Occupancy Model Fact Sheet
Occupancy Model Fact Sheet
include: Thus, we can write the probability of detection history (0000) as:
(1) Occupancy state is “closed.”
4
Species are present at occupied
sites for the duration of the sam-
Pr(H i = 0000) = ψ Π (1 – p ) + (1-ψ)
j=1
j
independent of detecting the spe- Maximum likelihood methods incorporated in program PRESENCE
cies at other sites. This might be or program MARK are used to obtain estimates of occupancy and
a problem if your sites are closely detectability (see MacKenzie et al. 2002).
spaced, allowing animals to move
Biases Caused by Assumption
Violations
If the assumptions mentioned
above are not met, estimates of oc-
cupancy and detectability can be
biased10 and inferences about fac-
tors that influence these parameters
may be flawed (e.g., Gu and Swihart
2004). If the target species is not
present at sites throughout the entire
study season, then estimates of oc-
cupancy may still be unbiased if the
species moves randomly in and out
of a sampling unit. Interestingly, the
interpretation of occupancy changes
to the proportion of sites used by the
Kenai National Wildlife Refuge
target species in this case. Likewise
the probability of detecting the spe-
cies at occupied sites is now a combi-
nation of two different components:
among sites and be detected at Sites should be chosen according the probability that the species was
multiple sites. to some type of probability-based present at the sampling unit and the
(3) No unexplained heterogene- sampling9 (e.g., simple-random probability of detecting the species
ity8 in occupancy. Probability of sample, stratified-random sample, given it was present. If movement in
occupancy is the same across sites etc.) to ensure that estimates of oc- and out of the sampling unit is not
or differences in occupancy can be cupancy apply to the area of interest. random, the occupancy estimator
explained with site characteristics Sampling may be standardized to try will likely be biased. For example,
(covariates) that have been quanti- to minimize differences in detection non-random movement occurs if the
fied for inclusion in the model. probability caused by variation in target species was not initially at the
(4) No unexplained heterogeneity in site when sampling commenced, then
environmental conditions or time of
detectability. Detectability at oc- moves into the sample unit and stays
day. Unfortunately, it is impossible
cupied sites is the same across all for the duration of the season. The di-
to control for all possible factors that
surveys and sites, or differences in rection of the occupancy bias depends
can affect detectability or occupancy.
detectability can be explained with on the direction of the movement (see
When design-based approaches
site or survey characteristics that Kendall 1999 for more details and
cannot reduce all of the variation
have been quantified for inclusion possible solutions).
(sometimes termed heterogeneity)
Little work has been done involv-
in the model. in either occupancy or detectability,
ing the impact of variation in occu-
model-based approaches might help.
pancy probability among sites that
Approaches to Meeting Investigators should collect informa-
cannot be associated with covariates
Assumptions tion about factors they believe could (occupancy heterogeneity). It is pos-
Investigators can use design- cause heterogeneity in either of these sible that the overall average occu-
based or model-based approaches to two parameters and then incorporate pancy estimate may still be unbiased,
meet these assumptions. A design- these covariates into the estimation but more research on the impacts of
based approach involves collecting process. For example, it may be im- occupancy heterogeneity is needed.
data in a way that assures the as- practical and wasteful to only sample Heterogeneity in detection prob-
sumptions will be met. For example, amphibian breeding ponds during ability will often result in occupancy
natural history information may aid sunny days, but larvae are more estimates that are low (negatively
in scheduling surveys during times difficult to see on cloudy days; thus, biased). This problem is further exac-
when the closure assumption is most investigators should record weather erbated in studies involving a small
likely to hold. In addition, a scientist conditions during each survey. The number of sites, few repeated surveys
may use existing movement informa- variation in detection probabilities at each site, or species with excep-
tion to ensure that sample sites are caused by cloud cover can easily tionally low detection probabilities.
dispersed across the sample area in a be incorporated to obtain unbiased Anticipating variation and minimiz-
manner that maintains independence. estimates of occupancy. ing its effects either through study
design or collecting relevant covari-
ates to model it is essential for good
performance of these methods. Box 2: Results of model selection for salamanders of
Finally, if detection is not inde- the Desmognathus imitator complex in Great Smoky
pendent among sites, the standard Mountains National Park
error estimates are usually too low. In
these instances, the number of inde- The models we considered assume occupancy was either
pendent sites is actually smaller than constant across sites ψ (.), or varied according to the sites’
the total number of sites surveyed. previous disturbance history, ψ (dist). Detectability was either
There are existing model-based constant across all surveys and sites p(.), or it varied among
methods that aid in detecting and cor- surveys p(t), or across sites according to previous disturbance
recting this problem (MacKenzie and history p(dist), or both p(dist+t). Model selection was based
on Akaike’s Information Criteria for small sample size (AICc),
Bailey 2004).
which selects the most parsimonious model, balancing model
Occupancy and Terrestrial fit and parameter precision. Models with lower AICc are con-
Salamanders sidered best. AICc is the difference between the AICc of the
best model and a subsequent model. As a general rule, models
Approximately 10% of the with AICc<2.0 have substantial support and should be consid-
world’s salamander species are found ered when making inferences or reporting parameter estimates.
in the southern Appalachian region, Another way of assessing model support is with Akaike weight
with 31 species occurring inside the (w), which can be interpreted as the weight of evidence for a
boundaries of Great Smoky Moun- particular model. All weights sum to 1.
tains National Park (GSMNP; Dodd
2003). Despite this rich diversity,
large-scale or long-term studies of Model Parameters AICc AICc Weight (w)
terrestrial salamanders in this re- ψ (dist) p (dist + t) 7 89.68 0.00 0.32
gion are almost nonexistent (but see ψ (dist) p (t) 6 90.27 0.59 0.24
Hairston and Wiley 1993). An area of ψ (dist) p (dist) 4 91.03 1.35 0.16
interest involves the impact of vari- ψ (dist) p (.) 3 91.47 1.78 0.13
ous forms of disturbance on terrestrial ψ (.) p (dist + t) 6 92.25 2.57 0.09
salamander populations (Petranka et ψ (.) p (dist) 3 93.54 3.86 0.05
al. 1993, Ash 1997, Petranka 1999, ψ (.) p (t) 5 100.36 10.68 0.00
Ash and Pollock 1999). Some areas ψ (.) p (.) 2 102.01 12.33 0.00
within GSMNP incurred heavy hu-
man use in the form of logging or
settlement prior to the park’s estab-
lishment in 1934. In this example 1.0
analysis, we ask if previous distur-
bance history affects the probability 0.8
of occurrence for salamanders of the
Desmognathus imitator complex in
Detectability
0.6
GSMNP. We caution that this analysis
is meant as an example only, and we 0.4
remind readers that there is always an
inherent danger in inferring biological 0.2 Undisturbed
And