Rim weighting
Rim weighting is used when:
▪you want to weight according to various characteristics, but do not know the relationship of the intersection of those characteristics, or
▪you do not have enough respondents to fill all the possible cells of the table if you were to weight the data using the multidimensional technique described above.
For example, you might want to weight by age, sex and marital status and know the weights for each category of those characteristics; for example, people aged 25 to 30, men, single people. However, you might not know the weights for, say, single men aged between 25 and 30, married women aged between 31 and 40, and so on.
On another study, you might need to weight by a large number of characteristics at the same time; for example, sex, age, race, occupation and income. Since each of these characteristics are broken down into categories, you will require a weighting matrix with many cells. You might not have enough information to write a standard multidimensional weighting matrix which defines weights for the intersection of all these characteristics. However, as long as you have information on each category individually (for example, male, female, 21-24, 25-30, and so on) you can perform the weighting required with rim weights.
Rim weighting tries to weight all characteristics at the same time. The accuracy of the weighting depends on how well the sample matches the known universe. If the sample is a good match, it is likely that Quantum can generate acceptable weights; if the sample is not a good match it is possible that the weights look perfectly acceptable when you look at the number of men or the number of married people, but look unacceptable when you look at the number of married men.
As the rim weighting process runs, it tries to distort each variable as little as possible while still trying to attain all the desired proportions among the characteristics. The root mean square figure which Quantum produces indicates how much distortion you have introduced, and therefore how reliable the sample is. The larger the number, the more the distortion and therefore the less accurate the sample is.
Rim weighting also rescales all the target values to the same base. For example, suppose you have a sample of 5,000 respondents. Your rim weighting matrix defines:
▪a weighted total (table base) of 10,000
▪weights for age in percentages
▪weights for sex in target numbers which add up to 758
▪weights for occupation in numbers which add up to 1134.
Quantum calculates the weights for these characteristics, using the figures given, and then adjusts them so that the total for the weighted table is 10,000. If you do not define a total, it adjust the weights to the total of the first variable defined in the matrix.
This example shows that rim weighting can be used when you have weights coming from different sources, and when those weights do not have a common form or total.
See also