Effects of imputation on correlation: Implications for analysis of mass spectrometry data from multiple biological matrices

With expanded access to, and decreased costs of, mass spectrometry, investigators are collecting and analyzing multiple biological matrices from the same subject such as serum, plasma, tissue and urine to enhance biomarker discoveries, understanding of disease processes and identification of therapeutic targets. Commonly, each biological matrix is analyzed separately, but multivariate methods such as MANOVAs that combine information from multiple biological matrices are potentially more powerful. However, mass spectrometric data typically contain large amounts of missing values, and imputation is often used to create complete data sets for analysis. The effects of imputation on multiple biological matrix analyses have not been studied. We investigated the effects of seven imputation methods (half minimum substitution, mean substitution, k-nearest neighbors, local least squares regression, Bayesian principal components analysis, singular value decomposition and random forest), on the within-subject correlation of compounds between biological matrices and its consequences on MANOVA results. Through analysis of three real omics data sets and simulation studies, we found the amount of missing data and imputation method to substantially change the between-matrix correlation structure. The magnitude of the correlations was generally reduced in imputed data sets, and this effect increased with the amount of missing data. Significant results from MANOVA testing also were substantially affected. In particular, the number of false positives increased with the level of missing data for all imputation methods. No one imputation method was universally the best, but the simple substitution methods (Half Minimum and Mean) consistently performed poorly.