Aspect-based sentiment analysis identifies the polarity of targeted words or visual regions expressing latent opinions, typically via large language models (LLMs) on semantic and context understanding. Anyway, LLMs tend to converge on plenty of underlying correlations that are spurious and insignificant. To this end, we propose a chain-of-thought prompting with causal intervention (CPCI) to exploit the causality correlations for multimodal aspect-based sentiment analysis. More specifically, CPCI introduces stochastic perturbations on the original sentences to generate augmented ones as correlation candidates. Furthermore, backdoor adjustment is devised in the causal intervention (CI) module by using the inverse attention matrix of transformer as confounder in the structural causal model. In particular, a random sampling is conducted on the confounder to be integrated with the attention features of the transformer to eliminate the influence of spurious correlations. Finally, a chain-of-thought prompting (CP) module is devised on the augmented sentences without spurious correlations to infer the underlying properties from the latent intent of opinion to implicit sentiment polarity gradually. Extensive experiments conducted on two public datasets show that CPCI outperforms state-of-the-art approaches.