How Does a Census Work? Counting Everyone, and Why It Is So Hard

The census answers questions that no other source can, because it attempts completeness rather than a sample. Its outputs are used constantly and mostly invisibly:

• •Apportionment: how many representatives each region sends to a national legislature, which in the United States is the constitutional purpose of the count and is why it is mandated every ten years
• •Drawing constituency boundaries, since equal-sized districts require knowing where people are
• •Distributing funding, with formulas for health, education, transport and local government grants built on population counts and their characteristics
• •Planning infrastructure, from school places and hospital capacity to water supply and transport, twenty years ahead
• •Providing the denominator for almost every other statistic, since a rate of anything per hundred thousand people requires knowing the hundred thousand
• •Anchoring surveys, which draw their samples from a known population structure and are reweighted against it

The mechanics have three phases. Before the count, the statistical agency builds an address register, listing every dwelling in the country, which is the single largest determinant of coverage, since a household at an address nobody knows about will not be contacted. On census day, forms are delivered or an online code is issued, and every household is legally required to respond, with a fine for refusal in most countries. Afterwards comes the follow-up, in which enumerators visit non-responding addresses repeatedly, and then the processing: forms are scanned, responses coded, contradictions resolved, and missing items filled by imputation, a statistical estimate based on similar households. The reference point is a single date so that people are counted once, and rules exist for students, people with two homes, prisoners, the homeless, visitors and those born or dying near the date.

Undercounting is universal and not random, which is what makes it a political problem rather than a technical one. The groups consistently missed are young men, renters, recent movers, people in shared or informal housing, ethnic minorities, undocumented migrants and the homeless, while the groups over-counted are people with two addresses, chiefly students and second-home owners. Because the missed groups are concentrated in particular places, the error shifts representation and money away from exactly the areas with most need. Statistical agencies deal with this openly: a post-enumeration survey re-canvasses a random sample of areas with great intensity shortly after the count, compares the results person by person, and produces an estimate of how many were missed and of what kind, which is then used to adjust the published figures. Britain and several other countries publish adjusted estimates as the official result; the United States Supreme Court ruled in 1999 that sampled adjustment cannot be used for apportionment specifically, so the raw count stands for that purpose.

Which questions appear is decided politically and each addition is fought over. Religion, ethnicity, language, sexual orientation, gender identity, income and immigration status have all been contested somewhere. The stakes are that a question creates an official count, which creates a recognised group, which creates a basis for claims; and that a question can also suppress response, which is the argument that defeated a proposed citizenship question in the United States in 2019, where the government's own statisticians estimated it would cause millions of people in immigrant households to avoid the count entirely. Confidentiality is the other permanent issue. Individual returns are protected by statute, typically sealed for a hundred years, and agencies apply disclosure control to published tables so that small cells cannot identify a person. The reason the guarantees are taken seriously is historical, since census and registration records were used to locate targeted populations in the Netherlands and elsewhere during the Second World War, and in Rwanda in 1994.

Several countries have stopped. The Nordic states abolished the traditional census decades ago and produce the same outputs from linked administrative registers of population, address, employment and education, which is cheaper, annual rather than decennial, and possible only because those countries maintain a universal population register that many others regard as politically unacceptable. The Netherlands, Austria and others have followed. Britain has debated the same move and, after the 2021 count, its statistical agency recommended a system based on administrative data with surveys attached; a decision to make 2021 the last traditional census has been repeatedly discussed. The trade-off is real: registers give timely data and lose the ability to ask a new question, capture self-defined identity, or reach people the state has no record of, who are precisely the people a count exists to find.

A census tries to count every resident on a single date, and its outputs determine parliamentary seats, funding formulas, infrastructure planning and the denominator of nearly every other statistic. It runs on an address register, a legally compulsory response, intensive follow-up of non-responders and statistical imputation of gaps, and it always undercounts the same groups, which agencies measure with a post-enumeration survey and correct for. Which questions appear is fought over politically, and several countries have replaced the exercise with linked administrative registers.