Abstract
This paper presents and contrasts procedures and conceptual underpinnings associated with statistical modelling and machine learning in the study of past locational patterns. This was done by applying the methods of logistic regression and random forest to a case study of coastal Mesolithic settlement patterns in southern Norway—a context that has not been subject to formal locational pattern analysis in the past. While the predictive accuracy of the the two methods was comparable, the different strengths and weaknesses associated with the methods offered a firmer foundation on which to both draw and moderate substantive conclusions. The main findings were that among the considered variables, the exposure of sites was the most important driver of Mesolithic site location in the region, and while there are some small indications of diachronic variation, the differences detected appear to both be of a different and far more modest nature compared to that which has previously been proposed. All employed data, the Python script used to run the analyses in GRASS GIS, as well as the R script used for subsequent statistical analysis is freely available in online repositories, allowing for a complete scrutiny of the steps taken.