Event Causality Identification (ECI) is the task of recognizing causal relationship between events mentioned in texts. Due to its applications, ECI has been explored extensively by the Information Extraction community. However, the existing works are limited to either sentence-level ECI or they employ limited word interactions/structures at document level. As such, we propose a novel and effective method to comprehensively model contextual structures for ECI at three different levels, i.e., syntax, semantics, and background knowledge. Specifically, the contextual structures are integrated at different levels of the input encoder. Structure-aware representations are also combined using the graph transformer architecture to induce richer representations for ECI. We extensively evaluate the proposed model on two different benchmark datasets for the general and cross-domain settings. The experiments show the effectiveness of the proposed method by establishing new SOTA performance.