Gaussian Copula-based Imputation Method for Mixed Data with Non-Ignorable Missing Values

issing data with non-ignorable mechanisms (MNAR) poses a significant challenge for the analysis of mixed-type datasets. To address this, we propose a Corrected Marginal Gaussian Copula (CM-GC) imputation framework. The method first employs an Auxiliary Empirical Likelihood (AEL) approach within a semiparametric exponential tilting model to consistently estimate the propensity scores under MNAR. These scores are used to construct a Horvitz-Thompson estimator, recovering unbiased marginal distributions for the missing variables. Subsequently, the corrected marginals are linked via a Gaussian copula, projecting the data into a latent multivariate normal space.