Due to Synthetic Aperture Radar (SAR) imaging characteristics, SAR vehicle recognition faces the problem of extracting discriminative and robust target features from a small dataset. Deep learning has shown impressive performance on the MSTAR dataset. However, data bias in a small dataset, such as background correlation, impairs the causality of these methods, i.e., discriminative features contain target and background differences. Moreover, different operating conditions of SAR lead to target signatures and background clutter variations in imaging results. However, many deep learning-based methods only verify robustness to target or background variations in the current experimental setting. In this paper, we propose a novel domain alignment framework named Hierarchical Disentanglement-Alignment Network (HDANet) to enhance features' causality and robustness. Concisely, HDANet consists of three parts: The first part uses data augmentation to generate signature variations for domain alignment. The second part disentangles the target features through a multitask-assisted mask to prevent non-causal clutter from interfering with subsequent alignment and recognition. Thirdly, a contrastive loss is employed for domain alignment to extract robust target features, and the SimSiam structure is applied to mitigate conflicts between contrastive loss and feature discrimination. Finally, the proposed method shows high robustness across MSTAR's multiple target, sensor, and environment variants. Noteworthy, we add a new scene variant to verify the robustness to target and background variations. Moreover, the saliency map and Shapley value qualitatively and quantitatively demonstrate causality. Our code is available in \url{https://github.com/waterdisappear/SAR-ATR-HDANet}.