Chapter 12. Analysis of Incomplete Data

Geert Molenberghs^[]

^[] Geert Molenberghs is Professor, Center for Statistics, Universiteit Hasselt, Belgium.

Caroline Beunckens^[]

^[] Caroline Beunckens is Research Assistant, Center for Statistics, Universiteit Hasselt, Belgium.

Herbert Thijs^[]

^[] Herbert Thijs is Postdoctoral Fellow, Center for Statistics, Universiteit Hasselt, Belgium.

Ivy Jansen^[]

^[] Ivy Jansen is Postdoctoral Fellow, Center for Statistics, Universiteit Hasselt, Belgium.

Geert Verbeke^[]

^[] Geert Verbeke is Professor, Biostatistical Center, Katholieke Universiteit Leuven, Belgium.

Michael Kenward^[]

^[] Michael Kenward is Professor, Medical Statistics Unit, London School of Hygiene and Tropical Medicine, United Kingdom.

Kristel Van Steen^[]

^[] Kristel Van Steen is Professor, Universitiet Gent, Belgium.

Relying on Rubin's standard missing-data taxonomy, and using simple algebraic derivations, this chapter argues that some methods that are commonly used to handle incomplete longitudinal data are based on poor principles and are unnecessarily restrictive. We define longitudinal clinical trial data as complete case analyses and methods based on last observation carried forward (LOCF), for which the missing completely at random (MCAR) assumption is required.

Because flexible software is available that can analyze longitudinal sequences of unequal length, this chapter proposes a shift to a likelihood-based ignorable analysis that is carried out using SAS software.