Project snapshot
Anyone joining an outside dataset to Overture divisions has to answer one
question first: which level of their own geography does an Overture
subtype correspond to? The convenient assumption is that a subtype means
one depth everywhere, so a division tagged county lines up
with counties. Measured against three national administrative
hierarchies, that assumption fails often enough to break joins without
any error appearing, because the subtype reflects OpenStreetMap tagging
practice and tagging practice varies by country and by province.
This project replaces the assumption with a measurement. Every country and subtype pair is compared against every hierarchy level available for that country, and the result is recorded with the release it was measured against.
Data and methods
Sources used in the analysis:
- Overture divisions, release 2026-07-22.0. 5.5 million features across twelve subtypes running from country down to microhood. This is the system being measured.
- Global administrative boundaries. A worldwide hierarchy covering admin 0 through admin 2, which gives every country a comparison baseline even where no deeper source exists.
- National administrative spines for the United States, Canada, and Australia. The deepest official hierarchy each country publishes, used for the deep comparison that a global baseline cannot support.
Shared area between two boundaries is observed and measured directly. The verdicts built on top of that measurement are derived from thresholds: a pair reaches strong fit at 0.90 median coverage with 90 percent mutual-best agreement, reaches usable at 0.70 and 70 percent, and falls to no fit below that. A verdict describes the relationship at those thresholds and carries no claim beyond them.
How the comparison works
Each candidate level produces two numbers. Mutual-best agreement counts how often an Overture unit and one of our units each pick the other as their closest match, which answers whether the two systems are describing the same places. Median coverage measures how much of one boundary sits inside the other, which answers whether they agree about the edges. Both numbers are needed, because two hierarchies can partition the same country at similar granularity for unrelated reasons.
A ranking across candidate levels always produces a winner, so the method needs a way to reject every candidate. Greenland's best-scoring level reached 0.0018 coverage on a single Overture unit. That is a measurement, and an explicit no-fit verdict keeps it out of the published crosswalk. A second guard rejects pairs whose unit counts are too far apart to compare at all: Greece offers 332 Overture counties against a 14-unit baseline, a ratio of 23.7 to 1, and is recorded as unmeasurable with its reason rather than scored.
Current findings
The global pass holds 353 country and subtype pairs: 252 reach strong fit, 62 reach usable, 39 return no fit, and one is unmeasurable. Coverage is strong enough across most of the world to support a published crosswalk, and the exceptions cluster in ways that turn out to be informative.
Subtype does not mean a fixed depth. Of Canada's 1,532 Overture units
tagged county, 1,358 correspond to census subdivisions and
174 correspond to census divisions, one level higher, because provinces
tag the underlying boundaries differently. A single mapping applied to
that country sends more than a tenth of the rows to the wrong tier of the
hierarchy.
The better source also changes by country and by subtype. Australia's Overture counties align with the global baseline, where admin 2 holds Local Government Areas, and score 0.0549 against the national statistical hierarchy, which divides the country into statistical areas that no Overture subtype describes. Canada runs the other way for the same subtype. Ranking across every hierarchy available for a country produces the right answer; preferring one source by policy does not.
Below county, no administrative level fits anywhere. Overture's
locality subtype returns no fit against the deep hierarchies
of all three countries, and the pattern is identical each time: agreement
climbs steadily as the comparison gets finer while coverage stays flat and
low. Measured against United States block groups, the median Overture
locality has 69 percent of its area inside its closest match while that
match has only 31 percent of its area inside the locality, and which
direction dominates depends on population density.
That signature identifies a family that crosses the hierarchy rather than a level chosen wrongly. Every deep hierarchy in this comparison is statistical, and Overture localities are populated places, which cross statistical boundaries by design. Canada is the clearest warning: 5,830 Overture localities against 5,161 census subdivisions is a ratio of 1.1 to 1, which looks like a match and loses to a level with eight times the unit count. Similar cardinality is a coincidence, and only measured overlap settles the question.
What this changes for anyone joining these systems
Measure the correspondence per country and per subtype, and store the resulting level alongside the release it was measured against, because both are properties of that country's mapping rather than global constants. Where the measurement returns no fit, treat the geography as its own family and connect it through shared places instead of forcing it into a hierarchy level it was never going to occupy.