DeepComplete: Generative Adversarial Autoencoders for Multiple Imputation

• In Event: Advances and Challenges in the Measurement of Latent Traits

Abstract

Despite data collection and quality advancements, multiple imputation remains a valuable tool for improving the effectiveness and reliability of statistical estimation. Unfortunately, most strategies for multiple imputation either do not accommodate non-linear data representations or do not work with non-tabular data structures like images and texts, which are increasingly relevant in political science research. This paper introduces a novel methodology for multiple imputations across tabular and non-tabular data to fill this gap. It combines a deep learning variational autoencoder with generative adversarial networks to effectively learn and reconstruct the data-generating process. The algorithm's efficacy is demonstrated across various data structures, including surveys, panel data, time series, text, and image classification tasks. This paper offers a versatile, easy-to-use methodology for diverse data structures implemented for Julia, Python, and R.

Author

• Umberto Mignozzetti, UC San Diego