The recursive model as a new approach to validate and monitor activity sensors

Activity sensors are increasingly being used to monitor animal activity but current methods, used to validate the relationship between the motion sensor information and the actual behavior of animals, have weaknesses. This study aims to improve the methods used to estimate activity level from dual axis activity sensors and to validate the Tellus activity sensor for reindeer (Rangifer tarandus). We developed a new approach, the recursive model (a recursive application of a logistic regression), to predict continuous values of activity without biased estimations or previous modifications of the dataset. We compared this new recursive model approach with two traditional approaches: the tree classification method and the standard model (based on simple logistic regression). Estimations from the tree classification and the standard model were dependent on the dataset used for validation, whereas the recursive model gave unbiased estimations. Estimations from standard and recursive models were also more accurate (lower average absolute errors) than those from the tree classification method and they had a slightly better discriminatory power (higher percentage of good classification). We successfully applied the recursive model for the first time and validated the Tellus activity sensor for reindeer. Any user can apply our methodology to obtain their own equations of the relationship between activity sensor values and the level of activity of the individual, and users monitoring reindeer activity with Tellus activity sensor can directly apply the provided equations under appropriate conditions.