Assessing Missing-Data Mechanisms in Longitudinal Studies: An Example Using a Smoking Cessation Trial

Longitudinal data analysis in the field of substance abuse studies commonly has two related problems: large numbers of repeated measures and incomplete data matrices. These problems are critical because the structures and mechanisms of the missing data are usually more complicated than other longitudinal studies with fewer repeated measures. For an incomplete longitudinal data set, different longitudinal modeling strategies are developed based on various assumptions on the mechanisms underlying the missingness. In this paper, we demonstrate a systematic method to test the missingness mechanisms underlying intermittent missing data (occasional missed data points) and dropouts (missing data due to premature withdrawal). For dropouts, we use a dimension-reduction technique based on bootstrap and multiple partial imputation. Based on the procedure demonstrated, proper longitudinal modeling choices can be recommended.