Anda di halaman 1dari 8

The Journal of Wildlife Management 79(8):1361–1368; 2015; DOI: 10.1002/jwmg.

943

Research Article

Occupancy Modeling of Bird Point Counts:


Implications of Mobile Animals
DANIEL B. HAYES,1 Department of Fisheries and Wildlife, Michigan State University, 480 Wilson Road, Michigan State University,
East Lansing, MI 48824-1222, USA
MICHAEL J. MONFILS, Michigan Natural Features Inventory, Michigan State University Extension, P.O. Box 13036, Lansing,
MI 48901-3036, USA

ABSTRACT Occupancy modeling has been applied to a wide variety of taxa and sampling methods,
including bird point counts. A critical assumption of basic occupancy models is that sites are occupied
throughout the duration of the study, which is unlikely to be true for typical bird point-count studies. As
such, we evaluated the implications of mobile animals on parameter estimates. We simulated the movement
and detection of individual birds using an individual-based simulation model. We fit the basic occupancy
model to data that represented a range of animal mobility, and determined the bias relative to known
parameters used in the simulation. Occupancy depends on the size of the site selected, with smaller sites
leading to lower occupancy for a given area and number of individuals present. At low animal density,
occupancy scales approximately linearly with the area of sites, but at very high density, occupancy asymptotes
at 1.0 across all site sizes. Even small amounts of movement lead to bias in estimates of occupancy and
detectability, and the typical size of bird home ranges can lead to highly biased parameters. Moreover,
variation in home range size over time or across habitats can lead to varying degrees of bias. Because of the
potential for large bias in occupancy estimates, and their sensitivity to behaviors of birds (e.g., home range
size), we recommend against applying current occupancy models to bird point-count data. Ó 2015 The
Wildlife Society.

KEY WORDS birds, closure assumption, detectability, home range size, individual-based models, movement
behavior, occupancy modeling, point count.

Occupancy modeling (e.g., MacKenzie et al. 2002,2006) has sites and surveys, or is a function of covariates; 4) detection of
been widely applied in studies of the demography of many species and detection histories at each location are
animal groups, including birds (e.g., Frey et al. 2011, Miller independent; and 5) a species is not falsely detected when
et al. 2011, Warren et al. 2013), mammals (Kalies et al. absent. Of these assumptions, the closure assumption seems
2012), fish (Falke et al. 2012), amphibians (MacKenzie et al. most likely to be violated for bird point-count samples and is
2002), and reptiles (Kery et al. 2009). This modeling is the focus of this investigation. Although methods have been
intended to address shortcomings of approaches that assume proposed to accommodate lack of closure (e.g., Farnsworth
complete detectability, an assumption that is rarely upheld in et al. 2002, Rota et al. 2009, Chandler et al. 2011, Amundson
natural populations. As MacKenzie et al. (2006) articulate, et al. 2014), these methods require additional data (such as
estimates of occupancy provide important insights into distance measurements, or identification of individual
various aspects of animal ecology, including habitat selection animals) or changes to sampling methods. Such an approach
and metapopulation dynamics. Because of the power and (sometimes termed a robust design; e.g., Kery et al. 2009,
flexibility of these models, their application has become Rota et al. 2009), although useful, still requires closure at
standard practice for many wildlife studies. some temporal or spatial scale.
Assumptions of the basic occupancy model (MacKenzie In our research applying these models to point counts of
et al. 2006) are 1) occupancy status at each site does not marsh birds (Monfils et al. 2014), we have become concerned
change over the survey season (closure); 2) the probability of that the assumptions of the basic occupancy model (i.e.,
occupancy is constant across sites, or is a function of MacKenzie et al. 2006) are often violated, particularly the
covariates; 3) the probability of detection is constant across assumption that the occupancy status of a site does not
change. Some authors have addressed this concern,
concluding that animal movement in and out of a site
Received: 27 February 2015; Accepted: 1 July 2015
Published: 13 August 2015
does not seriously bias results (e.g., MacKenzie et al. 2004,
Kendall and White 2009), whereas other authors have
1
E-mail: hayesdan@msu.edu concluded that such movement is an issue that requires

Hayes and Monfils  Occupancy Modeling of Mobile Animals 1361


changes in sampling design or analytical methods (e.g., Rota that distance from the site’s center point (e.g., Rota et al.
et al. 2009, Efford and Dawson 2012). We observe that the 2009, Monfils et al. 2014). For other taxa, such as small
basic occupancy model and assumptions continue to be mammals (e.g., deer mouse [Peromyscus maniculatus]) or
applied in many studies (e.g., Darrah and Krementz 2009, terrestrial amphibians (e.g., red-backed salamander [Pletho-
Bolenbaugh et al. 2011, Valente et al. 2011), possibly because don cinereus]), sites are often square regions where traps are
of the conflicting advice provided in theoretical papers. The set or the region searched (e.g., Otto and Roloff 2011). For
goal of this paper is to examine the closure assumption of larger mammals, camera traps are commonly used to detect
occupancy models and where violations of this assumption the presence of target species (e.g., Clare et al. 2015).
could lead to potentially erroneous results, with a focus on Finally, we note that detection probability can be
applications to bird point counts. Our main objective was to interpreted in several ways. We simulated the situation
determine the performance of the standard occupancy model where individual animals have a probability of being detected
and typical bird point-count data in light of the mobility of if present, and for simplicity, we treat this probability (p) as
birds as an example of mobile animals. being constant across all sites and sampling periods. In
We focused on the common sampling design where N sites contrast, the standard occupancy model is founded on a
are visited T times, assuming a constant probability of definition of detection probability where p is the probability
occupancy (c) and constant detection probability (p) across that a species will be detected at a site if present, regardless of
all sites and sampling visits. We note, however, that the the number of individuals present. Amundson et al. (2014)
terminology associated with occupancy modeling can at further distinguish the probability of detection of individual
times be confusing, and for clarity, we reiterate the definition animals within a site as being a function of availability (pa)
of terms commonly used in occupancy modeling. The first and perceptibility (pd). Availability (pa) is the probability that
critical term is occupancy itself. At the scale of an individual an animal is present at a site at the time of sampling, and
site, a site is occupied if it contains at least 1 or more perceptibility (pd) is the probability that an animal will be
individuals of the focal species across the survey season, detected if present. We note, however, that the standard
consistent with assumption 1 above. For a broader region, occupancy model, where individual animals are not distin-
occupancy is often defined as the proportion of sites guished and detection is recorded as a binary response for a
containing 1 or more individuals, or the probability that a site, does not distinguish these components.
site contains individual(s). If the closure assumption holds,
this definition also leads to the concept of proportion of area METHODS
occupied (PAO; MacKenzie and Royle 2005). For situations We constructed an individual-based model to simulate the
where mobile animals lead to a violation of the closure main biological processes involved in occupancy modeling.
assumption, Chandler et al. (2011) use the term temporary We created a 2,500-m  2,500-m region, containing 2,500
immigration or emigration to emphasize that movement sites, each 50 m  50 m in size. We used square sites in the
within an animal’s home range can lead to temporary simulation because circular sites do not provide complete
absences from an occupied site. MacKenzie and Royle (2005) coverage of a region (i.e., circular sites do not form a regular
suggest that the term use should replace occupy, leading to tessellation). Within this region, we randomly and indepen-
the concept of proportion of area used. When mobile animals dently distributed a total of M animals by generating a
are considered, Efford and Dawson (2012) further distin- random uniform number U(0, 2,500) for both an x and y
guish asymptotic occupancy, which they define as the position (Xc,Yc) that we treated as the centroid of the
accumulated area used over time, in contrast to instantaneous animal’s home range, which we then converted to a site
occupancy, which is the proportion of sites used at a specific location.
point in time. To ensure our estimation procedure replicated results from
Within this definition of occupancy, the concept of what previous studies, we included scenarios where animals were
constitutes a site is often flexibly defined across studies. Some immobile (i.e., home range is a point in space) as well as
applications of the occupancy model focus on meta- having larger home ranges. In initial simulations, animals
population dynamics, where a site is a patch large enough were not allowed to move from the centroid of their home
to contain a population, and where the boundaries of the site range, thereby meeting the assumptions of the basic
coincide with population boundaries. Clear examples of this occupancy model. In simulations where animals were mobile,
definition of a site are wetlands that contain breeding we treated the probability of detecting an individual animal if
subpopulations of amphibians (e.g., MacKenzie et al. 2002), present (p) as a constant across the entire study region and
or islands containing discrete populations. However, in many across all survey periods. When detection probability was less
other cases, the size of a site is controlled by the investigator, than 1.0 (i.e., incomplete detectability), the detection of
and is chosen for logistical concerns and not necessarily individual animals occurred at random, and we simulated
related to the biology of the species. Efford and Dawson detection by generating a U(0, 1) random variate, with
(2012) use the phrase “arbitrary plots in continuous habitat” detection occurring when this random variate was less than
to define this situation, which often occurs in studies of bird the detectability (p) for that simulation scenario. We
populations. For example, a common definition of site for aggregated observations of all animals occurring within a
bird point counts is a circle with a 50-m radius, where site within a survey period, and developed a detection history
detection depends on seeing or hearing an individual within for each site based on whether no (coded as 0), or 1 or more

1362 The Journal of Wildlife Management  79(8)


animals were detected (coded as 1) at the site. This Table 1. Relationship between home range multiplier (h), radius of a circle
encompassing 95% of simulated points, and area of the circle. P(central) is
formulation follows the conceptual model presented in Royle the proportion of detections that would take place within the site in which
and Nichols (2003). the circle’s centroid is located. We chose steps for h to represent 95% home
For simulations with mobile animals, we treated the range circles that are 10%, 50%, 100%, 2, 5, and 10 of the area of the
location generated above as the centroid or focal point of the 50  50-m site.
animal’s home range. For each sampling period, we h Radius (m) Area (m2) P (central)
simulated an animal’s position within its home range by 3.645 8.92 250 0.943
the following function: 8.152 19.95 1,250 0.874
11.527 28.21 2,500 0.824
ðXt ; Yt Þ ¼ ðXc þ Nð0; 1Þ  h; Yc þ Nð0; 1Þ  hÞ ð1Þ 16.302 39.90 5,000 0.758
25.775 63.08 12,500 0.631
36.451 89.21 25,000 0.503
where, (Xc, Yc) is the fixed centroid of the animal’s home
range, t is an index for survey period (e.g., whether it is the
first, second, or later visit to the site), and h is a multiplier
used to scale the size of an animal’s home range. This
implementation resulted in observations within the home provided us with a complete census of all possible sites,
range that were distributed as a bivariate normal variate allowing a direct evaluation of the bias and consistency of the
(Fig. 1). A consequence of this implementation is that home estimators (as defined by Cochran 1977) without the
ranges were circular in shape. We calculated the radius of a potential confounding effects of imprecision that would
circle enclosing 95% of generated points for 6 different occur with a subsample of sites. The topic of the precision of
scaling multipliers (h; Table 1). We then discretized the x,y estimates as a function of number of sites sampled, number of
positions of individual observations at time t to associate with visits, and detectability has been well addressed in the
a specific site, and summarized the data to produce a literature (e.g., Mackenzie et al. 2002), and as such was not a
detection history for each site (Fig. 2). focus of our simulations. Bias in occupancy and detection
Following data generation, we applied the basic occupancy probability was evaluated as a function of number of animals
model to estimate occupancy and detection probability across present, size of home range, number of repeat visits to each
the region simulated. We obtained estimates of occupancy site, and detectability. We evaluated 4 scenarios: 1) immobile
and detection probability by maximizing the likelihood animals with complete detectability; 2) immobile animals
function (Mackenzie et al. 2002, 2006): with incomplete detectability; 3) mobile animals with
complete detectability; and 4) mobile animals with incom-
Lðc;pÞ ¼ ðcn: Ppnt
t ð1  pt Þ
n:nt
Þ  ðcPð1  pt Þ þ ð1  cÞÞN n plete detectability.
ð2Þ

where t is the number of visits to a site, N is the number of


sites surveyed, and n. is the number of sites where at least 1
detection occurred. Because our focus was the bias of
estimators, we simulated the situation where data from all
sites in the region for all time periods were collected. This

Figure 2. Example of 10 detection points for an animal with a home range


multiplier (h) of 50 across an area broken into 50  50-m sites, with the 9
sites surrounding the centroid indicated by letters. The plus symbol denotes
the centroid of the animal’s home range, and the numbered points indicate
the location of the animal on each of 10 independent visits to the sample
area. Potential detections histories for the 9 sites surrounding the centroid
are: A (0000000000), B (0000000010), C (1000000000), D (0100000000),
Figure 1. Example of 500 detection points for an animal with a home range E (0000101100), F (0001000001), G (0000000000), H (0010010000), and I
multiplier (h) of 50 across a 50  50-m grid. (0000000000).

Hayes and Monfils  Occupancy Modeling of Mobile Animals 1363


RESULTS Comparison of results predicted by the binomial distribution
and the simulation produced equivalent results (Fig. 3).
Scenario 1: Immobile Animals With Complete Because the probability of success (p) in the binomial
Detectability depends on both the number of animals in a region, and the
As expected, occupancy within a region increased asymptot- number of sites that the region is broken into, these results
ically with increasing number of animals present (Fig. 3) have 2 important implications. Occupancy for randomly and
when animals do not exhibit temporary emigration or independently distributed individuals is inherently a density-
immigration and where they are detected with a probability dependent process, with a nonlinear relationship between
of 1.0. Clearly, if only 1 animal is present, then only 1 site can density and occupancy. Occupancy depends critically on the
be occupied. With N sites, this yields a probability of size of the site selected, with smaller sites leading to lower
instantaneous occupancy for any single site of 1/N. When occupancy for a given region and number of individuals
more than 1 animal is present, then each site may have 0, 1, 2, present. Further, the relationship between occupancy and
up to M animals present. The probability that a specific site size varies with the number of animals present. At low
number of animals is present at a given site, assuming animal density, occupancy scales approximately linearly with
animals are randomly and independently distributed can be the area of sites, with a doubling of site area halving
readily computed from the binomial distribution: occupancy. At very high density, occupancy asymptotes at 1.0
k!
in all cases (Fig. 4).
PðxÞ ¼ px qkx ð3Þ
x!ðkxÞ! Scenario 2: Immobile Animals With Incomplete
where k ¼ number of trials (i.e., number of animals present Detectability
¼ M); p ¼ 1/N; q ¼ 1  (1/N), and x ¼ number of animals Consistent with prior research, our results confirm that when
present at a site. Note that in the traditional formulation of the assumption of immobility is met, the methods outlined in
the binomial distribution used here, p refers to the MacKenzie et al. (2002, 2006) produce unbiased estimates of
probability of success, and not detection probability referred occupancy and detectability across a range of population sizes
to elsewhere in the paper. As an example, if 50 animals are and levels of detectability. Results of this scenario confirm
present within a region composed of 1,000 sites, the correct implementation of the maximum likelihood estima-
probability that a site will have exactly 1 animal is: tion procedure (Fig. 5).
Scenario 3: Mobile Animals With Complete
50!
Pðx ¼ 1Þ ¼ 0:0011 0:999ð501Þ ¼ 0:0476 ð4Þ Detectability
1!ð50  1Þ! Estimated occupancy was unbiased when home range
Using the binomial distribution as above, we can compute size ¼ 0 (as shown in scenario 2), but bias increased rapidly
the proportion of sites that are unoccupied as: as a function of the home range size (Fig. 6). As in the above
scenarios, occupancy was a function of the number of animals
k! present, and for a relatively small number of animals
Pð0Þ ¼ p0 qðk0Þ ¼ qk ð5Þ (M ¼ 100–300; occupancy for immobile animals
0!ðk  0Þ!
¼ 0.04–0.11), increases in occupancy with home range size
And the proportion of the sites that are occupied (i.e., x  1)
by its complement:

Pðx  0Þ ¼ 1  qk ð6Þ

Figure 3. Comparison of simulated occupancy (diamonds) and theoretical


occupancy predicted by the binomial distribution (solid line) as a function of Figure 4. Occupancy predicted by binomial distribution as a function of
animal abundance for a hypothetical region composed of 2,500 sites. animal abundance and size of site for a region of 2,500  2,500 m. A site size
Animals were immobile in this simulation and were assumed to be 100% of 50 m results in 2,500 total sites. The number of sites for a site size of 70,
detectable; we conducted 100 simulations for each level of abundance. 100, 141, and 204 m is approximately 1,250, 625, 312, and 150, respectively.

1364 The Journal of Wildlife Management  79(8)


Figure 5. Estimates of occupancy and site-level detectability for simulations
with no movement, 3 visits per site, and across a range of modeled
detectability (p) and number of animals present in the region. Points
represent mean of 1,000 simulations; confidence intervals are within symbol.

was approximately parallel (Fig. 6). Bias in occupancy


estimates occurred even for small amounts of movement; for
home ranges that were 10% of the area of the sampled site,
estimated occupancy was biased high by 22–24%. Home
range sizes that were equal to the site area resulted in
estimated occupancy 81–90% greater than that for immobile
animals.
Concurrent with increases in estimated occupancy,
estimated detectability decreased with increasing home
range size (Fig. 6). As expected, estimated detectability
was 100% when home range size was 0 across all values of M
Figure 6. Estimates of occupancy and site-level detectability for simulations
but declined precipitously with increasing home range size. including movement but with complete detectability (P ¼ 1.0). Simulations
Similar decreases in estimated detectability occurred across were conducted with 3 visits per site across a range of number of animals
all values of M (Fig. 6). As with estimated occupancy, present in the region (M). In the top panel, the true occupancy used in the
estimates of detectability were biased even for small amounts simulation was 0.039 for M ¼ 100, 0.077 for M ¼ 200, and 0.113 for
M ¼ 300. In the center panel, the true detectability was 1.0, and is
of movement; detectability was only about 0.82 for a home represented by a dashed line. Points represent mean of 1,000 simulations;
range size 10% of the site area. Estimated occupancy and confidence intervals are within symbol.
detectab6ility responded in opposite directions to movement,
and as such, were inversely related (Fig. 6). This relationship scenario 2) but was increasingly biased low as home range
was nonlinear, and depended on the number of animals size increased (Fig. 7), with convergent estimates as home
present (Fig. 6). range size reached 10 times the site size. Estimated
Scenario 4: Mobile Animals With Incomplete occupancy and estimated detectability were negatively
Detectability related in a nonlinear fashion (Fig. 7) that varied by the
Results of these simulations demonstrated that differing level of detectability used for data generation. The number of
levels of detectability had little effect on bias in occupancy as repeat visits to each site had less effect on estimates of
a function of home range size (Fig. 7), at least for a modest occupancy and detectability than home range size; however,
population size of 300 animals. Estimated detection estimated occupancy was consistently higher as the number
probability for a home range size of 0 was unbiased (as in of visits increased to 10 (Fig. 8).

Hayes and Monfils  Occupancy Modeling of Mobile Animals 1365


Figure 8. Estimates of occupancy and site-level detectability for simulations
with varying numbers of site visits (T). We conducted simulations with
M ¼ 300 animals, across a range of home range sizes and with detectability
(p) ¼ 0.5. The true value of occupancy (0.113) and detectability (0.5) are
represented by dashed lines in both panels. Points represent mean of 1,000
simulations; confidence intervals are within symbol.

or 250 m2 [0.03 ha]) of the plot size used in our simulations


(50  50 m), which is likely to be much smaller than the
home range of a typical songbird or marsh bird. For example,
the ovenbird (Seiurus aurocapillus) is a readily detectable
species with estimated breeding territory sizes ranging from
0.20 ha to 1.80 ha in the Midwest and Ontario (Hann 1937,
Figure 7. Estimates of occupancy and site-level detectability for simulations
with varying degrees of movement, expressed as home range size as a ratio to Stenger 1958, Wenny 1989), which is double the typical plot
the site area. Simulations were conducted with M ¼ 300 animals, and 3 visits size of 0.78 ha (50-m circular plot). Although the amount of
per site, across a range of detectability. In the top panel, the true occupancy bias depends on a number of factors, such as number of visits
used in the simulation was 0.113 and is represented by a dashed line. In the
and actual detectability, a home range size that is twice the
center panel, the true detectability (p) used is indicated in the legend. Points
represent mean of 1,000 simulations; confidence intervals are within symbol. area of the plot can lead to estimates of occupancy that
are more than twice the true level. For marsh birds, which
were the impetus for this work, the home range size is likely
DISCUSSION to be much larger. For example, Pickens and King (2013)
In bird point-count studies, the size of sites are generally report that the home range of king rail (Rallus elegans), as
chosen by the investigator, but do not typically correspond to determined by radio telemetry, was between 0.80 ha and
the boundaries of individual animal’s home ranges, 32.8 ha. Moreover, they found that although home range size
constituting what Efford and Dawson (2012) term arbitrary did not vary by sex or year, it was negatively correlated with
plots in continuous habitat. Researchers typically select site the amount of open water. As such, estimates of detection
sizes and sampling designs that address logistical and probability and occupancy could be confounded with habitat
funding constraints and ensure independence of sampling features through the effect on home range size instead of
units. Our results show that violations to the closure actual changes in the detectability of individual animals when
assumption caused by temporary emigration introduce located within a plot.
substantial bias to estimates of occupancy and detectability An additional issue that arises is that the home range size of
and that this bias is not reduced by increasing the number of birds typically varies seasonally and across habitats, potentially
site visits. Bias in estimated occupancy and detectability leading to systematic bias in estimates of detectability and
occurred for home ranges that are a small fraction (e.g., 10% occupancy across time or location. As such, the application of

1366 The Journal of Wildlife Management  79(8)


occupancy models to bird point-count data has the potential to the probability of having more than 1 animal present within a
lead to seriously biased results. The degree of bias will depend site is low, the 2 conceptualizations differ little in practical
on specific details of the biology of the species under terms. At higher densities, however, variation in the number
consideration and site size, and should be carefully considered of animals per site can result in heterogeneity in site-level
by the investigator. The examples above, however, suggest that detection probability without any variation in the detectabil-
50-m-radius plots that are commonly used are small relative to ity of individual animals (Royle and Nichols 2003).
the home range of many common bird species. In other Some authors have alluded to the issue of animal
applications, where the site boundaries coincide with natural movement by distinguishing between use and occupancy
boundaries that relate to the biology of the organism (e.g., (e.g., MacKenzie and Royle 2005, MacKenzie et al. 2006,
wetland anuran surveys), the assumptions of occupancy models Kendall and White 2009). Our reading and understanding of
are more likely to hold, leading to reasonable estimates of this distinction is unclear; MacKenzie and Royle (2005) state
detectability and occupancy. that “the proportion of area ‘used’ by a species will often be
Our results are in contrast to some previous authors but larger than the proportion of area where species physically
concur with others. MacKenzie et al. (2006) claim that “we occurs.” We fail to see the difference between physical
would thus expect the occupancy estimator to be unbiased if occurrence and usage. From the perspective of the data
species randomly moved in and out of a sampling unit”. entering occupancy models, the detection–no detection input
Likewise, Kendall and White (2009) suggest that mobile used by the models simply does not make a distinction
species can mitigate some problems of subsampling locations between animals that are occupying a site or animals that are
within a site. Consistent with Rota et al. (2009), we found using a site. As such, the basic occupancy model does not
that the implication of assuming immobility, when animals allow one to separate availability from detectability, as
are in fact mobile, is that estimates of occupancy are biased defined by Amundson et al. (2014). Further, the closure
high and estimates of detectability are biased low, with the assumption in the basic occupancy model effectively implies
degree of bias depending on the size of the animal’s range that availability is 100% throughout the survey period, and
relative to the size of the site used in the study. that all non-detections at sites where there is at least 1
We offer several reasons for the discrepancy between our detection are due to a lack of perceptibility.
simulation results and previous analyses or simulations Although not the initial focus of our investigation, we
concluding that violations to the closure assumption by found that the effects of site size are critical in our
random movement results in minimal bias. Conceptualizing understanding of occupancy, and relate closely to the
the problem in an individual-based framework highlights mobility of the organisms under study. We show that for
some of the differences; only 1 site can contain the core area a given number of animals (M) in a region, occupancy varies
or focal point of a bird’s home range (as seen in Fig. 2), and all widely for different site sizes. For example, a 50-m radius
other sites that are used have lower probabilities of use. plot has an area of 7,584 m2 and a 70-m radius plot has an
Because of this, all occupied sites do not have equal number area nearly twice as much at 15,394 m2, and the occupancy
of detections, even when detection probability is 100%. As for a given number of animals would also be nearly double for
such, simulations using an assumed level of occupancy across the 70-m radius plot. This highlights the sensitivity of
a region, and simply perturbing that level of occupancy occupancy estimates to even relatively small differences in
downward to account for random periods when a bird sampling design, or to misestimates of actual distance from
is absent misses a critical point, namely that mobile animals the plot center point. Issues of plot size also are important
create a situation where the occupancy status of adjacent considerations when a broader region is subsampled as
sites are not independent of one another. Similar to other Kendall and White (2009) discuss.
authors (e.g., Ramsey et al. 2015), we used a bivariate normal One design consideration to minimize the influence of
distribution to represent animal movement within their animal mobility would be to enlarge the size of the sampled
home range. Other statistical distributions are also plausible, sites, thereby reducing the ratio of the home range area to the
and would likely influence the degree to which estimates are size of sites (e.g., Rota et al. 2009). Three negative
biased. The critical element, however, is that any movement consequences of such a decision would be 1) occupancy
within the home range leads to some probability that animals would inherently increase as shown in our discussion of the
will not be present at a site when it assumed to be present, effects of site size; 2) larger sites are more difficult to
leading to bias in detection probability. completely assess (Kendall and White 2009), so detectability
Another way in which our simulation differs is how would likely decline for many species as costs of conducting
detection probability is conceptualized and represented. We the study would increase; and 3) larger sites are likely to
simulated the detection probability for individual animals encompass greater habitat heterogeneity, and thus, impor-
(i.e., the probability an individual will be detected if it is tant questions regarding the influence of site characteristics
present), as constant across space and time, similar to Royle on occupancy or use rates would likely be obscured.
and Nichols (2003), and aggregated detections of multiple Unfortunately, we do not have direct advice for addressing
individuals into a single measure. In contrast, the usual the issues we raise if the basic occupancy model and data
parameterization of occupancy models is that detection collection methods are used. Several authors have proposed
probability at the site level is fixed regardless of the number of methods for addressing violations to the closure assumption.
individuals present. When a population is at low density, and For example, Rota et al. (2009) recommend subdividing the

Hayes and Monfils  Occupancy Modeling of Mobile Animals 1367


sampling time to provide a robust design; this method Darrah, A. J., and D. G. Krementz. 2009. Distribution and habitat use of
assumes closure within a visit to a site, which is clearly king rails in the Illinois and Upper Mississippi River Valleys. Journal of
Wildlife Management 73:1380–1386.
violated when birds are seen entering or leaving the site Efford, M. G., and D. K. Dawson. 2012. Occupancy in continuous habitat.
during sampling. Other methods include measuring dis- Ecosphere 3:1–15.
tances to individual birds (Amundson et al. 2014), or using a Falke, J. A., L. L. Bailey, K. D. Fausch, and K. R. Bestgen. 2012.
removal sampling method where individual birds are counted Colonization and extinction in dynamic habitats: an occupancy approach
for a Great Plains stream fish assemblage. Ecology 93:858–867.
and density is estimated (e.g., Chandler et al. 2011). These Farnsworth, G. L., K. H. Pollock, J. D. Nichols, T. R. Simons, J. E. Hines,
methods hold promise for improving estimates of occupancy and J. R. Saur. 2002. A removal model for estimating detection
and detectability for mobile animals, but they require the probabilities from point-count surveys. Auk 119:414–425.
collection of auxiliary data, and as such, do not provide a Frey, S. J. K., A. M. Strong, and K. P. McFarland. 2011. The relative
contribution of local habitat and landscape context to metapopulation
means for addressing the lack of closure for commonly processes: a dynamic occupancy modeling approach. Ecography 35:581–589.
applied methods of conducting bird point counts where such Hann, H. W. 1937. Life history of the oven-bird in southern Michigan.
data are not collected. Although investigators formulating Wilson Bulletin 49:145–237.
new surveys should consider the use of these emerging Kalies, E. L., B. G. Dickson, C. L. Chamber, and D. W. Covington. 2012.
Community occupancy responses of small mammals to restoration
methods, analysts confronted with historical surveys without treatments in ponderosa pine forests, northern Arizona, USA. Ecological
such auxiliary data are left with little recourse. Although we Applications 22:204–217.
agree that multiple visits to individual sites are of value in Kendall, W. L., and G. C. White. 2009. A cautionary note on substituting
spatial subunits for repeated temporal sampling in studies of site
addressing imperfect detection, and estimating detectability
occupancy. Journal of Applied Ecology 46:1182–1188.
is of importance when interpreting study results, the Kery, M., R. M. Dorazio, L. Soldaat, A. van Strien, A. Zuiderwijk, A., and
incorporation of more biologically realistic assumptions J. A. Royle. 2009. Trend detection in populations with imperfect
into the models used is of greater importance. detection. Journal of Applied Ecology 46:1163–1172.
MacKenzie, D. I., L. L. Bailey, and J. D. Nichols. 2004. Investigating
MANAGEMENT IMPLICATIONS species co-occurrence patterns when species are detected imperfectly.
Journal of Animal Ecology 73:546–555.
We show that commonly used occupancy models address 1 MacKenzie, D. I., J. D. Nichols, G. B. Lachman, S. Droege, J. A. Royle, and
untenable assumption (i.e., complete detectability) by C. A. Langtimm. 2002. Estimating site occupancy rates when detection
probabilities are less than one. Ecology 83:2248–2255.
making another equally untenable assumption (i.e., immo- MacKenzie, D. I., J. D. Nichols, J. A. Royle, K. H. Pollock, L. L. Bailey, and J. E.
bility) when applied to bird point-count data, leading to Hines. 2006. Occupancy estimation and modeling: inferring patterns and
serious bias in estimated detection probability and occupan- dynamics of species occurrence. Academic Press, San Diego, California, USA.
cy. Moreover, the meaning and interpretation of occupancy MacKenzie, D. I., and J. A. Royle. 2005. Designing occupancy studies:
general advice and allocating survey effort. Journal of Applied Ecology
itself is unclear in situations where the home range of the 42:1105–1114.
study organism is larger than the size of the sample plot. As Miller, M. W., E. V. Pearlstine, R. M. Dorazio, and F. J. Mazzotti. 2011.
such, we recommend against the application of the basic Occupancy and abundance of wintering birds in a dynamic agricultural
occupancy model to bird point-count surveys where the only landscape. Journal of Wildlife Management 75:836–847.
Monfils, M. J., P. W. Brown, D. B. Hayes, G. J. Soulliere, and E. N. Kafcas.
data available are detection histories for individual sites. 2014. Breeding bird use and wetland characteristics of diked and undiked
coastal marshes in Michigan. Journal of Wildlife Management 78:79–92.
ACKNOWLEDGMENTS Otto, C. R., and G. J. Roloff. 2011. Using multiple methods to assess
detection probabilities of forest-floor wildlife. Journal of Wildlife
We thank G. Roloff for his helpful review of a draft of this Management 75:423–431.
manuscript. We thank the U.S. Fish and Wildlife Service, Pickens, B. A., and S. L. King. 2013. Microhabitat selection, demography
Michigan Natural Features Inventory, Michigan State and correlates of home range size for the king rail (Rallus elegans).
Waterbirds 36:319–329.
University Extension, Michigan State University Depart- Ramsey, D. S. L., P. A. Caley, and A. Robley. 2015. Estimating population
ment of Fisheries and Wildlife, and the Michigan density from presence-absence data using a spatially explicit model.
Department of Natural Resources for funding and time to Journal of Wildlife Management 79:491–499.
Rota, C. T., R. J. Fletcher, Jr., R. M. Dorazio, and M. G. Betts. 2009.
perform this research.
Occupancy estimation and the closure assumption. Journal of Applied
Ecology 46:1173–1181.
LITERATURE CITED Royle, J. A., and J. D. Nichols. 2003. Estimating abundance from repeated
presence-absence data or point counts. Ecology 84:777–790.
Amundson, C. L., J. A. Royle, and C. M. Handel. 2014. A hierarchical
Stenger, J. 1958. Food habits and available food of ovenbirds in relation to
model combining distance sampling and time removal to estimate
territory size. Auk 75:335–346.
detection probability during avian point counts. Auk 131:476–494.
Valente, J. J., S. L. King, and R. R. Wilson. 2011. Distribution and habitat
Bolenbaugh, J. R., D. G. Krementz, and S. E. Lehnen. 2011. Secretive
associations of breeding secretive marsh birds in Louisiana’s Mississippi
marsh bird species co-occurrences and habitat associations across the
Alluvial Valley. Wetlands 31:1–10.
Midwest, USA. Journal of Fish and Wildlife Management 2:49–60.
Warren, C. C., J. R. Ott, and J. A. Veech. 2013. Comparative occupancy and
Chandler, R. B., J. A. Royle, and D. I. King. 2011. Inference about density
habitat associations of black-and-white (Mniotilta varia) and golden-
and temporary emigration in unmarked populations. Ecology
cheeked warblers (Setophaga crysoparia) in the juniper-oak woodlands of
92:1429–1435.
Central Texas. American Midland Naturalist 169:382–397.
Clare, J. D. J., E. M. Anderson, and D. M. MacFarland. 2015. Predicting
Wenny, D. 1989. Population density and area requirements of three forest
bobcat abundance at a landscape scale and evaluating occupancy as a
interior warblers in central Missouri. Thesis, University of Missouri,
density index in central Wisconsin. Journal of Wildlife Management
Columbia, USA.
79:469–480.
Cochran, W. G. 1977. Sampling techniques, third edition. John Wiley and
Sons, New York, New York, USA. Associate Editor: Michael Morrison.

1368 The Journal of Wildlife Management  79(8)

Anda mungkin juga menyukai