Replication Method for Survey Estimation

zli034

Hi all,

What is a good way to use replication methods for estimation? For a dataset with no design information, only weighting remain. I know the basic principles of replication. Are there any considerations to use replication effectively?

There is a public data of 15,000 people from the Census I'm using to estimate its mean and standard error for some variables. The dataset has its survey weighting, but the survey design variables are not released to the public. The clustering and stratification stage of the survey were based on geographic information, I have no where to find them.

I do have GVF to compute variance of some variable estimations. After using some replication methods, my estimations are very different from the official GVF results.

Stephen Tashi

You are using terms unfamiliar to me. By "replication", do you mean "re-sampling" or "bootstrapping"?

What kind of information is the "survey weighting" of the data?

zli034

To do Jackknife, each replication need to be a pseudo-sample has same design as the original sample. We can call replications re-samples, it is also similar to bootstrapping. Because of the survey design, each observation unit in the sample has to be weighted, the data have weights, which is good thing. However the data do not have design variables, which means it is impossible to re-sample like the original sample did.

Need to make re-samples has similar design to the original sample.