A comparison study of prediction approaches for multiple training data sets and test data with block-wise missing values