Matthias Templ (FHNW) presented new work on robust imputation in August 2026, at the 9th
International Conference on Econometrics and Statistics
(EcoSta 2026), held 8–10 August at Ryukoku University
in Kyoto. The talk, “Imputation under cellwise contamination,” extends
VIM, the R package for visualizing and imputing missing
data that Templ co-maintains, to a harder problem: data where the outliers themselves are hidden
inside otherwise ordinary-looking rows.
Cellwise contamination is the more realistic cousin of the classical outlier problem: instead of a handful of entire records being corrupted, individual cells are, scattered across otherwise clean data — a mistyped lab value here, a faulty sensor reading there. The distinction sounds minor, but the arithmetic is not: at a modest 5 % per-cell contamination rate and 20 variables, about 64 % of all rows end up containing at least one bad cell. Any method that responds by downweighting or discarding whole rows (the standard robust-statistics playbook) throws away two-thirds of the sample to deal with a much smaller amount of actually-bad data.
The talk’s organizing idea is a reframe: a cellwise outlier is a missing value that hides. A
detector can flag suspicious cells and hand them to an imputation model as if they were NA — but
the detector’s own guess about which cells are contaminated is itself uncertain, and a hard
flag-then-delete pipeline discards that uncertainty instead of carrying it into the final result.
Templ’s alternative keeps the flag soft: every cell gets a weight between 0 and 1 from its residual
against a model fit on the other variables, with the weights and the imputations estimated jointly
rather than in two disconnected steps. Downweight, don’t delete.
Three estimators built on that idea now ship in the development version of VIM:
imputeCellIRMI() chains cell-weighted regressions across variables; imputeCellM() fits one
robust M-regression per variable and is provably consistent under the cellwise-contamination
model (the one Templ recommends in practice, with full convergence across the benchmark); and
imputeCellEM() treats contamination as a latent state estimated by an EM algorithm, which works
well at low contamination levels but grows unreliable past about 10 %. All three handle mixed
continuous and categorical data, where, perhaps counterintuitively, the untouched categorical
variables turn out to be useful anchors: they help unmask continuous outliers that would otherwise
disguise themselves by shifting together with correlated neighbors.
In simulation, at 20 variables and 20 % contamination, imputeCellM()’s regression estimates
landed closer to the true parameters than four competing pipelines, including the widely used
mice package. The method was also tested on a real dataset: 300 horses, 10 mixed-type variables,
a well-known missing-data problem in veterinary records. It flagged 123 of 3,000 cells as
contaminated, spread thinly across 110 rows at roughly one flagged cell each — exactly the
scattered pattern the approach is built for, rather than a handful of badly broken rows. In a
logistic-regression fit on the same data, a predictor (respiratory rate) that lost statistical
significance under the conventional flag-and-delete pipeline stayed significant using the new
method.
Templ was equally candid about where the new methods fall short. imputeCellEM()’s
convergence turned out to be fragile at higher contamination levels, and mice, a much simpler
existing method, proved a surprisingly strong baseline throughout. Robustness bounds how much
damage contamination does; it cannot recover information that contamination has already destroyed.
A fourth, more advanced method is next in line for the package: an MCD-based estimator for mixed
data, imputeCellMCD(), backed by a separate paper currently under review at Statistics and
Computing.
The cellwise-robust extensions are on GitHub now
(statistikat/VIM), with a CRAN release planned; the
approach traces back to Templ, Kowarik and Filzmoser’s 2011 paper on iterative stepwise regression
imputation in Computational Statistics & Data Analysis, and VIM itself has been maintained on
CRAN for well over a decade.