I am attempting to devise a replication-based survey design that closely replicates a taylor-series design used by the US Government publication - where I do not have access to all of the
The US Census Bureau publishes a triennial survey of New York City dwellings called the NYCHVS. In their official documentation, they recommend that analysts use a generalized variance formula, which is obviously a computational nightmare - since everything needs to be calculated by hand. However, they also publish a series of table packages that contain standard errors calculated through a Taylor-Series Linearization method. When I contacted them about how to reproduce their statistics, they said
The appropriate variables for him to use would be borough and segment.
borough is available in the data set, but the variable
segment is not - and they will not release it due to confidentiality concerns.
I am able to precisely reproduce their means and medians - the
hhweight variable is sufficient for that.
I have tried constructing my own replicate weights (just a matrix of
1), but my standard errors either come in far, far too big or basically no different from a
simple random sample (and so far too small).
Obviously, I will never match the US Government's statistics perfectly, but if I've only got one of two
CLUSTER variables, is there a decent way I could calibrate and create a survey design that would get me close to the right answer?