Accession Number : ADP007120

Title :   Robustness of Regression M-Estimators Over Complex-valued Distributions,

Corporate Author : MONTANA UNIV MISSOULA

Personal Author(s) : Ghosh, Krishnendu ; Heiberger, Richard M.

Report Date : 1992

Pagination or Media Count : 4

Abstract : Noisy complex-valued data, for which robust regression techniques are the natural analysis approach, arise in many physical fields. Evaluation of the efficiency of such techniques requires that their behavior be charted over a series of known reference distributions. We have defined several symmetric long-tailed complex distributions (e.g., complex slash, complex Cauchy, complex double exponential) based on complex normal distribution. We have compared via the maximin method the robustness of different regression M-estimators (as defined by their weight functions) over these distributions. The variances of the estimators of the regression coefficients are obtained by simulation over all the distributions and for all the weight functions. The relative efficiencies over each distribution are obtained and then these relative efficiencies are compared over different distributions to identify the best weight function. Three different samples sizes 5, 11 and 15 have been used for this purpose. We apply our estimators to the evaluation of the Magnetotelluric response function.

Descriptors :   *NORMAL DISTRIBUTION, *REGRESSION ANALYSIS, APPROACH, BEHAVIOR, COEFFICIENTS, DISTRIBUTION, EFFICIENCY, FUNCTIONS, RESPONSE, SIMULATION, WEIGHT.

Subject Categories : Statistics and Probability

Distribution Statement : APPROVED FOR PUBLIC RELEASE