Distribution-free multiple imputation in an interaction matrix through singular value decomposition

Abstract

Some techniques of multivariate statistical analysis can only be conducted on a complete data matrix, but the process of data collection often misses some elements. Imputation is a technique by which the missing elements are replaced by plausible values, so that a valid analysis can be performed on the completed data set. A multiple imputation method is proposed based on a modification to the singular value decomposition (SVD) method for single imputation, developed by Krzanowski. The method was evaluated on a genotype × environment (G × E) interaction matrix obtained from a randomized blocks experiment on Eucalyptus grandis grown in multienvironments. Values of E. grandis heights in the G × E complete interaction matrix were deleted randomly at three different rates (5%, 10%, 30%) and were then imputed by the proposed methodology. The results were assessed by means of a general measure of performance (Tacc), and showed a small bias when compared to the original data. However, bias values were greater than the variability of imputations relative to their mean, indicating a smaller accuracy of the proposed method in relation to its precision. The proposed methodology uses the maximum amount of available information, does not have any restrictions regarding the pattern or mechanism of the missing values, and is free of assumptions on the data distribution or structure.