Traffic flow reflects urban transportation health, and its prediction is vital for traffic management. However, current research overly emphasizes the spatiotemporal correlations of traffic flow, neglecting the physical concepts and causal relationships behind observational data generation. Under the influence of varying environmental factors, such as weather conditions, road construction, and traffic accidents, spatiotemporal correlations in traffic flow are considered unstable and spurious correlations may exist in the observations. To address the false correlation issue, we propose a causal inference-enhanced framework named CausMark that identifies genuine causal relationships in spatiotemporal data while mitigating spurious correlations through counterfactual reasoning. The model consists of a prior network and a posterior network. During latent variable inference, the posterior network separates conditional information from observed data and extracts the causal representation of the traffic features. To further explore possible causal links within this representation, we employ the causal propagation module. To make the model easier to understand and more reliable, we use a mutual supervision training method between the prior and posterior networks. Experiments with multimodal transportation data and weather data in Shanghai show that the model can effectively separate causal representations of concepts of interest, identify causal relationships, and accurately predict traffic characteristics.