Correlation Analysis in Contaminated Data by Singular Spectrum Analysis

Correlation analysis is one of the standard and most informative descriptive statistical tools when studying relationships between variables in bivariate and multivariate data. However, when data is contaminated with outlying observations, the standard Pearson correlation might be misleading and result in erroneous outcomes. In this paper, we propose three new approaches to find linear correlation based on the nonparametric method designed to analyse time series data, the singular spectrum analysis. In these proposals, the correlation is obtained after removing the noise from the data by using singular spectrum analysis based methods. The usefulness of our proposals in contaminated data is assessed by Monte Carlo simulation with different schemes of contamination, and with applications to real data on aluminium industry and synthetic sparse data. In addition, the model comparisons are made with robust hybrid filtering methods.