I have a question about how calibration weights can be used to sufficiently correct for unit non-response bias. Suppose the sample is s and the response set is r.
Calibration is applied to the response set r to generate weights that are as close as possible to the original sampling weights and meanwhile can reproduce the population totals for the selected auxiliary variables. If the auxiliary variables are (strongly) correlated with unit response probability and the study variables, it will help correct for unit non-response bias and reduce variance. However, I find it is not intuitive to understand this.
Another more intuitive approach: first, estimate the response probability using a logistic regression model, and update the sampling weights by being divided by estimated response probability, which correct for unit non-response bias; second, use calibration to further adjust the updated weights to match known population totals for certain variables, to reduce variance.
Overall, it seems that the calibration approach (the first method) simplifies the process quite a lot, which makes me wonder it might not be the ideal approach when dealing with unit non-response.
Any thoughts or comments would be greatly appreciated! :)
