Official Figures Keep Moving After Publication

Why a published figure is not a final one

Anyone quoting an official statistic is trusting something the table cannot confirm: that the number taken today is the number a reader will find tomorrow. Often it is not. Statistical offices revise, and the new value takes the old one's place in the same cell, with nothing to announce that anything moved.

So a reader arriving today cannot tell whether the figure before them has stood for years or was replaced yesterday: only someone who kept the earlier copy can see that anything changed, and almost nobody keeps it.

Why sources revise at all

Every statistical office works against the same conflict: the date it has promised a figure falls before the information needed to compute it arrives. Tax records, company accounts and the reports that offices collect themselves arrive on calendars of their own, annual accounts long after the period they describe.

An office could wait for all of it, and publish a very accurate figure about a period nobody is still arguing about. The trade is deliberately made the other way: an approximation on time, labelled an early estimate, replaced as the evidence lands. A revision is therefore not a repair but the second half of the bargain struck when the figure went out early.

A producer that never revised would be doing one of two things, both worse: publishing so late that its work is history rather than information, or declining to act on what it had learned since — the definition of a source not to trust.

The kinds of revision, from routine to rare

Most revisions are routine: the same figure recalculated as fuller returns arrive. Producers call the successive results vintages: the same measurement of the same period, as it stood at different dates. Which way the next one will go is not knowable in advance; if it were, the producer would have applied it already.

A second kind is larger and rarer: every so often an office changes something structural — a definition, the classification deciding which activity counts as which industry, or the base year an index is measured against, the year set equal to one hundred so later values read as movement away from it. Nothing about the economy changed; the ruler did.

Such a change cannot be applied from today onward alone: a series whose definition changed halfway along would show a jump reflecting nothing in the world, a break in the series rather than an event in the economy. So the earlier years are reworked on the new basis, and a table unchanged for decades is restated at once. Nothing about the past has been re-measured; it is being described in a new vocabulary, and a figure on the old basis is one the source no longer publishes.

The third kind, the outright error, is rare and usually obvious in size: a jump of an order of magnitude ordinarily means a decimal point or a unit slipped, not an event in the world.

Figures that move with no new data at all

Two ordinary treatments do this. Seasonal adjustment strips out the yearly rhythm of holidays, harvests and shopping seasons, so a real change is not mistaken for the time of year, and the model estimating that rhythm is re-fitted as each period arrives. Chain-linking, which joins stretches of a series priced in different years so quantities stay comparable, does the same.

A recent figure deserves more caution than an old one

The latest period is the one still being assembled, so its figure is the likeliest to move and moves furthest, while older periods settle. That inverts a natural instinct: the freshest number feels the most authoritative and is in fact the most provisional. And the gap between two vintages can be wider than the change being reported as news from it.

Before you quote a number

Record the vintage: which figure you took and the day you took it. A claim that names its source but not its date cannot be defended once the table has moved on.

Prefer the revision, and name it as one: where the earlier figure is already published, correct it in the open rather than quietly, because both numbers are in circulation and only you can connect them.

Keep the extract, not the link that produced it: a link resolves to today's table, and only a saved extraction can still show what you actually read.

Those are the rules this site applies to itself. Every figure published here is checked against an extraction archived on the day it was taken, a daily watch pulls a fresh copy whenever a declared source moves, and a superseded extraction is never deleted — which is what puts a second dated snapshot beside the first.

The revisions observatory this note accompanies compares those consecutive snapshots and lists what differs: the value before, the value after, and the dates of the two extractions. It records what this archive happened to photograph twice, not everything a statistical office has ever restated.