Statistical Analysis with Missing Data., 3rd Edition

Preface to the Third Edition

There has been tremendous growth in the literature on statistical methods for handling missing data, and associated software, since the publication of the second edition of “Statistical Analysis with Missing Data” in 2002. Attempting to cover this literature comprehensively would add excessively to the length of the book and also change its character. Therefore, our additions have focused mainly on work with which we have been associated and we can write about with some authority. The main changes from the second edition are as follows:

1. Concerning theory, we have changed the “obs” and “mis” notation for observed and missing data, which, though intuitive, caused some confusion because subscripting data by “obs” was not intended to imply conditioning on the pattern of observed values. We now use subscript (0) to denote observed values and subscript (1) to denote missing values, which is in fact similar to the notation employed by Rubin's original (1976a) paper. We have also been more specific about assumptions for ignoring the missing data mechanism for likelihood-based/Bayesian analyses and asymptotic frequentist analysis; the latter involves changing missing data patterns in repeated analysis. These changes reflect material in Mealli and Rubin (2015). A definition of “partially missing at random” and ignorability for parameter subsets has been added, based on Little et al. (2016a).
2. Data previously termed “not missing at random” are now called ...