Robust Multivariate Methods for Income Data

Abstract

Income inequality and poverty measures are central to the analysis of social welfare. However, recording and measurement errors, outlying observations exert strong influence on non-robust estimators of these measures. If the data cannot be purged of these, welfare conclusions drawn from the data can be seriously misleading. Moreover, these measures are computed on the basis of a univariate income variable, which is an aggregation of several distinct income sources or components. Notably outliers in several income components may severely affect the univariate income variable and thus the estimates. In addition, the aggregation process may propagate or mask outliers in the components. Therefore, instead of focusing on univariate robust estimators, propose to adopt truly multivariate outlier-detection and robust imputation methods. Both, outlier-detection- and imputation methods are adapted for the finite population sampling context and can cope with missing values and the multiple zero-inflation structure of income data. This kind of data.