Nonparametric Density Estimation Using Complex Survey Data

                          Trent D. Buskirk
                       Arizona State University

Abstract

In survey sampling, the finite population from which the
data are collected using the complex survey design can
be treated as one realization from an infinite super-population 
which itself is generated according to a particular
distribution (or density) function. In this case, it is 
generally assumed that the underlying distribution function
is continuous. Because the finite population (and hence
any data collected from it) will be at best discrete, and
because interest is given to estimating a continuous density 
function, interpolation methods which can smooth
the data are desired.

In this talk we present a new method for nonparametric
density and function estimation using complex survey
data. The methods are evaluated using both a design-based 
approach incorporating weights from the survey
design and a model-based superpopulation approach.
These methods extend the standard kernel smoothing
techniques which are currently used for nonparametric
regression and density estimation. The method will be
illustrated using data obtained from the National Crime
Victimization Survey.