a finite-state continuous-time approach for inferring regional migration and mortality rates from...
TRANSCRIPT
A Finite-State Continuous-Time Approach for Inferring Regional Migration and MortalityRates from Archival Tagging and Conventional Tag-Recovery ExperimentsAuthor(s): Timothy J. Miller and Per K. AndersenSource: Biometrics, Vol. 64, No. 4 (Dec., 2008), pp. 1196-1206Published by: International Biometric SocietyStable URL: http://www.jstor.org/stable/25502202 .
Accessed: 25/06/2014 01:29
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .http://www.jstor.org/page/info/about/policies/terms.jsp
.JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range ofcontent in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new formsof scholarship. For more information about JSTOR, please contact [email protected].
.
International Biometric Society is collaborating with JSTOR to digitize, preserve and extend access toBiometrics.
http://www.jstor.org
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
Biometrics 64, 1196-1206 DOI: 10.1111/j.l541-0420.2008.00996.x December 2008
A Finite-State Continuous-Time Approach for Inferring Regional Migration and Mortality Rates from Archival Tagging
and Conventional Tag-Recovery Experiments
Timothy J. Miller
Large Pelagics Research Center, Zoology Department, University of New Hampshire,
Durham, New Hampshire 03824, U.S.A.
Current address: National Marine Fisheries Service, Northeast Fisheries Science Center, 166 Water Street, Woods Hole, Massachusetts 02543, U.S.A.
email: timothy.j [email protected]
and
Per K. Andersen
Department of Biostatistics, University of Copenhagen, DK-1014 Copenhagen, Denmark
Summary. Spatially structured population dynamics models are important management tools for har
vested, highly mobile species and although conventional tag recovery experiments remain useful for esti
mation of various demographic parameters of these models, archival tagging experiments are becoming an
important data source for analyzing migratory behavior of mobile marine species. We provide a likelihood
based approach for estimating the regional migration and mortality rate parameters intrinsic to these mod
els that may use information obtained from conventional tag recovery and archival tagging experiments.
Specifically, we assume that the regional location and survival of animals through time is a finite-state
continuous-time stochastic process. The stochastic process is the basis of probability models for observa
tions provided by the different types of tags. Results from application to simulated tagging experiments for western Atlantic bluefin tuna show that maximum likelihood estimators based on archival tagging ob
servations and corresponding confidence intervals perform similar to conventional tagging observations for
a given number of tag releases and releasing tags in each region can improve the behavior of maximum
likelihood estimators regardless of tag type. We provide an example application with Atlantic bluefin tuna
released with conventional tags in 1990-1992.
Key words: Finite-state continuous-time process; Implanted archival tag; Markov process; Pop-up satellite
archival tag; Spatially structured population; Tag-recovery.
1. Introduction
Tagging experiments have proven to be powerful tools to gain
demographic information from animal populations (Pollock,
1991; Schwarz and Seber, 1999; Seber and Schwarz, 2002). Statistical models are available to estimate a variety of param
eters including abundance (Darroch, 1958), mortality (Hearn,
Sundland, and Hampton, 1987), size-specific harvest (Taylor,
Walters, and Martell, 2005), and growth rates (Laslett, Eve
son, and Polacheck, 2002). For populations of migratory an
imals that may be subjected to different levels of mortality in various regions, migration rates among adjacent regions
have also been estimated through tagging experiments (e.g.,
Hilborn, 1990; Hestbeck, Nichols, and Malecki, 1991; Brownie
et al., 1993; Schwarz, Schweigert, and Arnason, 1993), but the
models usually require assumptions that all tagged animals
of a particular release group are released at the same time,
recaptures are made in discrete time, and migration occurs
only once and instantly between intervals. The assumption of
migration time in discrete time models is known to provide bi
ased estimation of migration rates when mortality rates vary
across regions (Hestbeck, 1995). Furthermore, many popula
tion dynamics models used to manage fisheries are parame
terized with instantaneous mortality rates (e.g., Quinn and
Deriso, 1999) and it would appear natural to consider migra
tion in the same continuous time framework.
Tag-recovery experiments continue to be used widely to in
fer rates of movement and mortality of fish populations, but
the relatively new pop-up satellite and implanted archival tags
provide extensive observations of movements of pelagic mi
gratory species (e.g., Block et al., 2001; Bonfil et al., 2005)
and allow inference of location continuously in time (e.g.,
Sibert, Musyl, and Brill, 2003; Sibert et al., 2006). Because
many species that are the subject of archival tagging studies
may have been or continue to be the subject of conventional
1196 ? 2008, The International Biometric Society
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
Finite-State Continuous-Time Tagging 1197
tag-recovery studies as well, a common estimation framework
that can combine the information provided by these studies
would be useful for their management.
We provide likelihoods that (1) provide a basis for esti
mating instantaneous migration and mortality rates, (2) al
low conventional and archival tagging studies to be combined,
and (3) are unencumbered by the unrealistic assumptions of
separation of migration and survival processes in time and
of tagged animals being released in groups at the same time
that are required for most existing models to estimate migra
tion from tag-recovery experiments. To assess the behavior of
maximum likelihood estimators (MLEs), we provide simula
tion results for experiments using different types of tags in a
population of animals that migrates between two regions and
experiences different levels of natural and harvest mortality in either region. We also provide an example application to
Atlantic bluefin tuna (Thunnus thynnus) conventional tagging
studies.
2. Methods
2.1 Finite-State Continuous-Time Markov Processes
Stochastic processes that describe the transition among a fi
nite number of states in continuous time are referred to by
various names including multistate processes or finite-state
continuous-time (FSCT) processes (Karlin and Taylor, 1975;
Taylor and Karlin, 1984; Andersen et al, 1993). We use FSCT
to emphasize the continuous-time aspect of these models and
distinguish them from the discrete-time multistate processes
more commonly used to model mark-recapture experiments
(e.g., Nichols and Kendall, 1995; Lebreton and Pradel, 2002). The fundamental characteristics of FSCT processes are the
instantaneous intensities of transition between states h and
i at time t,ah,i(t). The intensities are formally defined as
ahii(t) =
\im6t->QPh,i(t, t + 8t)/8t where 0 < ahil(t) < oo and
Ph,i(si *) =
p(Y(t) = * I Y(s)
= h) is the probability that the
process is in state Y(-) = i at time t > s given it is in state
Y(-) = h at time s (Andersen and Keiding, 2002). In general,
these transition intensities may change through time (non
homogeneous) and depend on the previous transition history
of the process. We consider here the FSCT Markov process
that is a class of FSCT processes where the instantaneous
transition rates dh,i(s) depend only on the state at time s
(Markovian) and are constant through time (homogeneous) or are piecewise constant through time.
The H x H infinitesimal matrix of homogeneous transi
tion intensities, A = {a^}, has diagonal elements, a,h,h
?
~Yli=i i^hCLh>i' ^ne vames7 ~ah,h-> are often termed hazards
or forces of transition out of the corresponding states (Hoem,
1971). These requirements imply that each row of A sums to
zero. For homogeneous FSCT Markov processes, the probabil
ity transition matrix P(s, t) =
{phji(s, t)} is the exponential of the infinitesimal matrix
P(s, t) = exp{A(t -
?)} = V {Mt ~
S)r , (1) z?' XI x=()
where t > s (e.g., Cox and Miller, 1965, pp. 178-186). The
infinitesimal matrix can be diagonalized with its eigenvalues and eigenvectors so that the probability transition matrix can
be calculated as P(s, t) = Vexp{D(?
- s)}V-1 where D
is a diagonal matrix of the eigenvalues and V is the matrix
of eigenvectors (Kalbfleisch and Lawless, 1985; Commenges,
2002).
2.2 Defining States for Animal Populations In general, an animal may at any instant survive or die from
one of several causes, but these outcomes may be further de
tailed. For example, in many studies we discretize the space
where a migratory animal may be found so that the resulting
regions correspond to habitats that the animals use for dif
ferent purposes (e.g., breeding and feeding grounds) or areas
used to manage harvest of the species (e.g., area-specific har
vest limits). Within a discrete-space system, an animal may,
at any instant, remain alive in a region, die from one of sev
eral causes in that region, or move to any of the adjacent
regions. If H = {1,..., H} is the set of states that an animal
may exhibit, let R, T, and M represent the subsets of states
where the animal is alive and in one of R regions, dead due to
harvest in one of the R regions, and dead due to other causes
(i.e., natural mortality) in one of the R regions, respectively.
We will use these subsets of states in formulating observation
models for the three different types of tagging experiments
below.
The FSCT framework we consider is a generalization of
simpler FSCT Markov processes assumed for many animal
populations. For example, it is common in many population
dynamics models for fish populations, where migration is not
considered, to assume animals either survive or die from nat
ural or harvest mortality sources during a time interval r.
This equates to a process under our framework where there
is a single region and two sources of mortality in that region.
Letting F and M represent the instantaneous rates of transi
tion (within t) into the harvest and natural mortality states,
respectively, the infinitesimal matrix for this process in the
defined time interval is
(-(F + M) F MX
\ ? 0 0/ where 0 is a 2 x 1 column vector. The probability transition
matrix for the interval that provides the probabilities of being in the living and two different mortality states at the end of
the interval is
P = exp(Ar)
=
(-(F+M)t F fi -(F+M)t^ M r, -(F+M)tA
F + MX f F + M1 s \ 0
0 0 1 /
(2) When N0 animals are alive at the start of the interval, the
expected numbers in each state (product of Nq and the prob abilities in the first row of equation (2)) provide results widely known as the Baranov catch equations (Chapman, 1961;
Ricker, 1975; Seber, 1982, p. 329). In fact, equation (2) is the stochastic analog of the Baranov equations and these proba bilities are general results for competing risks models with ex
ponentially distributed life spans (Dupont, 1983; Kalbfleisch and Prentice, 2002, pp. 247-248). Similarly, for a population
occupying two regions with migration rates fii2 and p2\ from
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
1198 Biometrics, December 2008
regions 1 to 2 and 2 to 1, respectively, and no mortality during the interval r, the infinitesimal matrix,
A(r)=(-"u "12) V M21 -A*2l/
yields the probability transition matrix,
P = exp(Ar)
=
/ fl21 + g12e-^^)r m c_(u?+u?)T]
\
I M12 + M21 M12 + M21 |
I /X21 u
_ e-(^^n)r}
Ml2+M2ie-^+^)- I '
\/Xl2 + /i21 ^12+^21 /
where the diagonal provides the probabilities of an animal
being present in each region at the end of the interval given
its presence in that region at the beginning of the interval.
These probabilities were also deduced using different means
by Whitehead (2001, equation (5)) for a model to estimate abundance and migration rates for sperm whales. Analytic so
lutions for P = Vexp{Dr}V_1
are readily obtained when the
set of states are small or when there are no transient states,
but obtaining these solutions analytically becomes too cum
bersome for more complex scenarios and numerical methods
may be more appropriate.
2.3 Probability Models for Tagging Experiments The probability models we present below for observations de
rived from conventional tags, pop-up satellite archival tags
(PSATs), and implanted archival tags (IATs) account for the data provided by the corresponding observations as func
tions of the instantaneous migration and mortality parame
ters. These probability models can also be combined to make
likelihood-based estimates of the instantaneous parameters
for studies that use multiple types of tags. 2.3.1 Pop-up satellite archival tags. When PSATs are de
ployed on animals, they collect measurements of various en
vironmental characteristics including light intensity until the
tag releases from the animal and floats to the surface (pops
up) where it relays information via satellite to the researcher.
The tagged animal may theoretically be harvested prior to
the pop-up time or the tag may pop-up when a tagged an
imal stays at a constant depth for specified amount of time
because the animal is dead due to natural mortality sources.
The relayed measurements of environmental characteristics
can in turn be used to estimate geographic location of the
tagged animal at a given time during the interval that the
PS AT was attached to the animal (e.g., Sibert et al., 2003).
When the ocean is compartmentalized into regions, the lo
cations at time translate into regional occupation at time and
times of migration between regions. Although the locations
are estimated rather than known, it may be safe to ignore the
resulting uncertainty in times of migration between regions
when the regions are large relative to the uncertainty in the
locations. Times of entry into states corresponding to harvest
mortality are directly observed and not subject to measure
ment error. Likewise, when entry into natural mortality states
are observed by pop-off of the tag due to constant depth, the
times of transition into these states may be assumed known.
Here we consider the migration times as known so that the
data on a tagged animal k observed from the time the animal
is tagged ?0,fc to the pop-up time ta,k are the times of tran
sition, ti?fc =
?1)fc,..., tijk,..., tnk,k '
to,k <
ti,k <
ta,k, and the
states, Yi,fc =
Y(thk),..., Y(tz>fc),..., Y(tnktk) : Y(tlik) G H,
the animal takes on during the period the tag is operating,
Tk = ta,k
? ^o,fc- As these are data for a continuous-time ob
servation of a FSCT process (Andersen et al., 1993; Andersen
and Keiding, 2002; Commenges, 2002), the probability func tion for the set of observations from a PSAT is
Pi,k(Yi,k,t\,k | to,k,Y(tojk))
i=o (Jtik )
x exp I J aY{tnktk)iY{tnktk)(t)dt \ . (3)
Note in equation (3) that the first line represents the prod
uct of probabilities of remaining in states between transition
events and the transition to the next state. The second line
of equation (3) represents the probability of remaining in the
final observed state after the final observed transition. We as
sume that the release time (ta,k) is fixed or at least indepen
dent of the process Y(-) in equation (3) (i.e., an independent
censoring time; Andersen et al., 1993, Ch. III). When the transition intensities are constant within the in
terval rk, equation (3) reduces to
H
Vhk(Yi,k,ti,k \to,k,Y(t0,k)) =
Y[ea^T^ l\anh^k, (4) h=l i^h
where raMjfc =
XXV* I(Y(tijk) = h, Y(tl+hk)
= i) is the
number of transitions from state h to state i and Th,k =
{?r=5>~l7(y(*',*) =
>0(*i,fc -
*i+i,*)} + i(Y(tnk,k) =
h)(ta[k -
tnk,k) is the total amount of time spent in state h. Determin
ing the MLE for the transition rate ah^ is straightforward
(Albert, 1962) and asymptotic variance estimators have
also been derived (Andersen and Keiding, 2002). When a
piecewise homogeneous model is used where transition rates
are constant between time points T\,..., T3 ,,..., T/v, some or
all of the instantaneous rates may be interval specific.
2.3.2 Conventional tags. In conventional tag-recovery ex
periments, an animal is released with a unique tag and people
harvesting the species may recover the tagged animal during
the course of their efforts, so given the state at time of release
(Y(to,fc)) we observe the time of recovery (?r,fc) and states at
the instant prior to recovery (Y(tr,k?) G 1Z) and at recovery
(Y(?r>fc) G T) for harvested animals. For unharvested animals,
we know that the animal is either still alive or dead from nat
ural sources at the time of analysis (Y(ta^k) G {IZ, M}). Note
that Y(tr,k) =
Y(tUfk) for recovered animals. Thus, the data
are
({Y(tr,k-),Y(tr,k),tr,k} ifY(ta,k)eF {
u,k,tii,k}-^Y^^e{nM^ otherwise
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
Finite-State Continuous-Time Tagging 1199
The probability function for the set of observations from a
conventional tag is
^VfcCYii^, tnffc | t0,fc, ta,k, Y(tQ,k))
rD \i(Y(ta,k)er) = VY(t0,k),Y(tr,k-)aY(tr,k-),Y(tr.k)}
{\ HY(tatk)e{R,M})
^2 PY(t0,k),Y(ta,k) > , (5)
Y(ta,k)e{TZM} )
where I(Y(ta,k) T) and I(Y(ta,k) <E {11, M}) are indicators of whether the animal is in harvested and nonharvested states,
respectively, at the time of analysis. The first line in equa
tion (5) is the product of the probability of being alive in the
region of recovery just prior to harvest at time tr,k ?
given
Y(t0jk) and the instantaneous rate of harvest in the region
where recovery occurred. The probability Py(t0,fc),y(trifc-) is
the (y(t0,fc), Y(tr,k?)) element of the probability transition matrix P(*0,fc, tr,k-) and
ay(.rifc-),y(t0ffc) is the (Y(tr,k~),
Y(ta,k)) element of the infinitesimal matrix. The second line in equation (5) is the probability of being in any of the non
harvested states at the time of analysis given Y(tQ^k), which is
the sum of the elements of the probability transition matrix
P(to,fc, *<_,*) in row Y(*0,fc) where Y(ta,k) ? {11, M}. All tags are assumed to be reported when animals are har
vested, but incomplete tag reporting could be incorporated
into the model by expanding and redefining the possible states
(see Section 4 below for an example). In practice, we de
termine the maximum likelihood estimates using numerical
methods similar to Kalbfleisch and Lawless (1985) where we
determine the instantaneous rates and the probability transi
tion matrices using the eigendecomposition of the infinitesi
mal matrix and equation (1) such that the required elements
maximize equation (5). When a model is used where transition rates are piecewise
constant between time points T\,..., 7},..., TN and j0,fc =
min{j : T3 > ?0,j_} and jr>fc
= max{j
: 7} < ?r,fc}, the prob
ability transition matrices required are the product of the
interval-specific matrices,
P(?o,fc,?r,fc) =
P(tQ,k,TJok)P(TJok,Tjo+])
due to the Markov assumption. The probabilities necessary for an unrecovered tag are obtained by the product of the
interval-specific matrices where ta,k is used instead of trjk. 2.3.3 Implanted archival tags. Like conventional tags, IATs
will remain in tagged animals for the remainder of their lives and we only obtain postrelease information from animals that
are harvested. However, like PSATS, we can observe times of
transition and states before the time of tag battery expiration
(te,k) for a recovered animal (ti.fc =
*i>fc,..., tLk,..., tnk,k :
to,* <
U,k <
te,k, YI>fc =
Y(t_,fc), , Y(t.,fc), , Y(tnk,k)). If
the animal is recovered after the battery expires, we also have
an observation analogous to that of a conventional tag where
the state at the time of battery expiration (te,k) and states
just prior to and at the time of recovery (tr,k) are known.
Finally, we also have animals that are tagged and not recov
ered at the time of analysis (?a,fc) as in conventional tagging
experiments. Thus, the data for an I AT are
{Yni,fc, tin,fe}
{Yilfc,ti,fc if Y(te,k)ef
Y,lfc,ti>fc,Y(tc,fc), if Y(te,k)eU
Y(tr,k-),Y(tr,k),tr,k and Y(ta,k) G T '
[Y(ta,k) G {II, M}} otherwise
The probability function for the set of observations from
an IAT is
/^>iii,fc(Yn])/e, tin,*; | t0,k, ta,k, Y(t{)^k))
= ?u(Yi,*,ti,fc \tQ,k, Y(to,fc))/(y(t-fc)e^
rD i/(y(*e,fc)eW,y(tO|fc)GJ0 x vy(te../c)^(tr.fe-)ay(tr,fc-),y(tr,fe))
{^ \ HY{tatk)e{K,M})
Yl py^.khY(ta,k) \ > (6)
y(ta.fc)G{7e,A^} J
where 7\fc(Yi,fc, ti)fc | ?o,fc> Y(?o,fc)) is the component represent
ing continuous-time observation between times to,k and te,k
(i.e., equation (3)), the second line represents the observation
of recovery after battery expiration (analogous to the first line
in equation (5)), and the final line represents the observation
of an animal that remains unrecovered at the time of analysis
(analogous to the second line in equation (5)). As with the release time for PSATs, we assume that the battery life (tBjk) of an IAT is fixed or at least independent of the process Y(-) in equation (6).
3. Behavior of MLEs for Conventional
and Archival Tagging Experiments
Through simulation, we explored the behavior of MLEs for the
regional migration and mortality rate parameters based on the
three likelihoods presented above for conventional, PS AT, and
IAT observations. We intend the process observed through the
tagging experiments to reflect the life history of adult west
ern Atlantic bluefin tuna. The northern Atlantic Ocean is split into western and eastern regions at the 45?W meridian for in
ternational harvest management (e.g., ICCAT, 2007) and fish
may migrate between western and eastern regions of the At
lantic Ocean and die from harvest or natural causes in either
region.
We assumed instantaneous west-to-east (pw,e) and east
to-west (pe,w) migration rates, region-specific instantaneous
harvest (Fw and FE), and natural mortality (Mw and ME) rates are constant throughout the time of analysis for each
animal to reduce the complexity of the results and better ob
serve the behavior of the MLEs for the six instantaneous rate
parameters. Thus, the infinitesimal matrix we use throughout simulations is
(ai,i Vw,e Fw 0 Mw 0 \
Ve,w a2,2 0 FE 0 ME , (7) 0 0 0 0 0 0/
where aM =
-(pw,E + Fw + Mw),a2j2 =
~(pE,w + FE +
ME), and 0 is a 4 x 1 column vector. Note that states 1 and
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
1200 Biometrics, December 2008
2 (alive in the western and eastern Atlantic Ocean) comprise the subset IZ, states 3 and 4 (harvested in the western and
eastern Atlantic Ocean) comprise T, and states 5 and 6 (nat ural mortality in the western and eastern Atlantic Ocean)
comprise M. The values we assumed for harvest and natu
ral mortality (Fw =
0.18, FE =
0.35, Mw =
0.1, and ME =
0.15) are similar to those stock assessment scientists believe
to operate on adult Atlantic bluefin tuna in the western and
eastern Atlantic Ocean based on population dynamics mod
els (ICCAT, 2007). The western-to-eastern and eastern-to
western Atlantic migration rates we assumed (pw,E = 0.073
and Pe,w =
0.066, respectively) were estimated from prob abilities of western Atlantic bluefin tuna being in different
regions of the Atlantic at times subsequent to PSAT experi ments reported by Block et al. (2005). All rates are specified on the yearly time scale.
The general algorithm for simulating the transition history is the same for all tag types. The animal is released in a given
region (Y(tQjk) G IV) and the waiting time until the animal either leaves the region or dies in the region is an exponen
tial random variable with rate parameter ?
a^/i- Given that
the animal either leaves or dies, the (H ?
1) x 1 indicator
vector for the next state is a multinomial random variable
with probabilities ?
a/lji/a/l)/l, i / h (e.g., Taylor and Karlin,
1984, pp. 253-255). This procedure is repeated until the anal
ysis time is reached or the animal dies. As described above, each tag type yields observations with different information
on the process. We simulated 1000 experiments for each ex
periment type. Experiment type is defined by the type of tag used (PSAT, IAT, or conventional tag), the total number of
tags released, and the proportion released in each of the two
regions. We generated data and calculated likelihoods for each
experiment using C programs compiled for R (R Development Core Team, 2006).
For each simulated experiment we maximized the corre
sponding likelihood and determined whether parameters were
estimable. For parameters that were estimable, we calculated
MLEs, coefficients of variation (CVs), and 95% confidence in
tervals so that we could compare bias, precision, and 95%
confidence interval coverage, ^(0.95), of different experiment
types. As the possibility of making inferences is as important as (or more important than) the behavior of inference pro
cedures when it is possible to use them, we also calculated
the proportion of experiments where each parameter was es
timable ((p) for each experiment type.
We maximized the likelihoods using the Nelder-Mead algo rithm of the optim function in R (R Development Core Team,
2006). The maximizations were carried out with respect to the
log-transformed instantaneous rates (A^ =
log(aij:/)) to avoid
boundary problems and variance estimates were derived from
inverting numerically derived Hessian matrices in the corre
sponding space. We calculated asymmetric confidence inter
vals as exp{Ai?j ?
ZQ,975SE(Xij)}. There is a high incidence of release of PSATs prior to the
scheduled time from causes other than mortality (e.g., Sibert
et al., 2006), which we assume is independent of the FSCT
process. In our simulation of PSAT experiments, we modeled
the release time of the tag (?0,j_) using a delta-beta distri
bution: indication of tag release prior to the scheduled time
is a Bernoulli random variable with probability p and given
t I
5> \ ?5 " > _
\
q _ -n^^
p _ I_I_I_I_I_I_|_I_1 I >] S-1-1-1-1-H
0.0 0.2 0.4 0.6 0.8 1.0
Proportion of maximum tag life
Figure 1. Distribution of the scaled release times of pre
maturely released pop-up satellite archival tags scheduled to
release more than 250 days after deployment (n =
192). The curve represents a beta distribution with parameters (a
=
0.679,6 = 1.508) determined by fitting the scaled release
times. The probability of premature release is p = 0.255.
the tag releases early, the time of release scaled by the sched
uled release time is a beta random variable with parameters a and b. We used values for delta-beta parameters that we
estimated from release times of pop-up tags that have been
deployed by Lutcavage and colleagues (Figure 1). We assumed
that the scheduled time of release for tags is 0.83 years (10 months), which is similar to the intended release time of cur
rent PSATs. For IATs and conventional tags, we assumed the
time of analysis is 5 years after the time of release and the bat
tery life of IATs is similar to that of batteries currently used
(2 years).
3.1 Simulation Results
3.1.1 Effect of tag release size on MLE behavior. The prob
ability that assumed instantaneous rates are estimable con
verges to one for all tag types, as the total number of released
tags increases (Figure 2, row 1). The rate of convergence of
(j) is generally higher for harvest rate MLEs than migration and natural mortality rate MLEs. Bias of the MLEs dimin
ishes with total number of released tags for all tag types, but
the rate of convergence to unbiasedness is greatest for har
vest rate MLEs (Figure 2, row 2). The rate of convergence to
unbiasedness of migration rate MLEs is generally lowest for
PSATs. The precision of MLEs for all migration and mortal
ity parameters increases with total number of released tags
and is similar for all types of tags when estimation is possi
ble (Figure 2, row 3). For a given total number of released
tags, the CVs of estimators for migration and harvest mortal
ity parameters are lower for IATs and conventional tags than
PSATs and the CV of the natural mortality rate estimator in
the western region is lowest when PSATs are used. The cover
age of 95% confidence intervals converges to the appropriate
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
Finite-State Continuous-Time Tagging 1201
\^w,e ^e,w Fw FE Mw ME ? .-. .-~_-:-. .-. .-, .-. .-. . _ qoo-?>o-o - Q0'?9m..m?.-
~ o - ooooo ? o-o - ooooo ? o-o - q?~~?-?
~ o8?
? ?-?
- if f^ 1 I ?* ?!
~H ? - -
i -I d-
-J
- - -I
^ _ ? _ o 0 -i.-j-j-j-j-_j -i_-__-(-!-j-_j "Xj-j-j-j-j-_j -i_-j-|-!-]-.J -l.-j-j-j-(-_J "1-,-j-,-(-,-_J
i?TTT I -n\ I tn I fl I tl I 11 I >s- U -
5\ -i - -
k - s
I-?o\ -
c\ -
v, -i - v- - i
(So- lK.-:-S^^rr,8 - ?
o:^^-:? - S-8^--_Ji___B - ^8^8=^__g . \?--^^g
- "^^^z^.%
1 t-1-1-1-1-r t-1-1-1-1-r t-1-1-1-1-r h-1-1-1-1-r h-1-1-1-1-r t-1-1-1-1-r
oo . _ ; 6 . O ~ o. o :
~ o".
|j>
o _ _ _ _ ~~.0 _ ? -L--j-j-j-j--pi "__-j-,-j-,--J -tj-j-,-j-j--J -L.-,-(-(-(--J T-,-(-(-(-(--J "l--(-(-(-f--J
O_ ,_ _;_ _, _ ,_,_, ? _j : : I _j n _i : i j _j I _j
^-f?. -f^^ -|s>?. ?**? .^r^, .j_-_?.
ts:i : : :' T :? O ? 8 00 _
? n-1-1-1.i?r h-r?i-1-1-r n-1?-~i-1-1-r n-1-1-1-1-r n-1-1-1-1-r n-1-1-1-1-r 0 200 600 1000 0 200 600 1000 0 200 600 1000 0 200 600 1000 0 200 600 1000 0 200 600 1000
Number of tags released
Figure 2. Probability of estimation ((f)), percent relative bias, coefficient of variation (CV), and probability of approximate
95% confidence interval containing true parameter value (^(0.95)) for maximum likelihood estimators of the migration and
mortality rates (columns) when experiments use pop-up satellite archival (solid line), implanted archival (dashed line), or
conventional (dotted line) tags released in equal proportions in the western (W) and eastern (E) regions. Bias, CV, and
^(0.95) results are based on the subsets of 1000 simulations of tagging experiments for each tag type and sample size in which
parameters are estimable.
probability for all types of tagging experiments as the total
number of released tags increases (Figure 2, row 4). The rate
of convergence is highest for MLEs of harvest rates, but low
est for the MLE of the natural mortality rate in the western
region (Mw) using IATs and conventional tags in particular.
Because the relative bias, CV, and confidence interval cover
age for a given parameter are all based only on simulations
where the parameter is estimable, the values observed when
4> is low may be less informative than when <j> is high.
3.1.2 Effect of regional allocation of tag releases on MLE
behavior. For PS AT experiments, behavior of a given MLE
generally improved with the proportion of tags released in
the region corresponding to the state prior to transition
(Figure 3). For example, pw.E^Wi and Fw are more likely to be estimable and exhibit less bias, greater precision, and
better confidence interval coverage when more PSATs are re
leased in the western region. For studies using conventional
tags or IATs, the relationship of a region where all releases
occur to the instantaneous rate parameter is less predictable.
For example, the probabilities of Pw.e^e.w^w^ and ME
being estimable for IATs are low whether all tags are released
in the western or eastern region. When some tags are released
in both regions, behavior of MLEs is generally similar across
tag types with the exception that MLEs for migration rates
can be slightly more likely to be estimable and have slightly
less bias with IAT and conventional tag studies. As with the
results above on the effect of tag release size on MLE behav
ior, the values observed when <j> is low may be less informative
than when cf) is high.
3.1.3 Effect of release size and regional allocation on corre
lation of MLEs. The correlation of migration rate MLEs with
each other and other MLEs is negligible for all tag types when
tags are released in either region (Figure 4). For IAT and con
ventional tag studies, harvest mortality rate and natural mor
tality rate MLEs for a given region are positively correlated
but are negligibly correlated between regions. The correlation
of MLEs in conventional tag and IAT experiments can cause
confounding of parameters and the poor behavior of MLEs
we observe at low tag release sizes and releasing all tags in a
single region.
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
1202 Biometrics, December 2008
\*>w,e \^e,w Fw FE Mw ME - 0 - - - O - ? - @ -i-^? O - O ??ZTg?-
-* O -. - _ 0 ~0 - Qz-=^*~0 .o??o -o o o o.O -
_,q.^-^Q -o" - O O 0=^^-0.o 00 - / j/ \ - >' \
N - >'/ - V - i':? \ _ / V o ; ox \ / \ \ / / \\ /./ \ / \v ; j \ O/ O S O/ I \ \ /. / ^ f \ \
^_? / - \\ - / - \ -": - \? 0/ '
\ ? / ? / \
7 - \
-? - -/ ? o ~S-1-1-1-1-H ~h-1-1-1-1-r ~h-1-1-1-i-r ~h-1-1-1-1-r ~h-1-1-1-1-r ~H-1-1-1-1-r
^sn 1 f1 n n 11 n k^i i i n 1 . v .-" -;
A--'
"
/
" v
" I,'
^^_ - - - - _8" 1 t-1-1-1-1-r n-1-1-1-1-r n-1-i-1-1-r n-1-1-1-1-r n-1-r~?i-1-r i-1-1-1-1-r
CJ "L,-,-!-(-!-jJ -L,-,-j?-,-1-jJ -1^-j-,-!-!-,J "1,-(-,-,-,-^ "t,-,-,-,-,-,J "t,-,-,-,-,-,J o ,_, ,_, ,_, ,_;_, ,_, ,_, O -I I -I 51 -I I -I -fft
,J O ^ ?' ?. ^00 o ** ft -: - - a ̂ J-~~' 8_. 8 ?t- Q *- - O . . O O ?_n n _Q- ? '?. ._? -' '"' ? ̂ ̂ a_a-" ? s ' n
^ _ ,8 ?-^8=^- _0-?f
-8^- _ . ??or-0 -8--^e-^r=?-,,A,.,@ _ y o-o ^0. .o^-o-^e..,
,.<..-o
71...H-../,,..H. ... .\ o n-1-1-1-1-r t-1-1-1-r-r V?i-1-1-1-r n-1-1-1-1-r n?h-1-1-1-r n-1-1-1-1-ir
0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80 0 20 40 60 80
% tags released in western region
Figure 3. Probability of estimation (</>), percent relative bias, coefficient of variation (CV), and probability of approximate
95% confidence interval containing true parameter value (^(0.95)) for maximum likelihood estimators of the migration and
mortality rates (columns) when experiments use 200 pop-up satellite archival (solid line), implanted archival (dashed line), or conventional (dotted line) tags released in varying proportions in the western (W) and eastern (E) regions. Bias, CV, and
ij)(Q.95) results are based on the subsets of 1000 simulations of tagging experiments for each tag type and sample size in which
parameters are estimable.
4. An Example: Atlantic Bluefin Tuna
Conventional Tagging Data
Conventional tagging experiments for Atlantic bluefin tuna
have been conducted throughout the northern Atlantic Ocean
for several decades and the resulting data have been archived
by the International Commission for the Conservation of
Atlantic Tunas (ICCAT). We considered a subset of the archived data where tagged fish were released in 1990-1992
and recaptures were made prior to the beginning of 1995
(assumed end of study). We used the longitudes of tagged fish at release and recovery to determine the regional locations
at those times. The dates of release and recovery are used
to determine the times on study (ta^k ?
?o,& !__ 5 years). Of
the 2022 tagged fish released in the western region of the
Atlantic Ocean, 65 were recaptured in the same region
and 6 were recaptured in the eastern region. Of the 4288
tagged fish released in the eastern region, 72 were recap
tured in the same region and 6 were recaptured in the western
region.
For this example, we have the same instantaneous rates
as the simulation study above, but the probability of report
ing harvested tags is thought to be substantially less than 1.
Because there is no auxiliary information to estimate report
ing probability and it is completely confounded with natural
mortality rates, we assume a value for natural mortality rate
(Mw =
ME ?
0.15), similar to values believed by Atlantic
bluefin assessment scientists (ICCAT, 2007), and estimate re
porting probability (0 < p < 1) along with migration and harvest mortality rates. The resulting infinitesimal matrix is
(ai,i Pw,e PwFw 0 (1 -
pw)Fw + Mw 0 \
Pe,w a2i2 0 p^Fb 0 (1 -
pE)FE + ME\ , (8) 0 0 0 0 0 0 /
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
Finite-State Continuous-Time Tagging 1203
% tags released in western region 0 20 60 100 0 20 60 100 0 20 60 100 0 20 60 100 0 20 60 100
J_I_I_I_L_L_ J_1_I_I_1_L_ J_I_I_I_I_L_ J_I_I_I_I_L_ J_I_I_1_I_L p
\lWE -- ; ?^-- ^:- i^i^^irrr_
^^^??^
.
^A_~~-^- L^?^-]-?
?
T p_ I.. ,i_1_i_i., ,i i,, ., i_i i i i J_I ?,i, . , I_i_l_ I. I_i i i I q
g-^^. - ---
\*>E,W -y77^^-.-_ -^-^?^- ^?-**- ^--?y,:?
? "
? ~
4.-r?T------,-pi - - - -1--'
p _,_,_,_ _:____ J_I_I_I_1_L J_I I_1_I_L_ J_I_I_1_I_i__ p
? ^ -t,-,-,-._,-pj -_.-,-,-,-,-_j - -
-h^ ?_ O ,_?_. ,_ ,_._. J_1_I_I_I_L J_I_I_I_I_L O
s -1 i r ~
U q-t-_- -1^^.~~-^--~^_?_?: -u^^^^~^-^ J7j-, =^^~~~^z^iL._ ill_LL_.2_i___._q
? p --, -_.-___?j-j-,-pJ -4--,-,-,-,-pJ -L.-j-,-(-,--J
-- --.
p _._,_,_ _ J_I_I_I_I_L O
q _\z^?Lzjt^z^^=^ J^?^?___-__-______- -?c^-.--.--..-.->^--.. A/1\rr isl*^^.^^^^-^
? ~| ? --, H( t ,-f-(-!--J -4--___?|-]-!-pJ -__-(?__--j-(--J -Lj-j-j-j-j-pJ
- -p
O _,_,_ _,_:_ _;_ _ _
? ---
H_-,-!-,-,--J -4--,-,-,-,-jJ 4_-(-,-(?_--pJ -4--(-(-(-(--J -L.-,-,-(-pJ 0 2000 4000 0 2000 4000 0 2000 4000 0 2000 4000 0 2000 4000
Number of tags released
Figure 4. Correlation of maximum likelihood estimators of the migration (pw,E and /jLe,w) and harvest (Fw and Fe) and
natural mortality (A_V and ME) rates when experiments use 200 pop-up satellite archival (solid line), implanted archival
(dashed line), or conventional (dotted line) tags released in varying proportions in the western (W) and eastern (E) regions
(upper right) and released in equal proportions in the western (W) and eastern (E) regions (lower left). Results are based on
the subsets of 1000 simulations of tagging experiments for each tag type and sample size in which parameters are estimable.
where aM =
-(pw,E + Fw + Mw),a2j2 ?
-(ps,w + FE +
ME), and 0 is a 4 x 1 column vector. We fit a suite of nested
models where all rates are the same for each region (pw,E ?
(J>e,w,Ew ?
FE, and pw =
Pe\ mo)^ only harvest mortality rates are region specific (mi), harvest and migration rates
are region specific (m2), migration rates are region specific
and harvest rates are both region and year specific (m3), and
reporting probability is also region specific (m?). For each model, we maximized equation (5) with respect
to the migration and harvest mortality rates and reporting
probabilities. The probabilities necessary in equation (5) are
calculated using equation (1) and eigendecomposition of equa
tion (8) (year specific for models m% and m4). For the best
model, we estimated variances and calculated asymmetric
confidence intervals for the migration and mortality rates us
ing the same methods as the simulation study above, but a
logit transformation (rather than log transformation) of re
porting probability is estimated due to its constraints (i.e, 0 < p< 1).
We calculated x2 goodness-of-fit test statistics (x2gof)
f?r
each fitted model by comparing the observed numbers of tags
reported in each region each year for each region of release
and the observed numbers of tags never reported with the
corresponding expected values under each model. As there
are five years and two regions of reporting and one category
of tags never reported for each region of release, the degrees
of freedom for the goodness-of-fit test is dfcoF ? 20. We also
compared nested models sequentially using likelihood ratio
tests with degrees of freedom equal to the difference in the
number of parameters between models.
The goodness-of-fit tests show that models mQ,m\, and m^
are not plausible mechanisms for generating the pooled num
bers of reported tags in each region and year (Table 1). Both
models m3 and ra4 could not be rejected using the goodness-of fit test, but the model with region- and year-specific harvest
mortality rates and region-specific migration rates (m%) pro
vides a better fit to the Atlantic bluefin tuna data using the likelihood ratio test. Based on model m3, there is evidence of
a very low tag-reporting probability from harvested individu
als (~0.122) and estimates of harvest mortality are greater for
the western region than the eastern region within a given year,
but the estimates tend to decrease over time for a given region
(Table 2). The migration rate from the western region to east
ern region of the Atlantic Ocean is greater than the opposite
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
1204 Biometrics, December 2008
Table 1
Goodness-of-fit tests and sequential likelihood ratio tests of nested models fit to Atlantic bluefin tuna conventional tagging data. The fitted models have all migration rates, fishing mortality rates, and reporting
rates constant over years and regions (mo), harvest mortality is region specific (mi), migration rates are also
region specific (m2), harvest mortality rates are also year specific (ra3), and reporting rates are also region
specific (ra4). The goodness-of-fit statistic (Xgof) and corresponding p-values (Pgof), log likelihood at the maximum (log(L)), number of parameters (np), (log)likelihood ratio statistics (xIrt) and corresponding
degrees of freedom (dfrnr) and p-values (Plrt) are provided.
Model XcoFidfcoF =
20) PG0F log(?) np xIrt dfLRT PLRT
mo 90.57 <0.0001 -1786.65 3 mi 76.99 <0.0001 -1780.38 4 12.54 1 0.0004 ma 65.62 <0.0001 -1776.26 5 8.24 1 0.0041
m3 9.34 0.9786 -1752.03 13 48.46 8 <0.0001 m4 8.48 0.9881 -1750.81 14 2.44 1 0.1183
Table 2 Maximum likelihood estimates for migration rates (pw,E and
He,w), region- and year-specific harvest mortality rates (F) and reporting probability (p), and corresponding asymmetric
95% confidence intervals provided by model m3 from fitting Atlantic bluefin tuna conventional tagging data
Parameter Estimate SE CI
pw,E 0.1832 0.0750 (0.0821-0.4089) pE,w 0.0374 0.0158 (0.0164-0.0855)
^,1990 0.5168 0.2491 (0.2009-1.3292) Fe'mm 0.1773 0.0816 (0.0719-0.4369) Fvy,i99i 0.3946 0.1545 (0.1832-0.8498)
FE',i<m 0.1360 0.0538 (0.0627-0.2953)
FW1992 0.1255 0.0580 (0.0507-0.3107) FE',im 0.0370 0.0167 (0.0153-0.0895) F^i993 0.1256 0.0584 (0.0505-0.3123) F^'1993
0.0641 0.0279 (0.0273-0.1505)
Fw,m4 0.0910 0.0523 (0.0295-0.2809) FE]vm 0.0300 0.0157 (0.0107-0.0838)
p '
0.1222 0.0063 (0.0630-0.2238)
migration rate, which would result in a greater probability of
living individuals occurring in the eastern region over time.
5. Discussion
The results we present here demonstrate the potential utility
of the FSCT approach for combining information from the
different tagging studies to estimate migration and mortality rates. Although the simulation results assume instantaneous
rate parameters presumed to be operating within western At
lantic bluefin tuna populations, the general result of increased
precision, reduced bias, and accurate confidence interval cov
erage with increased numbers of released tags in each region
of interest should apply to any scenario.
There are other attributes of a tagging study we did not
consider that can affect the behavior of MLEs. The battery
life of archival tags, rates of migration and harvest mortality,
and duration of the study will also affect the behavior of the
MLEs for the various migration and mortality rates. In the
extreme case of zero harvesting in any region, none of the pa
rameters will be estimable with any number of conventional
tags or IATs released because no tags will be recovered. For
both types of archival tags, the longer the battery life, the
more continuous-time observations are made, which in turn
provides more information. When the expected duration in
a particular state is long with respect to a given tag battery
life, duration of study, and/or number of deployed tags, the
expected number of transitions to and probabilities of being in other states will be small and vice versa. Also, the longer the duration of the study given a set of migration and mor
tality rates, tag battery life, and number of deployed tags, the more transitions would be expected. The form of the
asymptotic variance estimator for MLE with PSAT experi ments (Andersen and Keiding, 2002) shows that the precision
increases with the number of transitions. That the same like
lihood is a component of the IAT experiments implies that
increased precision with more transitions would also, to per
haps a lesser extent, occur for those experiments. However, the interval censoring that occurs for IATs and conventional
tags makes the relationship less clear for these experiments.
In practice the rates of migration and mortality are unknown,
which would further argue for release of tags in all regions. Realities such as less than total tag reporting and tag shed
ding not considered here will also diminish the information
provided from all types of tagging experiments. In more realistic population models, the rates of migration
and mortality may change through time and the life history of the animal. For example, migration rates from the region of
origin may increase with age for Atlantic bluefin tuna or har
vest mortality rates may be zero outside of or may change dra
matically within harvesting seasons for many species. In these
cases, several sequential infinitesimal matrices with rates con
stant over shorter intervals may be considered. With a given
number of releases, duration of analysis, and tag battery life
(for IATs), the instantaneous rates for these more complex
life-history processes would also be more poorly estimated
than those driving the simpler stochastic process assumed in
the simulations. As such, more releases throughout each year
would provide better information for estimating these param
eters.
Although we partition space and parameterize the FSCT
process for fisheries management, the partitioning and
parametrization can also address behavioral or ecological
characteristics of the population of interest. With Atlantic
bluefin tuna, for example, we might further partition the space
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
Finite-State Continuous-Time Tagging 1205
to represent spawning and feeding grounds within each re
gion because adults are thought to migrate between these
areas during certain periods each year. We might also con
strain migration rates to the spawning and feeding grounds to be higher at the appropriate periods to reflect these life
history attributes of the population. The regional migration
and mortality model is also directly applicable to manage
ment regimes for many species that employ protected areas
where no harvest is allowed (e.g., Roberts et al., 2001; Field
et al., 2006). Harvest rates would be set equal to zero in the
protected areas and migration rates among the protected and
unprotected areas could be estimated from tagged animals.
Acknowledgements
We thank Daniel Commenges, Andy Cooper, Andy Rosen
berg, and John Sibert for productive discussions on ideas
involved in this work and Victor Restrepo and Papa Kebe
at ICCAT for help with conventional tagging data. We also thank Carl Schwarz for providing helpful comments
on an early draft and a referee and associate editor who
also provided comments that improved this manuscript. This
work was generously funded by contract NA04NMF4550391 awarded to M. Lutcavage by NOAA National Marine Fish
eries Service.
References
Albert, A. (1962). Estimating the infinitesimal generator of a
continuous time, finite state Markov process. Annals of
Mathematical Statistics 33, 727-753.
Andersen, P. K. and Keiding, N. (2002). Multi-state models
for event history analysis. Statistical Methods in Medical
Research 11, 91-115.
Andersen, P. K., Borgen, 0., Gill, R. D., and Keiding, N.
(1993). Statistical Models Based on Counting Processes.
New York: Springer-Verlag.
Block, B. A., Dewar, H., Blackwell, S. B., Williams, T. D.,
Prince, E. D., Farwell, C. J., Boustany, A., Teo, S. L. H.,
Seitz, A., and Walli, A. (2001). Migratory movements,
depth preference, and thermal biology of Atlantic bluefin
tuna. Science 293, 1310-1314.
Block, B. A., Teo, S. L. H., Walli, A., Boustany, A., Stokes
bury, M. J. W., Farwell, C. J., Weng, K. C, Dewar, H., and Williams, T. D. (2005). Electronic tagging and pop
ulation structure of Atlantic bluefin tuna. Nature 434, 1121-1127.
Bonfil, R., Meyer, M., Scholl, M. C, Johnson, R., O'Brien,
S., Oosthuizen, H., Swanson, S., Kotze, D., and Paterson, M. (2005). Transoceanic migration, spatial dynamics and
population linkages of white sharks. Science 310, 100
103.
Brownie, C, Hines, J. E., Nichols, J. D., Pollock, K. H., and Hestbeck, J. B. (1993). Capture-recapture studies
for multiple strata including non-Markovian transitions.
Biometrics 49, 1173-1187.
Chapman, D. G. (1961). Statistical problems in the dynam ics of exploited fisheries populations. Proceedings of the
Berkeley Symposium on Mathematical Statistics and Prob
ability 4, 153-168.
Commenges, D. (2002). Inference for multi-state models from
interval-censored data. Statistical Methods in Medical Re
search 11, 167-182.
Cox, D. R. and Miller, H. D. (1965). The Theory of Stochastic
Processes. New York: Wiley.
Darroch, J. N. (1958). Estimation of a closed population.
Biometrika 45, 343-359.
Dupont, W. D. (1983). A stochastic catch-effort method for
estimating animal abundance. Biometrics 39, 1021-1033.
Field, J. C, Punt, A. E., Methot, R. D., and Thompson, C. J. (2006). Does MPA mean "Major Problem for As
sessments?" Considering the consequences of place-based
management systems. Fish and Fisheries 7, 284-302.
Hearn, W. S., Sundland, R. L., and Hampton, J. (1987). Ro
bust estimation of the natural mortality rate in a com
pleted tagging experiment with variable fishing intensity.
Journal Du Conseil International Pour L'exploration De
La Mer 43, 107-117.
Hestbeck, J. B. (1995). Bias in transition-specific survival
and movement probabilities estimated using capture
recapture data. Journal of Applied Statistics 65, 429
433.
Hestbeck, J. B., Nichols, J. D., and Malecki, R. A. (1991). Es
timates of movement and site fidelity using mark-resight
data of wintering Canada geese. Ecology 72, 523-533.
Hilborn, R. (1990). Determination offish movement patterns
from tag-recoveries using maximum likelihood estima
tors. Canadian Journal of Fisheries and Aquatic Sciences
47, 635-643.
Hoem, J. M. (1971). Point estimation of forces of transition
in demographic models. Journal of the Royal Statistical
Society, Series B 33, 275-289.
ICCAT. (2007). Report of the 2006 Atlantic bluefin tuna stock assessment session. Collective Volume of Scientific Papers
ICCAT 60, 652-880.
Kalbfleisch, J. D. and Lawless, J. F. (1985). The analysis of
panel data under a Markov assumption. Journal of the
American Statisical Association 80, 863-871.
Kalbfleisch, J. D. and Prentice, R. L. (2002). The Statistical
Analysis of Failure Time Data. New York: Wiley.
Karlin, S. and Taylor, H. M. (1975). A First Course in Stochas
tic Processes. New York: Academic Press.
Laslett, G. M., Eveson, J. P., and Polacheck, T. (2002). A
flexible maximum likelihood approach for fitting growth curves to tag-recapture data. Canadian Journal of Fish
eries and Aquatic Sciences 59, 976-986.
Lebreton, J. D. and Pradel, R. (2002). Multistate recapture models: Modelling incomplete individual histories. Jour
nal of Applied Statistics 29, 353-369.
Nichols, J. D. and Kendall W. L. (1995). The use of multi
state capture-recapture models to address questions in
evolutionary ecology. Journal of Applied Statistics 22, 835-846.
Pollock, K. H. (1991). Modeling capture, recapture, and re
moval statistics for estimation of demographic parame
ters for fish and wildlife populations: Past, present, and
future. Journal of the American Statistical Association
86, 225-238.
Quinn, T. J. and Deriso, R. B. (1999). Quantitative Fish Dy namics. New York: Oxford University Press.
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions
1206 Biometrics, December 2008
R Development Core Team. (2006). R: A Language and Envi
ronment for Statistical Computing. Vienna: R Foundation
for Statistical Computing.
Ricker, W. E. (1975). Computation and the Interpretation of
Biological Statistics of Fish Populations. Bulletin of the Fisheries Research Board of Canada, Number 191, Ot
tawa, Canada.
Roberts, C. M., Bohnsack, J. A., Gell, F., Hawkins, J. P.,
and Goodridge, R. (2001). Effects of marine reserves on
adjacent fisheries. Science 294, 1920-1923.
Schwarz, C. J. and Seber, G. A. F. (1999). Estimating animal
abundance: Review III. Statistical Science 14, 427-456.
Schwarz, C. J., Schweigert, J. F., and Arnason, A. N. (1993).
Estimating migration rates using tag-recovery data. Bio
metrics 49, 177-193.
Seber, G. A. F. (1982). The Estimation of Animal Abundance and Related Parameters. London: C. Griffin & Co., Ltd.
Seber, G. A. F. and Schwarz, C. J. (2002). Capture-recapture:
Before and after EURING 2000. Journal of Applied Statistics 29, 5-18.
Sibert, J. R., Musyl, M. K., and Brill, R. W. (2003). Hori
zontal movement of bigeye tuna (Thunnus obesus) near
Hawaii determined by Kalman filter analysis of archival
tagging data. Fisheries Oceanography 12, 141-151.
Sibert, J. R., Lutcavage, M. E., Nielsen, A., Brill, R. W., and
Wilson, S. G. (2006). Interannual variation in large scale
movement of Atlantic bluefin tuna (Thunnus, thynnus) determined from pop-up satellite archival tags. Cana
dian Journal of Fisheries and Aquatic Sciences 63, 2154
2166.
Taylor, H. M. and Karlin, S. (1984). An Introduction to
Stochastic Modeling. Orlando, Florida: Academic Press.
Taylor, N. G., Walters, C. J., and Martell, S. J. D. (2005). A
new likelihood for simultaneously estimating von Berta
lanffy growth parameters, gear selectivity and natural
and fishing mortality. Canadian Journal of Fisheries and
Aquatic Sciences 62, 215-223.
Whitehead, H. (2001). Analysis of animal movement using opportunistic individual identifications: Application to
sperm whales. Ecology 82, 1417-1432.
Received April 2007. Revised November 2007.
Accepted November 2007.
This content downloaded from 185.2.32.121 on Wed, 25 Jun 2014 01:29:50 AMAll use subject to JSTOR Terms and Conditions