Checking an official spending figure: four places a citation quietly breaks

A national spending figure is the easiest kind of statistic to quote and the easiest to misquote. China's household survey puts per-person consumption expenditure at 29,476 yuan for 2025 (NBS release, 2026-01-19), a number that now appears in relocation guides, salary threads and market decks. I spent a day checking every clause I planned to use against the releasing body's own pages. Four of the checks changed what I wrote, and they repeat on almost any official series.

These are the working notes, not the article: what broke, and the check that caught it.

1. An annual table with a nine-month row in it

The 2025 annual release and the first three quarters of 2026 arrived in documents that look alike. The quarterly figure is 22,354 yuan per person (NBS release, 2026-10-19). It sits one row below an annual figure, in the same unit, at the same precision.

Annualising it is the obvious next step, and it is wrong. 22,354 times four thirds is 29,805, which is 329 yuan above the actual 2025 total — close enough to look like rounding, far enough to be a different number. Spending is seasonal: the fourth quarter carries holiday and heating costs, and the release never annualises the partial year.

The check I now run on every table: print the coverage window in the same cell as the value, or in the row directly above it, before comparing anything. Two numbers that differ by 1% are usually two different periods, not a revision.

2. Category shares only reconcile against the published total

The release breaks spending into eight categories. Adding them up is a cheap integrity check, and it caught an error of mine on the first pass.

The eight 2025 categories sum to exactly 29,476 yuan, and the eight for the first three quarters of 2026 sum to exactly 22,354. If a sum is off, the usual cause is a mixed pair of editions: one category renamed between the Chinese and English editions, and a row matched by name that should have been matched by value.

Name matching fails here for a second reason. Category labels are translated, not standardised, and two editions of the same release can carry slightly different wording for the same line. The reliable key is the pair (rank, value): sort both editions descending, compare the eight values, and only then attach the label. Any label that survives that test is safe to quote; any label that does not is a translation, not a fact.

3. The definition lives in a different document than the numbers

Knowing what counts as consumption expenditure decides whether the figure can be described as a living cost. The answer is in the indicator explanations that accompany the statistical yearbook, not in the release: cash spending plus in-kind consumption, plus an imputed rent for owner-occupied housing.

That imputation is the part most secondary write-ups drop. It means the residence line is not a rent payment and not a mortgage payment. A household that owns its flat outright still reports housing consumption, so the housing share is closer to what the market would charge than to what the household hands over each month.

There is a smaller honesty problem underneath. The explanation page the statistics bureau links to carries an older yearbook's URL. The substance matches what the current releases count, but I could not open the newest edition's equivalent page to compare it line by line. The defensible sentence is therefore "as defined in the indicator explanation I could reach on this date", with the date in it, rather than a claim about the current definition.

4. Province figures are four different document types, published months apart

The national release gives one number. Anything below national level comes from the provincial statistics bureau, and the formats do not match:

Three details break naive comparison. The communiqués appear two to four months after the national release, so a cross-section stitched together in January is incomplete. Rounding differs between a data page and a communiqué. And one provincial site served its document over plain HTTP, which is a fetching problem before it is a data problem.

One more trap sits inside the province table. The all-residents figure is not the midpoint of the urban and rural figures, and averaging them is wrong. In Henan the urban figure is 1.48 times the rural one; in Sichuan, 1.62; nationally, 1.77. The all-residents value is weighted by the resident population in each part, and the weights are not published in the table. Use the published all-residents number or don't use one.

5. Derived ratios inherit every mismatch upstream

Dividing spending by disposable income is the fastest way to say something about the cost of living, and the fastest way to build a number that is internally inconsistent. Both sides have to come from the same survey release and the same period. Mixing an income figure from a preliminary release with a spending figure from the annual one produces a ratio that looks precise and means nothing.

When both sides do match, the spread is large: spending is about 59.5% of disposable income in Shanghai and 72.2% in Sichuan, with the national figure at 68.0%. A province that spends little is partly a province that earns little, which is why a single spending number cannot separate prices from pay.

6. The conversion rate is a dated fact, not a constant

Converting to dollars needs one decision made explicit: the rate and its date. I used the central parity rate published for 2026-11-10, 6.7167 yuan to the dollar, and labelled the column with that date.

That conversion sizes a budget. It does not compare price levels: the same yuan buys different amounts in Shanghai and in an interior county, and a market rate does not correct for that. A purchasing-power comparison needs a different factor, and no factor of that kind appeared in any release I opened.

What I could not check

• Cross-country comparison. The US Bureau of Labor Statistics pages answered 403 to a non-browser request from this machine, and India's statistics ministry site did not resolve on the day. Both surveys also count different units — a consumer unit per year against a person per month — so lining them up needs work I could not do here.
• A purchasing-power factor. No primary series was reachable, so the dollar figures stay market-rate only.
• Household-level totals. The releases publish per-person figures; the household totals behind them are not in the public tables, so how spending scales with household size is untested here.
• The full year 2026. Only three quarters exist; the annual release is due in early 2027.

The check list I keep

Every number I publish now carries three fields that cost nothing to store: the period it covers, the unit it is counted in (person or household), and the date I read the source. Most of the errors above are a missing one of those three, not a bad source.

The other habit that paid off: recompute everything derived rather than copying a chart. Shares, ratios, sums and conversions took a few minutes and caught two mistakes — a mixed edition and an annualised partial year — that would have shipped silently.

Drafted with an AI assistant from my own reading of the primary documents; every clause number and link here was checked against the source.