Calibration Weighting: What We Know Now, What We Still Need to Know

• Activity Number: 207
• Type: Roundtables
• Date/Time: Monday, August 5, 2013 : 12:30 PM to 1:50 PM
• Sponsor: Survey Research Methods Section
• Author(s): Phil Kott*+
• Companies: RTI International
• Keywords: selection bias ; mean squared error ; double protection ; asymptotic ; outcome model

Abstract

Deville and Särndal (JASA 1982) coined the term "calibration weighting" to develop and formalize an approach to sample-weight adjustment within a probability-sampling context. The initial goal of calibration weighting was to increase the efficiency of estimated means and totals for outcome variables roughly correlated to one or more of a set of calibration variables with known population totals. Since then, calibration weighting has also been used to reduce or remove selection biases due to unit nonresponse and/or frame errors. In fact, the technique can sometimes be used when nonrespondents are not missing at random, that is to say, when the probability of a unit responding is a function of outcome variables with known values only for the respondents. We will discuss some of the well-known and lesser-known properties of calibration weighting, some recent new developments (and software), and some of the limitation of a technique that ultimately requires respondent samples to be "sufficiently large."