Model Free Multiple Imputation for Missing Data

• In Event: Missing Data and Data Augmentation

Abstract

Advances in methods for imputing missing data have led to substantial improvements in the validity of statistical inference in the presence of missing data. An important drawback of currently existing multiple imputation methods, however, is that they require the researcher to know a great deal of information about patterns of missingness in the data. As a result, coefficient point estimates and standard errors of regression models estimated under different assumptions about patterns of missingness can vary wildly, casting doubt on the choice of imputation model and tuning parameters used by the researcher. In this paper we introduce a model free, unsupervised learning algorithm for multiple imputation which employs a low-rank approximation theorem derived by Eckhart and Young (1936). Via an analytical proof and simulation studies, we demonstrate that our method improves upon state of the art missing data imputation methods with respect to both the bias and efficiency of parameter estimates in OLS, fixed effects and random effects models and across datasets with different patterns of missingness.

Authors

• Jason Anastasopoulos, University of Georgia

• Sarah Hunter, University of Colorado, Boulder

• Keith T. Poole, University of Georgia