Testing the Effectiveness of Outlier Detecting Methods in Property Classification

Abstract

The introduction of the property value tax in Poland may lead to an increase in the tax burden on real estate. Pilot studies may be carried out on samples and the results should feature a high degree of certainty as to the extrapolation of the results on populations (e.g. entire municipalities). Each study may, for various reasons, include outliers in the analyzed data sets. If their presence results from measurement errors or other reasons that cause such observations not to be the result of naturally occurring processes, they should be omitted in the calculations, because they interfere with the study of the occurring regularities. The study presents the results of statistical modelling carried out to determine whether individual objects (land properties), due to their attributes, are at risk of increasing the tax burden as a result of the introduction of ad valorem tax. First, logistic regression model estimation was carried out for the entire set of analyzed properties. Next, several methods of outlier detection were applied, and model estimation was repeated without the observations, i.e. real estates, pointed out as abnormal. The objective of the study is to verify the usefulness of outlier detecting methods in the context of improving the classification results of the analyzed properties.