Is it really enough to adjust the survey estimates with sampling weights to compare them with census estimates? Please suggest any good literature on this.
Sayantani -
Perhaps you are referring to "raking?" You might be interested in "calibration" estimation for surveys.
The following link may be of interest, and/or some of the references starting on page 26:
https://www.researchgate.net/publication/261508465_Use_of_Ratios_for_Estimation_of_Official_Statistics_at_a_Statistical_Agency
Cheers - Jim
Thank you so much. I will go through and let you know if I still have doubts.
If you use appropiate weights, you can estimate totals (with the corresponding margins of error), as long you have a random sample. A good book to read about survey estimation is Lohr's: "Sampling: design and analysis" (2009).
Relatedly, FYI, if you have a less frequent census survey for given data items, and more frequent sample surveys for those same data items, where you can use linear regression with x from the census, and a sample of the cases as y from the sample survey, then you can predict for the missing cases in the sample survey based on the, say "growth rate" (change between the period of the census, and the current sample, for a given "estimation group"). The link that follows shows how this is done for official statistics at the US Energy Information Administration (EIA):
https://www.researchgate.net/publication/319914742_Quasi-Cutoff_Sampling_and_the_Classical_Ratio_Estimator_-_Application_to_Establishment_Surveys_for_Official_Statistics_at_the_US_Energy_Information_Administration_-_Historical_Development
Weights and imputation are two different ways of handling the same situation.
If the sample was to be from the given census for some reason, with data recollected, then any difference here would be applied to the whole census and give you a slightly different result. You could divide by that change factor to get the same thing. Actually there could be many such change factors due to looking at each "estimation group" separately, where those are like strata where one regression model applies to each.
- Badges
- Science topic
- Similar topics
- Statistics
- Survey