I've been reading lots of commentary about the addition of the citizenship question to the Decennial Census. Instead of focusing on the impact the question may have on response rates, this thread examines the DOJ's arguments for why the question is needed. ht/ @colemanreedharr
I read the DOJ's letter (documentcloud.org/documents/4340…) and the Commerce Secretary's memo (documentcloud.org/documents/4424…). I want to offer a summary of the DOJ's argument and a rebuttal to aspects of that argument.
I offer this summary and rebuttal as someone who has worked with historical, small-area census data for the past 17 years @nhgis.
The DOJ argues that citizen voting-age population (CVAP) is required for census blocks, block groups, counties, and towns to enforce Section 2 of the Voting Rights Act of 1965.
The DOJ correctly states that the citizenship question was asked on the 1970-2000 long form of the Decennial Census, and that the 1970-2000 data on the citizenship was used to enforce Section 2.
In 2010, the Decennial census had no long form questionnaire. The former questions on the long form were now asked as part of the American Community Survey. Redistricting following the 2010 census required entities to blend decennial data on voting age population (VAP)
at the census block level with ACS data on CVAP at the census block group level. The smallest geographic unit in the ACS is the census block group - the sample density isn't high enough to produce reliable block-level estimates.
In their request, the DOJ argued that the block vs. block group discrepancy redistricting entities to "perform further estimates and inject further uncertainty in order to approximate citizen voting-age population at the level of the census block, which is the fundamental
building block of a redistricting plan." In my reading, the DOJ seems to imply that, prior to 2010, CVAP was available for census blocks. Thus, reinstating the the citizenship question allows the Census Bureau to again produce block-level CVAP counts.
The history of small area census data rebuts this implication. Census block-level CVAP counts have never been publicly available from the Census Bureau. In 1970 and 1980, the smallest geographic unit for which we have citizenship counts is the census tract.
In 1990 and 2000, the smallest geographic unit for which we have citizenship data is the census block group. Why is this the case? The long form in 1970-2000 was sent to 1 in 6 households in the US. This sample density was not
large enough to produce block-level estimates. The Bureau could produce tract and block group-level estimates from the long form. There have never been block-level CVAP counts published by the Census Bureau.
Thus, redistricting entities have always had to allocate CVAP from larger units to smaller units, and the DOJ has always used block-group or tract level data to enforce Section 2 of the VRA.
This isn't just because of the introduction of the ACS and the elimination of the long form. Citizenship data at the block level can't be computed reliably from a sample of households.
(Aside - census block data for the entire country was first published in 1990. That was the first year in which the entire country was subdivided into census blocks. So, for the first 25 years of the VRA, block level data on VAP was restricted to cities).
The DOJ also argued that uncertainty (as reported by the margin of error that accompanies every ACS estimate) in ACS estimates is large for small geographic units. The decennial census goes to every household and would then have less uncertainty.
By placing the citizenship question on the decennial questionnaire that goes to all households, the uncertainty would be less than the ACS. But, the long form data that was used from 1970-2000 was also uncertain! Those data were produced from a 1 in 6 household sample and not
a full enumeration of households. We can compute standard errors from sample data to get at the statistical accuracy of the estimate. The Census Bureau's technical documentation provides instructions for standard error computation (see
chapter 8 of nhgis.org/sites/www.nhgi… for more details). But, the Bureau didn't actually compute and publish the SEs for the long-form estimates. Data users had to do the computations themselves. Based on my experience, few users ever did.
In order to truly assess whether ACS margins of error are worse than long-form based standard errors, we would need to compute the standard errors and compare them to ACS margins of error.
I cannot tell if this was done. Without this analysis (which someone should do), I am unwilling to accept the argument that ACS block group estimates are worse than long-form block group estimates of the past.
To recap, the DOJ is requesting data (census block CVAP counts) that has never before been published. The ACS data is too uncertain to be used to enforce Section 2, but prior decennial data on CVAP was also based on sample data that was also uncertain.
The DOJ has used uncertain CVAP data to enforce Section 2 of the VRA since the early 1970s. How would block level CVAP help them better enforce Section 2? Their request provides no answers to this question. I would love to see more discussion of this question.
