A2045
Title: Imputation under cellwise contamination
Authors: Matthias Templ - University of Applied Sciences and Arts Northwestern Switzerland (Switzerland) [presenting]
Abstract: Cellwise contamination, in which outliers strike individual cells rather than whole observations, is pervasive in survey microdata, medical records and administrative registers. Even at a modest per-cell rate, most rows of a wide dataset contain at least one corrupted entry, so casewise-robust methods and listwise deletion fail; and when values are also missing, standard imputation propagates the contamination. Integrated procedures that jointly flag deviating cells and estimate model parameters exist but assume fully continuous, complete data; two-step detect-then-impute pipelines discard information and threshold hard at the boundary. A family of integrated cellwise-robust imputation methods is developed for mixed continuous and categorical data with missing values. Cell weights from robust conditional residuals enter an iteratively reweighted least-squares fit at the cell level. One variant performs per-variable M-regression and is provably consistent for the clean parameters under the independent cellwise model; another extends the cellwise minimum covariance determinant estimator, combining multivariate cell detection with cellwise-robust regression. Multivariate detection outperforms per-variable residuals when contamination is correlated within blocks of variables. Because cellwise breakdown points are intrinsically low, the methods complement casewise approaches; all are available in the R package VIM.
