The idea of informatively missing data, or informative dropout, is that the chance an observation is missing is related to its actual value. Examples of informatively missing data occur throughout Biostatistics, e.g., in cancer clinical trials with noncompliance, in longitudinal studies of numerical outcomes where the values of the numerical outcomes related to cancer and AIDS where the values of the numerical outcomes influence survival, etc. As the compliance example indicates, the field of informative missingness is closely linked to issues to causality. Informative missingness distorts standard analyses, and new approaches to statistical inference are required. There are three broad approaches to statistical inference in the presence of informatively missing data: (a) latent variable models; (b) pattern mixture models; and (c) selection models. In the pattern mixture model approach, separate models are fit for each pattern of missing data, and then through assumptions and/or sensitivity analysis, the disparate models are combined. In selection models and latent variable models, the missing data (or selection) mechanism is modeled directly in terms of the unobservables or in terms of latent variables. Pattern mixture models, latent variable models and selection models make assumptions that are not directly verifiable from data, and so sensitivity analysis is the norm. The purpose of this conference is to bring together some of the leading researchers in the area to present the newest statistical techniques. We plan a limited number of talks over a two-day span, with ample time for discussion and contrast of the methods.
