By Dr. William P. O’Hare
Standard Deviations blog posts represent the views of the author/organization, but not necessarily those of the Census Project.
I have been working with the Census Bureau and Census Bureau data since I started my PhD studies more than 50 years ago. I have written a couple of books on Census data accuracy along with numerous articles and reports on the Census. I have been on Census Bureau advisory committees and I have been a consultant to the Census Bureau.
Given that background, I was shocked by a recent suggestion that the Census Bureau and its long-standing imputation[1] program produce some sort of biased results, based on differences between the 2020 population estimates, and the actual census count.
There are several specific shortcomings in this line of thinking.
This perspective assumes the population estimates are more accurate than the census and should be used to determine how many members of congress each state deserves.
This is backwards. The Census Bureau’s population estimation program has been producing state population estimates every year for more than four decades. It is a partnership between states and the Census Bureau under the Federal-State Cooperative Program for Estimates (FSCPE). The estimates produced in this series are typically not thought of as being as accurate as the Census count. Every decade the Census Bureau, and demographers outside the Census Bureau – like those producing business demographics – use the decennial census count to assess the accuracy of their estimates and recalibrate their methodology based on the census count. It is not unusual for the estimates to differ from the Census count… that is one reason we take a census every ten years. No one should be surprised that the census counts and the population estimates do not match in the 2020 Census. This is not new…it happens every census. We take a census to recalibrate the population estimates series. If the estimates were as accurate as the census, the country could save billions of dollars by not conducting a census.
Some suggest a political motivation for the use of statistical imputation that sometimes produces a higher count in states they disfavor, while ignoring situations that are not consistent with their bias, inasmuch as imputation is used in blue and red states alike, adding more population than anticipated in some of each.[2]
A more reasonable explanation for why Census counts are higher than estimates in 2020 in some states has to do with the fact that the state governments in states (like California, New York and New Jersey) invested millions of dollars in outreach and state leaders did not join the Trump Administration’s effort to discourage immigrants from participating in the Census. That is consistent with the fact that census self-response rates in Hispanic-majority census tracts in Texas and Arizona are several percentage points below the national average while the self-response rates in California’s Hispanic-majority census tracts were several percentage points higher than the national average.[3]
Having worked with the Census Bureau staff for more than 50 years I feel very confident in saying it is the most non-partisan federal agency I know of. Any attempted politicization of the 2020 Census was due to the former administration which installed several unqualified political appointees at the last minute, tried to add a citizenship question to discourage immigrant participation in the Census, and tried to have undocumented immigrants removed from the apportionment count, among other things. Census Bureau professional staff tried their best to resist the partisan pressures and conduct a typical non-partisan census. [4]
Similarly, those who seek to cast doubt about imputation, a widely used statistical technique among data producers, show their own bias. Studies show using imputation makes data more accurate than assuming occupied housing units are vacant which is the alternative when people don’t respond to a survey or the census. [5]
Whole-person imputations are used where there is evidence of a person existing but no information about that person. For example, in the Census some people are included by proxy respondent like a landlord or neighbor, who may say there are 4 people who live in a housing unit but can’t provide any more information. At that point, the Census Bureau has two choices, they can decide to leave those four people out of the Census count, or they can impute them. Studies show that imputing such individuals results in more accurate data. The Census Bureau uses information about the neighbors to impute data for the four-missing people. For example, if 90 percent of the neighbors are Asian, they are likely to impute the race of the four missing people as Asian. That makes sense.
The National Research Council,[6] determined that the 1.2 million imputations in the 2000 Census were problematic but concluded that if they had not been imputed the census “would have undoubtedly underestimated the true number of household resident…” According to Cohn, the U.S. Supreme Court concluded, “Without imputation, the court stated, the results would be ‘a far less accurate assessment of the population’” [7]
In the 2010 Census there were 1.2 million Hispanic people and 1.2 billion Black people imputed compared to 3.1 million Non-Hispanic Whites.[8] A higher number of Non-Hispanic White people were imputed than Black people and Hispanic people.
Another frequent complaint is that counting the undocumented is some sort of partisan maneuver. That claim requires disregarding the U.S. Constitution (Article I and Amendment XIV) which mandates that the Census Bureau count everyone living in the country including undocumented immigrants. That is what the Supreme Court ruled.
I am also struck by facts that anyone claiming partisanship has to ignore in trying to build a case that the Census Bureau is adjusting the data in a partisan way to favor Democrats. The Decennial Census consistently overcounts the Non-Hispanic White population who are more right-leaning politically and undercounts minorities who are more left-leaning politically. In the 2010 Census, there were 6 million Non-Hispanic Whites double-counted while 1.7 million black people and 1.9 million Hispanic people were missed in the census. That is in my book… which perhaps new census critics should consult. [9]
It is not as if I have no complaints about the Census Bureau or concerns about the quality of the 2020 Census data (although the quality metrics I have seen so far, suggest the data from the 2020 Census is not likely to be as bad as some folks think). But the kind of baseless attacks and misrepresentations of the Census and the Census Bureau reflected in partisan attacks needlessly erodes public trust in one of our basic institutions. The Census not only provides the basic data for reapportionment and redistricting, but it also provides key information that allows government programs to provide assistance to communities in an efficient and effective way. And for businesses to make decisions that lead to a more efficient economy. These recent attacks undermine support for public policy decisions based on data, evidence, and science Maybe that is the point!
- Dr. O’Hare, a member of The Census Project Advisory Committee, is a social demographer who has spent forty years using data to increase public understanding of disadvantaged groups. For the past 25 years, he has been involved in the KIDS COUNT project at the Annie E. Casey Foundation. Bill has a Bachelor of Science, Master of Arts, and a Ph.D. from Michigan State University.
[1] Imputation is a statistical technic for estimating people and characteristics that are not reported by respondents. For example, if a respondent does not provide their race when they are responding to the census and 90 percent of the block where they live are Asians, the Census Burau is likely to impute Asian for that respondents’ race. This makes the dataset more accurate than leaving that item blank.
[2] U.S. Census Bureau (2021). A Preliminary Analysis of U.S. and State-Level Results from the 2020 Census, WP-104, U.S. Census Bureau, Washington DC.
[3] O’Hare, W. P. (2021). Measuring the Quality of the 2020 Census: What Do We Know Now? Presentation at the 2021 Population Association of American Conference, May 5, https://paa2021.secure-platform.com/a/gallery/rounds/9/details/1110
[4] O’Hare, W.P. (2020). “The Politicization of the 2020 Census,” PAA Affairs, Fall 2020, The Population Association of America, Washington DC. https://higherlogicdownload.s3.amazonaws.com/POPULATIONASSOCIATION/3e04a602-09fe-49d8-93e4-1dd0069a7f14/UploadedImages/Documents/PAA_Affairs/PAA-Fall_20_.pdf
Frey, W. H. (2020). “Trumps new plan to hijack the census will imperil America’s Future,” Brooking Institute, August 7, https://www.brookings.edu/blog/the-avenue/2020/08/07/trumps-new-plan-to-hijack-the-census-will-imperil-americas-future/
The Washington Post (2021. “Commerce department security Unit Evolved into counterintelligence-like operation, Washington Post examination found,” Shawn Boburg, May 24 https://www.washingtonpost.com/investigations/2021/05/24/commerce-department-monitoring-itms/
[6] National Research Council (2004) The 2000 Census Counting Under Adversity.
[7] Cohn, D. (2011). Imputation: Adding People to the Census, Pew Research Center, Washington DC. MAY 4, 2011
[8] U.S. Census Bureau (2012). DSSD 2010 Census Coverages Measurement Memorandum Series #2010-E-51, Table C, U.S. Census Bureau, Washington, DC.
[9] O’Hare W. P. (2019) Differential Undercounts in the U.S. Census: Who is Missing? Springer publishers, available open access at
O”Hare is right on all these points:
The culture of Census Bureau is essentially nonpartisan, statistical; its professional staff committed to professional, ethical practice. (see e.g. ASA Code of Ethics.)
Yes, there are differences in the 2020 Census headcount vs. earlier 2020 estimates.
Yes, some of the differences surprised us data-watchers.
But if there are differences, I (and O’Hare also) would trust the headcounts over the estimates. I wrote about this a month ago on twitter https://twitter.com/MetroGram/status/1393280571762085892
Where modeled estimation is weak: migration assumptions & migration data sources (USCB uses IRS summary counts of tax returns, misses people who live “under the radar”). That weakness seems confirmed by the “discrepancies” people are now balking about. If migration estimates calculations are flawed (because USCB uses IRS summary counts of tax returns?) then N.Y. will always be underestimated (underestimated in the estimates, even given a correct Decennial enumeration). Still other states will be overestimated: I’m thinking of states with people living “under the radar” (CA, AZ, TX). And also tax evasion havens, like FL. (A lot of people who “live” in FL for tax purposes… are not actually there when the Census-taker comes.)