Count before you administer

Adopt by the decision rule and publish in the code:

• What is asked, and the use of each question.
• That every household must answer, and the small, fixed penalty for refusing.
• That individual returns are used only to produce statistics — never for tax, conscription, policing, or any decision about a named person.
• That the returns are sealed: the UK keeps its census returns closed for 100 years, the US for 72. Nobody may open them early, including the people in charge.

The UN Fundamental Principles of Official Statistics state the rule in principle 6: individual data are strictly confidential and used exclusively for statistical purposes.

Every question costs answer quality on the others. A first census needs about ten, on one sheet per household:

• For each person: name (for checking only, never tabulated), relationship to the head of household, sex, age in completed years, birthplace, marital status, occupation, whether they can read and write.
• For the dwelling: number of rooms, wall and roof material, water source, latrine.

Write beside each question the decision it will inform — grain store size, school places, water points. A question with no decision beside it is cut. Where people do not know their age, estimate it from a list of remembered local events with known dates — a flood, a famine, a founding. Test the sheet on 20 households and reword every question someone misunderstood.

Divide the whole territory on a map into enumeration areas, each bounded by features anyone can find on the ground — streams, paths, walls — with no gaps between them and no overlaps.

Size each so one enumerator can visit every dwelling in the counting period. At about 20 minutes a household and 6 hours of visiting a day, one enumerator manages about 18 households a day — about 250 in two weeks; plan for 150–250 to leave time for return visits. Before the count, walk each area and list every building, numbered on the map. The building list is what the enumerator works through, and what the coverage check is measured against.

One enumerator per area, and a supervisor for each small group who re-visits a sample of each enumerator's first households, while mistakes can still be corrected. Train on the questions word for word, the definitions — who counts as a household, where a visitor is counted — and practice interviews, until each trainee completes a sheet without error.

Give each an identity card and a visible mark of the role, as in the photograph. An unidentified stranger asking questions does not get the door opened. Announce the count weeks ahead, at every gathering: the date, why it is taken, and that answers cannot be used against anyone.

Fix a reference date — census night — and record each person either where they were that night or where they usually live; choose one rule and keep it everywhere. Counting can take two weeks; every answer refers to the one night.

Count deliberately the people easiest to miss: travellers and herders on the move (where they spend census night), people without a dwelling, people in hospitals, prisons and boarding schools, newborns and the very old. Mark each building on the list as done, empty or refused, and return to every refusal with the supervisor.

No census is complete, and a total without its error is a guess with a decimal point. Within weeks of the count, a separate team re-counts a random sample of areas without seeing the census results, then matches the two lists person by person.

People found by the recount and missed by the census give the miss rate; publish it with the totals. Modern censuses run the same check, called a post-enumeration survey, as the UN census principles recommend; misses concentrate in infants, young men and people who move often, so check those groups first.

Copy the answers from each household sheet onto tally sheets by area, one line per answer category, and add every column twice, by different clerks. Area totals must sum to the territory total. Keep the sheets in the order of the building list, so any figure can be traced back to the households that produced it.

Hand tabulation is slow: the US Census Bureau records that the 1880 census took about eight years to process, which is why a tabulating machine was built for the next one — see audit, statistics and the published number.

Publish the tables, the questionnaire, the enumerator instructions and the miss rate together, within a year of the count, on one announced date, to everyone at the same moment — no early copy for those in charge. Post them where the published code is posted.

Count again at least every 10 years, the UN standard, on the same questions where possible, so growth and decline show. Between censuses, the register of births and deaths keeps the numbers current.