Not the Marriage Rate, but the Metadata: A Data-Provenance Crisis Exposed by Mexico's Statistical Release

On 29 September 2026, Mexico's National Institute of Statistics and Geography (INEGI) published press release 90/26. The document is dry, numerical, almost toneless. In 2026, 491,640 marriages were registered nationwide; the rate stood at 5.4 per 1,000 residents aged 18 and over. In Mexico City the rate was 3.0 — roughly 44 percent below the national figure. In Quintana Roo it was 8.6 — some 59 percent above it.

I studied civil engineering before turning to journalism, and I have spent more than thirty years listening to the sounds behind the numbers. The year I stopped calling the game and started listening to it, I settled on one rule: the number does not arrive first, the source of the number does. When this document entered an analytical pipeline, it carried a tag on its back: football. Yet across its 35 information points there is not one club, not one player, not one contract, not one league. The entire document is demography. And that is the real event here — not the marriage rate, but the metadata.

Context

The programme behind the release is EMAT — Estadística de Matrimonios, Mexico's official marriage statistics. It is a flow indicator: how many events were registered in a given year. A stock indicator is something else: what share of a population sits in a given state at a moment in time. EMAT does not measure the distribution of marital status; it measures a year's events.

Look at the series. In 2026 the rate was 6.6. In 2026 it collapsed to 3.8. Then 5.1 in 2026, 5.7 in 2026, 5.6 in 2026, 5.4 in 2026, 5.4 in 2026. The post-2026 recovery is real but partial — the series has slipped back from its 2026 peak and flattened. Volume reached 491,640, up 1.0 percent on 2026. An identical rate with a 1.0 percent rise in count mathematically implies roughly 1.0 percent adult population growth. Demographically, that is plausible. The internal arithmetic holds.

One age-related figure stands out: between 2026 and 2026 — nine years — average age at marriage rose 4.1 years for both sexes. That is about 0.46 years per calendar year, fast by international comparison. Part of that rise may be a composition effect: fewer early marriages mechanically lift the mean.

Education tells its own story: 81.1 percent of newlyweds had at least secondary schooling, and more than half married within the same education level. Same-sex marriages numbered 6,481 — about 1.3 percent of all unions, registered in 31 of 32 federal entities.

Core analysis: where the numbers say one thing and the headline another

The central problem in this document is not the statistics but the language wrapped around them. The headline claims that almost half of the capital's residents are single. No information point in the document supplies a marital-status distribution. A stock claim is riding on flow data. The two are different measurements. Mexico City's 3.0 is low not because marriage is vanishing there but because consensual unions — unión libre — are growing, and those are not counted in this indicator at all.

By the same logic, Quintana Roo's 8.6 is not purely a behavioural story. Quintana Roo is a destination-wedding jurisdiction. Non-residents register marriages there. An administrative ratio does not always measure resident behaviour.

An internal tension also remains. Women aged 15 to 19 account for 2.5 percent of marriages — roughly 12,300 events on a base of 491,640. Yet only two marriages involving a minor were recorded. The only coherent reconciliation is that the band is almost entirely composed of 18- and 19-year-olds, consistent with legal reforms raising the minimum marriage age to 18 in most states. Records with unspecified age number 2,623, or 0.53 percent of the total — too few to distort the distribution materially.

Now the tag. The document's type was labelled correctly — Data Insight. Its domain was labelled football. Type detection worked; domain detection failed. The fault is narrow, and therefore fixable — if anyone notices.

This is where the blockchain question becomes relevant, and it becomes relevant exactly where it is needed. In modern data governance, the most defensible use of a distributed ledger is not prediction but proof — an immutable log of who published a document, when, and in what state. What this record needed first was a chain of provenance: publication date, press release number, outlet, editing layer, each stage hashed and preserved. Had that existed, the primary source (INEGI press release 90/26) and the secondary gloss (an unnamed outlet, a stock-image caption) could never have merged into a single layer.

Notice the pattern. The document's paratext — title, subtitle, image caption — is almost entirely surplus. Its body is disciplined, numerical, careful. The headline speaks of a transformation of the capital involving its youth, something INEGI never says; the institute explicitly declines to speculate on causes. And the claim that almost half the population is single has not a single supporting figure in the release.

Contrarian angle: a blockchain verifies information, not meaning

Here is my second hesitation, and it marks the weakest point in this whole discussion. If a wrong label is written on-chain, it stops being wrong and becomes immutably wrong. A blockchain proves who wrote it, when, and whether it was altered. It does not prove the writing is true. Garbage in, garbage on-chain.

The reverse also holds. The person who dissents, who says your interpretation is wrong, is not easily discoverable on a ledger — because a wrong interpretation never gets a hash. That is precisely why body and paratext must stay separate, why source layers must stay separate, and why the institute's own toneless language must survive intact beside the printed figure. Only what the document states should travel under the institute's name; everything else belongs on a different layer.

A subtler lesson emerges. A mislabelled domain is not merely a bad tag — it is a pipeline signal. If a demography document enters as football, then next time a genuine football document may enter as demography. Dataset contamination is never a single event; it is a tendency.

In my own practice I have stood in an empty stadium with two recorders — 49,430 seats, zero spectators, and the loudest sound of the night a substitute's shout carrying from the bench in the 78th minute. I learned then that emptiness and absence of information are not the same thing. By the same token, a numerically impeccable document can be informationally incomplete — if it wears the wrong label.

Takeaway

Two signals remain. First: the rate has held at 5.4 from 2026 to 2026 — whether that is a floor or a pause will be answered by the 2026 vintage. Second: the 2026-to-2026 rebound carries the mark of deferred post-crisis marriage rather than structural recovery, because the series began falling again after its peak.

I no longer ask whether the marriage rate is declining. I ask whether the document carrying these numbers can carry its own identity. If it cannot, then some policymaker in the next decade will decide badly on the strength of these figures — and the blame will fall on the statistics, which were only quietly telling the truth.

The day a dataset forgets its own name, every true number inside it walks into the room of suspicion. So the question is not about marriage. The question is: how many documents in your pipeline are circulating under the wrong name today, and is anyone noticing?