Figure 2

Comparison of k nearest neighbor (KNN), Row Average, and Least Squares Methods (LSM), and NIPALS imputation methods. 500 simulations were performed, where each simulation generated a dataset containing 20% missing values by randomly removing spot values from the complete matrix of 70 protein spots. Missing values were imputed using row average (Row Ave), the KNN method with different k nearest neighbor values, or LSM method and the results compared using the normalized root mean square error (RMSE). One set of LSM options allow the user to choose a correlation between protein spots estimated via least squares (LSM.LS.G) or via the EM algorithm (LSM.EM.G). Another set of LSM options allows the user to choose a correlation between arrays estimated via least squares (LSM.LS.A) or via the EM algorithm (LSM.EM.A). The user is allowed a combined (array and spot) correlation (LSM.LS.C) and adaptive (LSM.LS.Ad) correlation procedure. The NIPALS methods are summarized by "nPR" which denotes the number of principal components used to impute the missing data.