Quality assurance on census data: catching errors while you can still fix them

Census errors are cheap to fix while the team is still in the field and impossible to fix afterwards. That asymmetry should determine how a survey is run, and usually does not: quality control is scheduled as a data-cleaning activity after enumeration closes, at which point the only available correction is deletion.

Three kinds of error, three different fixes

• Capture errors — a mistyped dimension, a wrong species, a missing field. Caught by validation at the point of entry.
• Systematic errors — an enumerator who consistently measures structures externally where the protocol says internally, or who classifies every mud-and-wattle wall as semi-permanent. Caught only by comparing enumerators to each other.
• Coverage errors — households, claims and assets that were never recorded. Caught only by an independent check against the ground.

The third is the most damaging and the least detectable in the data, because a missing record looks like nothing at all.[1]

A clean dataset and a complete one are different achievements. Most quality assurance measures the first and reports it as though it were the second.

Validation at capture

The instrument should refuse impossible entries rather than record them: areas outside a plausible range, structures with no dimensions, perennial crops with no maturity class, claimants with no contact detail, records with no coordinates. Each of these is a constraint that costs nothing and prevents a category of correction that would otherwise require a return visit.

Equally important is what the instrument permits: multiple claims against a parcel, individuals recorded within a household rather than a single head, and a route for the enumerator to flag an unresolved boundary or a disputed claim as such rather than being forced to record a resolved number.[2]

Daily supervision, not weekly

A supervisor reviewing each enumerator's records the same evening catches a misunderstanding on day two rather than in week six. That single practice removes most systematic error, and it requires the data to be visible centrally within hours — which is an argument for offline-capable digital capture with regular sync rather than for paper returned at the end of a phase.

The independent re-check

The check that matters most is also the one most often cut: an independent re-enumeration of a random sample by a different team, compared field by field against the original.

It should cover five to ten per cent of records, be selected randomly rather than by convenience, and report variance by field and by enumerator rather than as a single accuracy figure. Where a particular enumerator's structure areas run consistently high, or one team's tree counts run consistently low, that is a systematic error affecting every record they touched — and it can still be corrected while they are mobilised.[4]

Public display as the coverage check

No internal process detects a household that was never visited. The community does. Public display of the draft register in each settlement, with a stated objection period and a simple way to raise an omission, is the only coverage check that works.[3]

It has to happen before payments begin, because after that the incentive structure around objections changes entirely.

What to report

• Coverage: records completed against the enumerated area, with gaps named.
• Re-check variance by field and by enumerator, not as an aggregate.
• Objections received at public display, and how each was resolved.
• Records flagged as disputed, undocumented or succession-blocked, as a live worklist.

That last item is the one that connects quality assurance to the payment schedule: the flags raised at census are the cases that will still be open two years later unless somebody starts working them now.