PLACE:

SPEAKER:
Pierre Duchesne
University of Montreal

TITLE:
On robustness in survey sampling and the use of calibration estimators

ABSTRACT:

We begin in the first part of the talk by giving an overview of robustness in survey sampling.  In economic surveys, populations are often skewed, and extreme or influential units may occur.  In an ideal situation, the sampling design can control the problem of extreme units.  However, since this is not always possible, the use of robust estimators (M-estimators, GM-estimators, etc) became a possibility to obtain estimators less sensible to influential units. We will briefly review the pioneering work of Chambers (1986) and
others to explain the framework.

We will consider the use of calibration estimators in that context. These estimators are popular and used in important statistical agency, as for example Statistics Canada.  They are easy to interpret, since methodologists give a sense to the weights attached to the sampling units.

However, usual calibration estimators are not robust.  For example, the general regression estimator (GREG) is an important case, but it depends essentially on the generalised least squares estimator.  It is well known that least squares estimators can be affected by outliers.  We consider the construction of robust calibration
estimators.  Our approach consists essentially of a judicious choice of implicit robust weights, that we transform into calibrated weights. They are easy to interpret and to explain.  Thus we obtain robust calibration estimators for which both calibration constraints and range constraints (if desired) on the weights are satisfied.

COFFEE: 3:45 p.m., 104 Snedecor Hall