THE BEHAVIOR OF ROBUST ESTIMATORS ON DEPENDENT DATA,

Abstract

The report studies the effect of serial dependence on the efficiency of various robust estimators of the location parameter. In order to show that the asymptotic distribution of these estimators is a normal distribution a slightly stronger mixing condition than Rosenblatt's strong mixing is introduced and it is shown that the empiric c.d.f. formed from such a process approaches a Gaussian process. In particular, first order autoregressive processes with Gaussian, Cauchy and double-exponential marginal distributions are shown to obey the conditions. The behavior of robust estimators on Gaussian processes is studied in greater detail. One general result states that for any Gaussian process with serial correlation (rho sub k) > 0 and summation (rho sub k) < infinity, the efficiency of any linear combination of the order statistics relative to the sample mean is greater than its efficiency in the case of independent observations. The same result holds for the Hodges-Lehmann estimator. These results are applied to two models of contamination and show that the estimators which have been developed to be robust against outliers are robust against dependence. (Author)

Document Details

Document Type
Technical Report
Publication Date
Aug 01, 1969
Accession Number
AD0710751

Entities

People

  • Herman Rubin
  • Joseph L. Gastwirth

Organizations

  • Purdue University

Tags

Communities of Interest

  • Materials and Manufacturing Processes

DTIC Thesaurus Topics

  • Data Science
  • Efficiency
  • Estimators
  • Gaussian Processes
  • Information Science
  • Normal Distribution
  • Order Statistics
  • Statistical Analysis
  • Statistics

Fields of Study

  • Mathematics

Readers

  • Mathematical Modeling and Probability Theory.
  • Regression Analysis.
  • Statistical inference.