TabSODA Diffusion Model Cuts Imputation Errors by 23.7% on Survey Data

Missing data imputation in large surveys like PATH and NSDUH faces two overlooked challenges: structural skips (cells made inapplicable by questionnaire design) are often incorrectly imputed, and ordinal responses are treated as nominal via one-hot encoding, losing ordering information. To address this, Yuyu Chen and co-authors propose TabSODA (Tabular diffusion with Skip pattern detection and Ordinal Awareness), built on the Elucidated Diffusion Model (EDM). Using an Expectation-Maximization framework, TabSODA propagates structural skips through both the denoising loss and reverse-time sampler. It encodes ordinal variables with cumulative-probit scalar latents while retaining analog-bit encodings for nominal ones. When a codebook skip mask is available it uses it directly; otherwise, the TabSODA+SKIP variant estimates the mask using a CART-based miner that leverages raw responses and questionnaire order.

On the Population Assessment of Tobacco and Health (PATH) study and the National Survey on Drug Use and Health (NSDUH), TabSODA reduces ordinal MACE (mean absolute error for categories) by up to 23.7% and improves categorical accuracy by up to 9% over the strongest baseline across MCAR, MAR, and MNAR masking scenarios. The CART-based skip miner achieves near-perfect precision on both datasets, allowing TabSODA+SKIP to closely track the codebook-mask variant. These results demonstrate that explicitly modeling questionnaire structure and ordinal nature significantly improves imputation quality, with direct implications for public health and policy research relying on survey data.