“No” Is the Easiest Answer: Using Calibration to Assess Nonignorable Nonresponse in the 2002 Census of Agriculture

The 2002 Census of Agriculture adjusted for whole-farm nonresponse by dividing the potential farms on its list into size classes and then weighting the respondents within each class to account for the nonrespondents. Unfortunately, to assign the size-class memberships, a consistent measure of size was needed for all potential farms not just census respondents. Subsequently-collected census information sometimes contradicts these class assignments. By defining indicator variables for census respondents based on “corrected” class assignments, instrumental-variable calibration can be used to construct an alternative set of nonresponse weights. Assuming the response model underpinning these new weights is correct, the bias from using the original set of nonresponse weights can be assessed. The potential bias in the estimated farm count is caused by nonfarms on the list being more likely to respond to the census than farms.