It works like this:


Census ACS three-year estimates are period-based rather than representing a point in time, with a small number of samples drawn each month over a 36-month period.  Of course, when you have samples, they must be weighted to represent a control total.  That control total is the middle year of the period.  So far, Census has released one-year data for 2005 (households only, no group quarters), 2006, 2007, and the 3-year set for 2005-2007.  There is no ACS sample survey data yet for 2008 – only the annual control totals for each county, the Annual Estimates.  The 2008 ACS sample survey data will come out this fall.  Since there is no group quarters data for 2005, the three-year set has its GQ data from 2006 and 2007 weighted higher to account for the missing year. 


Here’s another complication: every year, Census revises the post-censal annual estimates/control totals for each previous year going back to the most recent decennial census, but not the ACS datasets controlled to them.  Example: the 2006 ACS sample data is controlled to the 2006 Annual Estimate, both published in 2007.  That 2006 Annual Estimate was updated and changed in 2008, while the 2006 ACS data remained the same.  I assume the 2005-2007 ACS 3-year dataset (published in 2008) will have been controlled to the 2008 version of the 2006 Annual Estimate, not the original 2007 version.  But I haven’t actually checked that.  If you want consistency between the ACS datasets and the constantly changing Annual Estimate control totals, you will have to adjust the ACS figures yourself.  What it all comes down to, is that the ACS figures are best used for data on proportions (e.g., percent SOV commuters) rather than numbers (e.g., number of SOV commuters).


Since the ACS data is a sample of only about 2.3%-2.5% (depending on how you want to count non-respondents), there is a pretty high margin of error for small universes.  If you use the 3-year dataset, it has the equivalent of about a 7%-7.5% sample size if you assume a stable population (which of course is never really true), so it offers a smaller margin of error in exchange for less timeliness.  There are a series of published studies on the Census website addressing the pros and cons of the shift from the decennial long form point-in-time data to the annual ACS period data.



Pete Swensson, Senior Planner

Thurston Regional Planning Council

2424 Heritage Ct. SW

Olympia, WA 98502

(360) 741-2530 (direct line)

(360) 956-7575 (main desk)

(360) 956-7815 (fax)


This e-mail and any attachments are for the use of the addressed individual.  If you have received this e-mail in error, please notify our systems manager.  TRPC has taken responsible precautions to ensure no viruses are present in this e-mail, however we do not accept responsibility for loss or damage arising from the use of this e-mail or attachments.


From: [] On Behalf Of Frank Lenk
Sent: Tuesday, April 21, 2009 9:10 AM
Subject: RE: [CTPP] Comparison of ACS 1-year, ACS 3-year & Annual Estimates


After more investigation, it appears that the 3-year estimates are, in fact, controlled to a moving average centered on the middle year. But you have to use the right 3-year estimates. When we did this with the most recent estimates (which now go to 2008), the three-year average centered on 2006 did NOT match, but when we looked at the historical release of 2007 estimates, and used those to calculate an 3-year average, the result DID match the 3-year ACS total.


I still find it confusing that the characteristics are NOT averaged, but the totals are?????




From: [] On Behalf Of Michael Cline
Sent: Tuesday, April 21, 2009 8:25 AM
Subject: RE: [CTPP] Comparison of ACS 1-year, ACS 3-year & Annual Estimates


Someone from the ACS discussed this in January at TRB (at the Census subcommittee meeting).  Unfortunately, I cannot find my notes, but I seem to recall that the 3-year estimates use a middle population estimate (i.e. 2006). 




Michael E. Cline
Research Associate
Institute for Demographic & Socioeconomic Research
The University of Texas at San Antonio
1 UTSA Circle
JPL 4.03.18A
San Antonio, TX 78249-0704

(210)458-6537 f(210)458-6541



From: [] On Behalf Of Frank Lenk
Sent: Monday, April 20, 2009 11:26 PM
Subject: [CTPP] Comparison of ACS 1-year, ACS 3-year & Annual Estimates


The attached spreadsheet was prepared by my staff, showing the comparison in total county-level population between the 2007 ACS 1-year estimates, the 2007 ACS 3-year estimates and the Census Bureau’s July 1, 2007 population estimates by county.  The latter are supposed to be the official population estimates to which ACS is controlled.  And, based on the attached spreadsheet, this appears to be true for (most) counties in the 1-year estimates.  But the total population in the 3-year ACS estimates is systematically biased downwards from the total population in the  1-year ACS estimates and/or the official estimates. 


Does anyone have a good idea why?


There is some vague language about differences in weighting in the Census Bureau’s documentation, but I can’t find a satisfying explanation. I do notice that the faster a county is growing the bigger the discrepancy between the 3-year and 1-year estimates of total population. This suggests that the 3-year estimates are being controlled to an average of the 3 years of official total population estimates (2005, 2006 and 2007).  But my understanding is that the3- year ACS estimates are not averaged.  Instead, they a represent a single sample taken over a 3-year period.  My expectation, then, is that this sample would be expanded to the same population as the 1-year estimates – The 3-year and 1-year estimates are, after all, identified by the same year (2007) while a 3-year estimate based on a 3-year moving average would be closer to 2006’s 1-year estimate. 


Any help in clarifying this issue would be greatly appreciated.






Frank Lenk

Director of Research Services

Mid-America Regional Council

600 Broadway, Suite 200

Kansas City, MO 64105