Incomplete hierarchical data

Abstract

The researcher collecting hierarchical data is frequently confronted with incompleteness. Since the processes governing missingness are often outside the investigator's control, no matter how well the experiment has been designed, careful attention is needed when analyzing such data. We sketch a standard framework and taxonomy largely based on Rubin's work. After briefly touching upon (overly) simple methods, we turn to a number of viable candidates for a standard analysis, including direct likelihood, multiple imputation and versions of generalized estimating equations. Many of these require so-called ignorability. With the latter condition not necessarily satisfied, we also review flexible models for the outcome and missingness processes at the same time. Finally, we illustrate how such methods can be very sensitive to modeling assumptions and then conclude with a number of routes for sensitivity analysis. Attention will be given to the feasibility of the proposed modes of analysis within a regulatory environment.