Abstract
Chinese tallow ( Triadica sebifera ) is a menacing invasive tree species in the southeastern United States (US), posing serious ecological and economic threats. Despite numerous regional-scale studies, little research has examined methods to understand localized drivers of tallow density, particularly in the pine flatwoods of coastal Mississippi. This study evaluated five statistical models, Linear Regression (LR), Spatial Autoregressive Error Model (SEM), Spatial Autoregressive Lag Model (SLM), Conditional Autoregressive Regression model (CAR), and Geographically Weighted Regression (GWR) with Gaussian and bisquare kernels, to compare their performance in predicting tallow density across a 4.75 km 2 area within the Mississippi Sandhill Crane National Wildlife Refuge. Field data from 219 quadrats were aggregated at the patch level. A significant spatial autocorrelation (Moran’s I = 0.12 , p -value = 0.005) justified the use of spatial models. GWR with a bisquare kernel yielded the best performance (AIC = 27.24), followed by the CAR model (AIC = 28.62), both of which effectively captured localized spatial patterns and reduced residual autocorrelation. Tallow density was associated with lower elevation, frequently moist soils, and moderate pine midstory, conditions favorable for seed dispersal and seedling establishment. Integrating local-scale spatial models with ecological field data can inform targeted management strategies in vulnerable coastal ecosystems.