Comment.

This article comments on a paper by Roderick J. A. Little which discussed a method of imputation that is a mixture of model-based and nearest neighbor ideas. The matching is based on the predictions of missing values, but the actual imputations are from the matching record identified. Problems arise with this idea because the match is done on the predictions, not the predictors, and implies a great deal of faith in the model that one is using. Certainly, one would have to check out the predictive efficiency of a model before proceeding with this method. One positive feature is that some of the predictors could be non-numeric, whereas the distance function is expressed in terms of the numeric predictions only. Another concern with this method is that the constraints on the data are not taken care of. Little does refer to this problem, but he appears to hope that it is a minor. Since the predicted variables are used only to define a match, the actual imputed variables are used only to define a match, the actual imputed values, since they come from a single valid record response, will be consistent among themselves. With the idea of multiple imputation comes the notion of repeating the analysis under different imputation scenarios. The sensitivity of the analysis to the missing data could then be assessed, assuming the means to generate appropriate measures of variability were available. The potential for generating vast quantities of paper is enormous. More seriously, analysts would have to learn to describe the procedures and report results in this multiple format.