Missing data is a persistent problem in biostatistics, and imputation—filling in plausible values—is a standard fix. The authors of a new paper propose MissARF, a method that uses adversarial random forests (ARF) for density estimation and data synthesis. For each missing value, the algorithm conditions on the non-missing entries of that observation and samples from the ARF-estimated conditional distribution.
The method is designed to be fast and easy to use, and it natively provides both single and multiple imputation. The authors report that MissARF performs comparably to state-of-the-art imputation approaches in terms of quality, while running quickly and requiring no additional cost for multiple imputation. The paper has been published in Statistics in Medicine and is available on arXiv.