Maximum Likelihood Methods for Nonignorable Missing Data

• Activity Number: 106
• Type: Contributed
• Date/Time: Monday, July 30, 2007 : 8:30 AM to 10:20 AM
• Sponsor: ENAR
• Author(s): Yan Zhou and Roderick J. Little and John D. Kalbfleisch*+
• Companies: University of Michigan and University of Michigan and University of Michigan
• Keywords: nonignorable missing ; maximum likelihood ; patter-mixture model ; factored likelihood

Abstract

Missing values are common issues in empirical studies. The performance of the methods analyzing missing data strongly depends on the missing-data mechanism. If missing mechanism is ignorable, the likelihood based inferences will only depend on the observed data. For two variables X and Y where both of X and Y are missing, however, missing-data mechanism may not be ignorable, since the missingness of Y can possibly depend on the values of X which are missing. We propose a nonignorable missing-data mechanism, in which X is MCAR and Y is MAR given the value and missingness of X. The non-iterative maximum likelihood estimates exist and data are excluded for estimating certain parameters. Extensions of this type of mechanism will be also discussed.