If I'm doing prediction with a generalized linear model and a new batch of inputs comes with some missing values, what strategies can I use to minimize the loss of information from the missing inputs?
Let's assume that some of the predictors may be correlated and that the missing inputs are random.
My first guess would be to use mean imputation from mean of the training data. I read some bad things about mean imputation, but they were all about when doing estimation, not prediction. Is it a better choice in this case?