Abstract:
|
A classical challenge faced by survey statisticians is how to reduce the impact of certain collected values on the survey estimates. Traditionally, outlier detection methods will focus on the values, or the weighted values, of the variable of interest. However, such approaches ignore the possibility that a typical value may still adversely impact the estimates as a result of the sample design employed, the nature of the parameter to be estimated or the estimator used. Conversely, outlier values may not be influential for a given sample design and estimation method. A more all-encompassing approach was sought, one that would reveal which units have the greatest influence over a given estimate and how exactly it exerts that influence.
To meet the challenge of producing robust estimates, the Survey of Household Spending (SHS) has looked into the notion of conditional bias which has just recently been proposed as a means of gauging a unit's overall influence on estimates. In this paper we describe this innovative approach as well as our results, along with the practical issues one may be facing when implementing this method in a complex survey.
|