Abstract

Covariate balancing methods have been widely applied to single or monotone missing patterns and have certain advantages over likelihood-based methods and inverse probability weighting approaches based on standard logistic regression. In this paper, we consider non-monotone missing data under the complete-case missing variable condition (CCMV), which is a case of missing not at random (MNAR). Using relationships between each missing pattern and the complete-case subsample, a weighted estimator can be used for estimation, where the weight is a sum of ratios of conditional probability of observing a particular missing pattern versus that of observing the complete-case pattern, given the variables observed in the corresponding missing pattern. Plug-in estimators of the propensity ratios, however, can be unbounded and lead to unstable estimation. Using further relations between propensity ratios and balancing of moments across missing patterns, we employ tailored loss functions each encouraging empirical balance across patterns to estimate propensity ratios flexibly using functional basis expansion. We propose two penalizations to separately control propensity ratio model complexity and covariate imbalance. We study the asymptotic properties of the proposed estimators and show that they are consistent under mild smoothness assumptions. Asymptotic normality and efficiency are also developed. Numerical simulation results show that the proposed method achieves smaller mean squared errors than other methods.