Analysis of Longitudinal Data with Nonmonotone, Nonignorable Missing Data

• Activity Number: 273
• Type: Contributed
• Date/Time: Tuesday, August 10, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Lin Wang*+ and Mari Palta and Jun Shao
• Companies: University of Wisconsin, Madison and University of Wisconsin, Madison and University of Wisconsin
• Keywords: longitudinal ; missing data ; IPW

Abstract

Nonmonotone missing outcomes present a difficult problem for longitudinal data analysis. When the missing probability depends on outcomes at previous time points, those outcomes could also be missing, and hence the missingness is nonignorable. Selection model approaches leading to weighted analyses have been proposed. Some of these exclude observations to make missingness monotone or assign individual level weights to subjects with complete observations only. We show that such methods can be quite inefficient. We extend Robins' method (1997), to the longitudinal case and apply separate inverse probability weights to each observation within an individual. We then use GEE with independent working correlation structure to obtain estimates. We also consider application of a regression-based, pattern mixture approach, where the regression parameters are estimated from a conditional mean function without specification of the missing mechanism.