Impact of Missing Data Imputation on The Structure of Real-World Clinical Datasets: A Comparison of Median, K-Nearest Neighbours, and MICE Approaches
Year: 2026
Type: preprint
Abstract: In clinical research, median substitution remains widely used for handling missing data despite its known tendency to compress variance and distort distributions. We compared median imputation, k-nearest neighbours (KNN), and multiple imputation by chained equations with predictive mean matching (MICE/PMM) on two independent clinical cohorts: a lung transplantation registry (n = 408; 52 continuous variables; 0.2–77.9% missingness) and an acute myocardial infarction dataset (n = 1700; 18 continuous variables; 0.24–63.29% missingness). Distributional fidelity was assessed using the Kolmogorov–Smirnov D-statistic, and multivariate structural preservation was quantified through correlation matrix comparison. KNN achieved the lowest mean KS statistics in both cohorts (0.036 and 0.032) compared with median imputation (0.076 and 0.074) and MICE (0.111 and 0.123). The proportion of variables with meaningful distortion (D > 0.10) was 9.6–11.1% for KNN, 22.2–25.0% for median, and 34.6–38.9% for MICE. KNN also showed the shallowest increase in distortion with increasing missingness. Multivariate correlations were well preserved by KNN and median imputation but substantially altered by MICE. KNN preserved the statistical structure of clinical data more closely than median substitution, with performance comparable to MICE, while requiring considerably less methodological specification.
Source: Research Square
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204123379