A deep-neural-network-based anomaly detection and disease diagnosis have been actively researched and the integration is considered to be promising. However, devised effective networks may not be effective in real-world situations. One of the main reasons for hindering the network's performance is domain diversity. We have proposed the multi-frequency-based (MFB) standardization method for the multi-institutional data harmonization of the chest radiograph. In this study, we analyze and compare three methods for data harmonization: Conventional normalization, conventional standardization, and proposed MFB standardization. For comprehensive data analysis, we tried two methods; First, we extracted the radiomic features of the images and examined the harmonization ability of radiomics. Second, we visualized the embedded deep features and inspected the possible out-of-distribution. Both radiomic features and deep representations are well harmonized for the proposed MFB standardization, while conventional methods showed suboptimal results. The proposed method of feature analysis may contribute to improve the reliability and explainability of domain transfer studies in the medical deep learning tasks.