The census that missed a street

A national census undercounted one city district by an estimated eleven per cent. The questionnaires were sound, the staff were trained, and the error was almost entirely about addresses.

The district contained a large number of subdivided houses: one front door, four households, no separate numbers. The address list, built from postal records, held one entry per door. Enumerators knocked, spoke to whoever answered, recorded a household, and moved on. Nothing in the procedure was wrong; the frame was simply incomplete, and no amount of care in the interview can recover a household the list never contained.

Undercount of this kind is not random. It concentrates among renters, recent arrivals and people sharing accommodation, which means the missing eleven per cent differed systematically from the counted eighty-nine. Because funding formulas for schools, clinics and transport use census population, the district received resources for a population it did not have, and the shortfall persisted for a decade until the next count.

Statistical agencies now treat coverage as a separate problem from response. A post-enumeration survey revisits a sample of areas with a fresh address list built from aerial imagery, utility connections and local knowledge, and the difference between the two counts estimates the undercount. The estimate is then published alongside the headline figure, with its own margin of error.

The habit this builds is worth borrowing. For any dataset, the first question is not how the measurements were taken but how the list of things to measure was assembled. Errors in measurement scatter and partly cancel. Errors in the frame point in one direction and are invisible to every check that works inside the data.