging. Eord, Bor hers & Byrom (2009a) found SECR estimators
of density to be robust to violation of the independent a tivity entre distri-
bution assumption in the form of lustering (overdispersion), and they may
also be robust to violation in the form of underdispersion.
Sin e no bait or lure was used, it is unlikely that the behavioural re-
sponse in g0 is a true trap-happiness ee t. The amera stations are lo ated
on trails resulting in a higher probability of dete ting the individuals that
habitually use those trails, and hen e the behavioural response parameter
22
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le may be a ting as a proxy for individual heterogeneity (Royle et al., 2009;
Soisalo & Caval anti, 2006).
Estimator properties
Like Eord, Dawson & Bor hers (2009b), we found that a SECR ount model
estimator with a single sampling interval (equivalent to our CT estimator
with onstant hazards) an give pre ise and unbiased density estimates. In
addition, we show that when there is inadequate spa ing between dete tors,
the estimator from a ount model with a single sampling interval performs
substantially better than binary model estimators in whi h the interval is
divided into multiple o asions.
Information loss with binary data
In general, information is lost if exa t dete tion times are dis arded, although
this is not the ase if the hazard of dete tion is onstant and ount data are
retained. There is a loss if ount data are redu ed to binary data although
Eord, Dawson & Bor hers (2009b) found the loss to be small in the s enarios
they onsidered. Similarly we found little dieren e between the perform-
an e of binary dis rete-o asion estimators and CT estimators with onstant
dete tion hazard when using dete tor spa ings of 1.25σ and 1.88σ. However,
when dete tor spa ing was redu ed to 0.625σ, the CT estimator performed
substantially better than the binary dis rete-o asion estimators.
To understand this behaviour, we investigated simulated datasets in whi h
no ount was greater than 1 in any interval. The binary and CT estimat-
ors produ ed dierent estimates in these ases, although they operate on
23
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le identi al data. This is be ause the binary likelihood models the probability
of at least one dete tion in an interval whereas the CT likelihood models the
probability of exa tly one dete tion, and the latter is more informative about
individual a tivity entre lo ation. For example, the fa t that an individual
was dete ted exa tly on e by the dete tor suggests that the entre was not
very lose to the dete tor (if it was, more than one dete tion would be likely),
whereas the fa t that an individual was dete ted at least on e in ludes the
possibility that it was dete ted many times, in whi h ase the entre ould
be very lose to the dete tor. Consequently, the binary data model tends to
attra t entres to dete tors, whi h an lead to negatively biased ee tive
sampling area estimates and positively biased density estimates (Figure 3).
This ee t is more pronoun ed for a tivity entres that are on the edge of
the dete tor array, whi h explains why the bias is worse for smaller dete tor
arrays. When an a tivity entre is surrounded by dete tors, the presen e of
other dete tors limits the extent to whi h the binary model an attra t the
entre to any dete tor (Figure 3).
Data aggregation
Ex ept in some spe ial ases, data aggregation involves information loss.
Aggregation an be useful when modelling unaggregated data is di ult,
but aggregation without good reason is not good pra ti e. When SECR
data are re orded in ontinuous time, we are not aware of any reason not
to use a CT model if dete tions are independent events. If the dete tion
hazard is onstant the CT model is identi al to a ount model spanning a
single o asion but the CT formulation an be extended to in orporate time-
24
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le varying hazards within o asion, and therefore enables extensions that are
not available without a CT formulation. When dete tion hazards depend on
observable time-varying ovariates (time of day, temperature or some other
time-varying environmental variables for example) there may be advantage
in modelling this using exa t dete tion times even if there is no inherent
interest in the nature of the varying hazard, and there is ertainly benet
in modelling the ee t of time-varying ovariates on dete tion hazard when
there is interest in how the hazard fun tion depends on su h ovariates.
When there is spatio-temporal orrelation in dete tions (e.g. due to move-
ment) and a design with adequate dete tor spa ing is used, there may be
merit in aggregating data a ross time and using binary dis rete-o asion
models with appropriate o asion lengths. However, our simulations suggest
that (a) the bias of binary dis rete-o asion density estimators an be quite
sensitive to the length of o asions, parti ularly when dete tors are too lose
together, and (b) that the CT point estimator of density is moderately robust
to inadequate dete tor spa ing and levels of spatio-temporal orrelation that
are substantially higher than is apparent in the jaguar eld data.
We have not in this paper onsidered the adequa y of interval estimators
based on the observed Fisher information matrix. If dete tions of the same
animal are not independent a ross dete tor lo ations and time (due to orrel-
ated animal movement, for example), varian e estimates may be biased and
onden e interval estimates may have in orre t overage probability. One
way of dealing with this would be to onsider the likelihood Equation (4)
to be an approximation to the orre t likelihood and use a version the ABC
bootstrap method of Efron (1987) that in orporates an animal movement
25
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le model (as per our movement simulation model, for example), to obtain or-
re ted point and interval estimates.
Con lusion
We have shown that CT SECR methods, unlike dis rete-o asion methods,
fully use the available data when dete tions are made ontinuously, and they
have a number of other advantages over methods that aggregate data into
dis rete o asions. We therefore re ommend that CT methods be onsidered
when dete tion times are available, and we re ommend olle ting dete tion
times whenever possible.
Although CT SECR methods are not yet as well developed as dis rete-
o asion methods (e.g. models in orporating individual ovariates or un-
observed individual-level heterogeneity remain to be developed), this is a
temporary obsta le that will soon be over ome. A more substantial obsta le
in the ase of amera-trap studies may be the development of CT models
that in orporate spatio-temporal dependen e in dete tions due to individual
movement. This seems hallenging but we see potential for CT models being
ombined with telemetry-based animal movement models to provide a ri her
inferen e framework than is urrently available.
A knowledgements
This work was part-funded by EPSRC grant EP/I000917/1. We would like to
thank the Editor and Asso iate Editor of MEE and three referees, in luding
Torbjørn Ergon, for useful suggestions that substantially improved the larity
26
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le of the paper.
Data A essibility
Data deposited in the Dryad repository: http://datadryad.org/resour e/doi:10.5061/dryad.mg5kv
Referen es
Barbour, A.B., Pon iano, J.M. & Lorenzen, K. (2013) Apparent survival estima-
tion from ontinuous mark-re apture/resighting data. Methods in E ology and
Evolution, 4, 846853.
Be ker, N.G. (1984) Estimating population size from apture-re apture experi-
ments in ontinous time. Australian Journal of Statisti s, 26, 17.
Be ker, N.G. & Heyde, C.C. (1990) Estimating population size from multiple re-
apture experiments. Sto hasti Pro esses and their Appli ations, 36, 7783.
Bor hers, D. & Eord, M. (2008) Spatially expli it maximum likelihood methods
for apture-re apture studies. Biometri s, 64, 377385.
Bor hers, D., Stevenson, B., Kidney, D., Thomas, L. & Marques, T. (2013) A
unifying model for apture-re apture and distan e sampling surveys of wildlife
populations. Journal of the Ameri an Statisti al Asso iation, (submitted).
Boy e, M.S., Ma kenzie, D.I., Manly, B.F., Haroldson, M.A. & Moody, D. (2001)
Negative binomial models for abundan e estimation of multiple losed popula-
tions. Journal of Wildlife Management, 65, 498509.
Caso, A., Lopez-Gonzalez, C., Payan, E., Eizirik, E., de Oliveira, T., Leite-Pitman,
R., Kelly, M. & Valderrama, C. (2008) Panthera on a. In: IUCN 2013. The
IUCN Red List of Threatened Spe ies. Version 2013.2 <www.iu nredlist.org>
Downloaded on 21 November 2013.
Chao, A. & Huggins, R. (2005) Modern losed-population apture-re apture mod-
els. S.C. Amstrup, T.L. M Donald & B.F.J. Manly, eds., Handbook of apture-
re apture analysis, hapter 4, pp. 5887. Prin eton University Press.
Chao, A. & Lee, S.M. (1993) Estimating population size for ontinuous-time
apture-re apture models via sample overage. Biometri al Journal, 35, 2945.
Codling, E.A., Plank, M.J. & Benhamou, S. (2008) Random walks in biology.
Journal of the Royal So iety Interfa e, 5, 813834.
27
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le Cook, R. & Lawless, J. (2007) The statisti al analysis of re urrent events. Springer.
Craig, C.C. (1953) On the utilization of marked spe imens in estimating popula-
tions of ying inse ts. Biometrika, 40, 170176.
Dawson, D. & Eord, M. (2009) Bird population density estimated from a ousti
signals. Journal of, 46, 12011209.
Eord, M.G., Bor hers, D.L. & Mowat, G. (2013) Varying eort in apture-
re apture studies. Methods in E ology and Evolution, pp. DOI: 10.1111/2041
210X.12049.
Eord, M. (2004) Density estimation in live-trapping studies. Oikos, 106, 598610.
Eord, M. (2013) se r: Spatially expli it apture-re apture models. R pa kage ver-
sion 2.5.0.
Eord, M.G., Bor hers, D.L. & Byrom, A.E. (2009a) Density estimation by
spatially expli it apture-re apture: Likelihood-based methods. D. Thomson,
E. Coo h & M. Conroy, eds., Modeling Demographi Pro esses in Marked Pop-
ulations, pp. 255269. Springer, New York, New York, USA.
Eord, M.G., Dawson, D.K. & Bor hers, D.L. (2009b) Population density estim-
ated from lo ations of individuals on a passive dete tor array. E ology, 90,
26762682.
Eord, B. (1987) Better bootstrap onden e intervals. Journal of the Ameri an
Statisti al Asso iation, 82, 171200.
Fewster, R. & Bu kland, S. (2004) Assessment of distan e sampling estimators.
S. Bu kland, D. Anderson, K. Burnham, J. Laake, D. Bor hers & L. Thomas,
eds., Advan ed Distan e Sampling, hapter 10, pp. 281306. Oxford University
Press, Oxford, UK.
Foster, R.J. & Harmsen, B.J. (2014) Data from: Continuous-time spatially expli it
apture-re apture models, with an appli ation to a jaguar amera-trap survey.
Methods in E ology and Evolution, doi:10.5061/dryad.mg5kv.
Foster, R.J. & Harmsen, B.J. (2012) A ritique of density estimation from amera-
trap data. The Journal of Wildlife Management, 76, 224236.
Foster, R.J., Harmsen, B.J. & Don aster, C.P. (2010) Habitat use by sympatri
jaguars and pumas a ross a gradient of human disturban e in belize. Biotropi a,
42, 724731.
Harmsen, B.J., Foster, R.J., Silver, S., Ostro, L. & Don aster, C.P. (2010a) Dif-
ferential use of trails by forest mammals and the impli ations for amera-trap
studies: A ase study from belize. Biotropi a, 42, 126133.
28
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le Harmsen, B.J., Foster, R.J., Silver, S.C., Ostro, L.E. & Don aster, C.P. (2010b)
The e ology of jaguars in the o ks omb basin wildlife san tuary, Belize. D.W.
Ma Donald & A.J. Loveridge, eds., The Biology and Conservation of Wild Felids,
hapter 18, pp. 403416. Oxford University Press.
Hwang, W. & Chao, A. (2002) Continuous-time apture-re apture models with
ovariates. Statisti a Sini a, 12, 11151131.
Hwang, W., Chao, A. & Yip, P. (2002) Continuous-time apture-re apture mod-
els with time variation and behavioural response. Australian and New Zealand
Journal of Statisti s, 44, 4154.
Langro k, R., King, R., Matthiopoulos, J., Thomas, L., Fortin, D. & Morales, J.M.
(2012) Flexible and pra ti al modeling of animal telemetry data: hidden markov
models and extensions. E ology, 93, 23362342.
Lin, D.Y. & Yip, P.S.F. (1999) Parametri regression models for ontinuous time
removal and re apture studies. Journal of the Royal Statisti al So iety: Series
B, 61, 401411.
Morales, J.M., Haydon, D.T., Frair, J., Holsinger, K.E. & Fryxell, J.M. (2004)
Extra ting more out of relo ation data: building movement models as mixtures
of random walks. E ology, 85, 24362445.
Nayak, T. (1988) Estimating population size by re apture sampling. Biometrika,
75, 113120.
R Core Team. (2013) R: A Language and Environment for Statisti al Computing.
R pa kage version 3.0.2
Royle, J.A., Karanth, K.U., Gopalaswamy, A.M. & Kumar, N.S. (2009) Bayesian
inferen e in amera trapping studies for a lass of spatial apture-re apture mod-
els. E ology, 11, 32333244.
Silver, S.C., Ostro, L.E.T., Marsh, L.K., Maei, L., Noss, A.J., Kelly, M.J., Wal-
la e, R.B., Gomez, H. & Ayala, G. (2004) The use of amera traps for estimating
jaguar panthera on a abundan e and density using apture/re apture analysis.
Oryx, 38, 148154.
Smouse, P.E., Fo ardi, Moor roft, P.R., Kie, J.G., Forester, J.D. & Morales, J.M.
(2010) Sto hasti modelling of animal movement. Philosophi al transa tions of
the royal so iety B, 365, 22012211.
Soisalo, M.K. & Caval anti, S.M.C. (2006) Estimating the density of a jaguar
population in the brazilian pantanal using amera-traps and apture-re apture
sampling in ombination with gps radio-telemetry. Biologi al Conservation, 129,
487496.
29
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le Stuart, S.N., Chanson, J., Cox, N., Young, B., Rodrigues, A., Fis hmann, D.
& Waller, R. (2004) Status and trends of amphibian de lines and extin tions
worldwide. S ien e, 306, 17831786.
Thompson, W. (2004) Sampling rare or elusive spe ies: on epts, designs, and
te hniques for estimating population parameters. Island Press, Washington, D.C.,
USA.
Tobler, M.W. & Powell, G.V.N. (2013) Estimating jaguar densities with amera
traps: Problems with urrent designs and re ommendations for future studies.
Biologi al Conservation, 159, 109118.
Wilson, K.R. & Anderson, D.R. (1995) Continuous-time apture-re apture popu-
lation estimation when apture probabilities vary over time. Environmental and
E ologi al Statisti s, 2, 5569.
Yip, P. (1989) An inferen e pro edure for a apture and re apture experiment with
time-dependent apture probabilities. Biometri s, 45, 471479.
Yip, P.S.F., Huggins, R.M. & Lin, D.Y. (1995) Inferen e for apture-re apture
experiments in ontinuous time with variable apture rates. Biometrika, 83,
477483.
Yip, P.S.F. & Wang, Y. (2002) A unied regression model for re apture studies
with random removals in ontinuous time. Biometri s, 58, 192199.
Yip, P., Zhou, Y., Lin, D. & Fang, X. (1999) Estimation of population size based on
additive hazards models for ontinuous-time re apture experimets. Biometri s,
55, 904908.
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le A Continuous-time proximity likelihood deriv-
ation
Consider a SECR survey of duration T with ontinuously-sampling prox-
imity dete tors and suppose that the ith dete ted individual is dete ted by
dete tor k at times tik = (tik1, . . . , tikωik), where ωik is the number of times
individual i is dete ted by dete tor k. Assuming that, onditional on its
a tivity entre lo ation (xi), the times of dete tions are independent, we an
model the dete tion times as a nonhomogeneous Poisson pro ess with pdf
fk(tik|xi; θ) = Sk(T,xi; θ)∏ωik
r=1 hk(tikr,xi; θ), where hk(t,xi; θ) is the non-
homogeneous event intensity or hazard at time t, θ is an unknown parameter,
and Sk(T,xi; θ) = e−∫
T
0hk(u,xi;θ) du
is the survivor fun tion for dete tor k
(the probability of individual i not being dete ted on the dete tor by time
T ).
We need to dene a few more things for onvenient expression of the likeli-
hood fun tion. Noting that ωi· =∑K
k=1 ωik is the individual's overall dete tion
frequen y and the event ωi·>0 indi ates dete tion by some dete tor, we let
ω· = ωi· (i = 1, . . . , n) and let ω·>0 indi ate dete tion by some dete tor
of individuals i = 1, . . . , n. t = tik (i = 1, . . . , n; k = 1, . . . , K) denotes the
set of all dete tion times. We also note that the ombined dete tion hazard
over all dete tors at time t is h·(t,xi; θ) =∑K
k=1 hk(t,xi; θ) and the prob-
ability of dete tion in (0, T ) is p·(xi; θ) = p(ωi·>0|xi; θ) = 1 − S·(T,xi; θ),
where S·(T,xi; θ) =∏
k Sk(T,xi; θ) = exp−∫
T
0h·(u,xi;θ)du
.
This likelihood derivation is very similar to that in Bor hers & Eord
(2008) and Equations (10) and (11) below are also derived in that paper.
31
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le The reader may nd it useful to refer to Bor hers & Eord (2008) when
reading this derivation. We assume that, onditional on a tivity entre lo a-
tions of dete ted individuals (X = (x1, . . . ,xn)), the times and lo ations of
dete tions of individuals are independent, and that individuals are dete ted
independently. Suppose also that a tivity entres are realisations of a non-
homogeneous Poisson pro ess (NHPP) in the plane, with intensity D(x;φ)
at x, and that φ is an unknown parameter ve tor. Then the likelihood for
the parameters φ of the NHPP governing a tivity entre lo ations and the
parameters θ of the dete tion pro ess, is obtained from the pdf of the number
of dete ted individuals n (P (n|φ, θ)), the pdf of the a tivity entre lo ations
given dete tion (fX(X|ω·>0;φ, θ)), and the pdf of dete tion times t, given
X and dete tion (ft(t|X,ω· > 0; θ)), after integrating over the unknown
lo ations X, as follows:
L(φ, θ|n, t) = P (n|φ, θ)ft(t|n, θ,φ)
= P (n|φ, θ)∫
AfX(X|ω·>0;φ, θ)ft(t|X,ω·>0; θ) dX (9)
where
P (n|φ, θ) =λ(θ,φ)ne−λ(θ,φ)
n!(10)
with λ(θ,φ) =∫
A D(x;φ)p·(x; θ) dx (where the integral is over all points x
in the area A in whi h a tivity entres o ur), and
fX(X|ω·>0;φ, θ) =n∏
i=1
D(xi;φ)p·(xi; θ)
λ(θ,φ)(11)
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le ft(t|X,ω·>0; θ) =n∏
i=1
∏Kk=1 fk(tik|xi; θ)
p·(xi; θ)(12)
=n∏
i=1
1
p·(xi; θ)
K∏
k=1
Sk(T,xi; θ)ωik∏
r=1
hk(tikr,xi; θ).
If we substitute Equations (11) and (12) into the integral in Equation (9),
note that
∏
k Sk(T,xi; θ) =∏
k e−∫
T
0hk(u,xi;θ) du = e−
∫
T
0h·(u,xi;θ) du = S·(T,xi; θ)
and an el p·(xi; θ) we get
ft(t|n, θ,φ) =∫
A
n∏
i=1
D(xi;φ)
λ(θ,φ)S·(T,xi; θ)
K∏
k=1
ωik∏
r=1
hk(tikr,xi; θ) dX
=n∏
i=1
∫
A
D(xi;φ)
λ(θ,φ)S·(T,xi; θ)
K∏
k=1
ωik∏
r=1
hk(tikr,xi; θ) dX.
(We an take the produ t over n outside the integral be ause the n xis are
independent.) Substituting this in Equation (9) and an elling λ(θ,φ) gives
Equation (4).
B Dis rete-o asion proximity likelihood de-
rivation
The likelihood for φ and θ for the dis rete-o asion ase in whi h there are S
o asions an be obtained in a similar way to that in whi h the ontinuous-
time (CT) likelihood was obtained, but repla ing the pdf of dete tion times,
given dete tion: ft(t|X,ω·>0; θ) with the pdf of dete tion histories, given
dete tion, PΩ(Ω|X,ω·>0; θ), where Ω = ωik (i = 1, . . . , n; k = 1, . . . , K)
and ωik = (ωik1, . . . , ωikS) is individual i's dete tion history on dete tor k
33
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le over the S o asions:
L(φ, θ|n,ωik) = P (n|φ, θ)∫
AfX(X|ω·>0;φ, θ)PΩ(Ω|X,ω·>0; θ) dX. (13)
This is the likelihood obtained by Bor hers & Eord (2008).
In 13 above PΩ(Ω|X; θ) is onditioned on being aught (ω·>0) and leads
to another term entering the likelihood p.(x, θ)−1. Two forms of PΩ(Ω|X; θ)
have been proposed for proximity dete tors: one in whi h ωiks is binary,
indi ating dete tion or not of individual i on dete tor k on o asion s, and
one in whi h ωiks is a ount of the number of times individual i was dete ted
by dete tor k on o asion s. Both have the form
PΩ(Ω|X; θ) =n∏
i=1
S∏
s=1
K∏
k=1
P (ωiks|xi; θ). (14)
The two forms of dis rete-o asion SECR model dier in the form they use
for P (ωiks|xi; θ). We show that when a CT dete tion pro ess is dis retised
into dis rete o asions this gives rise to a Bernoulli model in the ase of
binary data, and to binomial and Poisson SECR models in the ase of ount
data. (Binomial and Poisson models have been proposed for the ase in
whi h there is only one o asion (S = 1) but are easily extended to multi-
o asion s enarios, as we show below.) To do this, we divide the time interval
(0, T ) into S subintervals with interval s running from ts−1 to ts, of length
Ts = ts − ts−1, and with t0 = 0.
With a CT model, the probability of dete ting individual i at least
on e in dete tor k in this interval is pks(xi; θ) = 1 − e−Hks(xi;θ), where
34
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le Hks(xi; θ) =∫ tsts−1
hk(u,xi; θ)du. Spe ifying a model for pks(xi; θ) implies
a model for hk(t,xi; θ), although not a unique one. The mean value of the
dete tion hazard in interval s must be hks(xi; θ) = − log 1− pks(xi; θ) /Ts
and any hk(t,xi; θ) with this mean is onsistent with pks(xi; θ). When
the hazard is onstant in the interval there is a one-to-one relationship
between the dete tion hazard and the dete tion probability: hks(xi; θ) =
− log 1− pks(xi; θ) /Ts.
Bernoulli model This is obtained from a ontinuous-time model as PBern(ωiks|xi; θ) =
pks(xi; θ)ωiks 1− pks(xi; θ)
1−ωiks, with pks(xi; θ) = 1− e−Hks(xi;θ)
and bin-
ary ωiks.
Binomial ount model Eord, Dawson & Bor hers (2009b) proposed this
for the ase in whi h the S original intervals are ollapsed into S∗(< S) in-
tervals and ωiks∗ is the ount in the new interval s∗, whi h omprisesNs∗ adja-
ent original intervals. In this ase PBinom(ωiks∗|xi; θ) =(
Ns∗
ωiks∗
)
pks∗(xi; θ)ωiks∗ 1− pks∗(xi; θ)
1−ωik,
where pks∗(xi; θ) = 1− e−Hks∗(xi;θ), and Hks∗(xi; θ) =
∑Ns∗
s=1 Hks(xi; θ).
Poisson ount model A NHPP with intensity hk(t,xi; θ) at time t gives
rise to an event ount (ωiks) in a time interval (ts−1, ts) that has the Poisson
pmf PP (ωiks|xi; θ) = Hks(xi; θ)ωiks exp−Hks(xi;θ) /(ωiks!), whereHks(xi; θ) =
∫ tsts−1
hk(u,xi; θ)du and event ounts in non-overlapping intervals are inde-
pendent.
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le C La k of su ien y of Ω
To investigate the su ien y of Ω for the unknown parameters θ and φ, we
onsider the term fk(tik|xi; θ).
The onditional distribution of dete tion times, t(s)ik = t
(s)ik1, . . . , t
(s)ikωiks
, in
interval s, given that ωiks dete tions o urred in the interval is as follows:
fts(t(s)ik |ωiks; θ) = ωiks!
ωiks∏
r=1
hk(tikr,xi; θ)
Hks(xi; θ). (15)
We an now fa torise fk(tik|xi; θ) into the pmf for the ount ωiks in in-
terval s and the pdf for the dete tion times, given the ount, as follows:
fk(tik|xi; θ) =S∏
s=1
Hks(xi; θ)ωiks exp−Hks(xi;θ)
(ωiks!)
[
ωiks!ωiks∏
r=1
hk(tikr,xi; θ)
Hks(xi; θ)
]
=S∏
s=1
PP (ωiks|xi; θ)fts(t(s)ik |ωiks; θ) (16)
The likelihood Equation (4) an then be written as
L(φ, θ|n, t) =e−λ(φ,θ)
n!
n∏
i=1
∫
AD(xi;φ)
K∏
k=1
S∏
s=1
PP (ωiks|xi; θ)fts(t(s)ik |ωiks; θ) dx(17)
The likelihood for the Poisson ount model proposed by Eord, Dawson &
Bor hers (2009b) is this likelihood with a single o asion (S = 1) and without
fts(t(s)ik |ωiks; θ)). It ignores the times of dete tion and be ause fts(t
(s)ik |ωiks; θ)
involves θ, the dete tion frequen ies alone (without the times of dete tion)
are not in general su ient for θ. And be ause φ o urs in a produ t inside
the integral, we an't fa torise the likelihood into a omponent with φ and
without t(s)ik , so that (by the Fisher-Neyman Fa torization Theorem) Ω is
36
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le neither su ient for θ nor φ. Finally, be ause neither the binomial nor the
Bernoulli models involve the dete tion times, Ω is not su ient for θ and φ
with any of these models either.
There are two notable ex eptions. The rst is when hk(t,xi; θ) is on-
stant within intervals. In this ase fts(t(s)ik |ωiks; θ) = ωiks!/T
ωiks
s , whi h
does not involve θ (or the dete tion times) and so ounts with the Poisson
model are su ient for θ and φ. Note that in this ase, the multi-o asion
(S > 1) likelihood is identi al to a single-o asion likelihood with o asion
duration T =∑S
s=1 Ts (the only dieren e being the multipli ative onstant
ωiks!/Tωiks
s ). A onsequen e of this is that when dete tion hazards do not
depend on time, the notion of o asion is redundant when using a Poisson
ount model sin e the likelihood is identi al whether or not it involves
o asions.
The se ond is when density is onstant, i.e. D(x;φ) = D. In this ase D
an be fa torised out of the integral and n is onditionally su ient for φ,
given θ.
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le
Table 1 Movement model parameter values used in the movement simu-
lations. CRW refers to a Correlated Random Walk and BRW to a Biased
Random Walk. Only the Von Mises parameters are state dependent; 0o in-di ates movement in the same dire tion while φt represents a turning angle
from the urrent lo ation ba k towards the a tivity entre.
Parameter CRW (state s = 0) BRW (state s = 1)α 3 3
β -0.0012 -0.0012
a 750 750
b 450 450
κ(s) 6 0.5
µ(s) 0o φt
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le
Table 2 Model sele tion summary for the Belize jaguar data, (b) indi -
ates that a behavioural response was in luded. ∆AICc: dieren e in AICc
between the urrent and the best model (AICc is Akaike Information Cri-
terion orre ted for small sample size); wt: Akaike weight, ; npar: number
of estimated parameters.
Model ∆AICc wt npar
Dis rete-o asion models
D(.) g0(.) σ(.) 1.29 0.28 3
D(.) g0(b) σ(.) 0.00 0.53 4
D(.) g0(.) σ(b) 3.83 0.08 4
D(.) g0(b) σ(b) 3.04 0.12 5
Continuous-time models
D(.) g0(.) σ(.) 8.97 0.01 3
D(.) g0(b) σ(.) 0.00 0.81 4
D(.) g0(.) σ(b) 8.18 0.01 4
D(.) g0(b) σ(b) 3.16 0.17 5
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le
Table 3 Estimates from the sele ted SECR proximity model with Ts = 24hours and a ontinuous-time (CT) model tted to the male jaguar data.
Estimated density (D) is individuals per 100 km
2; g00 and g01 are estimated
dete tion fun tion inter epts before and after initial dete tion, respe tively;
σ (in km) is estimated dete tion fun tion s ale parameter. The CT model
used a hazard that is onsistent with a half normal dete tion fun tion shape
over an interval of 24 hours.
Ts D (95% CI) g00 (95% CI) g01 (95% CI) σ (95% CI)
24 3.6 (2.14 ; 6.05) 0.034 (0.016 ; 0.068) 0.067 (0.055 ; 0.094) 2.737 (2.419 ; 3.095)
CT 3.4 (2.02 ; 5.68) 0.028 (0.014 ; 0.055) 0.077 (0.061 ; 0.096) 2.942 (2.624 ; 3.299)
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le Table 4 %Bias and asso iated standard error of density and other estimat-
ors, from simulations with independent data, with dis rete-o asion models
with interval length Ts ranging from 6 to 120 hours, and the orrespond-
ing ontinuous-time (CT) model. `CT (t)' indi ates that a CT model with
dete tion hazard form that generates a half-normal dete tion shape over t
hours was used. True density was 4 individuals per 100 km
2, σ = 2, 400
m, g0 = 0.01 for Ts = 6, g0 = 0.05 for Ts = 24, g0 = 0.15 for Ts = 72,g0 = 0.25 for Ts = 120. Twenty dete tors spa ed at 1,500 m, 3,000 m and
4,500 m apart were used and the survey duration was T = 2, 160 hours. De-te tions is mean number of dete tions per survey, Unique is mean number
of individuals dete ted per survey and Reps is number of surveys that did
not result in estimation error or warning messages (su h estimates were ex-
luded).
Dete tor spa ing Ts %Bias D (se) %Bias g0 (se) %Bias σ (se) Dete tions Unique Reps
62.5% σ (1,500)
6 85.43 (2.51) -41.13 (0.69) -25.43 (0.44) 103.7 10.0 979
CT (6) 8.35 (1.31) -3.01 (0.75) -2.28 (0.46) 103.8 9.9 1000
24 69.35 (2.21) -36.94 (0.73) -21.62 (0.41) 130.6 10.6 988
CT (24) 8.69 (1.16) -4.10 (0.65) -2.49 (0.38) 132.3 10.5 998
72 36.15 (1.66) -21.17 (0.77) -13.49 (0.43) 129.6 10.4 995
CT (72) 6.64 (1.19) -3.40 (0.64) -1.92 (0.41) 134.7 10.4 999
120 21.75 (1.39) -11.99 (0.75) -8.01 (0.41) 130.0 10.5 998
CT (120) 6.42 (1.15) -3.55 (0.62) -1.27 (0.40) 139.1 10.5 1000
125% σ (3,000)
6 1.81 (0.81) -2.97 (0.53) -1.35 (0.25) 101.8 16.0 995
CT (6) -0.51 (0.79) -0.01 (0.54) 0.10 (0.25) 102.0 16.0 1000
24 0.01 (0.82) -0.31 (0.44) -0.68 (0.22) 128.1 16.7 991
CT (24) -0.95 (0.81) 0.61 (0.44) -0.14 (0.22) 129.5 16.6 1000
72 -0.84 (0.78) 0.42 (0.46) -0.05 (0.23) 127.6 16.7 981
CT (72) -1.14 (0.77) 0.17 (0.44) -0.06 (0.23) 132.4 16.6 1000
120 -0.24 (0.81) 0.23 (0.43) 0.15 (0.22) 128.0 16.8 950
CT (120) -0.86 (0.80) 0.24 (0.40) 0.06 (0.21) 136.6 16.7 1000
187.5% σ (4,500)
6 1.25 (0.67) 1.22 (0.52) -0.02 (0.23) 105.2 24.6 994
CT (6) 1.18 (0.67) 1.18 (0.51) -0.02 (0.23) 105.4 24.6 1000
24 0.89 (0.65) 0.74 (0.45) -0.09 (0.20) 130.9 25.6 982
CT (24) 0.84 (0.64) 0.84 (0.44) -0.11 (0.20) 132.8 25.6 1000
72 1.07 (0.68) -0.02 (0.45) 0.29 (0.19) 130.9 25.8 924
CT (72) 0.89 (0.65) 0.22 (0.41) 0.12 (0.18) 136.3 25.7 1000
120 1.51 (0.69) 1.29 (0.46) -0.20 (0.21) 131.6 25.9 865
CT (120) 1.15 (0.64) 1.16 (0.41) -0.25 (0.19) 141.0 25.8 1000
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le Table 5 % Bias and asso iated standard error of density estimators is
shown for two levels of PGC and for two dierent dete tor spa ings (in
meters), using half normal dete tion fun tions, from simulations with an an-
imal movement model, with o asion lengths ranging from Ts = 6 to Ts = 120hours spanning a survey duration of T = 2, 160 hours. Simulated survey
design is as per Table 4. PGC is the probability that an individual is dete ted
when it is in a 250m×250m ell ontaining a amera. Dete tions is mean
number of dete tions per survey, Unique is mean number of individuals
dete ted per survey and Reps is number of surveys that did not result in
estimation error or warning messages (su h estimates were ex luded). The
ontinuous-time (CT) model uses a hazard that is onsistent with a half
normal dete tion fun tion over an interval size of 24 hours.
Dete tor spa ing Ts %Bias D (se) Dete tions Unique Reps
3000 m
PGC = 0.9
6 35.33 (1.21) 367.17 19.81 990
24 13.76 (0.90) 332.07 19.77 997
72 3.82 (0.76) 303.64 19.78 995
120 0.82 (0.81) 283.49 19.77 865
CT (24) -4.33 (0.71) 468.17 19.77 1000
PGC = 0.3
6 8.00 (0.86) 143.20 17.24 992
24 2.48 (0.81) 138.09 17.25 998
72 1.49 (0.81) 133.16 17.25 995
120 1.50 (0.81) 129.35 17.26 990
CT (24) 0.38 (0.85) 153.63 17.28 898
4500 m
PGC = 0.9
6 5.39 (0.64) 361.73 29.15 997
24 -0.03 (0.60) 327.11 29.12 976
72 -0.12 (0.60) 299.25 29.16 987
120 -0.98 (0.62) 278.07 28.93 853
CT (24) -4.45 (0.57) 461.89 29.16 1000
PGC = 0.3
6 -0.91 (0.64) 142.02 25.56 1000
24 -0.26 (0.65) 137.00 25.65 960
72 -0.47 (0.64) 131.85 25.56 988
120 -0.47 (0.65) 128.05 25.56 969
CT (24) -1.83 (0.64) 153.64 25.56 986
42
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le
Fig. 1 Camera trap lo ations within Co ks omb Basin Wildlife San tuary,
Belize.
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le
X axis (km)
−10
−5
0
5
10
Y axis
(km
)
−10
−5
0
5
10
Density
0.000
0.005
0.010
0.015
0.020
Fig. 2 Example of simulated movement for a 90 day period in hourly steps,
showing the number of times ea h ell is visited in the period and smoothed
with a nonparametri smoother.
44
This article is protected by copyright. All rights reserved.
Acc
epte
d A
rtic
le
x axis (km)
y a
xis
(km
)
−5 0 5 10
−4
−2
02
46
810
0 0 0 0 1
0 0 0 0 0
0 0 0 1 0
0 0 0 0 1
True parameters(Binary=black; CT with constant hazard=blue)
Animal #4
x axis (km)
y a
xis
(km
)
−5 0 5 10
−4
−2
02
46
810
0 0 0 0 1
0 0 0 0 0
0 0 0 1 0
0 0 0 0 1
Estimated parameters(Binary=black; CT with constant hazard=blue)
Animal #4
x axis (km)
y a
xis
(km
)
−5 0 5 10
−4
−2
02
46
810
0 0 0 0 0
0 0 0 1 2
0 0 0 1 2
0 0 0 0 1
True parameters(Binary=black; CT with constant hazard=blue)
Animal #5
x axis (km)
y a
xis
(km
)
−5 0 5 10
−4
−2
02
46
810
0 0 0 0 0
0 0 0 1 2
0 0 0 1 2
0 0 0 0 1
Estimated parameters(Binary=black; CT with constant hazard=blue)
Animal #5
x axis (km)
y a
xis
(km
)
−5 0 5 10
−4
−2
02
46
810
0 4 2 3 2
0 1 4 7 3
1 2 5 4 3
0 0 2 1 5
True parameters(Binary=black; CT with constant hazard=blue)
Animal #8
x axis (km)
y a
xis
(km
)
−5 0 5 10
−4
−2
02
46
810
0 4 2 3 2
0 1 4 7 3
1 2 5 4 3
0 0 2 1 5
Estimated parameters(Binary=black; CT with constant hazard=blue)
Animal #8
Fig. 3 Probability ontours for the lo ation of a tivity entres, given the
spatial apture histories for sele ted individuals from a simulated dataset,
generated from both a binary likelihood and a ontinuous-time (CT) likeli-
hood with a onstant hazard. The left panels show the ontours when the
true dete tion fun tion parameters are used for both models, whereas the
right panels do the same using the estimated dete tion fun tion paramet-
ers. Numbers show lo ations and dete tion frequen ies for ea h individual
on ea h dete tor.
45
This article is protected by copyright. All rights reserved.