g., 30%�C50% bias) when kinase inhibitor Vandetanib the missing data are generated under a MAR process (Touloumi, Babiker, Pocock, & Darbyshire, 2001). Examining missing data patterns and testing assumptions of the missing data mechanisms is vital for the selection of an appropriate analytic technique to address missing data and minimize bias (Houck et al., 2004; Yang & Shoptaw, 2005). This article presented two alternative and accessible methods for examining longitudinal dichotomous data under the less stringent MAR assumption. Many missing data experts advocate use of MAR analysis as the default approach, unless there are strong reasons to support the MCAR assumption (e.g., Fitzmaurice, Laird, & Ware, 2004). In some cases, researchers may suspect that the MAR assumption is not reasonable and may suppose that missingness is related to what they would have observed (had they been able to do so).
Such a situation is labeled ��missing not at random�� (MNAR). It is impossible to distinguish between MAR and MNAR based on the observed data because the distinction involves the missing data. Nonetheless, if MNAR is strongly suspected, one can do a kind of sensitivity analysis using MNAR approaches, such as selection and pattern-mixture models (Little, 1995). Also, Hedeker, Mermelstein, and Demirtas (2007) describe a MNAR multiple imputation procedure for missing substance abuse data that allows varying degrees of association of missingness and the substance use outcome. One limitation of the current example is that we did not distinguish between intermittent missing data and missing data resulting from dropout.
Yang and Shoptaw (2005) differentiate missing data that are intermittently missing from missing data that are consistently missing due to dropout based on the notion that missing data due to dropout, compared with intermittent missing data, may be more related to study variables. Using this distinction can reduce the possible number of missing data patterns required for multiple imputation analyses via a method they labeled multiple partial imputation (Yang & Shoptaw, 2005). The present study also did not outline other methods of dealing with missing data, such as sequential imputation (Kong et al., 1994) and Bayesian quantile regression (Yuan & Yin, 2009), which can address non-ignorable missing data.
Given the ubiquitous nature of missing data in substance abuse trials and the potential bias resulting from mishandling of missing data, care is required when selecting an analytic plan. This example demonstrates the problem Anacetrapib of using an analytic technique, in this case GEE, without consideration of missing data patterns, and advocates testing the MCAR assumption. While GEE is an appropriate technique for analyzing data when the MCAR assumption is not violated, weighted GEE or mixed-effects regression is more appropriate when the missing data mechanism is not MCAR.