Speaker: Yanghyeon Cho

Title: Multivariate fractional imputation with non-monotone and nonignorable missing data

Abstract: We propose a method to infer parameters of interest with non-monotone incomplete data. Non-monotone missingness is frequently encountered in real-world problems, with reentering subjects who dropped out of longitudinal study serving as a classic example. There have been efforts to handle non-monotone missing patterns under the missing at random (MAR) condition. However, the MAR assumption under the non-monotone missingness may not be enough and the literature is insufficient to our knowledge. In this study, we employ the MAR assumption to the conditional odds between a pattern and its parent patterns given a pattern graph in Chen (2022) and then identify the nonignorable missing mechanism using this assumption. Next, we provide how to construct appropriate fractional weights with the identified missing mechanism for the parametric fractional imputation approach introduced by Kim (2011). Further, we illustrate the procedure under the multivariate categorical variables assuming the log-linear model. Extensive simulation studies empirically validate the proposed method.