Abraham Lincoln once expressed the desire, in a time of civil war, to preserve a government that was “of the people, by the people, for the people.” What he did not say was that such government has also always been of the data, by the data, and sometimes for the data. Democratic governance has been fundamentally data-driven for a very long time. Representation in the US depends on a constitutional requirement, instituted at the founding, for an “actual enumeration” of the population every 10 years: a census designed to ensure that the people are represented accurately, in their proper places, and in proportion to their relative numbers.
A complete national count is always a monumental task, but the most recent actual enumeration faced unprecedented challenges. The 2020 census had first to overcome the Trump administration’s ill-conceived effort to add a citizenship question. Then it spent half the year in the field straining to count every person during a pandemic that made knocking on strangers’ doors particularly difficult. A series of devastating hurricanes and wildfires added to the challenge. And yet, in late April 2021, the professional staff of the US Census Bureau managed to fulfill the constitution’s mandate and revealed state-level population totals, translating those into an apportionment of the 435 seats of the US House and a corresponding number of votes in the electoral college. (The apportionment occurred automatically according to an algorithm, called “equal proportions” or “Huntington-Hill,” that is prescribed by law.) Now, just last month, we learned that some of those numbers were, most likely, wrong.
The Census Bureau’s Post-Enumeration Survey (PES) went back out into the field, reinterviewing a sample of people from throughout the country, and then compared the new, more in-depth survey to the results of the census. Analyzing this comparison, the bureau now estimates that the 2020 census overcounted in eight states and undercounted in six. To give a sense of the scale of these errors, the PES reported with 90 percent confidence that New York’s state population was overcounted by anywhere from 400,000 to over 1 million additional people, or 1.89 to 4.99 percent of the population. Considering the circumstances of the count, such low error rates should be considered impressive, and yet such differences can have big consequences when the last seat in the US House has, since 1940, been decided by as few as 89 people and no more than 17,000. Much of the initial commentary on the PES results has focused on the horse-race implications of the errors, pointing out that more of the states that were overcounted were blue states, while more of those undercounted were red. The errors, apparently favoring one party over another, have even been labeled “a scandal” and the census written off as “a bust.”
These are overreactions, and yet the question remains: What should we do about these small, but both statistically and politically significant errors?
This is a conundrum that our nation’s leaders have wrestled with since the founding. Over the course of the last century, two distinct approaches have dominated. One depends on funneling money and energy into mobilizing more census takers and toward other systemic reforms that preemptively reduce error. The other involves statisticians who have worked to develop techniques that can measure error precisely and then make corrections to the census counts. Both of these approaches remain important, and yet the scale of the 2020 miscounts suggests that an older method for dealing with census error should be revived: We should expand the House and the electoral college, so that few or no states lose representation in the face of an uncertain count. We should try to count better and fix what errors we can, but our democracy will be more robust if we also lower the stakes of each census. Representation need not be a zero-sum game.
The earliest known reference to a census undercount came from Thomas Jefferson, then secretary of state, who wrote in 1791 about the prior year’s census, the nation’s first. Jefferson wrote his correspondents in Europe, assuring them that the American population was a few percentage points larger than officially declared. It’s hard to say if this was indeed the case, but the story makes clear that concerns about omissions and undercounts began more than two centuries ago. In subsequent decades, disasters and administrative failures caused serious omissions, such as when the official charged with counting Alabama’s residents died in office before completing his work on the 1820 census, or when many of California’s records (including the entirety of San Francisco County) burned after the 1850 census.
Most PopularBusinessThe End of Airbnb in New York
Amanda Hoover
BusinessThis Is the True Scale of New York’s Airbnb Apocalypse
Amanda Hoover
CultureStarfield Will Be the Meme Game for Decades to Come
Will Bedingfield
GearThe 15 Best Electric Bikes for Every Kind of Ride
Adrienne So
Around the turn of the 20th century, a newly professionalized Census Bureau began attempting to quantify miscounts to fend off perceptions of persistent mismanagement. An early estimate asserted an error rate of less than 1 percent for the 1890 census. But Black analysts working inside the segregated Census Bureau in the 1910s soon drew attention to alarming racial disparities in a series of 19th-century censuses, and the Howard University mathematician Kelly Miller published a 1922 article showing a substantial undercount of African Americans in 1920. An influential paper published in 1947 confirmed that omissions marred the count of nonwhite populations to an alarming degree, while also making the case that the entire census count was off by about 3 percent. By 1950, the Census Bureau was convinced it needed to institute its own regular measurement estimating flaws in its coverage.
One key step was to inaugurate the PES. For the 1950 census, its first, the PES determined that there was a “net” error rate of 1.4 percent. But net rates could disguise larger “gross” errors. The omission rate was 2.3 percent overall and reached 4.5 percent for the nonwhite population.
The 1950 census results were the first to be evaluated by a PES and also the first to be automatically translated into an apportionment of representatives by a new algorithmic system, one that remains in place today. Up to that point, the size of the House of Representatives and distributions of House seats for each decade had usually been established by Congress after an extensive debate grounded in the results of the most recent census. When Congress passed each new apportionment law, it usually ended up expanding the House (and so too the electoral college). Following this “count and increase” tradition, Congress often chose a size of the House that guaranteed that no state would lose a seat.
Then, in 1920, Republican leaders in Congress claimed that an enlarged House would be too costly and inefficient and determined to hold the House at 435 seats. Such efforts ended up torpedoing the entire apportionment process, and the US failed to ever use the 1920 census results to redistribute representation.
Those committed to preventing House expansion learned from this debacle and began constructing an algorithmic system that would take apportionment out of Congress’ hands. In 1929, Congress passed a new law, originally billed as a kind of “life insurance for the Constitution of the United States,” that would ensure that the results of the census would lead to apportionment even if Congress did not act during its subsequent lame-duck session. This back-up system presumed the House would be fixed at 435 seats, with the primary question before Congress being one of method: what algorithm would be employed to translate population totals into seat allocations?
There are many different ways to cut the representational cake, but each way results in some scraps and leftovers that have to be accounted for. Congress has used variations on at least four different methods in the past, and the debate over which method was best reached a fever pitch between 1920 and 1940. Ultimately, political considerations carried the day in 1941, and a Democratic Congress replaced the “major fractions” or “Webster” method, which had been the legal default, with the algorithm called “equal proportions” or “Huntington/Hill” because the latter handed the final seat in the House to solidly Democratic Arkansas instead of Republican-supporting Michigan. Every subsequent apportionment, beginning in 1951, would take place automatically without any Congressional involvement.
Most PopularBusinessThe End of Airbnb in New York
Amanda Hoover
BusinessThis Is the True Scale of New York’s Airbnb Apocalypse
Amanda Hoover
CultureStarfield Will Be the Meme Game for Decades to Come
Will Bedingfield
GearThe 15 Best Electric Bikes for Every Kind of Ride
Adrienne So
With the size of the House fixed, energies have been focused on two possible means of eliminating error within the data upon which the automated system relies. First, Congress and the bureau have, often in the face of complaints and lawsuits initiated by cities, states, and civil rights organizations, thrown bodies and money at the problem. Today, most households answer their census questions directly online, and yet the number of enumerators sent out into the field in 2020 was about three times the number sent out in 1950 when every household had to be counted directly by an enumerator.
The second approach grew out of the PES directly. In the closing decades of the 20th century, the Census Bureau developed a strategy of “statistical adjustment” that would use a very large, specially designed sample survey to produce revised numbers that could then be used in place of the original count when it came time to allocate House seats. Such efforts proved to be technically difficult, though—and, more importantly, politically untenable for apportionment. Politicians came to see adjustment as a partisan issue in the 1980s and '90s, as historian Margo Anderson explains, with Republicans stymieing the turn to sampling to make up for miscounts. Then, in 1999, the Supreme Court found that census laws forbid the use of sampling in apportionment, though it was silent as to whether a sampled and adjusted census could count as a constitutionally required “actual enumeration.” Until Congress summons the political will to revise standing laws and risks a constitutional challenge, House seat allocations cannot rest on data improved through sampling.
While both of these tracks seek to improve the census and its data, they can do nothing about the troubling sensitivity of the automatic apportionment system. It has become evident that small changes in population counts can swing a seat from one state to another. This makes it difficult even to determine how the errors reported by PES would change the apportionment, since the survey produces a wide range of estimates for each state—and since the PES doesn’t even include counts for “group quarters,” which are large institutional settings like nursing homes, prisons, or universities. (For this reason, I counsel significant skepticism when anyone claims to show how the PES would change apportionments.) It has also become clear that seats often swing due to changes from one valid method for calculating apportionments to another. For example, after the 2020 state totals were announced, I calculated the apportionment using the “method of major fractions” and found that it would have required swapping seats for two pairs of states. We have engineered a system within which minor changes to a host of different procedures or assumptions, as well as small (and to some degree inevitable) errors in a complicated data system, can deny a state and its residents a deserved representative or electoral college vote.
We’ve allowed our democracy to devolve into a game of musical chairs. The population of the United States has tripled, while states must fight over the same 435 seats and must see their fates determined by an algorithmic system plagued by arbitrary outcomes. It is time to enlarge the House of Representatives, a conclusion shared by a bipartisan committee of the American Academy of Arts and Sciences. That committee suggests an initial addition of 150 seats, which would certainly be a good start. Others favor instituting a new rule or method for determining the size, such as the cube-root law. I think the key factor is that Congress take back its constitutional authority to decide each apportionment, a move that would—if history can inform our predictions—result in a House that tends to grow in proportion with the population. The automated system has served primarily to lock the House at 435; if Congress reasserts its authority, the House will most likely grow every 10 years. This would be a good thing in itself, since it would make it more likely that each of our voices could be heard in Congress. Maybe it could also shake Washington out of perpetual gridlock. It is also a tried and true method for coping with uncertainty and error in our democracy’s data.
We must continue to plan for a better census count in 2030, with adequate funding, one that is more robust to unexpected events and better insulated from partisan pressure. Right now, a well-funded Census Bureau should also create revised population estimates to more fairly distribute federal funds to communities over the coming decade. But it is beyond time for Congress to decrease the injustice that arises from taking away seats from growing states based on tiny population differences using data that appears more precise than it is. Revelations of census error offer yet another argument that it is time to enlarge the US House.