Comparing False Crawls and Pseudo-Absence Points when Modeling Sea Turtle Nesting on Pensacola Beach
Abstract
Objectives
Loggerhead sea turtles (Caretta caretta) select nesting sites based on various environmental factors, yet the influence of beach morphology on nesting decisions remains unclear. In order to model nesting patterns, both observed nests and non-observed nests must be present in the data. The absence of nests can be represented with either pseudo absence points, which are randomly selected via ArcGIS, or false crawl points, which are observed and recorded by state employees while recording observed nests.
This study evaluated how different representations of “non-nesting” locations affect model estimation and classification performance when modeling loggerhead sea turtle (Caretta caretta) nesting on Pensacola Beach. Specifically, we assessed whether binary and multinomial logistic regression models produced different inferences about beach slope and foreshore slope on nesting likelihood, and whether the choice of absence data influenced prediction accuracy.
Methods
Presence (P) and false crawl (FC) locations were recorded by trained observers along Pensacola Beach, while pseudo-absence (PA) points were generated in ArcGIS by randomly sampling areas outside known nest buffers. Beach slope and foreshore slope were then calculated for each location.
Three models were fit:
- Binary logistic regression, comparing P vs. PA.
- Binary logistic regression, comparing P vs. FC.
- Multinomial logistic regression, simultaneously comparing P vs. PA and P vs. FC.
For each model, adjusted odds ratios (ORs), 95% confidence intervals, and Wald \(z\) tests assessed the significance of foreshore slope and beach slope. Statistical significance was defined as \(p < 0.05\).
To examine classification, predicted probabilities were used to generate confusion matrices, which identify the number of true positive, true negative, false positive, and false negative classifications. Using the entries in the confusion matrix, we calculated accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
Results
Slope variables had small, non-significant effects across all modeling approaches. In binary models, a 1° increase in beach slope decreased the odds of nesting by 5% when comparing to FC but incrased the odds of nesting by 2% when comparing to PA. Foreshore slope increased odds of nesting by 1–5%, depending on model, but none of these effects were significant (p > 0.15 for beach slope; p > 0.45 for foreshore slope). Multinomial models produced similar estimates and identical inferential conclusions.
Classification performance differed by absence type. Pseudo-absence models achieved moderate accuracy (≈56%) and modest sensitivity (0.55–0.65). Interestingly, False-crawl models failed to classify FC points as such, producing specificity near 1 and sensitivity of 0. Across all cases, binary and multinomial approaches yielded nearly identical accuracy, PPV, NPV, and classification patterns.
Conclusions
The choice between pseudo-absence points and observed false crawls affects model performance, but does not materially change estimated effects of environmental predictors. Beach slope and foreshore slope showed weak, non-significant associations with nesting probability, regardless of modeling approach. Because false crawls were never classified correctly, pseudo-absence points produced more stable and interpretable results for presence-absence modeling in this application. Both binary and multinomial logistic regression generated similar conclusions, suggesting little added benefit to using the more complex multinomial logistic regression model.