Quiñónez, AL, Fuller, TK and Randhir, TO. (2018). A Review of Otter Distribution Modeling: Approach, Scale, and Metrics. IUCN Otter Spec. Group Bull. 35 (2): 97 - 127
APPENDIX A2 – ADVANTAGES AND DISADVANTAGES OF MODELING APPROACHES
Presence-Absence phenomenological models/variable correlation methods
Advantages
The variety of tools available for this method (e.g. artificial neural networks, generalized linear models, generalized boosting models, generalized additive models, classification tree analysis, multi-adaptive regression splines, univariate and multivariate logistic regressions, stepwise discriminant analysis, flexible discriminant analysis, analyses for correlation, linear tend surface equation, partial regression analyses, piecewise linear functions, random forests, nearest-neighbor analyses, bootstrapping, variable transformations, correlation tests), allow us to take into consideration the information that is available and the selection of the most appropriate tools for the type of analysis that is being considered.
Can assess the relative importance of spatial, environmental and human factors that influence otter distribution (Barbosa et al., 2001).
Can be used to specify how much of the variance of the distribution of otters is due to different types of factors (variance partitioning), due to interactions between factors, and due to the combinations of factors (Barbosa et al., 2001).
These methods allow the establishment of negative or positive correlations, statistical significance of the relations found, and the level of confidence in the results obtained (Marcelli and Fusillo, 2009; Nel and Sommers, 2007).
Can be used to stablish ranges (thresholds) upon which the presence of certain factors exerts a greater influence (negative or positive) on the probabilities of observing a species (Marcelli and Fusillo, 2009; Nel and Sommers, 2007).
Distribution models based on presence probabilities, allow a more detailed knowledge of a species potential distribution when they are extrapolated to scales of finer resolution (Barbosa et al., 2003).
The sighting data used can take many forms, including presence-absence, sighting rate (de Oliveira et al. 2015), number of signs (Gori et al., 2013), proportion of positive sights (Barbosa et al., 2001), and frequency distribution (Thom et al., 1998).
These methods can be used to create habitat suitability maps, distribution of probability of occurrence maps, and presence-absence prediction maps which can be used to define action plans for conservation efforts (Gomez et al., 2014).
Disadvantages
Causal relationships among variables that are shown through the use of statistical regressions are not necessarily direct. A variable used could be an indicator or surrogate for a different unmeasured variable that does have an influence on the dependent variable (Barbosa et al., 2001; Barbosa et al., 2003).
It is important to understand that some of the factors we would like to consider can be correlated, and therefore the understanding of their individual effect may not be easy to interpret or define.
The variables chosen can provide randomness in the predictions made (Cianfrani et al., 2010; Santiago-Plata, 2013).
The quality of presence-absence HSM models should be carefully revised when the species-environment equilibrium assumption is not met (i.e. as in the case of species recolonization or expansion) (Cianfrani et al., 2010).
Since these methods use species presence/absence data for the analysis, false-positive, and false negative observations can affect the results obtained (Marcelli and Fusillo, 2009; Cianfrani et al., 2010).
Since the habitat suitability for otter species can be restricted to a 150m buffer around rivers or water bodies, it is important for the presence data to have high locational accuracy, particularly when calibrating the model (Cianfrani et al., 2013).
It is important to have enough otter presence/absence data to use some of the information in the calibration of the model and the other part in the validation of the model. If the same information is used for both processes, the model performance can be overestimated (Sepúlveda et al., 2009.)
Deductive Habitat Suitability methods
Advantages
In habitat suitability models, the classification of suitable and unsuitable areas can be made deductively based on the information known of the species of interest. A GIS overlay is a commonly used method (Ottaviani et al., 2009; Loy et al., 2009).
By providing information on overall habitat quality, HSMs provide an important base for determining potential habitats for species of interest (Ali et al., 2010).
Habitat suitability maps don’t require otter sighting data, but it can be used to validate the models ( Loy et al., 2009).
Disadvantages
When choosing areas for conservation prioritization based solely on ecological niche and habitat models, there is a high probability of including areas of low presence likelihood for the species of interest; i.e., overprediction (Cirelli and Cordero-Sanchez, 2009).
Wildlife habitat selection is affected by many factors and therefore no single theory is suitable for every animal since other factors that are not being considered or that have not been measure or determined could be limiting their distribution (Ali et al., 2010).
Because HSM are usually build based on the information that is available, they are commonly used under the assumption that wildlife-habitat relationships are consistent throughout different scales (Ali et al., 2010).
The variables chosen can provide randomness in the predictions made (Cianfrani et al., 2010; Santiago-Plata, 2013).
Occupancy methods
Advantages
Occupancy models use presence/absence data and the attributes of each site to define species-habitat relationship that are described as the probability of occupancy by a species (Santiago-Plata, 2013; Bieber, 2016).
To improve their assessment of species distribution and species-habitat relationships, occupancy models are now developed to account for mistakes in species detection by including estimates of detection probability. Detection probability reduces bias issues and allows for stronger inferences about species-habitat relationships (Jeffress et al., 2011; Bieber, 2016).
There are tools that are already developed for occupancy modeling, such as PRESENCE software and single season models.
One can assess the results obtained from the PRESENCE software with other statistical analyses to choose best model and define direction and relative effect size of variables used (Jeffress et al., 2011).
This approach can create ranges of occupancy estimates (Bieber, 2016).
Disadvantages
Requires more visits within the same site (like river stretch), which are necessary to allow for spatial replication, in order to determine the detection probability (Jeffress et al., 2011)
Substrate type can affect the detection probability (Jeffress et al., 2011).
In occupancy models, when the surveys for calculating detection probability are not the same day, it is recommended to survey the sites 3 or more times if the probability is > 0.5 (Bieber, 2016).
Presence-only methods
Advantages
These methods use incomplete information (presence-only data) to represent the ecological -niche of a species from the analysis of several variables and as a result produces a map of a species distribution probability (potential distribution) or habitat suitability within an area of interest (Santiago-Plata, 2013; Bieber, 2016).
There are tools that are already developed to create habitat suitability models, such as ENFA (Ecological Niche Factor Analyses), GARP, and MAXENT (Maximum entropy algorithm). These programs require presence data and thematic maps of the variables being considered.
Maxent includes the possibility using analyses that examine relative impact of each environmental variable (Jackknife test) and measures the fitness of the model (test of the area under the curve (AUC) in the receiver operating characteristic (ROC) plot) (Santiago-Plata, 2013; Bieber, 2016). It also has a generative approach, rather than a discriminative one, which prevents the over adjustment of the model when there is a reduced number of values (Santiago-Plata, 2013)
ENFA models compares the environmental characteristics of the sites occupied by the species to the characteristics of the whole area of interest (Cianfrani et al., 2010).
Absences may prevent models identifying areas that are suitable for a species to spread into. Therefore, when working with species that have an unstable distribution (recolonization a or expansion) presence-only models are more reliable (Cianfrani et al., 2010).
Distribution models based on presence probabilities, allow a more detailed knowledge of a species potential distribution when they are extrapolated to scales of finer resolution (Barbosa et al., 2003).
Disadvantages
Maxent does not have a rule of minimum of maximum number of data values require for it to provide an adequate analysis, therefore there is still some discrepancy regarding these values (Santiago-Plata, 2013). It is also sensible to the location of the presence data values; therefore, it may underestimate in areas where there are no observations registered, even when the region has suitable characteristics (Santiago-Plata, 2013).
Maxent produces three possible types of outputs: raw, log, and cumulative. This should be considered when choosing an output and interpreting the results (Santiago-Plata, 2013; Bieber, 2016).
When using Maxent for occupancy estimation, it may produce more liberal results than when using other occupancy estimation methods (Bieber, 2016). The estimates obtained through Maxent can vary a lot based on the occupancy threshold that is established (Bieber, 2016).
There is no limit to the variables used in these models, therefore one must be very careful with variables that are selected, since the quality of the model will depend on the quality of the predictors considered (Cianfrani et al., 2010; Santiago-Plata, 2013).
These methods function based on incomplete information, therefore underestimation and overestimation are still possible, depending on the quality of the data available. The results will be better when the presence data represents a wider variety of environmental condition under which the species can be found (Cianfrani et al., 2010).
Appendix A1 – Variables used for different otter species
Appendix A2 - Advantages and disadvantages of modeling approaches