Sampling methods for archaeological predictive modeling: Spatial autocorrelation and model performance

With a case study of a direct detection model (DDM) designed for Fort Irwin, in southern California, we compare models that are identical except for the sampling methods that draw the data used to train and test them. We find that models trained and tested with a random cell sampling strategy perform better than do models that implement a kind of leave-one-out cross-validation (LOOCV) that focuses on discrete sites and non-site areas. We argue that this difference in measured predictive performance is due to spatial autocorrelation, and that it underscores the importance of clarity and specificity in describing the sampling methods used in archaeological predictive modeling and site detection. Different sampling methods are not necessarily superior or inferior, but they generate models that may be more or less appropriate for different tasks. Different sampling methods can also yield calculations of predictive ability that over- or understate a model's performance at the task for which it was designed.