Abstract <p>Multivariate time series data collected extensively from the real world allows us to observe urban phenomena on an unprecedented scale. However, recovering the underlying causal relations from these observations remains a challenging task, as these causal relations tend to be time-varying. Previous methods have extracted a causal graph over a long period of observation, but cannot be directly applied to capture, interpret, and verify dynamic causal relations. In this paper, we propose a novel visual analysis method for in-depth analysis of dynamic causal relations in multivariate time series. To address the following three challenges: detecting causality, explaining dynamic causality, and uncovering questionable causality, we design and develop the interactive visual analysis system MTCausal. First, a causal detection framework based on Granger causality test is used to obtain the time-varying causal relations in multivariate time series. Then, a dynamic causal graph visualization is designed to explore and interpret these causal graphs over time. Finally, a set of novel visualizations and interactions are designed to support the validation and comparison of causal relations to improve the results of causal analysis. The effectiveness of MTCausal is evaluated through the case studies on the real-world air pollution dataset, which demonstrate that users can effectively explore and analyze dynamic causal relationships.</p> Graphic Abstract <p></p>