137 – Calibration
Combining the Nonresponse and Calibration Adjustments in Weighting
Stephen Ash
U.S. Census Bureau
There are two basic methods for the adjustment of non-interviews. In both you adjust the interviews for the non-interviews by applying a factor that is sometimes described as the inverse of the probability of completing an interview. One way to calculate the factor is directly as the ratio of the weighted count of eligible units (both completed interviews and non-interviews) with the weighted count of completed interviews. This is often done within cells to reduce bias, where cells are groups of similar units. A different method is to model the probability of completing an interview using a general linear model. One problem with a model is that the results may not be tied to the weighted totals of eligible units that includes both the completed interviews and non-interviews. That is, the resultant probability may produce weighted counts that are less than or greater than the weighted total of eligible units. This paper discusses using a logistic regression model that is constrained: either to the weighted totals of the eligible units, totals of the variables used in the model, or totals of