The 1918 border in today's figures, commune by commune

The wager: dead empires still show in statistics

Poland's election maps still seem to redraw the nineteenth-century partition borders, and the resemblance has become proverbial. For former Habsburg lands, a widely cited study finds higher trust in courts and police right beside the old frontier, in a sample that includes Romania.

Romania has its own vanished line. Until 1918 the frontier between the Kingdom and Austria-Hungary followed the Carpathians from Bukovina to Orșova. The explanation “that is how it is over there, it was the empire” is still heard daily. The test is simple: communes a few kilometres either side of the old line are compared as if only history separated them; if an indicator jumps exactly at the line, the line still means something. The assumption is checked on altitude, population and density — not on every possible difference, and the method section says what the test cannot prove.

The line reconstructed from the 1914 state polygons measures 1,185 km across today's Romania.

The vote does not redraw the empire

In the presidential run-off of 18 May 2025, communes beside the old border voted almost identically on both sides. At the line, the model places the former Habsburg side +0.8 percentage points from the Old Kingdom side, with a wide uncertainty interval: the data cannot tell the difference from zero, without proving it is exactly zero. Turnout behaves the same: -0.9 points. Weighted by population, the estimate becomes +14.1 points (±5.1; p = 0.005), but five large towns carry it: without them it falls to +3.5. Pairs of neighbouring communes across the line, 95 of them, in turn give +7.2 points. The three readings ask different things — what the typical commune looks like, what the communes' populations look like, what close neighbours look like — and do not replace one another; the main specification remains the typical commune.

An earlier manuscript by Roman Levkin (undated, publication status unverified; read here only in excerpts — the opened excerpt) found, on parliamentary elections of the 1990s and 2000s and a 60 km band, a surplus of about 3.5 points for the major right parties beside the same line. Different ballots, decades and bands: the two measurements cannot be overlaid, so neither shows that an effect “disappeared”. On the 2025 presidential data, with the 25 km band, no step measures.

At the line, the vote for Nicușor Dan starts from 36.6% on the Old Kingdom side; the estimated difference between the sides is +0.8 points.

Houses beside the line look alike

The 2021 census counts in every commune its buildings, dwellings, rooms and their area. Beside the old border, the share of occupied dwellings, rooms per dwelling and square metres of living space per resident cross the line without a step, in both directions and at every band tried.

Age structure and population change behave the same: the shares of children and of the elderly, and growth or decline against 2011 and 2002, show no clear step at the line in this comparison. One possible explanation is a century in one state, with the same schools and norms; a single snapshot in time, such as this one, cannot prove how the resemblance came about.

The census does not publish schooling or dwelling amenities (running water, sewerage, bathroom) by commune: those tables stop at county level. The full search log is in the method section.

Nor do national-exam grades

For schooling the census cannot help at commune level, but the Ministry of Education publishes every candidate's result with the school of origin. Aggregated by commune, mean grades at the 2026 national evaluation — the exam at the end of grade 8, marked from 1 to 10 with 5 as the pass mark — show no step at the old border: the estimated difference at the line is -0.01 grade points, with an uncertainty interval considerably wider than the estimate itself — absence of significance is not proof of equality.

The measure has a stated limit: grades sit on the school's commune, not the child's home commune, and commuting pupils move some results toward towns. How much commuting crosses the old line was not measured here.

143,251 candidates with a final mean, in 3,110 communes; the test uses the 2,445 communes with at least ten graded candidates (139,284 candidates).

Composition differs as a slope, not a step

Here the map really does split in two, but not at the line. The shares of declared Hungarians and of Hungarian mother tongue rise gradually from the crest toward the inner basins and reach almost the other side's level right beside the old border. At 25 km the fitted step at the line is -3.8 points and statistically unclear; widening the window to 40 km, the same step becomes +8.2 points. The estimate thus depends on the window's width — a sign that the geography behind the crest follows no straight line, and that the “slope” reading rests on the profile in the chart, not on a single coefficient.

The shares of declared Roman Catholics and Orthodox draw the same picture, and so, more weakly, does the 2025 vote. The differences between the country's two halves exist; beside the former frontier their shape is a continuous incline. The distinction matters: a step at the line would be compatible with a trace of the old administrations (without identifying it on its own), while a slope shows the geography of settlement, changing gradually in both directions.

Ethnic and religious composition is shown only as a distribution across communes. No sentence on this page characterises a group.

The mountain objection: how much of the line can be tested fairly

A crest frontier is a poor witness: “just across the border” often means “across a mountain” — different climate, roads and distance to towns. 75% of the line's length runs at 700 m or higher. Only 147 of the 379 communes in the test band face low stretches: the Bukovina hills, the passes and the Cerna valley.

That is why every result below is also recomputed for those low stretches alone. A jump that appears along the whole line but vanishes where the mountain separates nothing would be suspect; a jump that stays there deserves more belief.

The line's altitude is sampled from the Copernicus GLO-30 terrain model, kilometre by kilometre; its median lies above 1,100 m.

The exception: the paperwork of land

The national cadastre register tells another story — a less simple one than the first measurement suggested. In the baseline model, right at the old line, former Habsburg communes hold 80 more registered land properties per square kilometre than the modelled level on the Old Kingdom side of the line (66; that level is the fitted value exactly at the line, not the mean of all Old Kingdom communes), +0.76 more per dwelling, and 10.7 percentage points more registered area. Apartments do not enter this count: the source file reports them separately, and these indicators count registered land properties.

Three further tests move the conclusion. First, neighbouring communes resemble each other (spatial autocorrelation of the residuals, Moran's I = 0.50); with Conley spatial standard errors at 50 km, the pooled step along the whole line falls to p = 0.056, and after adjustment across the 29 indicators, q = 0.33 — suggestive, not established; a 50 km geographic block-bootstrap interval, [-21, +159], also covers zero. Second, the step is not spread evenly — without Suceava county it drops to +12 per square kilometre; on the Bukovina stretch (40 km), however, it is +155 and survives the spatial treatment (p = 0.0001). Third, the signal lives inside localities: registered properties within the built-up area, per dwelling, jump by +1.11 (spatial p = 0.009), while the outside-locality density does not (+4.4). Registration density is not a coverage rate: it can also reflect smaller parcels.

The 2016–2024 cadastre series changes the chronology as well: the per-square-kilometre step stood at +41 in December 2016, doubled by 2020 (+78) and levelled off there; on the Bukovina stretch it grew from +63 to +155. Levels rose on both sides, faster beyond the line. A difference that deepened during the years of systematic registration cannot be told purely as the passive survival of the land book: the legal inheritance (parcel-based land books beyond the line, person-indexed registers in the Old Kingdom; unification in 1938, applied in Transylvania in 1947; national framework in 1996) remains a documented hypothesis, but the recent programme's speed and the parcel structure are part of the story. The test measures the step and its movement; how much of it comes from 1918 and how much from 2016–2024 cannot be separated with these data.

The per-km² and area jumps survive the 15 and 40 km bands, density controls and dropping the communes the line runs through (38 of them); per dwelling changes sign under population weighting. Registered area can exceed 100% of a commune in the source file; 12 values above 150% are capped at 150% (inside the band: Victoria, Brașov county, ~171% in the source).

The test inside one county

The simplest rival explanation would be administrative: recent systematic-registration programmes run by county and by lot, and the old line often follows today's county borders. In the north, however, the frontier of Austrian Bukovina with Moldavia crosses the middle of today's Suceava county, over low hills.

Within this one county, on a 40 km band (its width calls for a wider one than the main test), former Austrian communes hold 135 more registered properties per square kilometre at the line than former Moldavian ones. A county-differentiated programme cannot by itself explain a step inside the same county; commune-level lots, parcel fragmentation and restitution can still differ and are discussed in the method section.

111 communes of Suceava county enter the test, on both sides of the former Bukovina line.

The counter-test: lines that never existed

The same test was repeated on five imaginary lines: the frontier shifted 30 and 60 km east and west and 40 km south. On each control line, between two and five of the 29 indicators pass the raw 5% threshold — more than pure chance would leave, because the shifted lines cross the same structured geography. For the cadastre, the shifted lines generally give smaller or sign-changing jumps, with one serious exception: for registered area, the line shifted 60 km west produces -13.0 points, larger in absolute value than the true line (+10.7). For properties per square kilometre, however, the true line (+80) stays above all five shifted lines (-47 … +20).

After the Benjamini–Hochberg correction for all 29 tests at once, a single indicator stays under the 10% threshold: registered properties per square kilometre. The other two cadastre measures have q = 0.213, so they do not pass the correction, though they carry the same sign in the main specifications — not in all (population weighting and the eastern segment, above). The result is stated here exactly as strongly as these tests support: a step measured at the line for registration density, supported more weakly by the family's other two measures, with a documented mechanism and its alternatives written beside it.

The full estimates, with standard errors and adjusted values, are in the results table below; nothing is selected merely because it “came out”.