RESUMEN
Predictive models are central to both archaeological research and cultural resource management. Yet, archaeological applications of predictive models are often insufficient due to small training data sets, inadequate statistical techniques, and a lack of theoretical insight to explain the responses of past land use to predictor variables. Here we address these critiques and evaluate the predictive power of four statistical approaches widely used in ecological modeling-generalized linear models, generalized additive models, maximum entropy, and random forests-to predict the locations of Formative Period (2100-650 BP) archaeological sites in the Grand Staircase-Escalante National Monument. We assess each modeling approach using a threshold-independent measure, the area under the curve (AUC), and threshold-dependent measures, like the true skill statistic. We find that the majority of the modeling approaches struggle with archaeological datasets due to the frequent lack of true-absence locations, which violates model assumptions of generalized linear models, generalized additive models, and random forests, as well as measures of their predictive power (AUC). Maximum entropy is the only method tested here which is capable of utilizing pseudo-absence points (inferred absence data based on known presence data) and controlling for a non-representative sampling of the landscape, thus making maximum entropy the best modeling approach for common archaeological data when the goal is prediction. Regression-based approaches may be more applicable when prediction is not the goal, given their grounding in well-established statistical theory. Random forests, while the most powerful, is not applicable to archaeological data except in the rare case where true-absence data exist. Our results have significant implications for the application of predictive models by archaeologists for research and conservation purposes and highlight the importance of understanding model assumptions.
