> >Technicalities aside, any imputation method implies a model for the >predictive distribution of the missing values. The important point is to >model that as well as possible, and then propagate the >uncertainty. Whether you are Bayesian or not, MI is a useful tool for >doing this. Rod Little >
Fine, that some discussion on properness is now here. Rod's above point (if I correctly understand it) is extremely important from the practitioner's point of view. My experience shows that the model should be very good if it may be nicely used as a model-donor imputation method. In most data I know, it is difficult to specify a model well. Instead, if certain conditions or contraints have been used, it may help (who has experiences about these?). Another point is that it is often (not always) easier to use the same model for a real-donor technique, and to obtain a more proper solution. This is not only the problem of MI. The same is concerned single imputation methods, which still today are (only?) used in NSI's. Seppo Laaksonen Statistics Finland
