Estimating Scale—Finite Sample Results.pdf

Robust Estimation and Testing by Robert G . Staudte and Simon J. Sheather Copyright © 1990 John Wiley Sons , Inc . C H A P T E R TW O Estimating Scale—Finite Sample Results In this chapter we will review some traditional ideas on scale parameter estimation, and in particular consider equivariant and unbiased estimators of the exponential scale parameter. Our goals are to illustrate the robust- ness or lack thereof of some of the classical estimators. We will introduce mixture models for contamination and the finite sample breakdown point to assist us in our analysis, and finally consider some different methods of estimating the standard error of our estimators, including the bootstrap. The difference between the location parameter and the scale parameter estimation problems is that in the former a shift of the distribution is pos- sible, while in the latter a fractional or multiplicative change is possible. Mathematically, location is modeled by adding the same unknown constant to each observation from a fixed distribution, while scale is modeled by multiplying each observation by an unknown positive constant. The reader should be wary that the sample mean, median, and some other estimators commonly used as location parameter estimators may also be reasonable scale parameter estimators in the scale context, at least when multiplied by suitable constants. The location problem is treated in Chapter 4. 2.1 EXAMPLES To begin we give examples where a scale parameter is of interest. In the first example an exponential model may be appropriate, and many charac- teri