Abstract:
|
In randomized trials of smoking cessation interventions, missing data on the primary outcome of interest, current smoking status, is handled by single-value, worst-case imputation: Everyone with missing data, such as non-respondents, is assumed to be smoking and included in the analysis coded as such. This is a long-standing, well-established practice in the field. Many researchers have long recognized this approach as being non-optimal. Recently, a working group from the Society for Research in Nicotine and Tobacco recommended alternative methods to analyze smoking status (a binary outcome) with missing data (Hall SM et al, Nicotine & Tobacco Research (2001) 3, 193-202). This paper reviews five methods recommended by the paper: GEE (when outcome is "covariate dependent," missing completely at random (MCAR)); mixed models (for MAR); multiple imputation (MAR); pattern mixture models (non-ignorable); and selection models (non-ignorable). The last two were referred to as "optimal missing data analysis strategies." We compare the five methods using simulations and actual data from smoking cessation randomized trials and comment on their strengths and weaknesses in various situations.
|