Is Your "National" Dataset Actually Regional? A Five-Minute Test

Our given-name data for the United Kingdom was built from British sources, labelled British, validated as British, and shipped as British. It was Scottish.

Not partly Scottish, not Scottish-flavoured: every counter in it came from births registered in Scotland, a country holding about 8% of the United Kingdom's population. The file had passed every validator we had. It failed the first arithmetic test anyone bothered to run on it, and the test took less time than reading this paragraph.

This article is that test, plus the three that back it up, plus the reason the mistake is far more common than it sounds. A "national" dataset is regional more often than it is wrong in any other way, because regional registries publish more readily than national ones, and the file that is easiest to obtain is the file that gets used.

Test 1: add everything up

Any frequency dataset built from counts is claiming a population. Sum the counters and compare the total to the official figure for whatever the file claims to describe. That is the whole test.

Our file was organised as birth decades from the 1940s to the 2000s. Summed:

Now the comparison. Scotland registers roughly 4.5 million births across those decades. The United Kingdom registers roughly twelve times that.

We can put a measured number on the multiplier rather than an assumed one. For the years 2001–2010, per-name birth counts are published by all three of the UK's registrars, so the ratio is not an estimate:

Scotland is 7.98% of the UK's boys and 8.19% of its girls — a factor of 12.5 and 12.2 respectively. A genuinely UK-wide file covering the same decades would therefore hold somewhere around 25–30 million boys' births, not 2.4 million.

That is the entire finding. A discrepancy of an order of magnitude is not an error bar. Data can be stale, truncated, mis-parsed or sampled and still land within a factor of two. Landing within a factor of twelve of a specific sub-territory, and nowhere near the country on the label, means the file is about that sub-territory.

Note what this test does not require: no knowledge of British names, no list of Scottish names, no idea what Eilidh is. Two sums and a division.

Test 1b: the sum matched to the digit

The order-of-magnitude test tells you something is wrong. The next step tells you exactly what, and it is worth doing because it converts a suspicion into a proof.

Take the file's internal groupings — decades, regions, years, whatever it has — and compare each one separately against the candidate source. Ours:

Five columns out of seven reproduce a single country's birth registrations to the unit, and the other two are off by one name each. The girls' file behaves identically, including the same two window anomalies at the same values (206,683 for 1946–1950 and 213,582 for 2001–2008).

A coincidence at that precision does not exist. This is not a file that leans Scottish; it is Scottish birth registrations with a British label on the folder.

And the test threw in two defects for free, both invisible to any other check:

• The column labelled 1941 actually holds 1946–1950 — half a decade, and the half that excludes the war years.
• The column labelled 2001 actually holds 2001–2008 — eight years of ten, so that decade was 21% light relative to its neighbours.

Neither error is detectable by looking at the names. Both fall out of comparing a subtotal against the source that supposedly produced it.

Test 2: disproportionate individual entries

The sum test proves the file is regional. This one proves it to a reader who does not want to do arithmetic, and it works even when you cannot get a clean national total.

Pick entries you expect to be geographically loaded, compute their share inside the file, and compute the same share in the national distribution. The ratio is the tell. Measured on births 2001–2010 across all three UK registrars:

Any single row could be explained away. Eleven rows pointing the same direction, with the magnitudes ordered exactly as regional distinctiveness would order them, cannot.

But the important half of this test is the other direction, and it is the half people skip. A regional file does not merely add local colour. It deletes the majority's names:

Oliver and Harry were among the most common boys' names in the country during that decade, and the file had them at roughly a third of their real weight. Mohammed was down by a factor of four, because the sub-territory the data actually came from has a much smaller South Asian population than the country as a whole.

This is the damage that matters. Over-representing Eilidh makes a dataset quaint. Under-representing the actual top of the distribution by 3× makes it wrong about the ordinary case, which is the case it will mostly be asked about.

Test 3: rank overlap — and why you must run it on every slice

The blunt version of test 2. Take the top 20 of the file and the top 20 of the national distribution and count the intersection.

Twelve of the twenty commonest boys' names in the file were not in the country's top twenty at all: Cameron, Liam, Jamie, Kyle, Adam, Dylan, Ben, Connor, Andrew, Aaron, Logan, Alexander. Conversely Thomas, Oliver, Harry, William, Samuel, Joseph, Charlie, Benjamin, Ethan, George, Alfie and Luke were all in the national top twenty and none of them was in the file's.

Now look at the asymmetry between the rows, because it is the methodological point. Boys 8/20 is a screaming failure. Girls 13/20 is the kind of number you shrug at. Same file, same defect, same source, and one slice of it would have passed a casual review.

If your data is split into slices — by sex, by year, by region, by category — run every check on every slice. A defect that is uniform in the data is not uniform in its visibility.

Why this happens so often

The reason is not carelessness. It is that the availability of data is inversely correlated with the size of the territory, and nobody notices because availability feels like a property of the world rather than a bias in the sample.

Here is the concrete shape of it for one country. Three registrars publish per-name birth counts for the UK:

The smallest of the three territories publishes the longest history, in one flat file, with no suppression at all. If you want per-name UK data before 1996, Scotland is not the convenient option — it is the only option. Every builder who has ever needed mid-century British given names has hit that wall, and the tempting move is to take the Scottish file, scale it up by 12, and call it Britain.

We very deliberately did not do that when rebuilding, and the reason is this article: scaling a regional file to national totals fixes the magnitude and preserves the composition. You get a file that passes test 1 and fails test 2 harder than before, because now Blair has national-sized numbers attached to it.

The same trap, in the same shape, elsewhere in our own data: the only per-name birth files we could obtain for Australia and Canada cover New South Wales and British Columbia — one state and one province, each around a quarter to a third of its country. Both are published cleanly, both are labelled by their region in the source, and both would be sitting in a folder named after a country if nobody had checked.

What we did about it, and what we could not

Only one decade could be rebuilt honestly. The 2001–2010 window is the only one where all three registrars publish, so that decade now aggregates all three. The result of the merge is worth showing, because it doubles as a check that the merge itself is not skewed:

Summing the raw counts over identical years gives each territory its true weight automatically; no manual weighting is applied, and the residual disagreement with population shares is at most 0.6 points. If you have to hand-weight the merge to make it look right, your years do not line up.

The decades from the 1940s to the 1990s remain Scottish, because no source exists that would make them anything else. That is now written down next to the data instead of being discovered by the next person. A limitation you have measured and labelled is a fact about your dataset. The same limitation undocumented is a defect that regenerates: we have watched the same phantom task get re-opened twice because a note recorded a fix without recording its scope.

The checklist

1. Sum the counters and compare to an official total. No domain knowledge required. Anything more than roughly 2× off needs an explanation before you use the file for anything; an order of magnitude means you are holding a different population.
2. Compare every subtotal separately, not just the grand total. Matching a candidate source to the unit on one slice converts a suspicion into proof, and misaligned slices fall out for free — ours was labelled as covering a decade and covered five years.
3. Verify the time window against the label. Not "does it cover 1941–2010" but "does each period contain the years it claims". Two of our seven periods did not, and the shortfall was 21% on one of them.
4. Test geographic markers in both directions. Over-representation identifies the region; under-representation measures the damage. Our regional file inflated a local name by 12× and deflated two of the country's commonest names to a third.
5. Count administrative units. A national file should contain roughly the number of regions, provinces or districts the country has. One region's worth of them, or a suspiciously round subset, is the same defect seen from a different angle.
6. Run every check on every slice. Ours failed at 8/20 on one half and 13/20 on the other. Checking one half is checking nothing.
7. Ask who publishes, before asking what is published. If one sub-territory has an unusually generous statistical agency, the world's supply of data about that country is quietly skewed toward it — and your dataset inherits the skew whether or not anyone chose it.

Data as of 2026-07-21

All figures were recomputed on 21 July 2026 directly from the published registrar files listed below, not from our own notes. "Share in Scotland" and "share in the UK" are shares of registered births in 2001–2010, the only decade covered by all three registrars; ratios are the quotient of those two shares. The 25–30 million estimate for UK boys' births across 1946–2008 is derived by scaling, not taken from a publication, and is used only as an order-of-magnitude comparison.

⚠ Birth registrations are not living bearers. Every figure here describes people born in a period, which is the strongest available basis for British given names — no UK census publishes given-name counts.

Sources:

• National Records of Scotland (NRS) — full given-name lists with counts, per year: top-100 tables for 1935–1973 and complete lists for 1974–2020. NRS publishes counts down to a single birth. <nrscotland.gov.uk/statistics-and-data/st…;
• Office for National Statistics (ONS) — Baby names in England and Wales, per-name counts 1996–2025; counts of 2 or fewer are suppressed. <ons.gov.uk/peoplepopulationandcom…;
• Northern Ireland Statistics and Research Agency (NISRA) — baby-name statistics, per-name counts 1997–2025; counts below 3 are suppressed. From 2008 the tables are broken down by local government district, with an unallocated residual row that must be added, not treated as a total. <nisra.gov.uk/statistics/births-deat…;
• UK population shares by constituent country, 2020, used only as a sanity check on the merge.

A companion piece on the same failure mode from the other end — datasets that are honest about their scope but get merged anyway — is Migration Layers in Birth Cohorts. For the general taxonomy of frequency-data traps, see How to Audit a Name-Frequency Dataset.