Early diagnosis of mild cognitive impairment (MCI) is crucial for the effective treatment and intervention of neurodegenerative diseases. Effective connectivity is one kind of brain network, which is helpful for analyzing the pathogenic mechanism of MCI. It is challenging to model causal relationships between brain regions from multimodal imaging data. This study proposes a new method for brain network causality modeling based on the structure-guided spatiotemporal diffusion model (SSDM), aiming to improve the accuracy of MCI diagnosis. By utilizing the advanced diffusion models, we introduced structural connectivity to guide the transformer-based network to learn topological and spatiotemporal features, which can better remove uncorrelated noise and improve effective connectivity estimation. The proposed model can not only generate temporal features of brain regions with individual differences but also construct discriminable effective connectivities. Experiments on the ADNI dataset demonstrate the effectiveness of our model, showing a certain improvement in diagnostic accuracy compared with competing methods. In addition, by analyzing the effective connectivities, our model predicts abnormal brain connections that are highly correlated with MCI. Overall, the framework proposed in this paper provides insights into the potential neurobiological mechanisms of MCI, which may promote early intervention strategies.