Abstract # 2370 Event # 60:

Scheduled for Friday, June 20, 2008 09:30 AM-11:30 AM: Session 5 (Meeting Room 1GHI) Workshop

The Ugly, the Bad, and the Good of Missing and Dropout Data in Analysis and Sample Size Selection

K. Muller
University of Florida, Gainsville, Division of Biostatistics, Gainsville, FL, USA
     Missing and dropout data cause all manner of mischief in data analysis and sample size selection. The problems range from minor annoyances to catastrophic failures. The problems and solutions vary with the type of study design and analysis. The overarching framework involves a four step planning process: 1) define the purpose; 2) choose a design; 3) define the analysis; 4) select a sample size. The Ugly of missing and dropout arises mostly from unbalanced repeated (and multivariate) measures, which makes estimation hard, and inference harder, especially in small samples. Problem data may be missing completely at random, missing at random, monotone, drop out (loss to follow-up for any reason), censored, or mistimed. The Bad of missing and dropout comes from partial fixes that can mislead by hiding problems. Missing data can do Good by defining new responses (such as time to event) and uncover lurking variables. A credible design process includes accommodations for missing and dropout data. Statisticians have created good but sometimes difficult to use methods for estimation with missing data. Much less has been done for hypothesis testing and other forms of inference, especially in small samples. The biggest threat to a defensible selection of sample size comes from misalignment of the power analysis with the data analysis, not from missing data. Applying simple strategies to a fully aligned power analysis provide reasonable approximations with standard software.