Influence Function and Its Application to Data Validation

摘要

Hampel's influence function has been used by Devlin, Gnanadesikan, and Kettenring to detect bivariate observations which have unusual influence on estimates of correlation. In the validation of energy data systems such observations may sometimes be considered to be outliers. The identification of such outliers may be valuable for the detection of errors in a data base. When data are used in regression equations, those observations which have the greatest effect on the multiple correlation coefficient or the regression coefficients are of interest. The contours of constant influence are derived for the multiple correlation coefficient in the case of regressing two variables on a third. In some problems the analytic form of the influence function may be difficult to derive. In such cases the empiric estimator of the influence function, as proposed by Mallows, may be useful for detecting outliers. For FPC form 4 power plant data, the correlation between generation and consumption is a parameter of interest to users of the data. Estimates of the contours of constant influence were determined and used to detect outliers with respect to bivariate correlation. (ERA citation 04:050920)