Read our COVID-19 research and news.

Residents in Greenville, Texas, are urged to fill out the 2020 census at a walk-in center.

LP Otero/AP

Will the 2020 census numbers be good enough, and how soon will we know?

With the clock winding down on the 2020 census, the U.S. statistical community is mounting a last-ditch campaign to prevent what it fears will be a seriously flawed head count.

Social scientists want Congress to create an independent oversight body to monitor the final weeks of the decennial census and to give the Census Bureau more time to count every U.S. resident. The proposals come in response to several recent actions by the Trump administration that experts fear will undermine the agency’s constitutionally mandated duty to produce a complete and accurate census. Those moves include the White House taking the unprecedented step of adding three high-level political appointees to the agency and throttling back the bureau’s massive effort, now underway, to track down the roughly one-third of all residents who failed to respond to repeated reminders to fill out and return the 10-question census.

Anything less than a final response of 99% will put the 2020 census in dangerous territory, statisticians say. A lower response rate will require the agency to use an unprecedented amount of data on file with other government agencies to infer the demographic characteristics of persons believed to be living at a particular address, a process called imputation. Even more worrisome, they say: Some demographic groups, including people of color, immigrants, and those with lower incomes, are both more likely to be missed and less likely to be captured in the existing administrative records used to plug holes in the count.

The census does not simply provide social scientists with a feast of fresh demographic data. Its primary purpose is to determine how many seats each state gets in the 435-member U.S. House of Representatives, a process called apportionment. Its data are also the basis for allocating $1.5 trillion per year in federal spending.

The 2020 count has been challenging. The coronavirus pandemic delayed the start of the bureau’s nonresponse follow-up (NRFU) campaign this spring and led to a request in April for a 4-month extension in submitting the state-by-state numbers used for apportionment. But the Trump administration rescinded that request, and on 3 August Census Director Steven Dillingham announced that field operations would end on 30 September, 4 weeks earlier than planned.

Restoring the additional time the bureau had said it needed to complete the count is a top goal for census advocates and many legislators. In May, the Democrat-controlled House endorsed the 4-month extension as part of a pandemic relief package. But that bill has stalled in the Republican-led Senate.

The curtailed schedule is affecting more than NRFU. Census officials also have canceled a follow-up exercise, known as count review operation (CRO), that was scheduled for mid-September. The exercise would have given state demographers the chance to review the bureau’s list of so-called group quarters—places that are home to large numbers of residents, including college dormitories, prisons, and nursing homes—to make sure NRFU enumerators didn’t miss such locations when they knocked on doors. But meeting the 31 December deadline “precludes conducting the [count review] event,” Census officials wrote in a 12 August email to state officials.

Metrics for success

Census advocates see the jettisoning of CRO as just the latest example of actions that are jeopardizing the bureau’s exhaustive plans to ensure a high-quality count. Those fears have led four former Census directors to ask Congress to create an outside body with the authority to monitor the final few months of the decennial census.

The panel, working with data provided by the bureau, would examine how well the agency is doing on a handful of important indicators. Those “metrics for success” could reveal in real time whether the 2020 count had resulted in an unacceptably high undercount among certain demographic groups or particular locations, says Kenneth Prewitt, who led the bureau during the 2000 census. The panel could also compare the 2020 numbers with current U.S. population estimates based on noncensus data such as births, deaths, and patterns of immigration/emigration, a time-honored technique for judging the accuracy of the decennial census.

Testifying before the House oversight committee on 29 July, Prewitt said that an expert panel assembled by the National Academies of Sciences, Engineering, and Medicine could use those metrics “to assess if the final 2020 numbers reasonably match what the bureau knows they should be and, if not, what steps the country should take.” Speaking recently with ScienceInsider, Prewitt added that “the federal standards for data quality are very high, and for uses like apportionment, you need to meet the highest standard.”

The head of the House oversight panel, Representative Carolyn Maloney (D–NY), supports the idea of an outside expert panel but hasn’t settled on a legislative strategy to make it happen. “The chairwoman is currently evaluating all options,” says a senior Democratic aide on the committee, which is also seeking to interview several top census administrators.

Keeping an eye on the data

Last week, the Census Bureau took a small step in the direction that experts want it to move. It announced it would provide daily updates on NRFU. That information, combined with numbers on self-response rates that it had been posting weekly since the census officially launched on 1 April, would present a running total of the percentage of addresses enumerated in each state.

The current tally reflects the tremendous variation that worries statisticians. Although the national self-response rate was 64%, for example, it was only 31% in Puerto Rico and a robust 73% in Minnesota. The NRFU rate also fluctuates, adding only three percentage points to the total in some states and more than 20 percentage points in others. And some states with low self-response rates have yet to benefit from NRFU: New Mexico’s subpar 54% self-response rate had grown by only six percentage points as of 20 August, for example, while Idaho’s high-end self-response rate of 68% had jumped another 23 percentage points.

But such aggregated, state-level data don’t tell the whole story, observers say. The biggest threat to a quality count, they note, are differing self-response and NRFU rates at very small geographic levels, such as several city blocks or a portion of a rural community. “To the extent there is variation,” says Robert Groves, a statistician and provost of Georgetown University, “it’s a red flag for the possibility of a disproportionate undercount.”

The nature of that undercount would be revealed through what Groves calls “process” indicators—metrics that an independent panel could use to evaluate whether the census is on track. One such metric could be the extent to which enumerators had to use proxies—interviewing a neighbor, for example—to find out whether a particular house or apartment was occupied and, if so, how many people were living at that address on 1 April.

Proxies are less reliable than self-responses, and Groves says “it would be good to know” how often the Census Bureau is relying on proxy responses to acquire the demographic information—race, sex, age, and size of a household—that is requested on the census form. Other helpful process indicators could be the share of returned forms that didn’t answer one or more questions, or that didn’t provide any useful information at all.

“Credibility requires transparency,” Groves says. “And the sooner the country can see multiple indicators of quality on the 2020 census, the sooner one can make the case for its use.”

The bureau still plans to field a post enumeration survey (PES), in which an elite cadre of enumerators knock on the doors of some 180,000 representative households around the country. Knowing the percentage of households that were not included in one or both exercises is another way for outsiders to evaluate the quality of the census, experts say. But the PES is typically done well after NRFU field operations have ended, and the analysis can take several additional months.

Too many cooks?

Another factor driving calls for outside oversight of the census is fear that three new political appointees to the Census Bureau could interfere with how the 2020 data are analyzed and eventually released to the public. Maloney’s House panel, for example, wants to speak with all of them, including statistician Benjamin Overholt, who on 17 August was unveiled as the agency’s new deputy director for data.

Census advocates worry that Overholt, an Army veteran with a 2013 Ph.D. in applied statistics and research methods who had been working for the federal Equal Employment Opportunity Commission, will at best be redundant and could even hamper census operations. They are also baffled by the timing of his appointment.

“Ensuring data quality is the job of everybody involved in the collection, curation, and dissemination of census products,” says biostatistician Thomas Louis, who was the bureau’s chief scientist from 2013 to 2015 and is an emeritus professor at Johns Hopkins University. “At this point, somebody like [Overholt] can only get in the way. Of all the things the Census Bureau doesn’t need at this point, a deputy director for data is at the top of my list.”

Advocates for the outside panel note that it can’t function without access to data available only from the bureau, which is bound by law to protect it from disclosures that could reveal someone’s identity. And the agency hasn’t indicated it would welcome such a body, much less provide it, in real time, with the needed information.

“There is precedent for this type of collaborative relationship,” says Katherine Wallman, a former chief statistician at the White House Office of Management and Budget. “But there’s nothing of this sort going on now. And I’m not optimistic that the bureau would do this voluntarily. I think it would require a mandate from Congress that would include asking a group like the National Academies to do it.”

Coming up short

Census experts fear the country will face a wrenching technical, political, and legal controversy if a sizable number of policymakers and the public conclude the 2020 count is flawed. And the depth of the dispute could depend on the magnitude of the potential undercount.

“The 2010 census was 99.6% complete” after follow-up door knocking, recalls Robert Rhatigan, the state demographer of New Mexico. “And the Census Bureau used imputation” to come up with information on the rest of the households on its master list of addresses. “But what do we do on 30 September if we’re at 95%, or even 98%?”

Prewitt’s guess is that the undercount will be much larger than that, with the census missing from 12% to 15% of U.S. residents. “And you can’t impute 15%,” he warns.

Such a large undercount is particularly problematic, he explains, because a census isn’t like an election, he explains. “An election works with a 40% turnout. But if you’re at 85%, it’s no longer a census.”

A hefty undercount will certainly trigger a heated discussion of whether the data can be used for apportionment (and, by extension, political redistricting at the state and local levels) and the allocation of federal dollars. And it’s not clear who would make the final decision.

“I think that the Census Bureau has the authority to say that the data are not good enough,” Prewitt says. “So I think at some point Dillingham needs to step up and tell it like it is.”

President Donald Trump has raised the political stakes with a 21 July executive order that asks for state-by-state numbers from the 2020 census that don’t include undocumented residents. Civil rights groups have sued to block implementation of that order, which they say violates language in the U.S. Constitution requiring the census to count every resident.

The decision-making process is further complicated by the fact that it is Dillingham’s boss, Commerce Secretary Wilbur Ross, who will actually deliver the apportionment numbers to the president, who then conveys them to Congress for adoption. In the past, the department released the state population totals at the same time they were delivered to the White House, in order to demonstrate that the data had not been subject to political influence. The chain of command was perfunctory; as one old census hand puts it, “the whole process is supposed to work automatically, without any political interference.”

But this year could be different, warns Terri Ann Lowenthal, a consultant on census matters and former staff director for a congressional panel overseeing the bureau. “This new policy breaks with many historical practices,” she notes. And she is “not confident” that Ross will follow precedent in releasing the state population data simultaneously.

The fraught political landscape likely reduces the chances that Congress will be able to pass legislation this fall requiring the Census Bureau to cooperate with an outside body, census advocates say. “Documenting where you are cutting corners doesn’t seem like a good strategy,” says one veteran census watcher about the expected resistance from the Trump administration. “And what are the odds that Congress could act that quickly?”

Over the longer term, however, the political gridlock now could boost support for a do-over later this decade. There’s nothing that prevents Congress from ordering up a new census before 2030, notes another census expert who requested anonymity to speak frankly because of professional commitments. If the 2020 count becomes a debacle, that expert predicts, “I think there will be a groundswell for a middecade census.”