
https://cpw.cvlcollections.org/files/original/edbc305e8910bb78576eb289d69c8ed2.pdf
411cde437a8cab39e9586f1b0dcd0a8e
PDF Text
Text
The research in this publication was partially or fully funded by Colorado Parks and Wildlife.
Dan Prenzlow, Director, Colorado Parks and Wildlife • Parks and Wildlife Commission: Marvin McDaniel, Chair • Carrie Besnette Hauser, ViceChair
Marie Haskett, Secretary • Taishya Adams • Betsy Blecha • Charles Garcia • Dallas May • Duke Phillips, IV • Luke B. Schafer • James Jay Tutchton • Eden Vardy
�CHAPTER 20
Density estimation...
Jake Ivan, Colorado Parks and Wildlife
Abundance is commonly sought after as a state variable for the study of populations. However,
density (number of animals per unit area) can be a more meaningful metric because it casts the state of
a population into a common currency. For example, using closed capture models from Chapter 14, we
may estimate 500 animals at site A and 200 animals at site B. Thus one conclusion we may reach is that
habitat management at site A has positively impacted the population there compared to site B. However,
if we know site A is 250 hectares and B is 100 hectares, then we realize that each has 2.0 animals/hectare.
That is, on a relative scale, the different management scheme at A had no effect compared to site B.
Conversely, we may estimate abundance at 2 sites to be similar and conclude management actions, or
habitat types, or harvest regulations, etc. are having a similar impact, but if the sites are different sizes,
then the impacts are actually quite different on a relative scale and our conclusion is erroneous. Thus,
while abundance can be a useful metric, estimating density can be helpful as well.
Chapter 14 covered in depth how one might use markrecapture techniques to obtain an estimate
of abundance from closed population samples. And in this age of remote imagery, GPS, and GIS, it is
relatively easy to delineate a study site and compute its area. Therefore, it should be easy to compute
the number of animals per unit area, right? Simply obtain an estimate of abundance using any one
(or several) of the multitude of models covered in Chapter 14, then divide that estimate by the area of
the study site. Certainly you can proceed in this straightforward manner, and sometimes people do.
However, your estimate will likely be biased. Why? Recall the assumptions of closed capture models.
Not only do we need the marks to be individually identifiable, readable, and permanent (at least over
the duration of the sampling), in order for our inference to be what we intend it to be, we need the
population we’re sampling to be closed both demographically and geographically.
Demographic closure can often be met by carefully considering the natural history of the species of
interest, then timing the sampling to correspond to periods when births and immigration/emigration
(e.g., dispersal) are unlikely. Sampling over a relatively short duration can give you some assurance that
deaths during the sampling period should be relatively unlikely as well. However, achieving geographic
closure is much more difficult. In fact, given that wildlife are mobile, we likely violate the geographic
closure assumption in most applications. If you happen to be sampling a terrestrial species on a small
island, you’re probably fine. Or, inversely, if you’re sampling fish in a pond, you probably meet the
assumption in that case as well. You may even be able to reasonably assume closure in other select
situations, such as sampling forest dwelling rodents in small woodland patches that are surrounded by
a matrix of agriculture fields.
However, in the vast majority of situations, we sample wildlife populations in relatively continuous
© Cooch & White (2019)
08.07.2019
�20  2
habitat such that some animal home ranges (simplistically depicted below as ovals) are completely
within the boundaries of our study site (i.e., the home range for animal 2 below falls completely within
the study site [rectangle]), but many others only partially overlap it (animals 1, 3, 5, 6, 7, 8, 9). If this is
the case, many individuals, especially those near the edges of the site, will move on and off of the area
of interest while markrecapture sampling is ongoing.
For example, during a traditional livetrapping study of small mammals in the fictitious situation
depicted above, animal 5 may be on the site and available for capture during the 1st day of trapping
(occasion 1), then off of the study area for day 2, back on for days 3 and 4, then off again on day 5. Assume
for a moment that capture probability is perfect (p � 1.0). That is, whenever this animal is on the study
area, it gets captured. Under this scenario, the encounter history for animal 5 would be ’10110’. A closed
capture model, however, assumes that if animal i was captured on the study area on any occasion, it was
available on the study area for all occasions (i.e., the study area is closed – once an animal is on it, it can’t
move off). Given this assumption, the ’10110’ capture history would indicate that p is closer to 0.6 (this
animal was available on 5 occasions and was detected on 3 of them). Thus, when we lack geographic
closure, closed capture models will underestimate capture probability. In fact, the estimate returned
from closed capture models is actually a product. It’s the product of the probability that animal i is
available for capture (a) and the probability that animal i is detected given it was available (p) such that
pestimated � (p × a). When we assume closure, we assume a � 1.0; when the assumption is violated
a < 1.0 and pestimated < p.
Now, recall from the beginning of Chapter 14 that the basic, heuristic estimator for abundance for a
single occasion is:
N̂ �
n
,
p̂
where n is the number of unique animals observed (the count statistic) and p̂ is the estimated probability
that an individual is detected. It stands to reason that if p̂ is underestimated when we lack closure, then
our estimate of N̂ will be inflated. Thus, when we fit a closed capture model to data from a design in
which the study site is not closed, we do not obtain an estimate of the number of animals on the study
site, which is what we want. Our estimate is larger than that. In fact, what we estimate is the number
of animals that could have used the study site during the course of sampling. We term this the super
Chapter 20. Density estimation...
�20  3
population (Schwarz & Arnason 1996, Kendall et al. 1997). Is this a problem? Maybe, maybe not. In
some cases you may elect to simply redefine what it is you’re estimating and report that number.
You can imagine, however, that if your goal is to estimate density, lack of geographic closure is
problematic. We have an estimate of abundance, but we do not know area to which that estimate applies.
Intuitively, the area used by the super population during the course of sampling was larger than the
study area itself, but how much larger? In other words, in the expression for density
D̂ �
N̂
,
A
what value do we use for A? The choice is not clear, yet the choice can have a big effect on your
estimate of density. Problem? Problem. What to do? Well, there are options. In fact, scientists have
been proposing solutions to this very problem since the inception of the field of wildlife ecology in the
1930s. These include using home range estimates to take a stab at the effective area sampled (Dice 1938),
exploiting differences in capture rates (or abundance estimates) between inner and outer detectors at
the site (e.g., Maclulich 1951, Hansson 1969, Otis et al. 1978), using assessment lines to determine the
reach of the initial sampling effort (e.g., Smith et al. 1971, Van Horne 1982), or computing the distances
moved by individuals between detection events (Otis et al. 1978, Wilson and Anderson 1985). However,
most of these approaches have fallen out of favor over time due to logistical issues, unrealistic data
requirements, or ad hoc rather than theoretical foundations. A notable exception is the method based on
mean maximum distance moved between trapping events (Otis et al. 1978, Wilson & Anderson 1985b),
which has received criticism as an ad hoc approach (Williams et al. 2002, Royle & Dorazio 2008), and
has shortcomings (Parmenter et al. 2003), but is still fairly popular (e.g., Converse et al. 2006, Zahratka
& Shenk 2008).
Spatially explicit capturerecapture (SECR; Efford 2004, Borchers & Efford 2008, Royle & Dorazio
2008, Efford et al. 2009, Royle et al. 2009) and trapping webs (Anderson et al. 1983, Link & Barker 1994)
are two contemporary density estimation approaches that circumvent the difficulties of estimating the
area used by the super population by estimating density directly. These approaches generally have a
stronger theoretical background behind them than the approaches mentioned above, but like any model,
they have their own sets of assumptions and thus advantages and disadvantages. See (Parmenter et al.
2003) and (Ivan et al. 2013) for discussions of pros and cons, advantages, and limitations.
Here we focus on the solution provided to you in Program MARK. It is called the “Density with
Telemetry” data type. From the name you can see right away that this approach requires auxiliary
information (telemetry locations). Thus, as with the Barker model and Burnham’s joint model, we are
combining markrecapture information with outside information to help solve a problem. The basic
approach is exactly opposite of that taken during earlier work described above. That is, rather than
estimating N, then trying to figure out the area to which the estimate applies, we first define the study
site of interest, and thus fix its area. We then attempt to estimate the total number of whole and partial
animals within its boundaries. In the case of our simple example above, this amounts to using telemetry
to estimate the shaded areas for each of the 9 animals that were part of the markrecapture data set. We
then sum these proportions (i.e., for animals 1 → 9, we might estimate there are a total of [0.47 + 1.00 +
0.08 + 0.00 + 0.37 + 0.93 + 0.09 + 0.82 + 0.07] � 3.83 animals on the study site) and divide by the area of
the study site to obtain an unbiased estimate of density.
The basic setup in the field is as follows. Assume that our sampling scenario is qualitatively similar
to the simplistic example above. That is, despite our effort to nicely demarcate the study site of interest
in a GIS, we know wildlife will go about their daily business largely ignoring our boundaries. Also
assume we sample animals at the study site during a time of year in which we are not concerned about
dispersal or births and we sample over a short enough time span that we’re not so much concerned with
deaths either (i.e., the population is demographically closed). During a sampling session, we capture,
Chapter 20. Density estimation...
�20  4
mark, and release individuals on multiple occasions, let’s say 5. Let’s also suppose we fit these captured
animals with a telemetry device (e.g., radio tag). We start collecting locations on telemetered individuals
after we’re done with the markrecapture sampling and we sample locations from these animals over
a relatively short period of time, say 2 weeks posttrapping. For each independent location, we note
whether it is within the study site, or outside the study site.
We know we can apply traditional markrecapture estimators to the markrecapture data collected
during the sampling session to estimate the super population (Schwarz & Arnason 1996, Kendall et al.
1997) of animals that used the site during sampling. Our goal is to use the telemetry data to estimate the
portion of the super population that occurred within the boundaries of the site to produce an estimate
of density corrected for the lack of geographic closure. Mathematically, we begin with the Huggins
(1989;1991) closed capture estimator for abundance:
N̂s p �
M
t+1
Õ
i�1
1
,
p̂ ∗i
where N̂s p is the abundance estimate that represents the super population of animals that could have
used the site during the trapping session, p̂ ∗i is the estimated probability that animal i is captured one
or more times during the sampling session (i.e., if p̂ i is an estimate of the probability animal that i is
detected on any given occasion, then p̂ ∗i � 1 − (1 − p̂ i )t , where t is the number of occasions), and M t+1 is
the total number of animals detected. This should look very familiar if you’ve read Chapter 14, with the
exception that we’re explicitly noting that in most cases, N̂ will be an estimate of the super population
rather than an estimate of the number of animals within the study site.
Next let’s make a substitution in the numerator of the Huggins estimator:
N̂s s �
�
M
t+1
Õ
i�1
�
p̃ˆ i
p̂ ∗i
,
where p̃ˆ is the estimated proportion of time (i.e., proportion of telemetry locations) animal i spent on the
study site. Those individuals that are always located on the study site contribute fully to the estimate
and are assigned p̃ˆ i � 1. That is, they are treated no differently than they are in the normal Huggins
implementation. Most other individuals receive a fractional p̃ˆ i because they spend some fraction of
their time on the site. Depending on the circumstances, some (many?) animals may be like animal 4
above. That is, they may be attracted to our detectors and become part of the markrecapture data set,
especially if we use bait or lures to improve our detection probability, but in reality their normal home
range doesn’t occur within the study site. These animals will be assigned p̃ˆ i � 0 and do not contribute
to the estimate.
So, functionally what we have here is an estimator that adjusts our count of animals detected upward
(i.e., divide by p̂ ∗i ) to accommodate imperfect detection, then ratchets that correction back down (i.e.,
multiply by p̃ˆ i ) to include only those animals within the study area. N̂s s , then, is the estimated number
of animals within the study site, which is what we’ve always wanted our abundance estimate to be.
At this point all we need to do is divide N̂s s by the area of the study site (e.g., area of the minimum
convex polygon encompassing all detectors. It is this polygon that is used to determine whether animals
are ‘in’ or ‘out’, which informs derivation of p̃) in order to obtain an estimate of density:
D̂ �
Chapter 20. Density estimation...
�
M
t+1
Õ
i�1
A
�
p̃ˆ i
p̂ ∗i
,
�20.1. Likelihood
20  5
where D̂ is estimated density (number of animals per unit area), A is the area of the study site, and p̃ˆ i , p̂ ∗i ,
and M t+1 are as defined previously. MARK takes care of the variance calculation for you. Note that you
do not necessarily have to put a telemetry device on each of the animals you sample via markrecapture.
If it is not possible to telemeter all individuals, we can use logistic regression to build a relationship
between p̃ˆ i and a suite of covariates for animals that do have a telemetry device. We can then use this
logistic model to estimate p̃ˆ i for those individuals that were not telemetered. See Ivan et al. (2013) for a
summary of estimator performance for different levels of telemetry effort.
20.1. Likelihood
The likelihood for this estimator is simply a combination of likelihoods you have seen before. Chapter
14 has extensive detail regarding the structure of the likelihood for various forms of the Huggins
estimator (i.e., the likelihood with respect to parameters p, c, and potentially π). Chapter 16 describes
the likelihood for known fate survival models, which is a simple binomial – the same model used for
logistic regression. In Chapter 16, the parameter we were trying to estimate was survival (S), success was
defined as surviving an interval, failure was dying, and N, the number of trials (not to be confused with
abundance, N̂), was the number of animals alive to start each interval. In the case of density estimation,
p̃ replaces S as the parameter of interest, a binomial “success” is locating an animal on the study site
(i.e., the polygon), a “failure” is locating it off of the study site, and the number of trials, N, is the number
of total locations collected (on an individual basis – the N’s need not be the same for each individual).
The two likelihoods can be written out separately, then multiplied together to form one big likelihood,
which is then maximized.
20.2. Implementation in MARK
With this conceptual and mathematical background in mind, let’s look at the mechanics of how to
implement the model in MARK. First, the input file. As you may have guessed, it looks generally like
a closed capture input file with a slight modification to accommodate the telemetry data. A piece of an
example input file appears below.
As per the usual closed capture format, each line begins with an optional comment (here an animal
ID #), followed by the encounter history (in this case, 5 occasions). Next are 2 new columns where we
input the telemetry data. The first of these new columns is the number of times animal i was located on
the study site; the second column is the total number of telemetry locations obtained on animal i. So,
in the example, animal 20 was located on the study site 100% of the time (10 locations on site out of 10
total locations collected), animal 24 was located on the site 50% of the time (5/10), and animal 26 was
never located within the bounds of the study site after having been captured and tagged there (0/10).
Note also that animals 27 and 28 were part of the markrecapture sample, but they were not fitted with
telemetry devices. For those animals, we enter “.” for each of the 2 telemetry columns as these animals
provide no information for that part of the model. The remaining columns are the same as they would
be for the usual closed capture input file. In this case, after the columns for the telemetry data, we have
2 groups (animals 2024 belong to group 1, animals 2528 belong to group 2), then a single covariate,
which we will discuss later.
/*20*/
/*21*/
/*22*/
/*23*/
/*24*/
10010
10000
00100
00001
00010
10
1
8
10
5
Chapter 20. Density estimation...
10
10
10
10
10
1
1
1
1
1
0
0
0
0
0
20;
0;
20;
20;
0;
�20.2. Implementation in MARK
/*25*/
/*26*/
/*27*/
/*28*/
00010
00010
00110
00001
20  6
5
0
.
.
10
10
.
.
0
0
0
0
1
1
1
1
20;
0;
25;
0;
To load the input file, open MARK, choose “File  New” and select the “Density with Telemetry”
data type near the bottom of the screen. A window appears in which you will have to select one of
3 possible parameterizations for the capture probability parameter(s). Note that these options are the
same as those available to you under the traditional Huggins closed capture data type (Chapter 14). For
the sake of this example, let’s keep it relatively simple and choose “Huggins p and c".
Once we’ve selected the parameterization we’d like to use to model capture probability, we need
select an input file and fill in the rest of the Specifications Window as per usual. Let’s title this data
“Example", and choose the “Density Estimation With Telemetry.inp” input file.
Chapter 20. Density estimation...
�20.2. Implementation in MARK
20  7
This is simulated data intended to mimic a study of small mammals, such as deer mice, sampled
in 2 sites (habitat types), A and B. Each habitat type was sampled with a (10 × 100) livetrapping grid
(10m trap spacing). There are 5 occasions. In addition to marking each mouse with an individually
identifiable ear tag, 50% of the individuals captured were fitted with a small VHF transmitter. These
radiotagged individuals were located once during the day and once at night for 5 days immediately
after markrecapture sampling (n = 10 locations total per animal) and each location was recorded as “in”
or “out” of the study site.
The single covariate we’ve recorded is the distance to the edge (DTE) of the of the site from the mean
trap location of each individual (i.e., compute the mean trap location for each individual captured ≥ 1
time, then compute the minimum distance from this mean location to the edge of the site). This is a
crude metric for characterizing whether an individual’s home range is near the edge or in the interior
of the site, which can be helpful in modeling both the p and p̃ parameters as we’ll see below.
Given this information, fill in the remainder of the “Specifications Window” as follows:
Label the 2 groups “Site A” and “Site B"; label the Individual Covariate “DTE". Note that with this
data type you are required to enter Group Labels. You cannot accept the defaults as you can with other
data types. Also, note that there is a new “Enter Areas” button at the bottom of this window. You must
click this button to specify the area of the study site for each group before you can proceed. The area
of the site is critical to the final computation, and it defaults to 1. It is also allowed to vary with each
group. Forcing you to double check the labels and corresponding areas before you proceed is MARK’s
way of imparting some quality control on your project!
You should enter the area for each group in whatever units you would like the density estimate
expressed. In our example (see top of the next page), each site (group) was sampled with a (10 × 10)
grid with 10m trap spacing and we want the answer to be in animals per hectare so we enter 0.81 for Sites
A and B (i.e., let’s define the study site as the minimum convex polygon around the traps so the site is
(90m × 90m). See  sidebar , below, for more discussion.). If we wanted the answer to be animals/m2 ,
we would enter 8,100 for each group; for animals per kilometer, we would enter 0.0081 for each group,
etc. If Site B was smaller, say 0.75 ha, we would enter 0.81 for the area of Site A and 0.75 for the area of
Site B.
Chapter 20. Density estimation...
�20.2. Implementation in MARK
20  8
begin sidebar
what is a ‘site’?
Note that the definition of the “site” is somewhat arbitrary. However, any workable definition should
ensure that animals within the site have a reasonable chance of detection. We suggest that a “reasonable
chance of detection” can be attained if the site is defined such that detectors are distributed at an
adequate and roughly consistent rate (i.e., 4 detectors per home range; Otis et al. 1978, p. 76) within it
so that there are no gaps in sampling effort.
We prefer to define the study site as the minimum convex polygon (MCP) encompassing the
detectors. Such a definition seems natural and trap density inside this polygon would meet our criteria
of being relatively high and consistent throughout. Alternatively, one could define the study site as
the MCP plus ½ trapwidth, or a full trapwidth, and still claim that trap density is relatively high and
consistent within the site.
However if we defined the site as the MCP plus 23 trapwidths or if detectors were distributed at a
rate of 2 per home range over part of the site, but 6 per home range over other parts, those definitions
would likely result in unequal sampling effort across the site and may lead to bias in the estimator. Note
that because we incorporate the proportion of time each individual spends on the site (i.e., telemetry
data), the estimator will be appropriate regardless of how the site is defined, as long as it is defined
following the guidance we provide here.
end sidebar
Once you’ve entered the area for each group, MARK will allow you to click ‘OK’ and proceed. In the
main MARK interface, open the PIM chart (shown at the top of the next page) to view the parameters
in the model. You will see the familiar p’s (5 for each group) and c’s (4 for each group) from the Huggins
portion of the model along with the new parameter we’re going to model using the telemetry data, p̃
(1 for each group).
Chapter 20. Density estimation...
�20.2. Implementation in MARK
20  9
Because the Huggins model is embedded within this model for density, all of the Huggins options
from Chapter 14 are available to you for modeling p and c. That is, you can specify that capture
probability is constant across animals and occasions (M0 ), or that it varies by occasion (M t ), or that it
differs depending on whether an animal has been captured previously (M b ), or that it differs generally
among individuals (M h ), or that it varies in relation to one or more individual covariates. Of course, you
can specify any combination of these basic model types as well. All of the nuances covered in Chapter 14
regarding model construction for closed abundance estimators (e.g., constraining the last p) apply here
as well.
You may initially be tempted to structure similar models for p̃, but think before you start filling in
your design matrix or PIM chart. For p̃ (proportion of time, or proportion of locations on the study site),
there is no notion of a behavioral effect, nor time effects (we generally sample through time to estimate
p̃, but p̃ is not connected to the markrecapture sampling occasions in any way). Heterogeneity in p̃ is
possible, and even likely, given that some animals will be located on the edge of the study site whereas
others have home ranges near the interior. However, we cannot model general heterogeneity using a
mixture model (π) like we do with capture probability. Instead, we’re limited to the use of covariates,
such as DTE. Thus, outside of covariates or groups, the structure we can apply to p̃ is limited.
Leave the PIMs in this general form, close the PIM chart, and open a ‘reduced’ design matrix with
2 columns. Let’s begin by building the simplest model possible: {p · p̃ · } (which we write in ASCII
as p(.)p∼(.)). For this first model, we’re specifying a constant detection probability (M0 ) and we’re
assigning each animal the average proportion of time on site across all animals and groups – akin
to the basic form of this estimator first introduced by White and Shenk (2001). This may not be the
most biologically realistic model we can think of, but it’s a basis for comparison and may be the most
supportable structure available if you have a sparse data set. The design matrix for this model is shown
at the top of the next page:
Chapter 20. Density estimation...
�20.2. Implementation in MARK
20  10
Run this model then examine the real parameters. We see that capture probability was estimated to
be 0.23 and p̃ was 0.66. So, on average, each animal was located on their respective study site about 2/3
of the time according to this model.
The density estimate for the 2 sites is a derived parameter. If we examine the derived estimates (below)
we notice that there appears to be some difference in density between the 2 study sites (21.63 animal/ha
Chapter 20. Density estimation...
�20.2. Implementation in MARK
20  11
in Site A, 14.80 animals/ha in Site B). Note, however, that because density is a derived parameter, we
cannot specifically test for differences in density between sites by specifying a model structure that
allows for differences and comparing it to one that does not. We can specify differences in p or p̃, but not
D̂. Instead we’re left to make inference by other means. That is, we have to examine the point estimates,
SEs, and confidence intervals, then make a judgment. In this case, the point estimates appear to be fairly
different, but there is substantial overlap in the 95% confidence intervals, so maybe we conclude there
is some evidence that density at Site A is higher than Site B, but the evidence is weak.
We might hypothesize differences in p and/or p̃ between the sites if the habitat imparts differential
home range sizes and/or differential availability of food resources that impact detection. Let’s go ahead
and build 2 models that reflect these hypotheses, that p or p̃ may differ by site. Recall from above, that
Site A (group 1) is represented by parameters 15 (p), 1114 (c), and 19 (p̃); Site B (group 2) is represented
by parameters 610 (p), 1518 (c), and 20 (p̃). Thus, a potential model that reflects the hypothesis that
p is different between sites might look like the design matrix on the left, whereas the one reflecting
differences in p̃ might look like the one on the right. Of course, the 2 structures could be combined as
well.
In this case, the model that allows p̃ to vary between sites has more support than the one that allows
p to vary between sites. If we examine the β estimates for the best fitting model thus far, we see that
with Site A as the reference point (intercept), Site B tended to have lower p̃. That is, animals on that site
tended to spend less time within the study boundaries compared to animals at the site represented by
the intercept (Site A).
Finally, let’s run a model that makes use of our DTE covariate. We might expect that if an animal is
trapped, on average, near the center of the study site, it may have a higher probability of detection than
animals trapped near the edge because there are more traps in its home range compared to the edge
animal. Similarly, we might also expect such an individual to be located on the study site more often
Chapter 20. Density estimation...
�20.2.1. Estimate proportion on site or use the data?
20  12
than an animal whose home range is near the edge of the site. So, {p(DTE) p̃(DTE) } (i.e., p(DTE)p∼(DTE))
is a candidate model we may want to run, and we expect the DTE covariate to be positively related to
both parameters. Since DTE is an individual covariate, we have to enter the covariate into the design
matrix (see Chapter 11). For our present example, our DM should look like the following:
If we run this model, we see that our hypothesis was well supported. This model is just over 50 AICc
units better than our previous best fitting model.
Also, as expected, the relationship between DTE and both p, and p̃ was positive. Note that this
covariate can be computed for most any closed capture or density estimation setup. Keep it in mind as
it often proves useful.
20.2.1. Estimate proportion on site or use the data?
To this point we’ve been accepting the default run options in MARK, such that p̃ is estimated using
logistic regression. This means that we’ve been specifying a model that relates the proportion of time
Chapter 20. Density estimation...
�20.2.1. Estimate proportion on site or use the data?
20  13
on site to an intercept, or an intercept plus group indicators, or an intercept plus covariates, etc. MARK
then uses this relationship to estimate p̃ for the animals that were sampled via markrecapture but not
telemetry (i.e., those individuals that had “. .” for the 2 telemetry columns in the input file) as well
as those that were sampled via telemetry. In other words, the resulting logistic model gets applied to
each of the i animals to derive an estimate of p̃ which is then summed over all individuals. You may
be wondering why you wouldn’t just use the telemetry data itself to estimate p̃ for those animals that
were telemetered, then use a logistic model for only those animals that weren’t telemetered.
Actually, you have that option. Retrieve model ‘p(DTE)p∼(DTE)’ in the browser and click the run
button. Notice that in the run window (shown below) you have an extra option for this data type. You
can check a box to “Use observed ptilde”. Let’s check the box this time and rerun the model. Add
the label “observed ptilde” to the name so we can keep it straight.
Notice (shown below) that this model has the exact same number of parameters, same log likelihood,
and same AICc as the original ‘p(DTE)p∼(DTE)’ model. This is because MARK still has to fit the logistic
model to your data and used the MLEs to estimate p̃ for those animals that weren’t telemetered. So,
it still estimates the same number of parameters, it just doesn’t use those parameters to estimate p̃ for
some individuals. Thus, it doesn’t make sense to compare the 2 approaches using AICc to see which
is a more appropriate. It’s completely up to the user and you should make a decision as to which way
you’d like to go before you start running models, then run each model the same way.
You might argue that since we’re making inference using a modelbased approach, we should be
consistent and apply the model to all individuals. Or, you might argue that there is no sense in estimating
Chapter 20. Density estimation...
�20.2.2. Threshhold Model
20  14
p̃ from a model when you can estimate it directly via sampling. Either approach is defensible. In practice,
this decision often has little consequence (especially if nearly all animals are telemetered) and estimates
are similar regardless of which flavor of model you prefer.
For example,if we compare the derived density estimates from the model we just ran to ‘p(DTE)p∼(DTE)’
we see that the estimates change slightly but we would probably make a similar inference regardless of
the approach.
20.2.2. Threshhold Model
Depending on the relationship between the size of your study site and the home range size of your
target species, you may find that you only want the DTE covariate to operate up to a certain threshold.
Consider the following example in which home ranges are now smaller than they were in our initial
figure above.
As we move through the home ranges from left to right, once we encounter the 3rd or 4th home
range, those animals are fully on the site and exposed to the same number of traps as those farther to
the right (more toward the center of the site). In this situation, p and/or p̃ may increase with DTE for a
Chapter 20. Density estimation...
�20.2.2. Threshhold Model
20  15
given distance, but after that, DTE is no longer a very good predictor of either parameter. Thus, it may
be useful to consider a “threshold model” for the estimation of p̃ and/or p such that the DTE covariate
is only meaningful to a point. After that p and/or p̃ are constant.
Mathematically, such a model can be represented as
�
�
��
logit p̃ˆ i � β̂1 + β̂2 min β̂3 , DTE ,
where β̂1 and β̂2 are the usual intercept and slope estimates from a logistic model using distance to
edge as a covariate, and β̂3 is the threshold parameter (i.e., the point at which the relationship between
p and/or p̃ asymptotes). Note however, that you cannot build a model exactly like this in MARK. MARK
is unable to estimate a parameter that is embedded within the ’min’ function. Instead, if you decide that
a threshold model may be appropriate for your situation, build several models in which you give MARK
an actual threshold value for DTE, then use AICc to help you figure out what the threshold should be.
In the example data set we’ve been analyzing, the trapping grid was (10 × 10) with 10m spacing.
Thus, the center of the grid is 45m from any edge when the edge of the grid is considered to be exactly
at the trap positions at the outer edge of the grid. Note that distance to the traps at the corners is in fact
√
( 2 × 45) from the center of the grid. Given that the edge of the grid is taken as the trap position on the
outer edge, the threshold values we consider should be less than 45 (hopefully you can see that if we
specify a model with a threshold of 45 in this specific example, we haven’t specified a threshold at all – p
or p̃ increase with DTE all the way to the center of the grid). Likewise, at the other end of the spectrum,
we expect the threshold to be greater than at least the 1st trap width (remembering there should be 4
traps per home range). Thus, let’s provide MARK with possible threshold values of 15, 25, and 35m for
the p parameter. To do this, simply translate the equation above into a design matrix, making use of the
design matrix function ‘min’ (see Chapter 11).
Thus, the design matrix for a threshold of 15m looks something like:
For other thresholds, just replace the ‘15’ in the design matrix expression with a different number
Chapter 20. Density estimation...
�20.3. Confidence Intervals
20  16
(25 or 35). If we examine the results browser (below) we see that the 15m threshold is the best fitting
model of the set we’ve constructed thus far:
20.3. Confidence Intervals
Recall from Chapter 14 that MARK computes logbased confidence intervals for abundance, partly due
to lack of asymptotic normality in point estimates and partly due to the idea that we want the lower
bound to be no less than M t+1 (makes no sense to have a lower confidence limit that is smaller than the
number of animals you know are out there!). Superficially, we’d like to invoke a similar approach with
density, and for the same reasons. Notice in section 14.1.10 that M t+1 (the number of animals capture or
the minimum bound on the abundance estimate) appears in the calculations as does f0 , which is simply
N̂ − M t+1 . You may be tempted to conclude that the analog to M t+1 for density estimation would the
Í
sum of all the p̃ i divided by the area of the study site. Then the analog to f0 would be D̂ − ( p̃ i )/A and
Í
all of the same computations would follow. However, p̃ i is not a known quantity as M t+1 was in the
Í
abundance world. It’s just an estimate (i.e., it’s really p̃ˆ i ).
Thus, instead of following the closed capture CI computation exactly, MARK uses logbased confidence intervals that ensure the CIs are > 0, which is the minimum condition that has to be true. As
such the 95% CIs are computed as:
"
± exp log(D̂) ± 1.96
�
SE(D̂)
D̂
�#
.
20.4. Assumptions
Recall from Chapter 14 (and earlier in this chapter) that closed capture models require 2 assumptions:
(1) the population is closed geographically (animals do not move on and off of the study site) and (2)
demographically (there are no births, deaths, immigration or emigration). Through the use of auxiliary
telemetry data we are able to relax the first of these assumptions when estimating density, but we still
need the population to be closed demographically. Fortunately, demographic closure is usually much
easier to attain and simply requires sampling for a relatively short duration at a time when you don’t
expect births or dispersal events. However, the density estimation model described here requires 3
additional assumptions beyond the usual closed capture ones:
1. The animals sampled with telemetry are representative of the population of animals that
use the study site.
2. Telemetry devices do not affect movements and there are no effects of markrecapture
sampling on animal movements.
Chapter 20. Density estimation...
�20.4. Assumptions
20  17
3. Error in telemetry locations is small relative to the size of the study site and assignment
(on/off) of locations near the edge of the site is unbiased.
The crux of assumption 1) is that you want to avoid oversampling “interior” animals while undersampling “edge” animals or viceversa. Hopefully if you think back the form of the estimator, you
can see why either of these cases would be problematic. If your severely undersample edge animals
(which should have relatively small values for p̃ i ), the summation of p̃ i in the numerator of our density
expression will be too large (all we sampled were animals with p̃ i near 1.0), and you will end up with
an estimate of density that is biased high to some degree. The reverse could also be true, and would
result in negatively biased results, although it is probably more likely that you would capture too many
interior animals compared to too many edge animals.
Note, however, we’ve also previously determined that DTE is often a good predictor of capture
probability. That is, interior animals may well have higher capture probabilities (p) than edge animals
because they tend to have more traps or detectors in their home range. If this relationship holds in
your study then it may be true that you tend to sample interior animals too often, which inflates the
numerator of our density expression, but we’re probably inflating the denominator at the same time,
thereby canceling at least some of the potential bias. Nevertheless, you want to provide yourself with
as much protection from bias as possible, and there are a couple of design features of markrecapture
sampling that may help.∗
For assumption 2, the main concern is if baits or lures were used to detect animals during the
markrecaptrue portion of your study, which is commonplace. Such attractants are designed to alter
the normal movements of animals (hopefully they entice animals to your detectors and increase your
capture probability!), but when we’re sampling animal movements using telemetry we’d like animal
movements to be “natural”. We have several recommendations for telemetry sampling when baits or
lures were used during markrecapture.
First, we recommend discarding from the analysis any telemetry locations obtained during the markrecapture session as it is very probable that baited detectors influenced animal movement. Second, it is
imperative that you the investigator ensure that every last morsel of bait has been removed from the site
at the end of the markrecapture session so there is no unnatural attractant left to influence movements
after the session. Third, we suggest that it may be appropriate to wait 1–2 days post markrecapture
before collecting location data to allow animals to revert to their normal activity patterns. Telemetry
sampling should, however, be completed within a reasonable time to avoid biasing estimates of p̃ i due
to seasonal movements, migration, or dispersal.
In summary, our experience suggests that bait and lures can have quite an impact on the movements of
animals during a markrecapture study, even enticing animals to make extensive movements to become
part of the markrecapture data set when their usual home range does not include the study site at all
(e.g., animal 4 in the original example). Following the guidelines outlined here, telemetry information
should result in p̃ i � 0 for these individuals, and they will not contribute to the density estimate, which
is appropriate.
∗
For traditional livetrapping studies in which telemetry devices are deployed on a subset of animals during the markrecapture
sampling, we recommend checking traps in the same order on each occasion, but selecting a different random starting point each
time. This strategy should help equalize the probability of capturing and telemetering edge vs. interior animals. In addition,
we suggest retaining some telemetry devices for deployment during the latter portion of a sampling session to facilitate the
inclusion of trapshy individuals in the radiotagged sample in addition to traphappy individuals that are captured early and
often.
All of this ‘hand waving’ is also a justification for using the logistic regression model for p̃ rather than the observed values,
because the logistic regression model would correct for the bias of a nonrandom sample of animals that were radioed, assuming
that at least a few edge animals (if not a random sample) made it into the telemetry sample.
Chapter 20. Density estimation...
�20.5. Summary
20  18
20.5. Summary
The summary for Chapter 14 states that “Despite a seemingly simple goal, estimating abundance can
be quite difficult...”. The same can be said for density estimation due to all of the numerous, subtle
complications embodied in the closed capture models along with some added complications necessary
to convert abundance to density. Despite these difficulties, the “Density with Telemetry” data type in
MARK provides a way forward and does so using auxiliary information that actually measures animal
movement on and off of the study site, which is the source of the density estimation problem. There are
other contemporary means of computing density as well, and we encourage readers to explore Ivan et
al. 2013 for a discussion of advantages and disadvantages of this approach compared to others.
20.6. References
Anderson, D. R., Burnham, K. P., White, G. C., and Otis, D. L. (1983) Density estimation of small mammal
populations using a trapping web design and distance sampling methods. Ecology, 64, 674680.
Borchers, D. L., and Efford, M. G. (2008) Spatially explicit maximum likelihood methods for capturerecapture studies. Biometrics, 64, 377385.
Converse, S. J., White, G. C., and Block, W. M. (2006) Small mammal responses to thinning and wildfire in
ponderosa pinedominated forests of the southwestern United States. Journal of Wildlife Management,
70, 17111722.
Dice, L. R. (1938) Some census methods for mammals. Journal of Wildlife Management, 2, 119130.
Efford, M. (2004) Density estimation in livetrapping studies. Oikos, 106, 598610.
Efford, M. G., Borchers, D. L., and Byrom, A. E. (2009) Density estimation by spatially explicit capturerecapture: likelihood based methods. Pages 255269 in D. L. Thomson, E. G. Cooch, and M. J.
Conroy, editors. Modeling Demographic Processes in Marked Populations. Springer Science, Boston,
Massachusetts, USA.
Huggins, R. M. (1989) On the statistical analysis of capturerecapture experiments. Biometrika, 76, 133140.
Huggins, R. M. (1991) Some practical aspects of a conditional likelihood approach to capture experiments. Biometrics, 47, 725732.
Ivan, J. S., White, G. C., and Shenk, T. M. (2013) Using simulation to compare methods for estimating
density from capturerecapture data. Ecology, 94, 817826.
Kendall, W. L.,Nichols, J. D., and Hines, J. E. (1997) Estimating temporary emigration using capturerecapture data with Pollock’s robust design. Ecology, 78, 563578.
Link, W. A., and Barker, R. J. (1994) Density estimation using the trapping web design: a geometric
analysis. Biometrics, 50, 733745.
Otis, D. L., Burnham, K. P., White, G. C., and Anderson D. R. (1978) Statistical inference from capture
data on closed animal populations. Wildlife Monographs, 7, 135.
Parmenter, R. R., Yates, T. L., Anderson, D. R., Burnham, K. P., Dunnum, J. L., Franklin, A B., Friggens,
M. T., Lubow, B. C., Miller, M., Olson, G. S., Parmenter, C. A., Pollard, J., Rexstad, E., Shenk, T., Stanley,
T. R., and White, G. C. (2003) Smallmammal density estimation: a field comparison of gridbased
vs. webbased density estimators. Ecological Monographs, 73, 126.
Royle, J. A., and Dorazio, R. M. (2008) Conceptual and philosophical considerations in ecology and
statistics. Pages 444 in J. A. Royle, and R. M. Dorazio, editors. Hierarchical Modeling and Inference
Chapter 20. Density estimation...
�20.6. References
20  19
in Ecology. Academic Press, Boston, Massachusetts, USA.
Royle, J. A.,Nichols, J. D., Karanth, K. U., and Gopalaswamy, A. M. (2009) A hierarchical model for
estimating density in cameratrap studies. Journal of Applied Ecology, 46, 118127.
Schwarz, C. J., and Arnason, A. N. (1996) A general methodology for the analysis of capturerecapture
experiments in open populations. Biometrics, 52, 860873.
White, G. C., and Shenk, T. M. (2001) Population estimation with radiomarked animals. Pages 329350
in J. J. Millspaugh, and J. M. Marzluff, editors. Radio tracking and animal populations. Academic Press,
San Diego, California, USA.
Williams, B. K., Nichols, J. D., and Conroy, M. J. (2002) Analysis and Management of Animal Populations.
Academic Press, New York, New York.
Wilson, K. R., and Anderson, D. R. (1985) Evaluation of two density estimators of small mammal
population size. Journal of Mammalogy, 66, 1321.
Zahratka, J. L., and Shenk, T. M. (2008) Population estimates of snowshoe hares in the southern Rocky
Mountains. Journal of Wildlife Management, 72, 906912.
Chapter 20. Density estimation...
�
Dublin Core
The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/.
Title
A name given to the resource
Journal Articles
Description
An account of the resource
CPW peerreviewed journal publications
Text
A resource consisting primarily of words for reading. Examples include books, letters, dissertations, poems, newspapers, articles, archives of mailing lists. Note that facsimiles or images of texts are still of the genre Text.
Dublin Core
The Dublin Core metadata element set is common to all Omeka records, including items, files, and collections. For more information see, http://dublincore.org/documents/dces/.
Title
A name given to the resource
Density estimation...
Description
An account of the resource
Abundance is commonly sought after as a state variable for the study of populations. However, density (number of animals per unit area) can be a more meaningful metric because it casts the state of a population into a common currency. For example, using closed capture models from Chapter 14, we estimate 500 animals at site A and 200 animals at site B. Thus one conclusion we may reach is that habitat management at site A has positively impacted the population there compared to site B. However, if we know site A is 250 hectares and B is 100 hectares, then we realize that each has 2.0 animals/hectare. That is, on a relative scale, the different management scheme at A had no effect compared to site B. Conversely, we may estimate abundance at 2 sites to be similar and conclude management actions, or habitat types, or harvest regulations, etc. are having a similar impact, but if the sites are different sizes, then the impacts are actually quite different on a relative scale and our conclusion is erroneous. Thus, while abundance can be a useful metric, estimating density can be helpful as well.
Creator
An entity primarily responsible for making the resource
Ivan, Jacob S.
Subject
The topic of the resource
Abundance
Density
Demographic closure
Sampling
Extent
The size or duration of the resource.
19 pages
Date Created
Date of creation of the resource.
20190807
Rights
Information about rights held in and over the resource
<div class="element">
<div class="elementtext"><a href="http://rightsstatements.org/vocab/InCNC/1.0/" target="_blank" rel="noreferrer noopener">In Copyright  NonCommercial Use Permitted</a></div>
</div>
Format
The file format, physical medium, or dimensions of the resource
application/pdf
Language
A language of the resource
English
Bibliographic Citation
A bibliographic reference for the resource. Recommended practice is to include sufficient bibliographic detail to identify the resource as unambiguously as possible.
Ivan, J. S. 2019. Density estimation. Pages 20.1–20.19 <em>in</em> E. G. Cooch and G. C. White, editors. Program MARK  'a gentle introduction'.
Type
The nature or genre of the resource
Article