Southeast Report
Waldemar Ortiz-Calo
2022-10-14
Intro/Description
This documents details the two models built for the Southeast study population which is also known as the cropland study population. This dataset consists of 77 white-tailed deer females. 74 of these individuals are classified as adults while the remaining 3 are considered juveniles.
Methods
Environmental Variables Considered
We developed a geographic information database consisting of 9 predictor variables describing the environmental characteristics of the study area. Each of these predictors was depicted as raster layers with a resolution of 30m. We conducted a moving window analysis in order to sampla area-based variables across the landscape. This moving window was based on a circular window with a 600m radius. Within this radius, all of our covariates were calculated and the resulting values were used to populate the resulting raster layers. We chose a 600m radius based on the mean available habitat across of our individual. This value was calculated using Durner’s available habitat methodology. The table below summarizes these variables.
| Variable | Description |
|---|---|
| Contagion | probability of two random cells belonging to the same class |
| Landscape Shape Index | ratio between the actual landscape edge length and the hypothetical minimum edge length |
| Mean Shape Index | ratio between the actual perimeter of the patch and the hypothetical minimum perimeter of the patch |
| Proportion of Developed | Proportion of developed land cover within 600m radius buffer |
| Proportion of Deciduous Forest | Proportion of deciduous Forest land cover within 600m radius buffer |
| Proportion of Evergreen Forest | Proportion of evergreen Forest land cover within 600m radius buffer |
| Proportion of Mixed Forest | Proportion of mixed Forest land cover within 600m radius buffer |
| Proportion of Grassland | Proportion of grassland land cover within 600m radius buffer |
| Proportion of Cropland | Proportion of cropland land cover within 600m radius buffer |
Logistic Regression
We calculated population-level resource selection pattern using a binomial logistic regression model. Prior to the analysis, we separated our data into 4 temporal bins that represented for the seasons of fall (September 1st – November 30th), winter (December 1st – February 28th or 29th), spring (March 1st – May 31st), and summer (June 1st – August 30th). Once these temporal bins were created, we then sampled for availability by estimating a population level polygon with a kernel density estimator and randomly sampling “available” locations within. Within this polygon, we sampled a used to available with a ratio of 1 to 5. Since the ratio between used and available can influence coefficient estimates and overall conclusions to resource selection studies, we calculated the optimal sampling ratio via a simulation of variable sampling intensities. An optimal sampling ratio ensures that’s that the number of available locations is sufficient to allow model convergence and accurately describes the distributions of covariates being used in the RSF analysis. Based on these simulations, we determined that a random sampling of five locations for each used location was appropriate. Once available locations were sampled, we extracted covariates values for the used and available locations. Then, we implemented a binomial logistic regression model to evaluate the patterns of selection in our study population. For these models, we implemented a backwards variable selection methodology, where we fit a global model and then progressively eliminated variables until all variables were below our designated threshold of significance. For this analysis, we considered a p-value of 0.05 as our significance threshold.
Discrete Choice
For our discrete choice models, we considered the same variables and seasons that were considered in our logistic regression models. However, we sampled for availability differently. We defined availability using the radius of available habitat method. Therefore, we established a buffer around each used location with a radius equal to c (a + 2b), where a, b, and c represent the mean hourly movement rate, the standard deviation of the movement rate, and the number of hours between visits to habitat units, respectively. When the used location occurred outside of the calculated buffer, we defined the radius as the straight-line distance between the previous and used location. Once the buffer was established, we randomly sampled available locations within said buffer. Based on the findings of the our simulations, we determined that a 1 to 5 ratio was the most appropriate sampling ratio for this analysis. We then fit our discrete-choice models and performed our variable selection. We implemented a backwards variable selection methodology, where we fit a global model and then progressively eliminated variables until all variables were below our designated threshold of significance. For this analysis, we considered a p-value of 0.05 as our significance threshold.
Logistic Regression Models
Fall
Winter
Spring
Summer
Model Interpretations
Across all four seasons, our models show that individuals strongly selected for areas with high proportions of cropland, deciduous forests, and grassland habitat types. These findings are not surprising considered white-tailed deer ecology and the findings of previous studies. These selection patterns remained relatively consistent across time, with slight differences between the seasonal models. For instance, in the fall, individuals did not demonstrate significant selection for mixed forests. Instead, they selected for cropland and deciduous forest. On the other hand mixed forests were selected for in the winter and spring seasons. We believe that these shifts are represent individuals selecting for landscape for thermal cover in the winter and forested habitat as refugia against predators in the spring during parturition.
Discrete Choice Models
Fall
Winter
Spring
Summer
Model Interpretations
The results of our discrete choice models are more ambiguous. These models demonstrate that individuals within our study area select for higher proportions of grassland and deciduous forests. Contrary to the logistic regression models, the proportion of cropland was selected against. However, the strength of these selection was relatively weak with broad standard errors around the mean. We believe that these results are a reflection of our sampling protocol in the discrete choice models. The sampling protocol that was implemented was a step-level scale. Considering the general movement patterns of our study population, individuals showed strong residential patterns by establishing relatively small home ranges and not moving great distances over time. Therefore, we believe, that our results reflect an over-sampling of the landscape with a limited amount of the landscape being considered since each choice-set was determined at a step-level.