Multiple imputation using predictive mean matching for canonical correlation analysis with block-wise missing data

Abstract

Block-wise missing data refers to instances where data is missing in one variable group within a dataset consisting of two variable groups. Such data commonly arises during the integration process of data from multiple sources into a unified dataset. Applying canonical correlation analysis, commonly used with complete data, to block-wise missing data poses inherent challenges. To address this issue, our study proposes a tailored imputation method called MCCO (Multiple Correlation Coefficient Order) for block-wise missing data. Our approach employs multiple correlation coefficients to determine the imputation order and introduces a similarity measure inspired by existing PMM algorithms to select the closest donors. Additionally, we propose methods to determine key algorithm parameters, such as the number of close donors and proximity parameters, using multiple correlation coefficients and k-fold cross-validation, reflecting the characteristics of the given data. We evaluate the performance of our proposed MCCO method through real data and simulations, observing its competitive performance compared to listwise deletion, a method commonly used with block-wise missing data, and various existing PMM algorithms.