Variable selection constitutes an essential step to reduce dimensionality and improve performance of fault detection and diagnosis in large scale industrial processes. For this reason, in this paper, variable selection approaches based on causality are proposed and compared, in terms of model adjustment of available data and fault detection performance, with several other filter-based, wrapper-based, and embedded-based variable selection methods. These approaches are applied in a simulated benchmark case and an actual oil and gas industrial case considering four different learning models. The experimental results show that obtained models presented better performance during the fault detection stage when variable selection procedures based on causality were used for purpose of model building.