Robust Correlation Measures

It is well established that the standard measure of correlation (Pearson’s product-moment) is very sensitive to outliers. It can give extremely misleading results when a few or even a single outlier is present. A number of robust correlation measures have been proposed. We do not consider estimators which require trimming (discarding) of some arbitrary fraction of the data, nor those which require iterative computation. Our overall aim is to find a practical and simple robust measure of correlation which can be recommended to practitioners alongside the classic Pearson and Spearman measures. The well known data sets of Anscombe are used to provide an initial assessment of these estimators. These four data sets were designed to have identical Pearson correlation coefficients as well as identical regression lines and other regression statistics. Nevertheless, visual inspection of their scatter-graphs indicates very different patterns. For data set C, there is a perfect linear relationship for all but one of the data points; whereas for data set D, apart from one outlier, all points have the same x-value and so there is essentially no co-variation or interdependence between the variables. We prefer a robust correlation measure to have a near-zero value for set D, and a high value for set C, with the other two data sets giving an intermediate value