THE SCALING OF DISPERSION AND CORRELATION - A COMPARISON OF LEAST-SQUARES AND ABSOLUTE-DEVIATION STATISTICS

In conventional measures of dispersion and correlation, the use of squ ared deviations from the mean causes a disproportionate influence of o utliers. We have compared various conventional and unconventional meas ures of dispersion and correlation through repeated computer trials, t aking large numbers of samples from populations of specified distribut ions. Under the normal distribution, the standard deviation and Pearso n correlation provide the most stable estimates of dispersion and corr elation. However, in other cases involving outliers within normal dist ributions, double exponential distributions, or skewed distributions, measures of dispersion and correlation involving absolute deviations f rom the mean are clearly more stable than are conventional formulae.