Leveraging variational autoencoders for multiple data imputation

TL;DR

This paper explores using variational autoencoders for multiple data imputation, highlighting their limitations and proposing $eta$-VAEs with cross-validation to improve uncertainty calibration and reduce false discoveries.

Contribution

It introduces the use of $eta$-VAEs for data imputation, addressing the calibration issues of standard VAEs and demonstrating improved downstream task reliability.

Findings

VAEs show poor empirical coverage for missing data.

$eta$-VAEs improve uncertainty calibration.

Proper $eta$ selection via cross-validation enhances imputation quality.

Abstract

Missing data persists as a major barrier to data analysis across numerous applications. Recently, deep generative models have been used for imputation of missing data, motivated by their ability to capture highly non-linear and complex relationships in the data. In this work, we investigate the ability of deep models, namely variational autoencoders (VAEs), to account for uncertainty in missing data through multiple imputation strategies. We find that VAEs provide poor empirical coverage of missing data, with underestimation and overconfident imputations, particularly for more extreme missing data values. To overcome this, we employ $\beta$-VAEs, which viewed from a generalized Bayes framework, provide robustness to model misspecification. Assigning a good value of $\beta$ is critical for uncertainty calibration and we demonstrate how this can be achieved using cross-validation. In downstream tasks, we show how multiple imputation with $\beta$-VAEs can avoid false discoveries that arise as artefacts of imputation.