Weighting that involves a numeric adjustment variable is often referred to as calibration, although the term calibration can sometimes be used to refer to what has been defined as the process of scaling a weight as well as to rim weighting.
Calibration is useful in the following circumstances:
- When numeric adjustment variables are required, they cannot be used with rim and cell weighting.
- When there is a desire to incorporate a pre-existing design weight into the new weight, such that the new weight is as close to the existing design weight as possible.
- When there is a desire to trim the weight, by specifying the minimum and maximum possible values.
A variety of different algorithms have been developed for calibration. They differ in terms of how they calculate the difference between the pre-existing design weight and the new weight. Where there is no design weight, one is created where the same value is assigned to all observations. Most commonly this weight is 1.
The most common algorithms are raking (which is so-called because it gets the same result as traditional raking when applied to categorical data), linear, and logistic. For more detail on algorithms, see Lumley, Thomas S. 2010. Complex Surveys: A Guide to Analysis Using R. Wiley Publishing.
There is no theory to suggest when to choose a particular algorithm nor when to trim. Further, experience suggests that they tend to have a marginal effect.
For more information, see See Deville, J.-C., and Särndal, C.-E. (1992). Calibration estimators in survey sampling. Journal of the American Statistical Association, 87, 376-382 and Deville, J.-C., Särndal, C.-E., and Sautory, O. (1993). Generalized raking procedures in survey sampling. Journal of the American Statistical Association, 88, 1013-1020.
Please sign in to leave a comment.