Optimism Bias in Ecological Modeling

Project Overview

This project explores the hazards of ignoring spatial autocorrelation in ecological modeling. Using the forested package and forest structure data from Washington State, this Quarto presentation demonstrates how standard random cross-validation yields overly optimistic performance estimates by allowing models to "cheat" via nearby neighbors. The analysis utilizes the spatialsample package to visualize and compare three distinct validation strategies—Random (the baseline), Spatial Blocking (geographic separation), and Environmental Clustering (ecological separation)—to establish robust, geographically transferable model performance metrics.