betaFeedback

WorksMathematics

Tag loss models for capture-recapture data without double tagging

Abstract

Standard capture-recapture (CR) models assume that marked individuals do not lose their tags during the sampling period. When this assumption is violated, the estimates of model parameters can be biased. Existing tag-loss models generally rely on double-tagging all or a subset of captured individuals, permanent marks, or auxiliary information on tag retention. However, double-tagging and permanent marking may be impractical, and reliable auxiliary information may be unavailable or difficult to obtain in many situations. In this paper, we develop a novel latent multinomial modelling framework for analysis of CR data with tag loss, without requiring double-tagging or permanent marking. The proposed model framework applies to both closed-population and open-population Cormack-Jolly-Seber (CJS) settings. Through simulation studies, we demonstrate that key model parameters can be reliably estimated in both settings. We illustrate the application of the method using a dataset from a long-term study on elephant seals. The results suggest that robust inference on tag loss and demographic parameters can be achieved without double tagging, thereby avoiding logistical challenges and broadening the framework's applicability to a wider range of studies.

Introduction

Capture-recapture (CR) data are often collected and analysed to estimate key demographic parameters of wildlife populations such as population size, survival, and recruitment. Estimation of these parameters requires individuals to be uniquely identifiable so that individual encounter histories can be constructed for CR models. Individual identification is typically achieved using: (i) physical, non-permanent tags, such as leg bands and neck collars for birds, ear tags for mammals, or PIT tags for reptiles and fish [22]; (ii) permanent marks such as brands or tattoos; or (iii) natural markings, i.e., unique physical features such as stripe patterns, scars, or pigmentation [23].

A critical assumption of most CR models [e.g. 4, 9, 21, 18] is that individuals do not lose their tags during the sampling period. However, in practice tag loss may occur for various reasons, including tag type and placement, individual animal characteristics [22], incorrect tag application, and tag malfunction [e.g. 5, 25]. If not accounted for, tag loss can bias estimates of population size in closed-population models [20] and affect parameter estimates and their standard errors in open-population models [e.g. 1, 13]. More specifically, the Jolly-Seber (JS) estimates of survival and birth, as well as the estimates of standard errors for birth, are biased when homogeneous tag loss occurs, i.e. when all marked animals have the same probability of tag loss between two sampling occasions [1], and the precision of all estimates is reduced. Under heterogeneous tag loss, the JS estimates of population sizes can be biased, particularly when capture probabilities are low [13].

Tag loss rates can be estimated through experiments in which animals are held in enclosures or tanks and tag retention is monitored directly over time (see e.g. Morgan and Walsh, 1993; Belser et al., 2017). Alternatively, tag loss can be investigated by double-tagging all or a subset of individuals in the population of interest. These studies account for tag loss either by applying ad-hoc corrections to parameter estimates post-analysis [e.g. 20, 1] or by incorporating tag loss directly into the model. For example, Hyun et al. [8] developed a closed-population model for a two-sample design in which tag loss is estimated by combining single- and double-tagged individuals, with double-tagged individuals carrying either two non-permanent tags or a non-permanent tag and a permanent one. Cowen and Schwarz [6] developed the Jolly-Seber tag-loss model (JSTL) for studies where a proportion of individuals are double-tagged with two non-permanent tags, allowing tag retention rates to vary by release group and time. For the Cormack-Jolly-Seber (CJS) model involving tag loss, Laake et al. [10] developed a hidden Markov model for tag loss that incorporates covariates associated with tag retention.

However, studies involving double-tagged individuals often make restrictive assumptions about the loss of non-permanent tags, such as assuming no complete tag loss occurs [e.g. 25] or assuming independent tag loss between the two tags (e.g. Cowen and Schwarz, [6]; Oosthuizen et al., [17]; Hyun et al., [8]; Malcolm-White et al., [12]. The assumption of independent tag loss may not hold in practice and is difficult to justify without empirical validation (e.g., using permanent marks to identify total tag loss). Studies validating this assumption found evidence of dependence between loss rates of the two tags. For example, in male black bears, the loss rate of a second ear tag depends on whether the first tag has been lost [7, 10]. A similar pattern was found in southern elephant seals, where individuals had, on average, a 28.1% greater likelihood of losing a second tag after losing the first one [15, 19]. Such dependence might arise from factors such as tag placement, differences in application between the two tags, or individual behaviour. When the independent tag loss assumption does not hold, survival estimates are negatively biased [19, 10]. Pairing a non-permanent tag with a permanent mark avoids these assumptions; however, permanent marks can be costly or difficult to apply, and often require specialised training [22].

In addition to requiring double tagging or permanent marks, existing tag loss models often assume that the effect of recycled individuals, i.e., those recaptured after losing both tags, is negligible. For example, the JSTL model [6] assumes that recycled individuals are either recognised upon recapture through permanent or natural marks and therefore not tagged again, or that such instances are very limited and have negligible effects on inference. However, Malcolm-White et al. [12] found that ignoring recycled individuals can lead to biased population size estimates, particularly when capture, survival, and tag loss probabilities are high, and this effect becomes more noticeable for long-term studies.

In this paper, we propose a novel method for modelling tag loss, which does not require individuals to be double-tagged or have permanent tags, avoiding the challenges and unrealistic assumptions associated with existing models. The new modelling framework is applicable to both closed-population and CJS problems. The proposed model falls within a general and broad class of latent multinomial models (LMMs, Link et al., [11]). The LMM framework is particularly well suited to situations where individual identification is latent and a single individual can contribute to multiple records in the observed data, but latent identifications cannot be determined from observed data (e.g. [3, 27]). This is exactly the case for tag loss problems. A single-tagged individual that loses its tag and is subsequently recaptured and re-tagged is essentially a recycled individual. It contributes two encounter histories in the observed data: one containing the partial history before tag loss, and the other one after tag loss. Within the LMM framework, a latent capture history combining the demographic, observational, and tag-loss processes is constructed for each individual, and the latent history is then deterministically linked to one or more observed histories. In this way, the likelihood of the observed data with a complex dependence structure can be specified. Model fitting can be achieved through the efficient saddlepoint approximation method of Zhang et al. [26].

Models and Methods

Assumptions and Notation

We develop tag loss models for two settings: i) closed-population studies under the classic model MtM_t [18], and ii) open-population studies under the CJS model. We denote the corresponding tag loss models as CPTL (closed-population tag loss) and CJSTL (CJS tag loss), respectively.

The new modelling framework explicitly accounts for tag loss while retaining assumptions similar to those of the corresponding models without tag loss. Similar to model MtM_t, the CPTL model assumes that the population is closed both geographically (i.e. no immigration or emigration) and demographically (i.e. no births or deaths). Capture probabilities vary across sampling occasions but are assumed equal for all individuals within each occasion. The CJSTL model follows the same population dynamics and observational process as the CJS model [4, 9, 21], with the likelihood constructed conditional on each individual’s initial capture and marking. The number of individuals first captured on each occasion is known under the standard CJS model. However, this number becomes uncertain for all occasions after the first, as individuals that have lost their tags may be recaptured and recorded as new individuals first marked on later occasions. As a result, the numbers of individuals first captured on all occasions except for the first must be estimated rather than treated as known in the CJSTL model. Under both models, we also assume that individuals may lose their tags at most once and that individuals are independent of each other.

Suppose that individuals are sampled from the population over TT capture occasions, indexed by t=1,…,Tt = 1, \ldots, T. We define the following model parameters for the CPTL model:

  • NN - population size, i.e. the total number of individuals in the population.

  • ptp_t - the probability that an individual is captured on occasion tt. We define the vector of capture probabilities across TT occasions as p=(p1,…,pT)T\boldsymbol{p} = (p_1, \ldots, p_T)^\mathsf{T}.

For the CJSTL model, we define the following parameters:

  • NtN_t - the number of individuals first captured and marked on occasion tt. Since tag loss can only occur after initial marking, N1N_1, the number of individuals first captured and marked on the first occasion, is known. Let N=(N2,…,NT−1)T\boldsymbol{N} = (N_2, \ldots, N_{T-1})^\mathsf{T} denote the vector of T−2T - 2 unknown parameters to estimate.

  • ptp_t - the probability that an individual is captured on occasion tt, given that it is in the population. We define the vector of capture probabilities across TT occasions as p=(p1,…,pT)T\boldsymbol{p} = (p_1, \ldots, p_T)^\mathsf{T}.

  • ϕt\phi_t - the probability that an individual survives from occasion tt to occasion t−1t-1. We define the vector of all T−1T - 1 survival probabilities as ϕ=(ϕ1,…,ϕT−1)T\boldsymbol{\phi} = (\phi_1, \ldots, \phi_{T-1})^\mathsf{T}.

For both the CPTL and CJSTL models, we define the tag loss parameter:

  • αt\alpha_t - the probability that a marked individual loses its tag between occasions tt and t+1t + 1, given that it retains its tag on occasion tt. We define the vector of all T−1T - 1 tag loss probabilities as α=(α1,…,αT−1)T\boldsymbol{\alpha} = (\alpha_1, \ldots, \alpha_{T-1})^\mathsf{T}.

Observed Data

We first consider the observed data for a standard CR study with no tag loss. Suppose there are nn individuals captured during the study. Let xi=(xi1,…,xiT)\boldsymbol{x}_i = (x_{i1}, \ldots, x_{iT}) denote the capture history of individual i=1,…,ni = 1, \ldots, n, where xit=1x_{it} = 1 if the individual is captured on occasion t=1,…,Tt = 1, \ldots, T and xit=0x_{it} = 0 otherwise. For example, xi=(1001)\boldsymbol{x}_i = (1001) represents an individual that was captured on the first occasion, not observed on the second and third occasions, and recaptured on the fourth occasion. Each xi\boldsymbol{x}_i belongs to the set of K=2T−1K = 2^T - 1 distinct observable capture histories, excluding the null history. The observed data y=(y1,…,yK)T\boldsymbol{y} = (y_1, \ldots, y_K)^\mathsf{T} consist of frequencies of all these capture histories, with yky_k, k=1,…,Kk = 1, \ldots, K denoting the number of times that the kk-th capture history is observed.

When tag loss occurs, the observed data retain the same format as in standard CR studies. However, the number of observed histories, nn, does not necessarily equal the number of individuals captured. This is because one individual may generate multiple observed histories. Specifically, a marked individual that loses its tag may be recaptured later and marked again as a new individual. Suppose an individual loses its tag between occasions ss and s+1s + 1 and is subsequently recaptured and marked as a new individual on occasion mm, where s+1≤m≤Ts + 1 \le m \le T. The individual then produces two separate observed histories:

  • xi(1)=(xi1,…,xis,0,…,0)\boldsymbol{x}_i^{(1)} = (x_{i1}, \ldots, x_{is}, 0, \ldots, 0), where xi1,…,xis∈{0,1}x_{i1}, \ldots, x_{is} \in\{0, 1\} for s<Ts < T; and

  • xi(2)=(0,…,0,1,xi,m+1,…,xiT)\boldsymbol{x}_i^{(2)} = (0, \ldots, 0, 1, x_{i,m+1}, \ldots, x_{iT}), where xi,m+1,…,xiT∈{0,1}x_{i,m+1}, \ldots, x_{iT} \in\{0, 1\} for m≥s+1m \ge s + 1.

For example, suppose we have a study with T=8T = 8 capture occasions in which some individual ii is captured on occasions 1 and 2, loses its tag between occasions 2 and 3, is later recaptured on occasion 6 as a new individual, and is then captured on occasions 7 and 8. This individual would contribute to the observed data with two distinct histories: xi(1)=(11000000)\boldsymbol{x}_i^{(1)} = (11000000) and xi(2)=(00000111)\boldsymbol{x}_i^{(2)} = (00000111).

The capture histories observed under tag loss have a complicated dependence structure, as multiple histories may be generated by the same individual, but their common origin cannot be identified from the observed data. Thus, it becomes challenging to obtain the likelihood for the data.

Latent States

To account for the uncertainty in individual identification arising from tag loss, we formulate the model in terms of an underlying latent process, that jointly describes capture, survival, and tag-loss dynamics. We introduce a set of latent states to describe an individual’s true but unobserved capture events over TT sampling occasions, as shown in Table 1. Note that not all possible combinations of the latent states above are valid latent histories. tories. For example, an individual that has entered state 5 cannot return to other states on later occasions. This will be described in more detail through the transition matrices introduced below.

EventTypeInterpretation
0Pre-tag lossNot released; CJSTL model only
1Pre-tag lossNot captured
2Pre-tag lossCaptured
3Post-tag lossNot captured
4Post-tag lossCaptured
5Pre- or post-tag lossNot survived; CJSTL model only

Table 1. Definitions of latent states under the CPTL and CJSTL models.

Let JJ denote the number of all valid latent histories, and let ωj=(ωj1,…,ωjT)\boldsymbol{\omega}_j = (\omega_{j1}, \ldots, \omega_{jT}) denote the jj-th latent history for j=1,…,Jj = 1, \ldots, J, where ωjt∈{0,1,2,3,4,5}\omega_{jt} \in\{0, 1, 2, 3, 4, 5\} under the CJSTL model and ωjt∈{1,2,3,4}\omega_{jt} \in\{1, 2, 3, 4\} under the CPTL model. For example, for a study with T=8T = 8 capture occasions, the latent history ωj=(12124455)\boldsymbol{\omega}_j = (12124455) represents an individual that was first captured on occasion 2, was not observed on occasion 3, was recaptured on occasion 4, lost its tag between occasions 4 and 5, was subsequently captured as a new individual on occasion 5, was observed again on occasion 6, and did not survive from occasion 6 to 7. This latent history produces two observed histories: (01010000) and (00000100), as only states 2 and 4 produce a capture under the tag loss model.

Transition Matrices

To formulate the tag-loss models, we derive the probability of each valid latent history in terms of model parameters. Let πj=P(ωj∣θ)\pi_j = P(\boldsymbol{\omega}_j \mid\boldsymbol{\theta}) denote the probability of latent history ωj\boldsymbol{\omega}_j, where θ=(p,α)\boldsymbol{\theta} = (p, \boldsymbol{\alpha}) under the CPTL model and θ=(p,α,ϕ)\boldsymbol{\theta} = (p, \boldsymbol{\alpha}, \boldsymbol{\phi}) under the CJSTL model. We use two transition matrices to define the probabilities of moving between the latent state from occasion tt and t+1t+1, ΓtCP\boldsymbol{\Gamma}^{CP}_t for the CPTL model and ΓtCJS\boldsymbol{\Gamma}^{CJS}_t for the CJSTL model:

ΓtCP=((1−αt)(1−pt+1)(1−αt)pt+1αt(1−pt+1)αtpt+1(1−αt)(1−pt+1)(1−αt)pt+1αt(1−pt+1)αtpt+1001−pt+1pt+1001−pt+1pt+1)(1)\boldsymbol{\Gamma}^{CP}_t = \begin{pmatrix} (1-\alpha_t)(1-p_{t+1}) & (1-\alpha_t)p_{t+1} & \alpha_t(1-p_{t+1}) & \alpha_t p_{t+1} \\ (1-\alpha_t)(1-p_{t+1}) & (1-\alpha_t)p_{t+1} & \alpha_t(1-p_{t+1}) & \alpha_t p_{t+1} \\ 0 & 0 & 1-p_{t+1} & p_{t+1} \\ 0 & 0 & 1-p_{t+1} & p_{t+1} \end{pmatrix} \tag*{(1)}
ΓtCJS=(ϕt(1−αt)(1−pt+1)ϕt(1−αt)pt+1ϕtαt(1−pt+1)ϕtαtpt+11−ϕtϕt(1−αt)(1−pt+1)ϕt(1−αt)pt+1ϕtαt(1−pt+1)ϕtαtpt+11−ϕt00ϕt(1−pt+1)ϕtpt+11−ϕt00ϕt(1−pt+1)ϕtpt+11−ϕt00001)(2)\boldsymbol{\Gamma}_{t}^{CJS} = \begin{pmatrix} \phi_t(1-\alpha_t)(1-p_{t+1}) & \phi_t(1-\alpha_t)p_{t+1} & \phi_t\alpha_t(1-p_{t+1}) & \phi_t\alpha_t p_{t+1} & 1-\phi_t \\ \phi_t(1-\alpha_t)(1-p_{t+1}) & \phi_t(1-\alpha_t)p_{t+1} & \phi_t\alpha_t(1-p_{t+1}) & \phi_t\alpha_t p_{t+1} & 1-\phi_t \\ 0 & 0 & \phi_t(1-p_{t+1}) & \phi_t p_{t+1} & 1-\phi_t \\ 0 & 0 & \phi_t(1-p_{t+1}) & \phi_t p_{t+1} & 1-\phi_t \\ 0 & 0 & 0 & 0 & 1 \end{pmatrix} \tag*{(2)}

Note that the transition matrices 1 and 2 apply to occasions fj≤t≤T−1f_j \le t \le T-1, where fjf_j denotes the occasion on which the individual with latent history ωj\omega_j is first captured and marked. The (r,s)(r,s) entry γrs(t)\gamma_{rs}^{(t)} of ΓtCP\boldsymbol{\Gamma}_t^{CP} and ΓtCJS\boldsymbol{\Gamma}_t^{CJS} gives the probability of transitioning from state rr on occasion tt to state ss on occasion t+1t+1, i.e. γrs(t)=P(ωj,t+1=s∣ωjt=r)\gamma_{rs}^{(t)}=P(\omega_{j,t+1}=s \mid\omega_{jt}=r), where r,s∈{1,2,3,4}r,s \in\{1,2,3,4\} for the CPTL model, r,s∈{1,2,3,4,5}r,s \in\{1,2,3,4,5\} for the CJSTL model.

The probability πj\pi_j of latent history ωj=(ωj1,…,ωjT)\omega_j=(\omega_{j1},\ldots,\omega_{jT}) under the CPTL model is given as

πj=∏t=1fj−1(1−pt)×pfj×∏t=fjT−1γωjt,ωj,t+1(t).(3)\pi_j=\prod_{t=1}^{f_j-1}(1-p_t)\times p_{f_j}\times\prod_{t=f_j}^{T-1}\gamma_{\omega_{jt},\omega_{j,t+1}}^{(t)}. \tag*{(3)}

where the term ∏t=1fj−1(1−pt)\prod_{t=1}^{f_j-1}(1-p_t) represents the probability of not being observed before occasion fjf_j. For example, the probability of latent history ωj=(11212434)\omega_j=(11212434) is

P(11212434)=[(1−p1)(1−p2)]p3[γ2,1(3)γ1,2(4)γ2,4(5)γ4,3(6)γ3,4(7)]=[(1−p1)(1−p2)]p3[(1−α3)(1−p4)(1−α4)p5α5p6(1−p6)p7].\begin{aligned} P(11212434)=[(1-p_1)(1-p_2)]p_3\left[\gamma_{2,1}^{(3)}\gamma_{1,2}^{(4)}\gamma_{2,4}^{(5)}\gamma_{4,3}^{(6)}\gamma_{3,4}^{(7)}\right] \\ &=[(1-p_1)(1-p_2)]p_3[(1-\alpha_3)(1-p_4)(1-\alpha_4)p_5\alpha_5p_6(1-p_6)p_7]. \end{aligned}

As the CJSTL model conditions on the first capture of each individual, we only consider capture events after occasion fjf_j. The probability of latent history ωj\omega_j under the CJSTL model then becomes

πj=∏t=fjT−1γωjt,ωj,t+1(t).\pi_j=\prod_{t=f_j}^{T-1}\gamma_{\omega_{jt},\omega_{j,t+1}}^{(t)}.

For example,

P(11212434)=γ2,1(3)γ1,2(4)γ2,4(5)γ4,3(6)γ3,4(7)=ϕ3(1−α3)(1−p4)ϕ4(1−α4)p5ϕ5α5p6ϕ6(1−p7)ϕ7p8.\begin{aligned} P(11212434)=\gamma_{2,1}^{(3)}\gamma_{1,2}^{(4)}\gamma_{2,4}^{(5)}\gamma_{4,3}^{(6)}\gamma_{3,4}^{(7)} \\ &=\phi_3(1-\alpha_3)(1-p_4)\phi_4(1-\alpha_4)p_5\phi_5\alpha_5p_6\phi_6(1-p_7)\phi_7p_8. \end{aligned}

The latent states defined in § 2.3 produce a total of 4T4^T possible latent histories under the CPTL model and 6T6^T under the CJSTL model. However, the number JJ of valid latent histories is smaller than these maximum numbers, as the zero probabilities in the transition matrices impose constraints on the order of states in a valid latent history. For example, once an individual enters state 3 or 4 it cannot return to state 1 or 2. This reflects the assumption that tag loss occurs at most once. Under the CJSTL model, if fj>1f_j>1 then ωjt=0\omega_{jt}=0 for 1≤t≤fj1\le t\le f_j, indicating that the individual has not been captured and marked. And once an individual enters state 5 it remains in that state afterwards.

Tag Loss Models

Let z=(z1,…,zJ)T\mathbf{z} = (z_1, \ldots, z_J)^{\mathsf{T}} denote the vector of latent history frequencies for the CPTL model, where zjz_j is the frequency of latent history j=1,…,Jj = 1, \ldots, J. Under the latent process, each latent history corresponds to exactly one individual so that ∑i=1Jzj=N\sum_{i=1}^{J} z_j = N. Assuming independence between individuals, we have

z∼Multinomial⁡(N,π),(4)\mathbf{z} \sim\operatorname{Multinomial}(N, \boldsymbol{\pi}), \tag*{(4)}

with π=(π1,…,πJ)T\boldsymbol{\pi} = (\pi_1, \ldots, \pi_J)^{\mathsf{T}}.

Under the CJSTL model, let zt=(zt1,…,ztJt)T\mathbf{z}_t = (z_{t1}, \ldots, z_{tJ_t})^{\mathsf{T}}, t=1,…,T−1t = 1, \ldots, T - 1, where JtJ_t denotes the total number of latent histories with the first capture and marking on occasion tt, and ztjtz_{tj_t} denotes the frequency of latent history jt=1,…,Jtj_t = 1, \ldots, J_t. It is straightforward that ∑t=1T−1Jt=J\sum_{t=1}^{T-1} J_t = J. Similar to the standard CJS model, the independence between individuals yields that the distribution of the NtN_t latent histories for individuals first marked on occasion tt is multinomial:

zt∼Multinomial⁡(Nt,πt),(5)\mathbf{z}_t \sim\operatorname{Multinomial}(N_t, \boldsymbol{\pi}_t), \tag*{(5)}

where πt=(πt1,…,πtJt)T\boldsymbol{\pi}_t = (\pi_{t1}, \ldots, \pi_{tJ_t})^{\mathsf{T}}. It follows that

z=(z1T,…,zT−1T)T(6)\mathbf{z} = (\mathbf{z}_1^{\mathsf{T}}, \ldots, \mathbf{z}_{T-1}^{\mathsf{T}})^{\mathsf{T}} \tag*{(6)}

follows a product multinomial distribution for the CJSTL model.

The vector z\mathbf{z} is unobservable for both models, so we cannot use (4) directly for inference. The observed data vector y\mathbf{y} is essentially a linear transformation of latent vector z\mathbf{z}. As discussed in § 2.3, each latent history can generate one or two observed histories, depending on if tag loss occurs. Since both observed and latent histories are known and fixed for a given TT, their mapping can be explicitly specified. More specifically, we have

y=Az,(7)\mathbf{y} = A\mathbf{z}, \tag*{(7)}

where A=(akj)A = (a_{kj}) is a known K×JK \times J matrix with akj=1a_{kj} = 1 if latent history jj results in observed history kk, and akj=0a_{kj} = 0 otherwise. The CPTL model falls with in the class of latent multinomial models (LMMs; Link et al. [11]). The CJSTL model is essentially a latent product-multinomial model, as considered in Zhang et al. [28].

Parameter Estimation

The likelihood of the models specified in (7) can be obtained by summing over all possible latent configurations for z\mathbf{z} consistent with the observed data y\mathbf{y}:

L(θ∣y)=∑z:y=AzPr⁡(z∣θ).(8)L(\theta\mid\mathbf{y}) = \sum_{\mathbf{z}: \mathbf{y}=A\mathbf{z}} \Pr(\mathbf{z} \mid\theta). \tag*{(8)}

However, it is in general computationally infeasible to evaluate this likelihood directly as it requires enumerating the set of all latent vectors consistent with the observed data, which is enormous in almost all relevant scenarios (Link et al. [11]). Instead, we apply the maximum likelihood approach of Zhang et al. [26] based on the saddlepoint approximation (SPA), which shows great efficiency for different types of LMMs and latent product-multinomial models [28, 27, 24]. The SPA method approximates the probability density or mass distribution of a random variable using its moment generating function (MGF). For the tag loss models considered in this paper, their MGFs can be obtained analytically; see Zhang et al. [26] for more details including the implementation. Maximum likelihood estimates are then obtained through numerical optimisation, with standard errors estimated from the inverse of the observed information matrix.

Simulation Study

We conducted a simulation study to evaluate the performance of the proposed tag-loss models. We explored various settings of parameter values and generated 1,000 datasets for each setting. We used N=500N = 500 and T=5T = 5 capture occasions for all simulations. In the closed-population setting, datasets were generated under a 323^{2} experimental design with capture probabilities and tag loss rates set to low (0.2), medium (0.5), or high (0.8) values. We fitted the CPTL model with time-varying capture probabilities and a constant tag loss rate. For convenience, we denote this model by ptαcp_t\alpha_c, where subscripts ‘t’ and ‘c’ denote time-varying and constant parameters, respectively. In the CJS setting, capture probabilities and tag loss rates were again set to low (0.2), medium (0.5), or high (0.8) values, while survival probabilities were set to either medium (0.5) or high (0.8). We considered three different CJSTL models: ϕtptαc\phi_t p_t\alpha_c, ϕtpcαc\phi_t p_c\alpha_c, and ϕcpcαc\phi_c p_c\alpha_c. Results for the CPTL model ptαcp_t\alpha_c and the CJSTL model ϕtpcαc\phi_t p_c\alpha_c with ϕ=0.8\phi= 0.8 are presented below, while the remaining results are provided in Section A of the Supplementary Material.

Across all parameter settings for the CPTL model ptαcp_t\alpha_c, the estimators of all capture probabilities in p\mathbf{p} are unbiased. The corresponding 95% confidence interval (CI) coverages fall within the 2.5-th and 97.5-th percentiles of the binomial distribution, based on true coverage of 95%. Mean 95% CI widths for capture probabilities in p\mathbf{p} range from 0.07 to 0.13. Similarly, the estimators of α\alpha are also unbiased and produce close-to-nominal CI coverages across almost all settings. One exception is that when the tag loss rate is high and the capture rate is low (i.e. α=0.8\alpha= 0.8 and p=0.2p = 0.2), the corresponding CI coverage rate (98%) is above the nominal value. Mean 95% CI widths for α\alpha range from 0.05 (α=0.2\alpha= 0.2, p=0.8p = 0.8) to 0.28 (α=0.5\alpha= 0.5, p=0.2p = 0.2).

Figure 1 presents the estimation results for population size NN under various settings of tag loss rates and capture probabilities for the CPTL model. The estimators of NN show negligible bias across all parameter settings. For a given tag loss rate, the precision of population size estimation increases (i.e. mean CI widths decrease) with increasing capture probabilities, which is not surprising as more data are obtained. When capture probabilities are medium or high (i.e. 0.5 or 0.8), the precision of estimation for NN also increases as tag loss rate increases. The precision does not show obvious variation across different tag loss rates when capture probabilities are relatively low (p=0.2p = 0.2). Across all scenarios, 95% CI coverage rates generally remain close to the nominal level. Overall, these results suggest that, for a closed-population CR study with tag loss, the CPTL model provides nearly unbiased parameter estimates with good precision across a range of scenarios even without double-tagging.

Figure 1. Boxplots of population size estimates from the CPTL model ptαcp_{t\alpha_c}, based on 1,000 simulated datasets under varying tag loss rates α=0.2,0.5,0.8\alpha= 0.2, 0.5, 0.8 and capture probabilities p=0.2,0.5,0.8p = 0.2, 0.5, 0.8, with T=5T = 5. The horizontal solid lines within each boxplot represent mean estimates and the red dashed line indicates the true population size N=500N = 500. Numbers above and below the boxes indicate the mean 95% CI widths and coverage rates, respectively.

We also assessed the performance of the CJSTL model with time-dependent survival probabilities, a constant capture probability, and a constant tag loss rate, i.e. the ϕtpcαc\phi_{t p_c \alpha_c} model. Survival probability estimators are unbiased in most scenarios. Minor negative bias was noticed in two settings: i) high tag loss and low capture probability (α=0.8\alpha= 0.8, p=0.2p = 0.2), where the relative bias was −6.45%-6.45\% for ϕ1\phi_1, −2.80%-2.80\% for ϕ2\phi_2, and 2.84%2.84\% for ϕ4\phi_4; and ii) medium tag loss and low capture probability (α=0.5\alpha= 0.5, p=0.2p = 0.2), where the relative bias was −1.64%-1.64\% for ϕ1\phi_1 and 1.53%1.53\% for ϕ4\phi_4. As shown in Figure 2, the 95% CIs exhibit some undercoverage for the survival probability under the two low-capture settings, and tend to be wide. At α=0.8\alpha= 0.8 and p=0.2p = 0.2, coverage for ϕ1\phi_1 drops by up to 10%, while the coverage rates for ϕ2\phi_2 and ϕ3\phi_3 also exhibit some undercoverage. Apart from these two cases, some very minor degree of undercoverage for ϕ1\phi_1 is observed in the setting of α=0.5\alpha= 0.5 and p=0.5p = 0.5. For the remaining parameter settings, coverage rates generally lie within the theoretical bounds, indicating that survival can be reliably estimated under the CJSTL model in most scenarios.

Figure 2. 95% CI coverage rates and mean 95% CI widths for survival probabilities from the CJSTL model ϕtpcαc\phi_{t p_c \alpha_c} fitted to 1000 datasets under varying tag loss rates (α=0.2,0.5,0.8\alpha= 0.2, 0.5, 0.8) and capture probabilities (p=0.2,0.5,0.8p = 0.2, 0.5, 0.8), with ϕ=0.8\phi= 0.8, Nt=500N_t = 500, and T=5T = 5. The solid horizontal line indicates the nominal coverage level of 95%, and the dashed lines represent the 2.5-th and 97.5-th percentiles of the binomial distribution for a nominal 95% coverage level.

We also checked the estimation results of other model parameters in the two settings discussed above, i.e. α=0.8\alpha= 0.8, p=0.2p = 0.2 and α=0.5\alpha= 0.5, p=0.2p = 0.2. Estimators of capture probabilities are generally unbiased and achieve nominal coverage across all settings, exexcept that for the setting with p=0.2,α=0.8p = 0.2, \alpha= 0.8, the 95% CI coverage rate (92%) falls slightly below the theoretical lower bound. The tag loss estimators show slight negative bias (−2.44% and −1.52%) and below-nominal coverage rates (87% and 92%) when p=0.2,α=0.8p = 0.2, \alpha= 0.8 and p=0.2,α=0.5p = 0.2, \alpha= 0.5, respectively, but are otherwise unbiased and achieve nominal coverage across the remaining scenarios. The population sizes NtN_t are generally estimated without bias and achieve the nominal coverage, with minor undercoverage (93%) in settings when α=0.5,p=0.2\alpha= 0.5, p = 0.2, and α=0.5,p=0.5\alpha= 0.5, p = 0.5.

The simulation results demonstrate that both the CPTL and the CJSTL model exhibit promising performance in estimating parameters of interest, such as survival probabilities and population size. The slight negative bias and undercoverage observed for ϕ1\phi_1 and ϕ2\phi_2 when α=0.8\alpha= 0.8 and p=0.2p = 0.2 may be attributed to the relatively low capture probability and high tag loss rate considered in this scenario. Under these conditions, individuals released during the first two occasions have a high probability of losing their tags before subsequent recapture occasions, reducing the amount of information available for estimating their survival probabilities. The resulting limited recapture information may lead to less stable parameter estimates and affect the accuracy of the associated CIs, potentially contributing to the observed bias and undercoverage. When tag loss rate is lower or capture probabilities become higher, more recaptures data are available to separate survival from tag loss, making estimation more stable and precise.

ParameterEstimate (95% CI)
ϕ1\phi_10.86 (0.80, 0.90)
ϕ2\phi_20.93 (0.87, 0.97)
ϕ3\phi_30.67 (0.64, 0.71)
ϕ4\phi_40.64 (0.60, 0.67)
ϕ5\phi_50.60 (0.57, 0.63)
ϕ6\phi_60.63 (0.60, 0.66)
ϕ7\phi_70.72 (0.69, 0.75)
pp0.67 (0.65, 0.68)
α\alpha0.32 (0.30, 0.34)

Table 2. Estimates of survival (ϕt\phi_t), capture probability (pp), and tag loss rate (α\alpha), along with the corresponding 95% CIs, obtained from fitting the CJSTL model ϕtpcαc\phi_t p_c \alpha_c to the elephant seal dataset.

Case Study

We apply the CJSTL model to the elephant seal dataset collected as part of a long-term study conducted between 1993 and 2000 on Macquarie Island, Australia [14]. In the study, approximately 6,000 individuals were double-tagged with uniquely numbered plastic flipper tags within 24 hours of birth and subsequently hot-iron branded on their flanks at weaning.

Because individuals were double-tagged and branded in the original dataset, we modified the data to apply the single-tag CJSTL model. We assumed individuals could not be recognised by their brands. Individuals captured and recognised based on brands were therefore treated as new individuals assigned with a new tag. The dataset did not contain information on which of the two non-permanent tags was lost, and thus we could not simply select a particular one for the analysis while ignoring the other. Individuals recaptured with only one non-permanent tag were therefore treated as new individuals that were newly tagged.

Table 2 presents estimation results of fitting the CJSTL model to the modified seals dataset. We fitted the CJSTL model ϕtpcαc\phi_t p_c \alpha_c, with time-varying survival probabilities, constant capture probability, and constant tag loss rate. The results indicate an estimated tag loss rate of 0.32 (95% CI=0.30, 0.34) in this study. Survival probability estimates vary from 0.60 (ϕ5\phi_5; between 1997 and 1998) to 0.93 (ϕ2\phi_2; between 1994 and 1995). While these results are not directly comparable to previous analyses of the seals dataset due to the modifications made to the data, they are broadly consistent with previous findings. McMahon and White [15] reported that, assuming dependence between tags, the probabilities of reducing from two tags to one or from one to zero in their double-tagging experiment were approximately between 0.20 and 0.40, depending on age of the individual. Malcolm-White et al. [12] reported a constant survival rate of 0.76 (95% CI=0.75, 0.77). The capture probability estimate from the CJSTL model (0.67; 95% CI=0.65, 0.68) is also consistent with that reported in Malcolm-White et al. [12] (0.68; 95% CI=0.67, 0.69).

Discussion

This paper presents a novel method of modelling tag loss in CR data under both closed-population and the open-population CJS settings. Unlike existing approaches, the proposed framework estimates tag loss without requiring double-tagging, permanent marks, or external tag-retention information, and thus avoids the associated practical difficulties or costs and some restrictive model assumptions. In addition, our approach naturally incorporates recycled individuals without using permanent marks. As the model is applicable to standard single-tag datasets, it can be used directly to analyse existing studies and does not require any design modifications in future studies.

Simulation studies demonstrated that model parameters can be estimated reliably across a wide range of scenarios under both the CPTL and CJSTL models. We observed reduced performance for the CJSTL model when it combines low capture probabilities, high survival probabilities, and moderate to high tag loss rates. In this scenario, lack of recaptures data makes it challenging to separate survival and tag loss parameters. This is not surprising, as biased inferences also exist in this scenario under the JSTL model even when all individuals are double-tagged [6]. This is why Malcolm-White et al. [12] emphasized that durable tags should be used whenever possible to minimise the probability of tag loss.

A limitation of the proposed approach is that it assumes that tag loss is homogeneous, i.e. all animals in the population have the same tag loss rate between two consecutive sampling occasions. Tag loss rates have been shown to depend on time since tagging and may differ between individuals depending on characteristics such as size, age or sex [1]. Future work could involve incorporating time since tagging as well as covariates thought to affect tag retention into the models by modelling the logit of the tag loss rate as a linear function of the covariates.

Finally, the proposed framework does not account for the possibility that individuals may lose their tags multiple times throughout the study. This may occur when tags have low retention rates, are attached poorly, or for biological reasons related to age, size, or behaviour of the animals. Multiple tag loss can be easily incorporated into the framework by modifying the transition matrices.

Acknowledgements

This work was funded by the Carnegie Trust for the Universities of Scotland. We are grateful to Dr. Jafet Belmont Osuna and Dr. Craig Wilkie for their helpful comments on earlier versions of this manuscript.

Supplementary Materials for “Tag loss models for capture-recapture data without double tagging” by Katarina Skopalova and Wei Zhang.

Web Appendix A: Additional simulation results for the CJSTL model

Model ϕcpcαc\phi_c p_c \alpha_c

Figure 3. 95% CI coverage rates and mean 95% CI widths for survival probabilities from the CJSTL model ϕcpcαc\phi_c p_c \alpha_c fitted to 1000 datasets under varying tag loss rates (α=0.2,0.5,0.8\alpha= 0.2, 0.5, 0.8) and capture probabilities (p=0.2,0.5,0.8p = 0.2, 0.5, 0.8), with ϕ=0.6\phi= 0.6, Nt=500N_t = 500, and T=5T = 5. The solid horizontal line indicates the nominal coverage level of 95%, and the dashed lines represent the 2.5-th and 97.5-th percentiles of the binomial distribution based on true coverage of 95%.

Figure 4. 95% CI coverage rates and mean 95% CI widths for survival probabilities from the CJSTL model ϕcpcαc\phi_c p_c \alpha_c fitted to 1000 datasets under varying tag loss rates (α=0.2,0.5,0.8\alpha= 0.2, 0.5, 0.8) and capture probabilities (p=0.2,0.5,0.8p = 0.2, 0.5, 0.8), with ϕ=0.8\phi= 0.8, Nt=500N_t = 500, and T=5T = 5. The solid horizontal line indicates the nominal coverage level of 95%, and the dashed lines represent the 2.5-th and 97.5-th percentiles of the binomial distribution based on true coverage of 95%.

Figure 5. 95% CI coverage rates and mean 95% CI widths for survival probabilities from the CJSTL model ϕtpcαc\phi_t p_c \alpha_c fitted to 1000 datasets under varying tag loss rates (α=0.2,0.5,0.8\alpha= 0.2, 0.5, 0.8) and capture probabilities (p=0.2,0.5,0.8p = 0.2, 0.5, 0.8), with ϕ=0.6\phi= 0.6, Nt=500N_t = 500, and T=5T = 5. The solid horizontal line indicates the nominal coverage level of 95%, and the dashed lines represent the 2.5-th and 97.5-th percentiles of the binomial distribution based on true coverage of 95%.

Figure 6. 95% CI coverage rates and mean 95% CI widths for survival probabilities from the CJSTL model ϕtptαc\phi_t p_t \alpha_c fitted to 1000 datasets under varying tag loss rates (α=0.2,0.5,0.8\alpha= 0.2, 0.5, 0.8) and capture probabilities (p=0.2,0.5,0.8p = 0.2, 0.5, 0.8), with ϕ=0.8\phi= 0.8, Nt=500N_t = 500, and T=5T = 5. The solid horizontal line indicates the nominal coverage level of 95%, and the dashed lines represent the 2.5-th and 97.5-th percentiles of the binomial distribution based on true coverage of 95%. Note that ϕ4\phi_4 could not be estimated separately due to confounding with p5p_5 and is therefore not shown.

Figure 7. 95% CI coverage rates and mean 95% CI widths for survival probabilities from the CJSTL model ϕtptαc\phi_t p_t \alpha_c fitted to 1000 datasets under varying tag loss rates (α=0.2,0.5,0.8\alpha= 0.2, 0.5, 0.8) and capture probabilities (p=0.2,0.5,0.8p = 0.2, 0.5, 0.8), with ϕ=0.8\phi= 0.8, Nt=500N_t = 500, and T=5T = 5. The solid horizontal line indicates the nominal coverage level of 95%, and the dashed lines represent the 2.5-th and 97.5-th percentiles of the binomial distribution based on true coverage of 95%. Note that ϕ4\phi_4 could not be estimated separately due to confounding with p5p_5 and is therefore not shown.

References

  1. [1]A. N. Arnason and K. H. Mills. 1981. Bias and loss of precision due to tag loss in Jolly–Seber estimates for mark–recapture experiments. Canadian Journal of Fisheries and Aquatic Sciences, 38(9):1077–1095.DOI
  2. [3]Simon J Bonner and Jason Holmberg. 2013. Mark-recapture with multiple, non-invasive marks. Biometrics, 69(3):766–775.
  3. [4]Richard M Cormack. 1964. Estimates of survival from the sighting of marked animals. Biometrika, 51(3/4):429–438.DOI
  4. [5]Laura Cowen and Carl J Schwarz. 2005. Capture–recapture studies using radio telemetry with premature radio-tag failure. Biometrics, 61(3):657–664.pubmed.ncbi.nlm.nih.gov/16135016
  5. [6]Laura Cowen and Carl J. Schwarz. 2006. The Jolly–Seber model with tag loss. Biometrics, 62(3):699–705.pubmed.ncbi.nlm.nih.gov/16984310
  6. [7]Duane R Diefenbach and Gary L Alt. 1998. Modeling and evaluation of ear tag loss in black bears. The Journal of Wildlife Management, 62(4):1292–1300.DOI
  7. [8]Saang-Yoon Hyun, Joel H. Reynolds, and Peter F. Galbreath. 2012. Accounting for tag loss and its uncertainty in a mark–recapture study with a mixture of single and double tags. Transactions of the American Fisheries Society, 141(1):11–25.DOI
  8. [9]George M. Jolly. 1965. Explicit estimates from capture-recapture data with both death and immigration-stochastic model. Biometrika, 52(1/2):225–247.pubmed.ncbi.nlm.nih.gov/14341276
  9. [10]Jeffrey L. Laake, Devin S. Johnson, Duane R. Diefenbach, and Mark A. Ternent. 2014. Hidden Markov model for dependent mark loss and survival estimation. Journal of Agricultural, Biological, and Environmental Statistics, 19(4):522–538.DOI
  10. [11]William A Link, Jun Yoshizaki, Larissa L Bailey, and Kenneth H Pollock. 2010. Uncovering a latent multinomial: Analysis of mark–recapture data with misidentification. Biometrics, 66(1):178–185.DOI
  11. [12]Emily Malcolm-White, Clive R. McMahon, and Laura L. E. Cowen. 2020. Complete tag loss in capture–recapture studies affects abundance estimates: An elephant seal case study. Ecology and Evolution, 10(5):2377–2384.
  12. [13]Trent L. McDonald, Steven C. Amstrup, and Bryan F. J. Manly. 2003. Tag loss can bias Jolly-Seber capture-recapture estimates. Wildlife Society Bulletin (1973-2006), 31(3):814–822.
  13. [14]Clive R McMahon, Harry R Burton, John Van Den Hoff, Rupert Woods, and Corey JA Bradshaw. 2006. Assessing hot-iron and cryo-branding for permanently marking southern elephant seals. Journal of Wildlife Management, 70(5):1484–1489.DOI
  14. [15]Clive R McMahon and Gary C White. 2009. Tag loss probabilities are not independent: Assessing and quantifying the assumption of independent tag transition probabilities from direct observations. Journal of Experimental Marine Biology and Ecology, 372(1–2):36–42.DOI
  15. [17]W. Chris Oosthuizen, P. J. Nico De Bruyn, Marthán N. Bester, and Marc Girondot. 2010. Cohort and tag-site-specific tag-loss rates in mark–recapture studies: A southern elephant seal cautionary case. Marine Mammal Science, 26(2):350–369.DOI
  16. [18]David L Otis, Kenneth P Burnham, Gary C White, and David R Anderson. 1978. Statistical inference from capture data on closed animal populations. Wildlife Monographs, (62):3–135.
  17. [19]L. K. Schwarz, M. A. Hindell, C. R. McMahon, and D. P. Costa. 2012. The implications of assuming independent tag loss in southern elephant seals. Ecosphere, 3(9):art81.DOI
  18. [20]G. A. F. Seber and R. Felton. 1981. Tag loss and the Petersen mark-recapture experiment. Biometrika, 68(1):211–219.DOI
  19. [21]George A. F. Seber. 1965. A note on the multiple-recapture census. Biometrika, 52(1/2):249–259.pubmed.ncbi.nlm.nih.gov/14341277
  20. [22]George A. F. Seber and Matthew R. Schofield. 2019. Capture-Recapture: Parameter Estimation for Open Animal Populations. Statistics for Biology and Health. Springer International Publishing, Cham.DOI
  21. [23]Nova J. Silvy, Roel R. Lopez, and Markus J. Peterson. 2020. Techniques of marking wildlife. In Nova J. Silvy, editor, The Wildlife Techniques Manual, 8 edition, volume 1. Johns Hopkins University Press, Baltimore, MD.
  22. [24]Katarina Skopalova, Jafet Osuna, and Wei Zhang. 2025. Estimating temporary emigration from capture-recapture data in the presence of latent identification. arXiv preprint arXiv:2510.06755.arxiv.org/abs/2510.06755
  23. [25]Giacomo Tavecchia, Jaume Adrover, Antoni Muñoz Navarro, and Roger Pradel. 2012. Modelling mortality causes in longitudinal data in the presence of tag loss: application to raptor poisoning and electrocution. Journal of Applied Ecology, 49(1):297–305.DOI
  24. [26]W Zhang, M V Bravington, and R M Fewster. 2019. Fast likelihood-based inference for latent count models using the saddlepoint approximation. Biometrics, 75(3):723–733.DOI
  25. [27]Wei Zhang, Simon J. Bonner, and Rachel S. McCrea. 2023. Latent multinomial models for extended batch-mark data. Biometrics, 79(3):2732–2742.DOI
  26. [28]Wei Zhang, Steven J Price, and Simon J Bonner. 2021. Maximum likelihood inference for the band-read error model for capture-recapture data with misidentification. Environmental and Ecological Statistics, 28(2):405–422.DOI

Paper details

vlnc:W_ev3n9a9m
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents