On the treatment of influential observations in household surveys / Jean-François Beaumont and Asma Alavi.: CS11-619/2001-2E-PDF

"Household expenditure or income surveys often deal with highly skewed distributions, which potentially lead to samples with some extreme observations. The problem is aggravated by the fact that there usually is a low amount of useful auxiliary information available at the design stage and that the sampling design is complex most of the time, leading to widely dispersed design weights. Therefore, it could happen that a large value be associated with a large design weight and that this combination have a great influence on the estimates produced by the survey. Design consistent estimators, such as the Generalized REGression (GREG) estimator, are usually highly variable in the presence of influential observations but they have a low bias whereas model-based estimators are more stable but they are generally not consistent and more biased. In this paper, a compromise between these two types of estimators is proposed and a simulation study shows that it performs well with respect to the bias and mean squared error (MSE) criteria in comparison with some other robust estimators. Conditions under which the compromise should have a small design bias are also given"--Abstract.

Permanent link to this Catalogue record:
publications.gc.ca/pub?id=9.841601&sl=0

• MARC XML format
• MARC HTML format

Request alternate formats