Prediction to the Max(Net): Comparing Logistic Regression Models to Maximum Entropy Models for Predicting Archaeological Sites within Pitt County North Carolina

Abstract

Archaeological predictive models are typically used to estimate the likelihood of encountering a prehistoric site within a given area. Traditionally, these models are created using logistic regression, however this requires the use of presence and absence data. This is an issue in archaeology because we work with presence only data. Maximum Entropy models offer an alternative by using presence only data which reduces the uncertainty introduced by randomly generated absence points. This thesis compares logistic regression and maximum entropy (MaxNet) models to determine if there is a difference between their precision and/or accuracy in estimating prehistoric archaeological site locations within Pitt County, North Carolina. Using the same environmental variables, these predictive models were generated for three time periods: a general prehistoric model, an Archaic sites model, and a Woodland sites model. These models were evaluated using P(S|M), precision ratios, and AUC scores. Across nearly all metrics, the Maximum Entropy models preformed as well as or slightly better than the logistic regression models in terms of accuracy. In terms of precision, the Maximum Entropy models preformed significantly better than logistic regression across all metrics. These findings suggest that Maximum Entropy models are on par with logistic regression models in terms of accuracy, and preform significantly better with precision. These difference in precision can have massive impact on the overall cost of archaeological surveys.