Imputation in Experimental Results

Missing values in experimental results are a frequent occurrence, and cause problems because statistical methods assume that the same variables are determined for all samples. The aim of this article is to evaluate the performance of four methods of data imputation: mean, autoencoder, clustering, and c-means. The tests were made on a database of 140 samples of archaeological ceramic fragments in which mass fractions were determined by instrumental neutron activation analysis. The tests showed differences in the Normalized Root Mean Squared Error values, however, data imputation had no impact on the determination of groups by the clustering algorithms.