In survey weighting, I've seen that first a design weight adjustment is made (to reflect over- or under-sampling), and then a non-response adjustment is made (to reflect groups that are less likely to respond), and then finally to adjust for non-coverage "raking" or iterative proportional fitting is used.
The seed values for raking are often the weighted-persons after the design and non-response adjustments. If I use all 1s, for example, I will get very different results.
Numerous previous questions here have noted that the odds ratios in the initial seed weights are maintained in the final weights:
Iterative Proportional fitting starting values
Iterative proportional fitting in R
Why does maintaining the odds ratios mean that the final values adjust for all three sources of error, rather than just non-coverage?
A complementary question is... why do we even do this in stages? Why not just throw raking at the original number of respondents rather than adjusting for non-response and survey design first?
