Improving the coherence of sequential imputation via calibration
*Recai Murat Yucel, State University of New York at Albany 

Keywords: missing data, congeniality, multiple imputation, calibration, rounding

Inference by multiple imputation(MI) is a popular method for handling missing data in many scenarios. One of the biggest obstacles facing practitioners of MI is the choice of the appropriate imputation model, particularly in applications where missingness is seen in variables of different nature. Sequential imputation or variable-by-variable imputation has been increasingly used in such settings. A major concern with this technique is the potential incoherence of the imputations with the underlying posterior predictive distribution. In this talk, I consider a calibration technique to improve the coherence of the sequential imputation. Calibration is a technique designed to preserve marginal distributions in the imputed data. Specifically, we impose a misspecified imputation model such as multivariate normal distribution and correct the adverse impact of this misspecification via calibration.