ACS population substantially higher than decennial census

Hi all - 

ACS population estimates for Flint, MI, and other cities, are much higher than the actual populations are likely to be - see the charts below. I'm concerned about whether the ACS estimates for other variables and for subpopulations might be impacted by what is likely not a uniform pattern of difference across these variables. I'd appreciate any thoughts about this and how to respond to it, particularly in interpreting ACS data from the past decade.

Of course COVID probably played some role in data completeness in the census. The Flint water crisis anecdotally led to an increase in the outflow of people leaving and a decrease in the influx of new residents.. However, that only partially responsible for the difference with the ACS numbers. Flint's decrease in decadal numbers matches the very strong linear trend over the previous 50 years almost perfectly, while after 2011 the ACS estimates decreased more slowly. Also, a similar pattern exists for similar nearby cities, but not for the state as a whole.

Has anyone observed a similar pattern elsewhere?  Do you have any data to describe subpopulations - in terms of race, income, family characteristics, etc. - and how much their numbers declined?

The top chart also includes the State of Michigan's annual population estimate   It's interesting to note that it shows an almost identical trend to the ACS.

Finally, the data strongly suggest a change in the ACS population estimate methodology.  Does anyone have information about any changes?

Thanks so much - 


  • Hello Jon -

    That's an excellent question. ACS data are benchmarked to the Census Bureau's population estimates for a given year, and those estimates are benchmarked to the most recent census... so the last year of a decade can often differ quite a bit from the decennial count.

    And yes, the differences will likely not be consistent across age, race/ethnicity, gender, etc... In this case, for example, I would expect the notable differences in multiracial population given what we've seen in national and state 2020 Census results.

    There are a couple of common ways to deal with this in the short term (each with pros and cons):

    Option 1: Use ACS, and note the break-in-series at 2020. (This is not a preferred approach, but is often used in the period between when the decennial is released and when revised historical estimates are available, see Option 2 below.)

    Option 2: Create an independent set of revised intercensal estimates (or work with a known set--the State of Michigan's demographic team will likely be a useful resource,5885,7-339-73970_2944_5325---,00.html). This is the preferred approach BUT note that the ACS data itself will not be re-benchmarked. So if you go with revised estimates, you should not report count estimates from ACS side-by-side with those revised estimates. (It is ok to report revised population counts from a new set of independent estimates alongside *rates* or *percentages* from ACS.)

    Good luck!

  • My understanding (and I am not a demographer unlike Beth) is that the Census population projections do miss the local population dynamics when it is caused by very localized effects. I remember a presentation at one of the ACS conferences that showed how the projections, and hence the ACS estimates, missed the inflows of temporary labor for a dam construction project, and for fracking. There are ways to adjust the demographic estimates to death records, so the (total) mortality from COVID would eventually be accounted for, but there is little way to adjust them for migration within the U.S. (until it is too late as the numbers have to come from the ACS migration questions themselves) unless you try to track every person like some communist country (I was born and raised in USSR... which folks may or may not remember existing... where everybody had to be registered with the local government, local vital records and local police).

    FWIW revisions to the published estimates do happen from time to time in the practice of official statistics. If you envision that better data will come with the decennial census and the ACS that are pegged to it (after 2022), maybe you can give yourself some slack, publish estimates as preliminary, and come back to the issue in a couple of years. (Note that the 2020 ACS is being released as "experimental" meaning that the data qualify suffered as the result of COVID, and they were not able to achieve the response rates that would get the stamp of quality.)