Principled methods for analyzing missing values, based chiefly on multiple imputation, have become increasingly popular yet can struggle to handle the kinds of large and complex data that are also becoming common. We propose an accurate, fast, and scalable approach to multiple imputation, which we call MIDAS (Multiple Imputation with Denoising Autoencoders). MIDAS employs a class of unsupervised neural networks known as denoising autoencoders, which are designed to reduce dimensionality by corrupting and attempting to reconstruct a subset of data. We repurpose denoising autoencoders for multiple imputation by treating missing values as an additional portion of corrupted data and drawing imputations from a model trained to minimize the reconstruction error on the originally observed portion. Systematic tests on simulated as well as real social science data, together with an applied example involving a large-scale electoral survey, illustrate MIDAS’s accuracy and efficiency across a range of settings. We provide open-source software for implementing MIDAS.

This letter compares the performance of multiple imputation and listwise deletion using a simulation approach. The focus is on data that are “missing not at random” (MNAR), in which case both multiple imputation and listwise deletion are known to be biased. In these simulations, multiple imputation yields results that are frequently more biased, less efficient, and with worse coverage than listwise deletion when data are MNAR. This is the case even with very strong correlations between fully observed variables and variables with missing values, such that the data are very nearly “missing at random.” These results recommend caution when comparing the results from multiple imputation and listwise deletion, when the true data generating process is unknown.