To the editor: Our editorial mainly focused on the shortcomings of the last observation carried forward approach and highlighted analysis in the presence of dropouts, depending on how the dropout occurred (missing completely at random, missing at random, or missing not at random). We were not asked to write an editorial discussing the many methods of handling or imputing missing data in general, and, as a result of space limitations, we could not cover all aspects of missing data in the detail some might argue such an important subject warrants. Therefore, we welcome the letter by Drs. Potkin and Siu, who have highlighted two very important issues in relation to missing data, especially dropout rate.
First, unfortunately we often have a missing data rate >20%, and thus it is crucial to make sure that the results allow for the missing data in an appropriate way. As Drs. Potkin and Siu state, “the key issue is whether those who drop out differ from those who remain in the study and how this relates to individuals’ treatment outcomes.” Hence, understanding whether the data are missing completely at random, missing at random, or missing not at random (in an informative way) allows much better understanding of the results. Second, Drs. Potkin and Siu use, as an example, their 40-week randomized study of ziprasidone and haloperidol (1). In their study, the missing completely at random assumption was indeed violated, and the pattern they observed may not have been even missing at random, since the missing pattern may depend on the outcome that is not available from the dropouts. If this is the case, then the pattern of missing data is informative, and as we mentioned in our editorial, using pattern-mixture models may be appropriate. Subjects are first divided into groups based on their missing data patterns and parameters estimated, and then results are aggregated. Although this approach is intuitively attractive, it can be difficult to apply in practice, since it is necessary to have adequate numbers with each pattern to adequately apply this methodology. Further, the data may lack power to detect the extent to which the missing assumption arises informatively, hence the need for other approaches, such as shared parameter models and definitely sensitivity analysis.
The author’s disclosures accompany the original editorial.
This letter (doi: 10.1176/appi.ajp.2009.09070959r) was accepted for publication in August 2009.