Self-Organising Map for Data Imputation and Correction in Surveys

This paper is dedicated to erroneous data detection and imputation methods in surveys. We describe experiments conducted under the scope of a European project for studying new statistical methods based on neural networks. We show that the self-organising map can be used successfully for these tasks. A self-organising map is calibrated according to the available observations, described through a set of correlated variables handled together. The map can then be used both to detect erroneous data and to impute values to partial observations. We apply these principles to a real size transport survey database. We show that the performance of our imputation model compares well to other classical methods, and that the use of a self-organising map for data correction provides a performing system fordata validation, data correction and data analysis.