![IconGems-Print](images/IconGems-Print.png)
Calibration Weighting for Nonresponse with a Flawed but Survey-Corrected Frame Variable
Calibration Weighting for Nonresponse That Is Not Missing at Random: Allowing More Calibration Than Response-Model Variables
Phillip Kott
RTI International
Dan Liao
RTI International
Sometimes in survey sampling we have access to a frame variable that is imperfectly measured. For example, the frame may contain an imperfect indicator of whether a housing unit is owned or rented. Although (we will assume) the error in this variable can be corrected on the survey itself, using a corrected-frame values as a calibration variable will generally bias the resulting estimates. We will show how to avoid that source of bias when adjusting for unit nonresponse through calibration weighting. This can be done by treating the flawed-frame variable as a shadow variable to the corrected-frame variable in the weight-adjustment function. In other words, by calibrating on the flawed version of the variable while assuming, more reasonably, that whether or not a sample unit responds is a function of the corrected version. Since only the respondents are reweighted, we only needed to have corrected versions of the respondents' values in the weight-adjustment function. Some simple simulations will show the effectiveness this weighting approach.