PCA model building with missing data: New proposals and a comparative study

This paper introduces new methods for building principal component analysis (PCA) models with missing data: projection to the model plane (PMP), known data regression (KDR), KDR with principal component regression (PCR), KDR with partial least squares regression (PLS) and trimmed scores regression (TSR). These methods are adapted from their PCA model exploitation version to deal with the more general problem of PCA model building when the training set has missing values. A comparative study is carried out comparing these new methods with the standard ones, such as the modified nonlinear iterative partial least squares (NIPALS), the iterative algorithm (IA), the data augmentation method (DA) and the nonlinear programming approach (NLP). The performance is assessed using the mean squared prediction error of the reconstructed matrix and the cosines between the actual principal components and the ones extracted by each method. Four data sets, two simulated and two real ones, with several percentages of missing data, are used to perform the comparison.