ACS Data Users Group

Hello all,

Does anyone know of any resources that examine measurement error in the decennial long form? I know the Census did not publish margins of error per se.

Thanks for any assistance.

There's been some attempt at finding noise injected at the block level in the name of privacy just by looking for oddities like people without housing units, children without adults, and Steve Ruggles at IPUMS has been following that pretty closely @HistDem Of course there's no longer a long form census (2000 was the last one) and the decennial census is not a survey so there's no sampling error, but there will be studies later to see if people were undercounted -- some details on that here

They did publish info on the MOE (as they call them). They use the normal approximation for the error term, which frankly does not provide a very good estimate of error. This is also true of the ACS, where they still have MOE's that push counts negative or over 100%. They published some random replicate error estimates, that are done correctly from a frequentist perspective and are generally much smaller than their published MOE's. But in the published documentation they did have some info on the errors, but it is quite primitive actually. If anyone is interested in this please correspond with me off line. We were forced to put up error bands for a project that the Bureau supported with the advent of the 2008-2012 ACS, for 2000, 1990 and 1980. andy@socialexplorer.com

I think the short answer is that the decennial long form survey is not subject to sampling error because it is a true census, ie. it queries every member of the population. Of course the decennial is not actually able to collect data from everyone, but this is due to non-sampling error and not random sampling. My understanding is that non-sampling error can not usually be expressed as a standard deviation or margin or error. The results of the 2020 Census had noise injected in line with differential privacy techniques and that I understand can be modeled and perhaps reported as a margin of error. However, I do not think the the CB has released the material needed to make those calculations, at least not in a straight forward way. Clarifying Tim's remarks, what we call the decennial today is the direct descendant of the last long form census undertaken in 2000, since split into the decennial and the ACS.

Hi Tim and Cliff--

I think you are confusion the 2020 (which has no long form) with 1970, 1980, 1990 and 2000, which did have a long form. The ACS replaced the long form. The long form was a sample, the shot form was a census. They began publishing what they called MOE's when the ACS went live. Up until that point, they had only produced memos about MOEs. I hope that clarifies things.

Now the 2020 (short form) has injected noise. It also has an awful lot of imputation. However, they have so far refused to publish distribute any real band of uncertainty or error. Van Riper of the NHGIS and PPMF fame is trying to compute something based upon the demo product. So far they have said blocks are bad, and anything small maybe bad.

Andy

Andy

You are correct. I should have waited for the caffeine to kick in!

Cliff

I am on my third cup

As I wrote earlier "Of course there's no longer a long form census (2000 was the last one) " I don't think that's a confusion. If you want to say the ACS replaced the long form (with a much smaller sample) and that the ACS publishes a MOE, fine. But the question was about a "decennial long form census" and that no longer exists. I think Cliff and I were right to clarify some of apparent confusion in the question ---

Tim--I think you were not answering the question of the person who asked it. They publish margins on the ACS, but never did on the Long Form. But they did discuss the likely margins of error in general. Actually the long form sampled about one out of six. The total ACS over 5 years is slightly smaller, but they had gotten a sample size bump. The Long Form had a different sampling approach, and in larger areas had a smaller sample proportionately, while in some places it sampled at 50% so that its estimates were reasonably accurate.

The error structure of the new DAS, which makes much of the decennial effectively synthetic, is such that it puts roughly the same actual count error into small counts and large counts. See this recent tweet by Steve Ruggles on Carrolton MS https://twitter.com/HistDem

Exactly. No confusion here!

And Carrollton is exactly what I'm referring to in my original note mentioning Ruggles. Maybe you were interrupted by coffee while reading it, since you're using so much of the same information as if it were new to me. By the way if we're going to suggest that the ACS is the new decennial long form, the Census has actually replaced the one-year estimates for 2020 with experimental estimates, citing counting difficulties in the pandemic -- www.census.gov/.../changes-2020-acs-1-year.html